2019-06-04 11:11:32 +03:00
// SPDX-License-Identifier: GPL-2.0-only
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
/*
* Copyright ( C ) 2009 Red Hat , Inc .
*/
2014-06-05 03:06:58 +04:00
# define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
# include <linux/mm.h>
# include <linux/sched.h>
2021-05-05 04:34:23 +03:00
# include <linux/sched/mm.h>
2017-02-08 20:51:30 +03:00
# include <linux/sched/coredump.h>
2017-02-08 20:51:31 +03:00
# include <linux/sched/numa_balancing.h>
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
# include <linux/highmem.h>
# include <linux/hugetlb.h>
# include <linux/mmu_notifier.h>
# include <linux/rmap.h>
# include <linux/swap.h>
2012-12-13 01:51:06 +04:00
# include <linux/shrinker.h>
2011-01-14 02:46:58 +03:00
# include <linux/mm_inline.h>
2016-01-16 03:54:10 +03:00
# include <linux/swapops.h>
2022-06-29 03:15:29 +03:00
# include <linux/backing-dev.h>
2015-09-09 00:58:45 +03:00
# include <linux/dax.h>
2011-01-14 02:46:58 +03:00
# include <linux/khugepaged.h>
2011-01-14 02:47:10 +03:00
# include <linux/freezer.h>
2016-01-16 03:56:43 +03:00
# include <linux/pfn_t.h>
2011-01-14 02:47:17 +03:00
# include <linux/mman.h>
2016-01-16 03:56:55 +03:00
# include <linux/memremap.h>
2012-10-15 15:44:56 +04:00
# include <linux/pagemap.h>
2016-01-16 03:54:40 +03:00
# include <linux/debugfs.h>
2012-11-02 15:33:45 +04:00
# include <linux/migrate.h>
2013-02-23 04:32:27 +04:00
# include <linux/hashtable.h>
2015-09-05 01:46:20 +03:00
# include <linux/userfaultfd_k.h>
mm: introduce idle page tracking
Knowing the portion of memory that is not used by a certain application or
memory cgroup (idle memory) can be useful for partitioning the system
efficiently, e.g. by setting memory cgroup limits appropriately.
Currently, the only means to estimate the amount of idle memory provided
by the kernel is /proc/PID/{clear_refs,smaps}: the user can clear the
access bit for all pages mapped to a particular process by writing 1 to
clear_refs, wait for some time, and then count smaps:Referenced. However,
this method has two serious shortcomings:
- it does not count unmapped file pages
- it affects the reclaimer logic
To overcome these drawbacks, this patch introduces two new page flags,
Idle and Young, and a new sysfs file, /sys/kernel/mm/page_idle/bitmap.
A page's Idle flag can only be set from userspace by setting bit in
/sys/kernel/mm/page_idle/bitmap at the offset corresponding to the page,
and it is cleared whenever the page is accessed either through page tables
(it is cleared in page_referenced() in this case) or using the read(2)
system call (mark_page_accessed()). Thus by setting the Idle flag for
pages of a particular workload, which can be found e.g. by reading
/proc/PID/pagemap, waiting for some time to let the workload access its
working set, and then reading the bitmap file, one can estimate the amount
of pages that are not used by the workload.
The Young page flag is used to avoid interference with the memory
reclaimer. A page's Young flag is set whenever the Access bit of a page
table entry pointing to the page is cleared by writing to the bitmap file.
If page_referenced() is called on a Young page, it will add 1 to its
return value, therefore concealing the fact that the Access bit was
cleared.
Note, since there is no room for extra page flags on 32 bit, this feature
uses extended page flags when compiled on 32 bit.
[akpm@linux-foundation.org: fix build]
[akpm@linux-foundation.org: kpageidle requires an MMU]
[akpm@linux-foundation.org: decouple from page-flags rework]
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Reviewed-by: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Greg Thelen <gthelen@google.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-10 01:35:45 +03:00
# include <linux/page_idle.h>
2016-07-27 01:25:51 +03:00
# include <linux/shmem_fs.h>
2017-08-19 01:16:15 +03:00
# include <linux/oom.h>
2019-03-06 02:42:58 +03:00
# include <linux/numa.h>
2019-08-25 03:54:59 +03:00
# include <linux/page_owner.h>
2022-03-23 00:46:27 +03:00
# include <linux/sched/sysctl.h>
2022-08-18 16:10:41 +03:00
# include <linux/memory-tiers.h>
2012-12-13 01:51:06 +04:00
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
# include <asm/tlb.h>
# include <asm/pgalloc.h>
# include "internal.h"
2022-05-10 04:20:47 +03:00
# include "swap.h"
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
2022-03-25 04:09:58 +03:00
# define CREATE_TRACE_POINTS
# include <trace/events/thp.h>
2011-01-14 02:46:58 +03:00
/*
2017-05-17 22:19:21 +03:00
* By default , transparent hugepage support is disabled in order to avoid
* risking an increased memory footprint for applications that are not
* guaranteed to benefit from it . When transparent hugepage support is
* enabled , it is for all mappings , and khugepaged scans all mappings .
2013-11-13 03:07:16 +04:00
* Defrag is invoked by khugepaged hugepage allocations and by page faults
* for all hugepage allocations .
2011-01-14 02:46:58 +03:00
*/
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
unsigned long transparent_hugepage_flags __read_mostly =
2011-01-14 02:47:07 +03:00
# ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
2011-01-14 02:46:58 +03:00
( 1 < < TRANSPARENT_HUGEPAGE_FLAG ) |
2011-01-14 02:47:07 +03:00
# endif
# ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
( 1 < < TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG ) |
# endif
mm: thp: set THP defrag by default to madvise and add a stall-free defrag option
THP defrag is enabled by default to direct reclaim/compact but not wake
kswapd in the event of a THP allocation failure. The problem is that
THP allocation requests potentially enter reclaim/compaction. This
potentially incurs a severe stall that is not guaranteed to be offset by
reduced TLB misses. While there has been considerable effort to reduce
the impact of reclaim/compaction, it is still a high cost and workloads
that should fit in memory fail to do so. Specifically, a simple
anon/file streaming workload will enter direct reclaim on NUMA at least
even though the working set size is 80% of RAM. It's been years and
it's time to throw in the towel.
First, this patch defines THP defrag as follows;
madvise: A failed allocation will direct reclaim/compact if the application requests it
never: Neither reclaim/compact nor wake kswapd
defer: A failed allocation will wake kswapd/kcompactd
always: A failed allocation will direct reclaim/compact (historical behaviour)
khugepaged defrag will enter direct/reclaim but not wake kswapd.
Next it sets the default defrag option to be "madvise" to only enter
direct reclaim/compaction for applications that specifically requested
it.
Lastly, it removes a check from the page allocator slowpath that is
related to __GFP_THISNODE to allow "defer" to work. The callers that
really cares are slub/slab and they are updated accordingly. The slab
one may be surprising because it also corrects a comment as kswapd was
never woken up by that path.
This means that a THP fault will no longer stall for most applications
by default and the ideal for most users that get THP if they are
immediately available. There are still options for users that prefer a
stall at startup of a new application by either restoring historical
behaviour with "always" or pick a half-way point with "defer" where
kswapd does some of the work in the background and wakes kcompactd if
necessary. THP defrag for khugepaged remains enabled and will enter
direct/reclaim but no wakeup kswapd or kcompactd.
After this patch a THP allocation failure will quickly fallback and rely
on khugepaged to recover the situation at some time in the future. In
some cases, this will reduce THP usage but the benefit of THP is hard to
measure and not a universal win where as a stall to reclaim/compaction
is definitely measurable and can be painful.
The first test for this is using "usemem" to read a large file and write
a large anonymous mapping (to avoid the zero page) multiple times. The
total size of the mappings is 80% of RAM and the benchmark simply
measures how long it takes to complete. It uses multiple threads to see
if that is a factor. On UMA, the performance is almost identical so is
not reported but on NUMA, we see this
usemem
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean System-1 102.86 ( 0.00%) 46.81 ( 54.50%)
Amean System-4 37.85 ( 0.00%) 34.02 ( 10.12%)
Amean System-7 48.12 ( 0.00%) 46.89 ( 2.56%)
Amean System-12 51.98 ( 0.00%) 56.96 ( -9.57%)
Amean System-21 80.16 ( 0.00%) 79.05 ( 1.39%)
Amean System-30 110.71 ( 0.00%) 107.17 ( 3.20%)
Amean System-48 127.98 ( 0.00%) 124.83 ( 2.46%)
Amean Elapsd-1 185.84 ( 0.00%) 105.51 ( 43.23%)
Amean Elapsd-4 26.19 ( 0.00%) 25.58 ( 2.33%)
Amean Elapsd-7 21.65 ( 0.00%) 21.62 ( 0.16%)
Amean Elapsd-12 18.58 ( 0.00%) 17.94 ( 3.43%)
Amean Elapsd-21 17.53 ( 0.00%) 16.60 ( 5.33%)
Amean Elapsd-30 17.45 ( 0.00%) 17.13 ( 1.84%)
Amean Elapsd-48 15.40 ( 0.00%) 15.27 ( 0.82%)
For a single thread, the benchmark completes 43.23% faster with this
patch applied with smaller benefits as the thread increases. Similar,
notice the large reduction in most cases in system CPU usage. The
overall CPU time is
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
User 10357.65 10438.33
System 3988.88 3543.94
Elapsed 2203.01 1634.41
Which is substantial. Now, the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 128458477 278352931
Major Faults 2174976 225
Swap Ins 16904701 0
Swap Outs 17359627 0
Allocation stalls 43611 0
DMA allocs 0 0
DMA32 allocs 19832646 19448017
Normal allocs 614488453 580941839
Movable allocs 0 0
Direct pages scanned 24163800 0
Kswapd pages scanned 0 0
Kswapd pages reclaimed 0 0
Direct pages reclaimed 20691346 0
Compaction stalls 42263 0
Compaction success 938 0
Compaction failures 41325 0
This patch eliminates almost all swapping and direct reclaim activity.
There is still overhead but it's from NUMA balancing which does not
identify that it's pointless trying to do anything with this workload.
I also tried the thpscale benchmark which forces a corner case where
compaction can be used heavily and measures the latency of whether base
or huge pages were used
thpscale Fault Latencies
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean fault-base-1 5288.84 ( 0.00%) 2817.12 ( 46.73%)
Amean fault-base-3 6365.53 ( 0.00%) 3499.11 ( 45.03%)
Amean fault-base-5 6526.19 ( 0.00%) 4363.06 ( 33.15%)
Amean fault-base-7 7142.25 ( 0.00%) 4858.08 ( 31.98%)
Amean fault-base-12 13827.64 ( 0.00%) 10292.11 ( 25.57%)
Amean fault-base-18 18235.07 ( 0.00%) 13788.84 ( 24.38%)
Amean fault-base-24 21597.80 ( 0.00%) 24388.03 (-12.92%)
Amean fault-base-30 26754.15 ( 0.00%) 19700.55 ( 26.36%)
Amean fault-base-32 26784.94 ( 0.00%) 19513.57 ( 27.15%)
Amean fault-huge-1 4223.96 ( 0.00%) 2178.57 ( 48.42%)
Amean fault-huge-3 2194.77 ( 0.00%) 2149.74 ( 2.05%)
Amean fault-huge-5 2569.60 ( 0.00%) 2346.95 ( 8.66%)
Amean fault-huge-7 3612.69 ( 0.00%) 2997.70 ( 17.02%)
Amean fault-huge-12 3301.75 ( 0.00%) 6727.02 (-103.74%)
Amean fault-huge-18 6696.47 ( 0.00%) 6685.72 ( 0.16%)
Amean fault-huge-24 8000.72 ( 0.00%) 9311.43 (-16.38%)
Amean fault-huge-30 13305.55 ( 0.00%) 9750.45 ( 26.72%)
Amean fault-huge-32 9981.71 ( 0.00%) 10316.06 ( -3.35%)
The average time to fault pages is substantially reduced in the majority
of caseds but with the obvious caveat that fewer THPs are actually used
in this adverse workload
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Percentage huge-1 0.71 ( 0.00%) 14.04 (1865.22%)
Percentage huge-3 10.77 ( 0.00%) 33.05 (206.85%)
Percentage huge-5 60.39 ( 0.00%) 38.51 (-36.23%)
Percentage huge-7 45.97 ( 0.00%) 34.57 (-24.79%)
Percentage huge-12 68.12 ( 0.00%) 40.07 (-41.17%)
Percentage huge-18 64.93 ( 0.00%) 47.82 (-26.35%)
Percentage huge-24 62.69 ( 0.00%) 44.23 (-29.44%)
Percentage huge-30 43.49 ( 0.00%) 55.38 ( 27.34%)
Percentage huge-32 50.72 ( 0.00%) 51.90 ( 2.35%)
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 37429143 47564000
Major Faults 1916 1558
Swap Ins 1466 1079
Swap Outs 2936863 149626
Allocation stalls 62510 3
DMA allocs 0 0
DMA32 allocs 6566458 6401314
Normal allocs 216361697 216538171
Movable allocs 0 0
Direct pages scanned 25977580 17998
Kswapd pages scanned 0 3638931
Kswapd pages reclaimed 0 207236
Direct pages reclaimed 8833714 88
Compaction stalls 103349 5
Compaction success 270 4
Compaction failures 103079 1
Note again that while this does swap as it's an aggressive workload, the
direct relcim activity and allocation stalls is substantially reduced.
There is some kswapd activity but ftrace showed that the kswapd activity
was due to normal wakeups from 4K pages being allocated.
Compaction-related stalls and activity are almost eliminated.
I also tried the stutter benchmark. For this, I do not have figures for
NUMA but it's something that does impact UMA so I'll report what is
available
stutter
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Min mmap 7.3571 ( 0.00%) 7.3438 ( 0.18%)
1st-qrtle mmap 7.5278 ( 0.00%) 17.9200 (-138.05%)
2nd-qrtle mmap 7.6818 ( 0.00%) 21.6055 (-181.25%)
3rd-qrtle mmap 11.0889 ( 0.00%) 21.8881 (-97.39%)
Max-90% mmap 27.8978 ( 0.00%) 22.1632 ( 20.56%)
Max-93% mmap 28.3202 ( 0.00%) 22.3044 ( 21.24%)
Max-95% mmap 28.5600 ( 0.00%) 22.4580 ( 21.37%)
Max-99% mmap 29.6032 ( 0.00%) 25.5216 ( 13.79%)
Max mmap 4109.7289 ( 0.00%) 4813.9832 (-17.14%)
Mean mmap 12.4474 ( 0.00%) 19.3027 (-55.07%)
This benchmark is trying to fault an anonymous mapping while there is a
heavy IO load -- a scenario that desktop users used to complain about
frequently. This shows a mix because the ideal case of mapping with THP
is not hit as often. However, note that 99% of the mappings complete
13.79% faster. The CPU usage here is particularly interesting
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
User 67.50 0.99
System 1327.88 91.30
Elapsed 2079.00 2128.98
And once again we look at the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 335241922 1314582827
Major Faults 715 819
Swap Ins 0 0
Swap Outs 0 0
Allocation stalls 532723 0
DMA allocs 0 0
DMA32 allocs 1822364341 1177950222
Normal allocs 1815640808 1517844854
Movable allocs 0 0
Direct pages scanned 21892772 0
Kswapd pages scanned 20015890 41879484
Kswapd pages reclaimed 19961986 41822072
Direct pages reclaimed 21892741 0
Compaction stalls 1065755 0
Compaction success 514 0
Compaction failures 1065241 0
Allocation stalls and all direct reclaim activity is eliminated as well
as compaction-related stalls.
THP gives impressive gains in some cases but only if they are quickly
available. We're not going to reach the point where they are completely
free so lets take the costs out of the fast paths finally and defer the
cost to kswapd, kcompactd and khugepaged where it belongs.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-18 00:19:23 +03:00
( 1 < < TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG ) |
2012-12-13 01:51:12 +04:00
( 1 < < TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG ) |
( 1 < < TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG ) ;
2011-01-14 02:46:58 +03:00
2016-01-16 03:54:17 +03:00
static struct shrinker deferred_split_shrinker ;
2011-01-14 02:47:04 +03:00
2012-12-13 01:51:06 +04:00
static atomic_t huge_zero_refcount ;
2015-02-12 02:24:51 +03:00
struct page * huge_zero_page __read_mostly ;
2021-06-16 04:23:49 +03:00
unsigned long huge_zero_pfn __read_mostly = ~ 0UL ;
thp: huge zero page: basic preparation
During testing I noticed big (up to 2.5 times) memory consumption overhead
on some workloads (e.g. ft.A from NPB) if THP is enabled.
The main reason for that big difference is lacking zero page in THP case.
We have to allocate a real page on read page fault.
A program to demonstrate the issue:
#include <assert.h>
#include <stdlib.h>
#include <unistd.h>
#define MB 1024*1024
int main(int argc, char **argv)
{
char *p;
int i;
posix_memalign((void **)&p, 2 * MB, 200 * MB);
for (i = 0; i < 200 * MB; i+= 4096)
assert(p[i] == 0);
pause();
return 0;
}
With thp-never RSS is about 400k, but with thp-always it's 200M. After
the patcheset thp-always RSS is 400k too.
Design overview.
Huge zero page (hzp) is a non-movable huge page (2M on x86-64) filled with
zeros. The way how we allocate it changes in the patchset:
- [01/10] simplest way: hzp allocated on boot time in hugepage_init();
- [09/10] lazy allocation on first use;
- [10/10] lockless refcounting + shrinker-reclaimable hzp;
We setup it in do_huge_pmd_anonymous_page() if area around fault address
is suitable for THP and we've got read page fault. If we fail to setup
hzp (ENOMEM) we fallback to handle_pte_fault() as we normally do in THP.
On wp fault to hzp we allocate real memory for the huge page and clear it.
If ENOMEM, graceful fallback: we create a new pmd table and set pte
around fault address to newly allocated normal (4k) page. All other ptes
in the pmd set to normal zero page.
We cannot split hzp (and it's bug if we try), but we can split the pmd
which points to it. On splitting the pmd we create a table with all ptes
set to normal zero page.
===
By hpa's request I've tried alternative approach for hzp implementation
(see Virtual huge zero page patchset): pmd table with all entries set to
zero page. This way should be more cache friendly, but it increases TLB
pressure.
The problem with virtual huge zero page: it requires per-arch enabling.
We need a way to mark that pmd table has all ptes set to zero page.
Some numbers to compare two implementations (on 4s Westmere-EX):
Mirobenchmark1
==============
test:
posix_memalign((void **)&p, 2 * MB, 8 * GB);
for (i = 0; i < 100; i++) {
assert(memcmp(p, p + 4*GB, 4*GB) == 0);
asm volatile ("": : :"memory");
}
hzp:
Performance counter stats for './test_memcmp' (5 runs):
32356.272845 task-clock # 0.998 CPUs utilized ( +- 0.13% )
40 context-switches # 0.001 K/sec ( +- 0.94% )
0 CPU-migrations # 0.000 K/sec
4,218 page-faults # 0.130 K/sec ( +- 0.00% )
76,712,481,765 cycles # 2.371 GHz ( +- 0.13% ) [83.31%]
36,279,577,636 stalled-cycles-frontend # 47.29% frontend cycles idle ( +- 0.28% ) [83.35%]
1,684,049,110 stalled-cycles-backend # 2.20% backend cycles idle ( +- 2.96% ) [66.67%]
134,355,715,816 instructions # 1.75 insns per cycle
# 0.27 stalled cycles per insn ( +- 0.10% ) [83.35%]
13,526,169,702 branches # 418.039 M/sec ( +- 0.10% ) [83.31%]
1,058,230 branch-misses # 0.01% of all branches ( +- 0.91% ) [83.36%]
32.413866442 seconds time elapsed ( +- 0.13% )
vhzp:
Performance counter stats for './test_memcmp' (5 runs):
30327.183829 task-clock # 0.998 CPUs utilized ( +- 0.13% )
38 context-switches # 0.001 K/sec ( +- 1.53% )
0 CPU-migrations # 0.000 K/sec
4,218 page-faults # 0.139 K/sec ( +- 0.01% )
71,964,773,660 cycles # 2.373 GHz ( +- 0.13% ) [83.35%]
31,191,284,231 stalled-cycles-frontend # 43.34% frontend cycles idle ( +- 0.40% ) [83.32%]
773,484,474 stalled-cycles-backend # 1.07% backend cycles idle ( +- 6.61% ) [66.67%]
134,982,215,437 instructions # 1.88 insns per cycle
# 0.23 stalled cycles per insn ( +- 0.11% ) [83.32%]
13,509,150,683 branches # 445.447 M/sec ( +- 0.11% ) [83.34%]
1,017,667 branch-misses # 0.01% of all branches ( +- 1.07% ) [83.32%]
30.381324695 seconds time elapsed ( +- 0.13% )
Mirobenchmark2
==============
test:
posix_memalign((void **)&p, 2 * MB, 8 * GB);
for (i = 0; i < 1000; i++) {
char *_p = p;
while (_p < p+4*GB) {
assert(*_p == *(_p+4*GB));
_p += 4096;
asm volatile ("": : :"memory");
}
}
hzp:
Performance counter stats for 'taskset -c 0 ./test_memcmp2' (5 runs):
3505.727639 task-clock # 0.998 CPUs utilized ( +- 0.26% )
9 context-switches # 0.003 K/sec ( +- 4.97% )
4,384 page-faults # 0.001 M/sec ( +- 0.00% )
8,318,482,466 cycles # 2.373 GHz ( +- 0.26% ) [33.31%]
5,134,318,786 stalled-cycles-frontend # 61.72% frontend cycles idle ( +- 0.42% ) [33.32%]
2,193,266,208 stalled-cycles-backend # 26.37% backend cycles idle ( +- 5.51% ) [33.33%]
9,494,670,537 instructions # 1.14 insns per cycle
# 0.54 stalled cycles per insn ( +- 0.13% ) [41.68%]
2,108,522,738 branches # 601.451 M/sec ( +- 0.09% ) [41.68%]
158,746 branch-misses # 0.01% of all branches ( +- 1.60% ) [41.71%]
3,168,102,115 L1-dcache-loads
# 903.693 M/sec ( +- 0.11% ) [41.70%]
1,048,710,998 L1-dcache-misses
# 33.10% of all L1-dcache hits ( +- 0.11% ) [41.72%]
1,047,699,685 LLC-load
# 298.854 M/sec ( +- 0.03% ) [33.38%]
2,287 LLC-misses
# 0.00% of all LL-cache hits ( +- 8.27% ) [33.37%]
3,166,187,367 dTLB-loads
# 903.147 M/sec ( +- 0.02% ) [33.35%]
4,266,538 dTLB-misses
# 0.13% of all dTLB cache hits ( +- 0.03% ) [33.33%]
3.513339813 seconds time elapsed ( +- 0.26% )
vhzp:
Performance counter stats for 'taskset -c 0 ./test_memcmp2' (5 runs):
27313.891128 task-clock # 0.998 CPUs utilized ( +- 0.24% )
62 context-switches # 0.002 K/sec ( +- 0.61% )
4,384 page-faults # 0.160 K/sec ( +- 0.01% )
64,747,374,606 cycles # 2.370 GHz ( +- 0.24% ) [33.33%]
61,341,580,278 stalled-cycles-frontend # 94.74% frontend cycles idle ( +- 0.26% ) [33.33%]
56,702,237,511 stalled-cycles-backend # 87.57% backend cycles idle ( +- 0.07% ) [33.33%]
10,033,724,846 instructions # 0.15 insns per cycle
# 6.11 stalled cycles per insn ( +- 0.09% ) [41.65%]
2,190,424,932 branches # 80.195 M/sec ( +- 0.12% ) [41.66%]
1,028,630 branch-misses # 0.05% of all branches ( +- 1.50% ) [41.66%]
3,302,006,540 L1-dcache-loads
# 120.891 M/sec ( +- 0.11% ) [41.68%]
271,374,358 L1-dcache-misses
# 8.22% of all L1-dcache hits ( +- 0.04% ) [41.66%]
20,385,476 LLC-load
# 0.746 M/sec ( +- 1.64% ) [33.34%]
76,754 LLC-misses
# 0.38% of all LL-cache hits ( +- 2.35% ) [33.34%]
3,309,927,290 dTLB-loads
# 121.181 M/sec ( +- 0.03% ) [33.34%]
2,098,967,427 dTLB-misses
# 63.41% of all dTLB cache hits ( +- 0.03% ) [33.34%]
27.364448741 seconds time elapsed ( +- 0.24% )
===
I personally prefer implementation present in this patchset. It doesn't
touch arch-specific code.
This patch:
Huge zero page (hzp) is a non-movable huge page (2M on x86-64) filled with
zeros.
For now let's allocate the page on hugepage_init(). We'll switch to lazy
allocation later.
We are not going to map the huge zero page until we can handle it properly
on all code paths.
is_huge_zero_{pfn,pmd}() functions will be used by following patches to
check whether the pfn/pmd is huge zero page.
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@linux.intel.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-12-13 01:50:47 +04:00
2022-07-07 02:59:25 +03:00
bool hugepage_vma_check ( struct vm_area_struct * vma , unsigned long vm_flags ,
bool smaps , bool in_pf , bool enforce_sysfs )
2018-12-28 11:38:21 +03:00
{
2022-06-16 20:48:37 +03:00
if ( ! vma - > vm_mm ) /* vdso */
return false ;
2022-06-16 20:48:38 +03:00
/*
* Explicitly disabled through madvise or prctl , or some
* architectures may disable THP for some mappings , for
* example , s390 kvm .
* */
if ( ( vm_flags & VM_NOHUGEPAGE ) | |
test_bit ( MMF_DISABLE_THP , & vma - > vm_mm - > flags ) )
2022-06-16 20:48:37 +03:00
return false ;
2022-06-16 20:48:38 +03:00
/*
* If the hardware / firmware marked hugepage support disabled .
*/
if ( transparent_hugepage_flags & ( 1 < < TRANSPARENT_HUGEPAGE_NEVER_DAX ) )
mm: thp: fix false negative of shmem vma's THP eligibility
Commit 7635d9cbe832 ("mm, thp, proc: report THP eligibility for each
vma") introduced THPeligible bit for processes' smaps. But, when
checking the eligibility for shmem vma, __transparent_hugepage_enabled()
is called to override the result from shmem_huge_enabled(). It may
result in the anonymous vma's THP flag override shmem's. For example,
running a simple test which create THP for shmem, but with anonymous THP
disabled, when reading the process's smaps, it may show:
7fc92ec00000-7fc92f000000 rw-s 00000000 00:14 27764 /dev/shm/test
Size: 4096 kB
...
[snip]
...
ShmemPmdMapped: 4096 kB
...
[snip]
...
THPeligible: 0
And, /proc/meminfo does show THP allocated and PMD mapped too:
ShmemHugePages: 4096 kB
ShmemPmdMapped: 4096 kB
This doesn't make too much sense. The shmem objects should be treated
separately from anonymous THP. Calling shmem_huge_enabled() with
checking MMF_DISABLE_THP sounds good enough. And, we could skip stack
and dax vma check since we already checked if the vma is shmem already.
Also check if vma is suitable for THP by calling
transhuge_vma_suitable().
And minor fix to smaps output format and documentation.
Link: http://lkml.kernel.org/r/1560401041-32207-3-git-send-email-yang.shi@linux.alibaba.com
Fixes: 7635d9cbe832 ("mm, thp, proc: report THP eligibility for each vma")
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Acked-by: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-19 01:57:27 +03:00
return false ;
2022-06-16 20:48:38 +03:00
/* khugepaged doesn't collapse DAX vma, but page fault is fine. */
2022-06-16 20:48:37 +03:00
if ( vma_is_dax ( vma ) )
2022-06-16 20:48:38 +03:00
return in_pf ;
/*
* Special VMA and hugetlb VMA .
* Must be checked after dax since some dax mappings may have
* VM_MIXEDMAP set .
*/
if ( vm_flags & VM_NO_KHUGEPAGED )
mm: thp: fix false negative of shmem vma's THP eligibility
Commit 7635d9cbe832 ("mm, thp, proc: report THP eligibility for each
vma") introduced THPeligible bit for processes' smaps. But, when
checking the eligibility for shmem vma, __transparent_hugepage_enabled()
is called to override the result from shmem_huge_enabled(). It may
result in the anonymous vma's THP flag override shmem's. For example,
running a simple test which create THP for shmem, but with anonymous THP
disabled, when reading the process's smaps, it may show:
7fc92ec00000-7fc92f000000 rw-s 00000000 00:14 27764 /dev/shm/test
Size: 4096 kB
...
[snip]
...
ShmemPmdMapped: 4096 kB
...
[snip]
...
THPeligible: 0
And, /proc/meminfo does show THP allocated and PMD mapped too:
ShmemHugePages: 4096 kB
ShmemPmdMapped: 4096 kB
This doesn't make too much sense. The shmem objects should be treated
separately from anonymous THP. Calling shmem_huge_enabled() with
checking MMF_DISABLE_THP sounds good enough. And, we could skip stack
and dax vma check since we already checked if the vma is shmem already.
Also check if vma is suitable for THP by calling
transhuge_vma_suitable().
And minor fix to smaps output format and documentation.
Link: http://lkml.kernel.org/r/1560401041-32207-3-git-send-email-yang.shi@linux.alibaba.com
Fixes: 7635d9cbe832 ("mm, thp, proc: report THP eligibility for each vma")
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
Acked-by: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-19 01:57:27 +03:00
return false ;
2022-06-16 20:48:37 +03:00
2022-06-16 20:48:38 +03:00
/*
* Check alignment for file vma and size for both file and anon vma .
*
* Skip the check for page fault . Huge fault does the check in fault
* handlers . And this check is not suitable for huge PUD fault .
*/
if ( ! in_pf & &
! transhuge_vma_suitable ( vma , ( vma - > vm_end - HPAGE_PMD_SIZE ) ) )
2022-06-16 20:48:37 +03:00
return false ;
2022-06-16 20:48:38 +03:00
/*
* Enabled via shmem mount options or sysfs settings .
* Must be done before hugepage flags check since shmem has its
* own flags .
*/
if ( ! in_pf & & shmem_file ( vma - > vm_file ) )
mm/shmem: add flag to enforce shmem THP in hugepage_vma_check()
Patch series "mm: add file/shmem support to MADV_COLLAPSE", v4.
This series builds on top of the previous "mm: userspace hugepage
collapse" series which introduced the MADV_COLLAPSE madvise mode and added
support for private, anonymous mappings[2], by adding support for file and
shmem backed memory to CONFIG_READ_ONLY_THP_FOR_FS=y kernels.
File and shmem support have been added with effort to align with existing
MADV_COLLAPSE semantics and policy decisions[3]. Collapse of shmem-backed
memory ignores kernel-guiding directives and heuristics including all
sysfs settings (transparent_hugepage/shmem_enabled), and tmpfs huge= mount
options (shmem always supports large folios). Like anonymous mappings, on
successful return of MADV_COLLAPSE on file/shmem memory, the contents of
memory mapped by the addresses provided will be synchronously pmd-mapped
THPs.
This functionality unlocks two important uses:
(1) Immediately back executable text by THPs. Current support provided
by CONFIG_READ_ONLY_THP_FOR_FS may take a long time on a large
system which might impair services from serving at their full rated
load after (re)starting. Tricks like mremap(2)'ing text onto
anonymous memory to immediately realize iTLB performance prevents
page sharing and demand paging, both of which increase steady state
memory footprint. Now, we can have the best of both worlds: Peak
upfront performance and lower RAM footprints.
(2) userfaultfd-based live migration of virtual machines satisfy UFFD
faults by fetching native-sized pages over the network (to avoid
latency of transferring an entire hugepage). However, after guest
memory has been fully copied to the new host, MADV_COLLAPSE can
be used to immediately increase guest performance.
khugepaged has received a small improvement by association and can now
detect and collapse pte-mapped THPs. However, there is still work to be
done along the file collapse path. Compound pages of arbitrary order
still needs to be supported and THP collapse needs to be converted to
using folios in general. Eventually, we'd like to move away from the
read-only and executable-mapped constraints currently imposed on eligible
files and support any inode claiming huge folio support. That said, I
think the series as-is covers enough to claim that MADV_COLLAPSE supports
file/shmem memory.
Patches 1-3 Implement the guts of the series.
Patch 4 Is a tracepoint for debugging.
Patches 5-9 Refactor existing khugepaged selftests to work with new
memory types + new collapse tests.
Patch 10 Adds a userfaultfd selftest mode to mimic a functional test
of UFFDIO_REGISTER_MODE_MINOR+MADV_COLLAPSE live migration.
(v4 note: "userfaultfd shmem" selftest is failing as of
Sep 22 mm-unstable)
[1] https://lore.kernel.org/linux-mm/YyiK8YvVcrtZo0z3@google.com/
[2] https://lore.kernel.org/linux-mm/20220706235936.2197195-1-zokeefe@google.com/
[3] https://lore.kernel.org/linux-mm/YtBmhaiPHUTkJml8@google.com/
[4] https://lore.kernel.org/linux-mm/20220922222731.1124481-1-zokeefe@google.com/
[5] https://lore.kernel.org/linux-mm/20220922184651.1016461-1-zokeefe@google.com/
This patch (of 10):
Extend 'mm/thp: add flag to enforce sysfs THP in hugepage_vma_check()' to
shmem, allowing callers to ignore
/sys/kernel/transparent_hugepage/shmem_enabled and tmpfs huge= mount.
This is intended to be used by MADV_COLLAPSE, and the rationale is
analogous to the anon/file case: MADV_COLLAPSE is not coupled to
directives that advise the kernel's decisions on when THPs should be
considered eligible. shmem/tmpfs always claims large folio support,
regardless of sysfs or mount options.
[shy828301@gmail.com: test shmem_huge_force explicitly]
Link: https://lore.kernel.org/linux-mm/CAHbLzko3A5-TpS0BgBeKkx5cuOkWgLvWXQH=TdgW-baO4rPtdg@mail.gmail.com/
Link: https://lkml.kernel.org/r/20220922224046.1143204-1-zokeefe@google.com
Link: https://lkml.kernel.org/r/20220907144521.3115321-2-zokeefe@google.com
Link: https://lkml.kernel.org/r/20220922224046.1143204-2-zokeefe@google.com
Signed-off-by: Zach O'Keefe <zokeefe@google.com>
Reviewed-by: Yang Shi <shy828301@gmail.com>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Chris Kennelly <ckennelly@google.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: James Houghton <jthoughton@google.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Pasha Tatashin <pasha.tatashin@soleen.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rongwei Wang <rongwei.wang@linux.alibaba.com>
Cc: SeongJae Park <sj@kernel.org>
Cc: Song Liu <songliubraving@fb.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-23 01:40:37 +03:00
return shmem_huge_enabled ( vma , ! enforce_sysfs ) ;
2022-06-16 20:48:37 +03:00
2022-07-07 02:59:25 +03:00
/* Enforce sysfs THP requirements as necessary */
if ( enforce_sysfs & &
( ! hugepage_flags_enabled ( ) | | ( ! ( vm_flags & VM_HUGEPAGE ) & &
! hugepage_flags_always ( ) ) ) )
2022-06-16 20:48:37 +03:00
return false ;
/* Only regular file is valid */
2022-06-16 20:48:38 +03:00
if ( ! in_pf & & file_thp_enabled ( vma ) )
2022-05-20 00:08:49 +03:00
return true ;
2018-12-28 11:38:21 +03:00
2022-06-16 20:48:37 +03:00
if ( ! vma_is_anonymous ( vma ) )
return false ;
if ( vma_is_temporary_stack ( vma ) )
return false ;
/*
* THPeligible bit of smaps should show 1 for proper VMAs even
* though anon_vma is not initialized yet .
2022-06-16 20:48:38 +03:00
*
* Allow page fault since anon_vma may be not initialized until
* the first page fault .
2022-06-16 20:48:37 +03:00
*/
if ( ! vma - > anon_vma )
2022-06-16 20:48:38 +03:00
return ( smaps | | in_pf ) ;
2022-06-16 20:48:37 +03:00
return true ;
2018-12-28 11:38:21 +03:00
}
2021-05-05 04:33:55 +03:00
static bool get_huge_zero_page ( void )
2012-12-13 01:51:06 +04:00
{
struct page * zero_page ;
retry :
if ( likely ( atomic_inc_not_zero ( & huge_zero_refcount ) ) )
2021-05-05 04:33:55 +03:00
return true ;
2012-12-13 01:51:06 +04:00
zero_page = alloc_pages ( ( GFP_TRANSHUGE | __GFP_ZERO ) & ~ __GFP_MOVABLE ,
thp: huge zero page: basic preparation
During testing I noticed big (up to 2.5 times) memory consumption overhead
on some workloads (e.g. ft.A from NPB) if THP is enabled.
The main reason for that big difference is lacking zero page in THP case.
We have to allocate a real page on read page fault.
A program to demonstrate the issue:
#include <assert.h>
#include <stdlib.h>
#include <unistd.h>
#define MB 1024*1024
int main(int argc, char **argv)
{
char *p;
int i;
posix_memalign((void **)&p, 2 * MB, 200 * MB);
for (i = 0; i < 200 * MB; i+= 4096)
assert(p[i] == 0);
pause();
return 0;
}
With thp-never RSS is about 400k, but with thp-always it's 200M. After
the patcheset thp-always RSS is 400k too.
Design overview.
Huge zero page (hzp) is a non-movable huge page (2M on x86-64) filled with
zeros. The way how we allocate it changes in the patchset:
- [01/10] simplest way: hzp allocated on boot time in hugepage_init();
- [09/10] lazy allocation on first use;
- [10/10] lockless refcounting + shrinker-reclaimable hzp;
We setup it in do_huge_pmd_anonymous_page() if area around fault address
is suitable for THP and we've got read page fault. If we fail to setup
hzp (ENOMEM) we fallback to handle_pte_fault() as we normally do in THP.
On wp fault to hzp we allocate real memory for the huge page and clear it.
If ENOMEM, graceful fallback: we create a new pmd table and set pte
around fault address to newly allocated normal (4k) page. All other ptes
in the pmd set to normal zero page.
We cannot split hzp (and it's bug if we try), but we can split the pmd
which points to it. On splitting the pmd we create a table with all ptes
set to normal zero page.
===
By hpa's request I've tried alternative approach for hzp implementation
(see Virtual huge zero page patchset): pmd table with all entries set to
zero page. This way should be more cache friendly, but it increases TLB
pressure.
The problem with virtual huge zero page: it requires per-arch enabling.
We need a way to mark that pmd table has all ptes set to zero page.
Some numbers to compare two implementations (on 4s Westmere-EX):
Mirobenchmark1
==============
test:
posix_memalign((void **)&p, 2 * MB, 8 * GB);
for (i = 0; i < 100; i++) {
assert(memcmp(p, p + 4*GB, 4*GB) == 0);
asm volatile ("": : :"memory");
}
hzp:
Performance counter stats for './test_memcmp' (5 runs):
32356.272845 task-clock # 0.998 CPUs utilized ( +- 0.13% )
40 context-switches # 0.001 K/sec ( +- 0.94% )
0 CPU-migrations # 0.000 K/sec
4,218 page-faults # 0.130 K/sec ( +- 0.00% )
76,712,481,765 cycles # 2.371 GHz ( +- 0.13% ) [83.31%]
36,279,577,636 stalled-cycles-frontend # 47.29% frontend cycles idle ( +- 0.28% ) [83.35%]
1,684,049,110 stalled-cycles-backend # 2.20% backend cycles idle ( +- 2.96% ) [66.67%]
134,355,715,816 instructions # 1.75 insns per cycle
# 0.27 stalled cycles per insn ( +- 0.10% ) [83.35%]
13,526,169,702 branches # 418.039 M/sec ( +- 0.10% ) [83.31%]
1,058,230 branch-misses # 0.01% of all branches ( +- 0.91% ) [83.36%]
32.413866442 seconds time elapsed ( +- 0.13% )
vhzp:
Performance counter stats for './test_memcmp' (5 runs):
30327.183829 task-clock # 0.998 CPUs utilized ( +- 0.13% )
38 context-switches # 0.001 K/sec ( +- 1.53% )
0 CPU-migrations # 0.000 K/sec
4,218 page-faults # 0.139 K/sec ( +- 0.01% )
71,964,773,660 cycles # 2.373 GHz ( +- 0.13% ) [83.35%]
31,191,284,231 stalled-cycles-frontend # 43.34% frontend cycles idle ( +- 0.40% ) [83.32%]
773,484,474 stalled-cycles-backend # 1.07% backend cycles idle ( +- 6.61% ) [66.67%]
134,982,215,437 instructions # 1.88 insns per cycle
# 0.23 stalled cycles per insn ( +- 0.11% ) [83.32%]
13,509,150,683 branches # 445.447 M/sec ( +- 0.11% ) [83.34%]
1,017,667 branch-misses # 0.01% of all branches ( +- 1.07% ) [83.32%]
30.381324695 seconds time elapsed ( +- 0.13% )
Mirobenchmark2
==============
test:
posix_memalign((void **)&p, 2 * MB, 8 * GB);
for (i = 0; i < 1000; i++) {
char *_p = p;
while (_p < p+4*GB) {
assert(*_p == *(_p+4*GB));
_p += 4096;
asm volatile ("": : :"memory");
}
}
hzp:
Performance counter stats for 'taskset -c 0 ./test_memcmp2' (5 runs):
3505.727639 task-clock # 0.998 CPUs utilized ( +- 0.26% )
9 context-switches # 0.003 K/sec ( +- 4.97% )
4,384 page-faults # 0.001 M/sec ( +- 0.00% )
8,318,482,466 cycles # 2.373 GHz ( +- 0.26% ) [33.31%]
5,134,318,786 stalled-cycles-frontend # 61.72% frontend cycles idle ( +- 0.42% ) [33.32%]
2,193,266,208 stalled-cycles-backend # 26.37% backend cycles idle ( +- 5.51% ) [33.33%]
9,494,670,537 instructions # 1.14 insns per cycle
# 0.54 stalled cycles per insn ( +- 0.13% ) [41.68%]
2,108,522,738 branches # 601.451 M/sec ( +- 0.09% ) [41.68%]
158,746 branch-misses # 0.01% of all branches ( +- 1.60% ) [41.71%]
3,168,102,115 L1-dcache-loads
# 903.693 M/sec ( +- 0.11% ) [41.70%]
1,048,710,998 L1-dcache-misses
# 33.10% of all L1-dcache hits ( +- 0.11% ) [41.72%]
1,047,699,685 LLC-load
# 298.854 M/sec ( +- 0.03% ) [33.38%]
2,287 LLC-misses
# 0.00% of all LL-cache hits ( +- 8.27% ) [33.37%]
3,166,187,367 dTLB-loads
# 903.147 M/sec ( +- 0.02% ) [33.35%]
4,266,538 dTLB-misses
# 0.13% of all dTLB cache hits ( +- 0.03% ) [33.33%]
3.513339813 seconds time elapsed ( +- 0.26% )
vhzp:
Performance counter stats for 'taskset -c 0 ./test_memcmp2' (5 runs):
27313.891128 task-clock # 0.998 CPUs utilized ( +- 0.24% )
62 context-switches # 0.002 K/sec ( +- 0.61% )
4,384 page-faults # 0.160 K/sec ( +- 0.01% )
64,747,374,606 cycles # 2.370 GHz ( +- 0.24% ) [33.33%]
61,341,580,278 stalled-cycles-frontend # 94.74% frontend cycles idle ( +- 0.26% ) [33.33%]
56,702,237,511 stalled-cycles-backend # 87.57% backend cycles idle ( +- 0.07% ) [33.33%]
10,033,724,846 instructions # 0.15 insns per cycle
# 6.11 stalled cycles per insn ( +- 0.09% ) [41.65%]
2,190,424,932 branches # 80.195 M/sec ( +- 0.12% ) [41.66%]
1,028,630 branch-misses # 0.05% of all branches ( +- 1.50% ) [41.66%]
3,302,006,540 L1-dcache-loads
# 120.891 M/sec ( +- 0.11% ) [41.68%]
271,374,358 L1-dcache-misses
# 8.22% of all L1-dcache hits ( +- 0.04% ) [41.66%]
20,385,476 LLC-load
# 0.746 M/sec ( +- 1.64% ) [33.34%]
76,754 LLC-misses
# 0.38% of all LL-cache hits ( +- 2.35% ) [33.34%]
3,309,927,290 dTLB-loads
# 121.181 M/sec ( +- 0.03% ) [33.34%]
2,098,967,427 dTLB-misses
# 63.41% of all dTLB cache hits ( +- 0.03% ) [33.34%]
27.364448741 seconds time elapsed ( +- 0.24% )
===
I personally prefer implementation present in this patchset. It doesn't
touch arch-specific code.
This patch:
Huge zero page (hzp) is a non-movable huge page (2M on x86-64) filled with
zeros.
For now let's allocate the page on hugepage_init(). We'll switch to lazy
allocation later.
We are not going to map the huge zero page until we can handle it properly
on all code paths.
is_huge_zero_{pfn,pmd}() functions will be used by following patches to
check whether the pfn/pmd is huge zero page.
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@linux.intel.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-12-13 01:50:47 +04:00
HPAGE_PMD_ORDER ) ;
2012-12-13 01:51:09 +04:00
if ( ! zero_page ) {
count_vm_event ( THP_ZERO_PAGE_ALLOC_FAILED ) ;
2021-05-05 04:33:55 +03:00
return false ;
2012-12-13 01:51:09 +04:00
}
2012-12-13 01:51:06 +04:00
preempt_disable ( ) ;
2013-04-30 02:08:44 +04:00
if ( cmpxchg ( & huge_zero_page , NULL , zero_page ) ) {
2012-12-13 01:51:06 +04:00
preempt_enable ( ) ;
2014-10-30 00:50:26 +03:00
__free_pages ( zero_page , compound_order ( zero_page ) ) ;
2012-12-13 01:51:06 +04:00
goto retry ;
}
2021-06-16 04:23:49 +03:00
WRITE_ONCE ( huge_zero_pfn , page_to_pfn ( zero_page ) ) ;
2012-12-13 01:51:06 +04:00
/* We take additional reference here. It will be put back by shrinker */
atomic_set ( & huge_zero_refcount , 2 ) ;
preempt_enable ( ) ;
2022-09-09 05:16:53 +03:00
count_vm_event ( THP_ZERO_PAGE_ALLOC ) ;
2021-05-05 04:33:55 +03:00
return true ;
thp: huge zero page: basic preparation
During testing I noticed big (up to 2.5 times) memory consumption overhead
on some workloads (e.g. ft.A from NPB) if THP is enabled.
The main reason for that big difference is lacking zero page in THP case.
We have to allocate a real page on read page fault.
A program to demonstrate the issue:
#include <assert.h>
#include <stdlib.h>
#include <unistd.h>
#define MB 1024*1024
int main(int argc, char **argv)
{
char *p;
int i;
posix_memalign((void **)&p, 2 * MB, 200 * MB);
for (i = 0; i < 200 * MB; i+= 4096)
assert(p[i] == 0);
pause();
return 0;
}
With thp-never RSS is about 400k, but with thp-always it's 200M. After
the patcheset thp-always RSS is 400k too.
Design overview.
Huge zero page (hzp) is a non-movable huge page (2M on x86-64) filled with
zeros. The way how we allocate it changes in the patchset:
- [01/10] simplest way: hzp allocated on boot time in hugepage_init();
- [09/10] lazy allocation on first use;
- [10/10] lockless refcounting + shrinker-reclaimable hzp;
We setup it in do_huge_pmd_anonymous_page() if area around fault address
is suitable for THP and we've got read page fault. If we fail to setup
hzp (ENOMEM) we fallback to handle_pte_fault() as we normally do in THP.
On wp fault to hzp we allocate real memory for the huge page and clear it.
If ENOMEM, graceful fallback: we create a new pmd table and set pte
around fault address to newly allocated normal (4k) page. All other ptes
in the pmd set to normal zero page.
We cannot split hzp (and it's bug if we try), but we can split the pmd
which points to it. On splitting the pmd we create a table with all ptes
set to normal zero page.
===
By hpa's request I've tried alternative approach for hzp implementation
(see Virtual huge zero page patchset): pmd table with all entries set to
zero page. This way should be more cache friendly, but it increases TLB
pressure.
The problem with virtual huge zero page: it requires per-arch enabling.
We need a way to mark that pmd table has all ptes set to zero page.
Some numbers to compare two implementations (on 4s Westmere-EX):
Mirobenchmark1
==============
test:
posix_memalign((void **)&p, 2 * MB, 8 * GB);
for (i = 0; i < 100; i++) {
assert(memcmp(p, p + 4*GB, 4*GB) == 0);
asm volatile ("": : :"memory");
}
hzp:
Performance counter stats for './test_memcmp' (5 runs):
32356.272845 task-clock # 0.998 CPUs utilized ( +- 0.13% )
40 context-switches # 0.001 K/sec ( +- 0.94% )
0 CPU-migrations # 0.000 K/sec
4,218 page-faults # 0.130 K/sec ( +- 0.00% )
76,712,481,765 cycles # 2.371 GHz ( +- 0.13% ) [83.31%]
36,279,577,636 stalled-cycles-frontend # 47.29% frontend cycles idle ( +- 0.28% ) [83.35%]
1,684,049,110 stalled-cycles-backend # 2.20% backend cycles idle ( +- 2.96% ) [66.67%]
134,355,715,816 instructions # 1.75 insns per cycle
# 0.27 stalled cycles per insn ( +- 0.10% ) [83.35%]
13,526,169,702 branches # 418.039 M/sec ( +- 0.10% ) [83.31%]
1,058,230 branch-misses # 0.01% of all branches ( +- 0.91% ) [83.36%]
32.413866442 seconds time elapsed ( +- 0.13% )
vhzp:
Performance counter stats for './test_memcmp' (5 runs):
30327.183829 task-clock # 0.998 CPUs utilized ( +- 0.13% )
38 context-switches # 0.001 K/sec ( +- 1.53% )
0 CPU-migrations # 0.000 K/sec
4,218 page-faults # 0.139 K/sec ( +- 0.01% )
71,964,773,660 cycles # 2.373 GHz ( +- 0.13% ) [83.35%]
31,191,284,231 stalled-cycles-frontend # 43.34% frontend cycles idle ( +- 0.40% ) [83.32%]
773,484,474 stalled-cycles-backend # 1.07% backend cycles idle ( +- 6.61% ) [66.67%]
134,982,215,437 instructions # 1.88 insns per cycle
# 0.23 stalled cycles per insn ( +- 0.11% ) [83.32%]
13,509,150,683 branches # 445.447 M/sec ( +- 0.11% ) [83.34%]
1,017,667 branch-misses # 0.01% of all branches ( +- 1.07% ) [83.32%]
30.381324695 seconds time elapsed ( +- 0.13% )
Mirobenchmark2
==============
test:
posix_memalign((void **)&p, 2 * MB, 8 * GB);
for (i = 0; i < 1000; i++) {
char *_p = p;
while (_p < p+4*GB) {
assert(*_p == *(_p+4*GB));
_p += 4096;
asm volatile ("": : :"memory");
}
}
hzp:
Performance counter stats for 'taskset -c 0 ./test_memcmp2' (5 runs):
3505.727639 task-clock # 0.998 CPUs utilized ( +- 0.26% )
9 context-switches # 0.003 K/sec ( +- 4.97% )
4,384 page-faults # 0.001 M/sec ( +- 0.00% )
8,318,482,466 cycles # 2.373 GHz ( +- 0.26% ) [33.31%]
5,134,318,786 stalled-cycles-frontend # 61.72% frontend cycles idle ( +- 0.42% ) [33.32%]
2,193,266,208 stalled-cycles-backend # 26.37% backend cycles idle ( +- 5.51% ) [33.33%]
9,494,670,537 instructions # 1.14 insns per cycle
# 0.54 stalled cycles per insn ( +- 0.13% ) [41.68%]
2,108,522,738 branches # 601.451 M/sec ( +- 0.09% ) [41.68%]
158,746 branch-misses # 0.01% of all branches ( +- 1.60% ) [41.71%]
3,168,102,115 L1-dcache-loads
# 903.693 M/sec ( +- 0.11% ) [41.70%]
1,048,710,998 L1-dcache-misses
# 33.10% of all L1-dcache hits ( +- 0.11% ) [41.72%]
1,047,699,685 LLC-load
# 298.854 M/sec ( +- 0.03% ) [33.38%]
2,287 LLC-misses
# 0.00% of all LL-cache hits ( +- 8.27% ) [33.37%]
3,166,187,367 dTLB-loads
# 903.147 M/sec ( +- 0.02% ) [33.35%]
4,266,538 dTLB-misses
# 0.13% of all dTLB cache hits ( +- 0.03% ) [33.33%]
3.513339813 seconds time elapsed ( +- 0.26% )
vhzp:
Performance counter stats for 'taskset -c 0 ./test_memcmp2' (5 runs):
27313.891128 task-clock # 0.998 CPUs utilized ( +- 0.24% )
62 context-switches # 0.002 K/sec ( +- 0.61% )
4,384 page-faults # 0.160 K/sec ( +- 0.01% )
64,747,374,606 cycles # 2.370 GHz ( +- 0.24% ) [33.33%]
61,341,580,278 stalled-cycles-frontend # 94.74% frontend cycles idle ( +- 0.26% ) [33.33%]
56,702,237,511 stalled-cycles-backend # 87.57% backend cycles idle ( +- 0.07% ) [33.33%]
10,033,724,846 instructions # 0.15 insns per cycle
# 6.11 stalled cycles per insn ( +- 0.09% ) [41.65%]
2,190,424,932 branches # 80.195 M/sec ( +- 0.12% ) [41.66%]
1,028,630 branch-misses # 0.05% of all branches ( +- 1.50% ) [41.66%]
3,302,006,540 L1-dcache-loads
# 120.891 M/sec ( +- 0.11% ) [41.68%]
271,374,358 L1-dcache-misses
# 8.22% of all L1-dcache hits ( +- 0.04% ) [41.66%]
20,385,476 LLC-load
# 0.746 M/sec ( +- 1.64% ) [33.34%]
76,754 LLC-misses
# 0.38% of all LL-cache hits ( +- 2.35% ) [33.34%]
3,309,927,290 dTLB-loads
# 121.181 M/sec ( +- 0.03% ) [33.34%]
2,098,967,427 dTLB-misses
# 63.41% of all dTLB cache hits ( +- 0.03% ) [33.34%]
27.364448741 seconds time elapsed ( +- 0.24% )
===
I personally prefer implementation present in this patchset. It doesn't
touch arch-specific code.
This patch:
Huge zero page (hzp) is a non-movable huge page (2M on x86-64) filled with
zeros.
For now let's allocate the page on hugepage_init(). We'll switch to lazy
allocation later.
We are not going to map the huge zero page until we can handle it properly
on all code paths.
is_huge_zero_{pfn,pmd}() functions will be used by following patches to
check whether the pfn/pmd is huge zero page.
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@linux.intel.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-12-13 01:50:47 +04:00
}
2016-10-08 03:00:08 +03:00
static void put_huge_zero_page ( void )
thp: huge zero page: basic preparation
During testing I noticed big (up to 2.5 times) memory consumption overhead
on some workloads (e.g. ft.A from NPB) if THP is enabled.
The main reason for that big difference is lacking zero page in THP case.
We have to allocate a real page on read page fault.
A program to demonstrate the issue:
#include <assert.h>
#include <stdlib.h>
#include <unistd.h>
#define MB 1024*1024
int main(int argc, char **argv)
{
char *p;
int i;
posix_memalign((void **)&p, 2 * MB, 200 * MB);
for (i = 0; i < 200 * MB; i+= 4096)
assert(p[i] == 0);
pause();
return 0;
}
With thp-never RSS is about 400k, but with thp-always it's 200M. After
the patcheset thp-always RSS is 400k too.
Design overview.
Huge zero page (hzp) is a non-movable huge page (2M on x86-64) filled with
zeros. The way how we allocate it changes in the patchset:
- [01/10] simplest way: hzp allocated on boot time in hugepage_init();
- [09/10] lazy allocation on first use;
- [10/10] lockless refcounting + shrinker-reclaimable hzp;
We setup it in do_huge_pmd_anonymous_page() if area around fault address
is suitable for THP and we've got read page fault. If we fail to setup
hzp (ENOMEM) we fallback to handle_pte_fault() as we normally do in THP.
On wp fault to hzp we allocate real memory for the huge page and clear it.
If ENOMEM, graceful fallback: we create a new pmd table and set pte
around fault address to newly allocated normal (4k) page. All other ptes
in the pmd set to normal zero page.
We cannot split hzp (and it's bug if we try), but we can split the pmd
which points to it. On splitting the pmd we create a table with all ptes
set to normal zero page.
===
By hpa's request I've tried alternative approach for hzp implementation
(see Virtual huge zero page patchset): pmd table with all entries set to
zero page. This way should be more cache friendly, but it increases TLB
pressure.
The problem with virtual huge zero page: it requires per-arch enabling.
We need a way to mark that pmd table has all ptes set to zero page.
Some numbers to compare two implementations (on 4s Westmere-EX):
Mirobenchmark1
==============
test:
posix_memalign((void **)&p, 2 * MB, 8 * GB);
for (i = 0; i < 100; i++) {
assert(memcmp(p, p + 4*GB, 4*GB) == 0);
asm volatile ("": : :"memory");
}
hzp:
Performance counter stats for './test_memcmp' (5 runs):
32356.272845 task-clock # 0.998 CPUs utilized ( +- 0.13% )
40 context-switches # 0.001 K/sec ( +- 0.94% )
0 CPU-migrations # 0.000 K/sec
4,218 page-faults # 0.130 K/sec ( +- 0.00% )
76,712,481,765 cycles # 2.371 GHz ( +- 0.13% ) [83.31%]
36,279,577,636 stalled-cycles-frontend # 47.29% frontend cycles idle ( +- 0.28% ) [83.35%]
1,684,049,110 stalled-cycles-backend # 2.20% backend cycles idle ( +- 2.96% ) [66.67%]
134,355,715,816 instructions # 1.75 insns per cycle
# 0.27 stalled cycles per insn ( +- 0.10% ) [83.35%]
13,526,169,702 branches # 418.039 M/sec ( +- 0.10% ) [83.31%]
1,058,230 branch-misses # 0.01% of all branches ( +- 0.91% ) [83.36%]
32.413866442 seconds time elapsed ( +- 0.13% )
vhzp:
Performance counter stats for './test_memcmp' (5 runs):
30327.183829 task-clock # 0.998 CPUs utilized ( +- 0.13% )
38 context-switches # 0.001 K/sec ( +- 1.53% )
0 CPU-migrations # 0.000 K/sec
4,218 page-faults # 0.139 K/sec ( +- 0.01% )
71,964,773,660 cycles # 2.373 GHz ( +- 0.13% ) [83.35%]
31,191,284,231 stalled-cycles-frontend # 43.34% frontend cycles idle ( +- 0.40% ) [83.32%]
773,484,474 stalled-cycles-backend # 1.07% backend cycles idle ( +- 6.61% ) [66.67%]
134,982,215,437 instructions # 1.88 insns per cycle
# 0.23 stalled cycles per insn ( +- 0.11% ) [83.32%]
13,509,150,683 branches # 445.447 M/sec ( +- 0.11% ) [83.34%]
1,017,667 branch-misses # 0.01% of all branches ( +- 1.07% ) [83.32%]
30.381324695 seconds time elapsed ( +- 0.13% )
Mirobenchmark2
==============
test:
posix_memalign((void **)&p, 2 * MB, 8 * GB);
for (i = 0; i < 1000; i++) {
char *_p = p;
while (_p < p+4*GB) {
assert(*_p == *(_p+4*GB));
_p += 4096;
asm volatile ("": : :"memory");
}
}
hzp:
Performance counter stats for 'taskset -c 0 ./test_memcmp2' (5 runs):
3505.727639 task-clock # 0.998 CPUs utilized ( +- 0.26% )
9 context-switches # 0.003 K/sec ( +- 4.97% )
4,384 page-faults # 0.001 M/sec ( +- 0.00% )
8,318,482,466 cycles # 2.373 GHz ( +- 0.26% ) [33.31%]
5,134,318,786 stalled-cycles-frontend # 61.72% frontend cycles idle ( +- 0.42% ) [33.32%]
2,193,266,208 stalled-cycles-backend # 26.37% backend cycles idle ( +- 5.51% ) [33.33%]
9,494,670,537 instructions # 1.14 insns per cycle
# 0.54 stalled cycles per insn ( +- 0.13% ) [41.68%]
2,108,522,738 branches # 601.451 M/sec ( +- 0.09% ) [41.68%]
158,746 branch-misses # 0.01% of all branches ( +- 1.60% ) [41.71%]
3,168,102,115 L1-dcache-loads
# 903.693 M/sec ( +- 0.11% ) [41.70%]
1,048,710,998 L1-dcache-misses
# 33.10% of all L1-dcache hits ( +- 0.11% ) [41.72%]
1,047,699,685 LLC-load
# 298.854 M/sec ( +- 0.03% ) [33.38%]
2,287 LLC-misses
# 0.00% of all LL-cache hits ( +- 8.27% ) [33.37%]
3,166,187,367 dTLB-loads
# 903.147 M/sec ( +- 0.02% ) [33.35%]
4,266,538 dTLB-misses
# 0.13% of all dTLB cache hits ( +- 0.03% ) [33.33%]
3.513339813 seconds time elapsed ( +- 0.26% )
vhzp:
Performance counter stats for 'taskset -c 0 ./test_memcmp2' (5 runs):
27313.891128 task-clock # 0.998 CPUs utilized ( +- 0.24% )
62 context-switches # 0.002 K/sec ( +- 0.61% )
4,384 page-faults # 0.160 K/sec ( +- 0.01% )
64,747,374,606 cycles # 2.370 GHz ( +- 0.24% ) [33.33%]
61,341,580,278 stalled-cycles-frontend # 94.74% frontend cycles idle ( +- 0.26% ) [33.33%]
56,702,237,511 stalled-cycles-backend # 87.57% backend cycles idle ( +- 0.07% ) [33.33%]
10,033,724,846 instructions # 0.15 insns per cycle
# 6.11 stalled cycles per insn ( +- 0.09% ) [41.65%]
2,190,424,932 branches # 80.195 M/sec ( +- 0.12% ) [41.66%]
1,028,630 branch-misses # 0.05% of all branches ( +- 1.50% ) [41.66%]
3,302,006,540 L1-dcache-loads
# 120.891 M/sec ( +- 0.11% ) [41.68%]
271,374,358 L1-dcache-misses
# 8.22% of all L1-dcache hits ( +- 0.04% ) [41.66%]
20,385,476 LLC-load
# 0.746 M/sec ( +- 1.64% ) [33.34%]
76,754 LLC-misses
# 0.38% of all LL-cache hits ( +- 2.35% ) [33.34%]
3,309,927,290 dTLB-loads
# 121.181 M/sec ( +- 0.03% ) [33.34%]
2,098,967,427 dTLB-misses
# 63.41% of all dTLB cache hits ( +- 0.03% ) [33.34%]
27.364448741 seconds time elapsed ( +- 0.24% )
===
I personally prefer implementation present in this patchset. It doesn't
touch arch-specific code.
This patch:
Huge zero page (hzp) is a non-movable huge page (2M on x86-64) filled with
zeros.
For now let's allocate the page on hugepage_init(). We'll switch to lazy
allocation later.
We are not going to map the huge zero page until we can handle it properly
on all code paths.
is_huge_zero_{pfn,pmd}() functions will be used by following patches to
check whether the pfn/pmd is huge zero page.
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@linux.intel.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-12-13 01:50:47 +04:00
{
2012-12-13 01:51:06 +04:00
/*
* Counter should never go to zero here . Only shrinker can put
* last reference .
*/
BUG_ON ( atomic_dec_and_test ( & huge_zero_refcount ) ) ;
thp: huge zero page: basic preparation
During testing I noticed big (up to 2.5 times) memory consumption overhead
on some workloads (e.g. ft.A from NPB) if THP is enabled.
The main reason for that big difference is lacking zero page in THP case.
We have to allocate a real page on read page fault.
A program to demonstrate the issue:
#include <assert.h>
#include <stdlib.h>
#include <unistd.h>
#define MB 1024*1024
int main(int argc, char **argv)
{
char *p;
int i;
posix_memalign((void **)&p, 2 * MB, 200 * MB);
for (i = 0; i < 200 * MB; i+= 4096)
assert(p[i] == 0);
pause();
return 0;
}
With thp-never RSS is about 400k, but with thp-always it's 200M. After
the patcheset thp-always RSS is 400k too.
Design overview.
Huge zero page (hzp) is a non-movable huge page (2M on x86-64) filled with
zeros. The way how we allocate it changes in the patchset:
- [01/10] simplest way: hzp allocated on boot time in hugepage_init();
- [09/10] lazy allocation on first use;
- [10/10] lockless refcounting + shrinker-reclaimable hzp;
We setup it in do_huge_pmd_anonymous_page() if area around fault address
is suitable for THP and we've got read page fault. If we fail to setup
hzp (ENOMEM) we fallback to handle_pte_fault() as we normally do in THP.
On wp fault to hzp we allocate real memory for the huge page and clear it.
If ENOMEM, graceful fallback: we create a new pmd table and set pte
around fault address to newly allocated normal (4k) page. All other ptes
in the pmd set to normal zero page.
We cannot split hzp (and it's bug if we try), but we can split the pmd
which points to it. On splitting the pmd we create a table with all ptes
set to normal zero page.
===
By hpa's request I've tried alternative approach for hzp implementation
(see Virtual huge zero page patchset): pmd table with all entries set to
zero page. This way should be more cache friendly, but it increases TLB
pressure.
The problem with virtual huge zero page: it requires per-arch enabling.
We need a way to mark that pmd table has all ptes set to zero page.
Some numbers to compare two implementations (on 4s Westmere-EX):
Mirobenchmark1
==============
test:
posix_memalign((void **)&p, 2 * MB, 8 * GB);
for (i = 0; i < 100; i++) {
assert(memcmp(p, p + 4*GB, 4*GB) == 0);
asm volatile ("": : :"memory");
}
hzp:
Performance counter stats for './test_memcmp' (5 runs):
32356.272845 task-clock # 0.998 CPUs utilized ( +- 0.13% )
40 context-switches # 0.001 K/sec ( +- 0.94% )
0 CPU-migrations # 0.000 K/sec
4,218 page-faults # 0.130 K/sec ( +- 0.00% )
76,712,481,765 cycles # 2.371 GHz ( +- 0.13% ) [83.31%]
36,279,577,636 stalled-cycles-frontend # 47.29% frontend cycles idle ( +- 0.28% ) [83.35%]
1,684,049,110 stalled-cycles-backend # 2.20% backend cycles idle ( +- 2.96% ) [66.67%]
134,355,715,816 instructions # 1.75 insns per cycle
# 0.27 stalled cycles per insn ( +- 0.10% ) [83.35%]
13,526,169,702 branches # 418.039 M/sec ( +- 0.10% ) [83.31%]
1,058,230 branch-misses # 0.01% of all branches ( +- 0.91% ) [83.36%]
32.413866442 seconds time elapsed ( +- 0.13% )
vhzp:
Performance counter stats for './test_memcmp' (5 runs):
30327.183829 task-clock # 0.998 CPUs utilized ( +- 0.13% )
38 context-switches # 0.001 K/sec ( +- 1.53% )
0 CPU-migrations # 0.000 K/sec
4,218 page-faults # 0.139 K/sec ( +- 0.01% )
71,964,773,660 cycles # 2.373 GHz ( +- 0.13% ) [83.35%]
31,191,284,231 stalled-cycles-frontend # 43.34% frontend cycles idle ( +- 0.40% ) [83.32%]
773,484,474 stalled-cycles-backend # 1.07% backend cycles idle ( +- 6.61% ) [66.67%]
134,982,215,437 instructions # 1.88 insns per cycle
# 0.23 stalled cycles per insn ( +- 0.11% ) [83.32%]
13,509,150,683 branches # 445.447 M/sec ( +- 0.11% ) [83.34%]
1,017,667 branch-misses # 0.01% of all branches ( +- 1.07% ) [83.32%]
30.381324695 seconds time elapsed ( +- 0.13% )
Mirobenchmark2
==============
test:
posix_memalign((void **)&p, 2 * MB, 8 * GB);
for (i = 0; i < 1000; i++) {
char *_p = p;
while (_p < p+4*GB) {
assert(*_p == *(_p+4*GB));
_p += 4096;
asm volatile ("": : :"memory");
}
}
hzp:
Performance counter stats for 'taskset -c 0 ./test_memcmp2' (5 runs):
3505.727639 task-clock # 0.998 CPUs utilized ( +- 0.26% )
9 context-switches # 0.003 K/sec ( +- 4.97% )
4,384 page-faults # 0.001 M/sec ( +- 0.00% )
8,318,482,466 cycles # 2.373 GHz ( +- 0.26% ) [33.31%]
5,134,318,786 stalled-cycles-frontend # 61.72% frontend cycles idle ( +- 0.42% ) [33.32%]
2,193,266,208 stalled-cycles-backend # 26.37% backend cycles idle ( +- 5.51% ) [33.33%]
9,494,670,537 instructions # 1.14 insns per cycle
# 0.54 stalled cycles per insn ( +- 0.13% ) [41.68%]
2,108,522,738 branches # 601.451 M/sec ( +- 0.09% ) [41.68%]
158,746 branch-misses # 0.01% of all branches ( +- 1.60% ) [41.71%]
3,168,102,115 L1-dcache-loads
# 903.693 M/sec ( +- 0.11% ) [41.70%]
1,048,710,998 L1-dcache-misses
# 33.10% of all L1-dcache hits ( +- 0.11% ) [41.72%]
1,047,699,685 LLC-load
# 298.854 M/sec ( +- 0.03% ) [33.38%]
2,287 LLC-misses
# 0.00% of all LL-cache hits ( +- 8.27% ) [33.37%]
3,166,187,367 dTLB-loads
# 903.147 M/sec ( +- 0.02% ) [33.35%]
4,266,538 dTLB-misses
# 0.13% of all dTLB cache hits ( +- 0.03% ) [33.33%]
3.513339813 seconds time elapsed ( +- 0.26% )
vhzp:
Performance counter stats for 'taskset -c 0 ./test_memcmp2' (5 runs):
27313.891128 task-clock # 0.998 CPUs utilized ( +- 0.24% )
62 context-switches # 0.002 K/sec ( +- 0.61% )
4,384 page-faults # 0.160 K/sec ( +- 0.01% )
64,747,374,606 cycles # 2.370 GHz ( +- 0.24% ) [33.33%]
61,341,580,278 stalled-cycles-frontend # 94.74% frontend cycles idle ( +- 0.26% ) [33.33%]
56,702,237,511 stalled-cycles-backend # 87.57% backend cycles idle ( +- 0.07% ) [33.33%]
10,033,724,846 instructions # 0.15 insns per cycle
# 6.11 stalled cycles per insn ( +- 0.09% ) [41.65%]
2,190,424,932 branches # 80.195 M/sec ( +- 0.12% ) [41.66%]
1,028,630 branch-misses # 0.05% of all branches ( +- 1.50% ) [41.66%]
3,302,006,540 L1-dcache-loads
# 120.891 M/sec ( +- 0.11% ) [41.68%]
271,374,358 L1-dcache-misses
# 8.22% of all L1-dcache hits ( +- 0.04% ) [41.66%]
20,385,476 LLC-load
# 0.746 M/sec ( +- 1.64% ) [33.34%]
76,754 LLC-misses
# 0.38% of all LL-cache hits ( +- 2.35% ) [33.34%]
3,309,927,290 dTLB-loads
# 121.181 M/sec ( +- 0.03% ) [33.34%]
2,098,967,427 dTLB-misses
# 63.41% of all dTLB cache hits ( +- 0.03% ) [33.34%]
27.364448741 seconds time elapsed ( +- 0.24% )
===
I personally prefer implementation present in this patchset. It doesn't
touch arch-specific code.
This patch:
Huge zero page (hzp) is a non-movable huge page (2M on x86-64) filled with
zeros.
For now let's allocate the page on hugepage_init(). We'll switch to lazy
allocation later.
We are not going to map the huge zero page until we can handle it properly
on all code paths.
is_huge_zero_{pfn,pmd}() functions will be used by following patches to
check whether the pfn/pmd is huge zero page.
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@linux.intel.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-12-13 01:50:47 +04:00
}
2016-10-08 03:00:08 +03:00
struct page * mm_get_huge_zero_page ( struct mm_struct * mm )
{
if ( test_bit ( MMF_HUGE_ZERO_PAGE , & mm - > flags ) )
return READ_ONCE ( huge_zero_page ) ;
if ( ! get_huge_zero_page ( ) )
return NULL ;
if ( test_and_set_bit ( MMF_HUGE_ZERO_PAGE , & mm - > flags ) )
put_huge_zero_page ( ) ;
return READ_ONCE ( huge_zero_page ) ;
}
void mm_put_huge_zero_page ( struct mm_struct * mm )
{
if ( test_bit ( MMF_HUGE_ZERO_PAGE , & mm - > flags ) )
put_huge_zero_page ( ) ;
}
2013-08-28 04:18:15 +04:00
static unsigned long shrink_huge_zero_page_count ( struct shrinker * shrink ,
struct shrink_control * sc )
thp: huge zero page: basic preparation
During testing I noticed big (up to 2.5 times) memory consumption overhead
on some workloads (e.g. ft.A from NPB) if THP is enabled.
The main reason for that big difference is lacking zero page in THP case.
We have to allocate a real page on read page fault.
A program to demonstrate the issue:
#include <assert.h>
#include <stdlib.h>
#include <unistd.h>
#define MB 1024*1024
int main(int argc, char **argv)
{
char *p;
int i;
posix_memalign((void **)&p, 2 * MB, 200 * MB);
for (i = 0; i < 200 * MB; i+= 4096)
assert(p[i] == 0);
pause();
return 0;
}
With thp-never RSS is about 400k, but with thp-always it's 200M. After
the patcheset thp-always RSS is 400k too.
Design overview.
Huge zero page (hzp) is a non-movable huge page (2M on x86-64) filled with
zeros. The way how we allocate it changes in the patchset:
- [01/10] simplest way: hzp allocated on boot time in hugepage_init();
- [09/10] lazy allocation on first use;
- [10/10] lockless refcounting + shrinker-reclaimable hzp;
We setup it in do_huge_pmd_anonymous_page() if area around fault address
is suitable for THP and we've got read page fault. If we fail to setup
hzp (ENOMEM) we fallback to handle_pte_fault() as we normally do in THP.
On wp fault to hzp we allocate real memory for the huge page and clear it.
If ENOMEM, graceful fallback: we create a new pmd table and set pte
around fault address to newly allocated normal (4k) page. All other ptes
in the pmd set to normal zero page.
We cannot split hzp (and it's bug if we try), but we can split the pmd
which points to it. On splitting the pmd we create a table with all ptes
set to normal zero page.
===
By hpa's request I've tried alternative approach for hzp implementation
(see Virtual huge zero page patchset): pmd table with all entries set to
zero page. This way should be more cache friendly, but it increases TLB
pressure.
The problem with virtual huge zero page: it requires per-arch enabling.
We need a way to mark that pmd table has all ptes set to zero page.
Some numbers to compare two implementations (on 4s Westmere-EX):
Mirobenchmark1
==============
test:
posix_memalign((void **)&p, 2 * MB, 8 * GB);
for (i = 0; i < 100; i++) {
assert(memcmp(p, p + 4*GB, 4*GB) == 0);
asm volatile ("": : :"memory");
}
hzp:
Performance counter stats for './test_memcmp' (5 runs):
32356.272845 task-clock # 0.998 CPUs utilized ( +- 0.13% )
40 context-switches # 0.001 K/sec ( +- 0.94% )
0 CPU-migrations # 0.000 K/sec
4,218 page-faults # 0.130 K/sec ( +- 0.00% )
76,712,481,765 cycles # 2.371 GHz ( +- 0.13% ) [83.31%]
36,279,577,636 stalled-cycles-frontend # 47.29% frontend cycles idle ( +- 0.28% ) [83.35%]
1,684,049,110 stalled-cycles-backend # 2.20% backend cycles idle ( +- 2.96% ) [66.67%]
134,355,715,816 instructions # 1.75 insns per cycle
# 0.27 stalled cycles per insn ( +- 0.10% ) [83.35%]
13,526,169,702 branches # 418.039 M/sec ( +- 0.10% ) [83.31%]
1,058,230 branch-misses # 0.01% of all branches ( +- 0.91% ) [83.36%]
32.413866442 seconds time elapsed ( +- 0.13% )
vhzp:
Performance counter stats for './test_memcmp' (5 runs):
30327.183829 task-clock # 0.998 CPUs utilized ( +- 0.13% )
38 context-switches # 0.001 K/sec ( +- 1.53% )
0 CPU-migrations # 0.000 K/sec
4,218 page-faults # 0.139 K/sec ( +- 0.01% )
71,964,773,660 cycles # 2.373 GHz ( +- 0.13% ) [83.35%]
31,191,284,231 stalled-cycles-frontend # 43.34% frontend cycles idle ( +- 0.40% ) [83.32%]
773,484,474 stalled-cycles-backend # 1.07% backend cycles idle ( +- 6.61% ) [66.67%]
134,982,215,437 instructions # 1.88 insns per cycle
# 0.23 stalled cycles per insn ( +- 0.11% ) [83.32%]
13,509,150,683 branches # 445.447 M/sec ( +- 0.11% ) [83.34%]
1,017,667 branch-misses # 0.01% of all branches ( +- 1.07% ) [83.32%]
30.381324695 seconds time elapsed ( +- 0.13% )
Mirobenchmark2
==============
test:
posix_memalign((void **)&p, 2 * MB, 8 * GB);
for (i = 0; i < 1000; i++) {
char *_p = p;
while (_p < p+4*GB) {
assert(*_p == *(_p+4*GB));
_p += 4096;
asm volatile ("": : :"memory");
}
}
hzp:
Performance counter stats for 'taskset -c 0 ./test_memcmp2' (5 runs):
3505.727639 task-clock # 0.998 CPUs utilized ( +- 0.26% )
9 context-switches # 0.003 K/sec ( +- 4.97% )
4,384 page-faults # 0.001 M/sec ( +- 0.00% )
8,318,482,466 cycles # 2.373 GHz ( +- 0.26% ) [33.31%]
5,134,318,786 stalled-cycles-frontend # 61.72% frontend cycles idle ( +- 0.42% ) [33.32%]
2,193,266,208 stalled-cycles-backend # 26.37% backend cycles idle ( +- 5.51% ) [33.33%]
9,494,670,537 instructions # 1.14 insns per cycle
# 0.54 stalled cycles per insn ( +- 0.13% ) [41.68%]
2,108,522,738 branches # 601.451 M/sec ( +- 0.09% ) [41.68%]
158,746 branch-misses # 0.01% of all branches ( +- 1.60% ) [41.71%]
3,168,102,115 L1-dcache-loads
# 903.693 M/sec ( +- 0.11% ) [41.70%]
1,048,710,998 L1-dcache-misses
# 33.10% of all L1-dcache hits ( +- 0.11% ) [41.72%]
1,047,699,685 LLC-load
# 298.854 M/sec ( +- 0.03% ) [33.38%]
2,287 LLC-misses
# 0.00% of all LL-cache hits ( +- 8.27% ) [33.37%]
3,166,187,367 dTLB-loads
# 903.147 M/sec ( +- 0.02% ) [33.35%]
4,266,538 dTLB-misses
# 0.13% of all dTLB cache hits ( +- 0.03% ) [33.33%]
3.513339813 seconds time elapsed ( +- 0.26% )
vhzp:
Performance counter stats for 'taskset -c 0 ./test_memcmp2' (5 runs):
27313.891128 task-clock # 0.998 CPUs utilized ( +- 0.24% )
62 context-switches # 0.002 K/sec ( +- 0.61% )
4,384 page-faults # 0.160 K/sec ( +- 0.01% )
64,747,374,606 cycles # 2.370 GHz ( +- 0.24% ) [33.33%]
61,341,580,278 stalled-cycles-frontend # 94.74% frontend cycles idle ( +- 0.26% ) [33.33%]
56,702,237,511 stalled-cycles-backend # 87.57% backend cycles idle ( +- 0.07% ) [33.33%]
10,033,724,846 instructions # 0.15 insns per cycle
# 6.11 stalled cycles per insn ( +- 0.09% ) [41.65%]
2,190,424,932 branches # 80.195 M/sec ( +- 0.12% ) [41.66%]
1,028,630 branch-misses # 0.05% of all branches ( +- 1.50% ) [41.66%]
3,302,006,540 L1-dcache-loads
# 120.891 M/sec ( +- 0.11% ) [41.68%]
271,374,358 L1-dcache-misses
# 8.22% of all L1-dcache hits ( +- 0.04% ) [41.66%]
20,385,476 LLC-load
# 0.746 M/sec ( +- 1.64% ) [33.34%]
76,754 LLC-misses
# 0.38% of all LL-cache hits ( +- 2.35% ) [33.34%]
3,309,927,290 dTLB-loads
# 121.181 M/sec ( +- 0.03% ) [33.34%]
2,098,967,427 dTLB-misses
# 63.41% of all dTLB cache hits ( +- 0.03% ) [33.34%]
27.364448741 seconds time elapsed ( +- 0.24% )
===
I personally prefer implementation present in this patchset. It doesn't
touch arch-specific code.
This patch:
Huge zero page (hzp) is a non-movable huge page (2M on x86-64) filled with
zeros.
For now let's allocate the page on hugepage_init(). We'll switch to lazy
allocation later.
We are not going to map the huge zero page until we can handle it properly
on all code paths.
is_huge_zero_{pfn,pmd}() functions will be used by following patches to
check whether the pfn/pmd is huge zero page.
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@linux.intel.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-12-13 01:50:47 +04:00
{
2013-08-28 04:18:15 +04:00
/* we can free zero page only if last reference remains */
return atomic_read ( & huge_zero_refcount ) = = 1 ? HPAGE_PMD_NR : 0 ;
}
2012-12-13 01:51:06 +04:00
2013-08-28 04:18:15 +04:00
static unsigned long shrink_huge_zero_page_scan ( struct shrinker * shrink ,
struct shrink_control * sc )
{
2012-12-13 01:51:06 +04:00
if ( atomic_cmpxchg ( & huge_zero_refcount , 1 , 0 ) = = 1 ) {
2013-04-30 02:08:44 +04:00
struct page * zero_page = xchg ( & huge_zero_page , NULL ) ;
BUG_ON ( zero_page = = NULL ) ;
2021-06-16 04:23:49 +03:00
WRITE_ONCE ( huge_zero_pfn , ~ 0UL ) ;
2014-10-30 00:50:26 +03:00
__free_pages ( zero_page , compound_order ( zero_page ) ) ;
2013-08-28 04:18:15 +04:00
return HPAGE_PMD_NR ;
2012-12-13 01:51:06 +04:00
}
return 0 ;
thp: huge zero page: basic preparation
During testing I noticed big (up to 2.5 times) memory consumption overhead
on some workloads (e.g. ft.A from NPB) if THP is enabled.
The main reason for that big difference is lacking zero page in THP case.
We have to allocate a real page on read page fault.
A program to demonstrate the issue:
#include <assert.h>
#include <stdlib.h>
#include <unistd.h>
#define MB 1024*1024
int main(int argc, char **argv)
{
char *p;
int i;
posix_memalign((void **)&p, 2 * MB, 200 * MB);
for (i = 0; i < 200 * MB; i+= 4096)
assert(p[i] == 0);
pause();
return 0;
}
With thp-never RSS is about 400k, but with thp-always it's 200M. After
the patcheset thp-always RSS is 400k too.
Design overview.
Huge zero page (hzp) is a non-movable huge page (2M on x86-64) filled with
zeros. The way how we allocate it changes in the patchset:
- [01/10] simplest way: hzp allocated on boot time in hugepage_init();
- [09/10] lazy allocation on first use;
- [10/10] lockless refcounting + shrinker-reclaimable hzp;
We setup it in do_huge_pmd_anonymous_page() if area around fault address
is suitable for THP and we've got read page fault. If we fail to setup
hzp (ENOMEM) we fallback to handle_pte_fault() as we normally do in THP.
On wp fault to hzp we allocate real memory for the huge page and clear it.
If ENOMEM, graceful fallback: we create a new pmd table and set pte
around fault address to newly allocated normal (4k) page. All other ptes
in the pmd set to normal zero page.
We cannot split hzp (and it's bug if we try), but we can split the pmd
which points to it. On splitting the pmd we create a table with all ptes
set to normal zero page.
===
By hpa's request I've tried alternative approach for hzp implementation
(see Virtual huge zero page patchset): pmd table with all entries set to
zero page. This way should be more cache friendly, but it increases TLB
pressure.
The problem with virtual huge zero page: it requires per-arch enabling.
We need a way to mark that pmd table has all ptes set to zero page.
Some numbers to compare two implementations (on 4s Westmere-EX):
Mirobenchmark1
==============
test:
posix_memalign((void **)&p, 2 * MB, 8 * GB);
for (i = 0; i < 100; i++) {
assert(memcmp(p, p + 4*GB, 4*GB) == 0);
asm volatile ("": : :"memory");
}
hzp:
Performance counter stats for './test_memcmp' (5 runs):
32356.272845 task-clock # 0.998 CPUs utilized ( +- 0.13% )
40 context-switches # 0.001 K/sec ( +- 0.94% )
0 CPU-migrations # 0.000 K/sec
4,218 page-faults # 0.130 K/sec ( +- 0.00% )
76,712,481,765 cycles # 2.371 GHz ( +- 0.13% ) [83.31%]
36,279,577,636 stalled-cycles-frontend # 47.29% frontend cycles idle ( +- 0.28% ) [83.35%]
1,684,049,110 stalled-cycles-backend # 2.20% backend cycles idle ( +- 2.96% ) [66.67%]
134,355,715,816 instructions # 1.75 insns per cycle
# 0.27 stalled cycles per insn ( +- 0.10% ) [83.35%]
13,526,169,702 branches # 418.039 M/sec ( +- 0.10% ) [83.31%]
1,058,230 branch-misses # 0.01% of all branches ( +- 0.91% ) [83.36%]
32.413866442 seconds time elapsed ( +- 0.13% )
vhzp:
Performance counter stats for './test_memcmp' (5 runs):
30327.183829 task-clock # 0.998 CPUs utilized ( +- 0.13% )
38 context-switches # 0.001 K/sec ( +- 1.53% )
0 CPU-migrations # 0.000 K/sec
4,218 page-faults # 0.139 K/sec ( +- 0.01% )
71,964,773,660 cycles # 2.373 GHz ( +- 0.13% ) [83.35%]
31,191,284,231 stalled-cycles-frontend # 43.34% frontend cycles idle ( +- 0.40% ) [83.32%]
773,484,474 stalled-cycles-backend # 1.07% backend cycles idle ( +- 6.61% ) [66.67%]
134,982,215,437 instructions # 1.88 insns per cycle
# 0.23 stalled cycles per insn ( +- 0.11% ) [83.32%]
13,509,150,683 branches # 445.447 M/sec ( +- 0.11% ) [83.34%]
1,017,667 branch-misses # 0.01% of all branches ( +- 1.07% ) [83.32%]
30.381324695 seconds time elapsed ( +- 0.13% )
Mirobenchmark2
==============
test:
posix_memalign((void **)&p, 2 * MB, 8 * GB);
for (i = 0; i < 1000; i++) {
char *_p = p;
while (_p < p+4*GB) {
assert(*_p == *(_p+4*GB));
_p += 4096;
asm volatile ("": : :"memory");
}
}
hzp:
Performance counter stats for 'taskset -c 0 ./test_memcmp2' (5 runs):
3505.727639 task-clock # 0.998 CPUs utilized ( +- 0.26% )
9 context-switches # 0.003 K/sec ( +- 4.97% )
4,384 page-faults # 0.001 M/sec ( +- 0.00% )
8,318,482,466 cycles # 2.373 GHz ( +- 0.26% ) [33.31%]
5,134,318,786 stalled-cycles-frontend # 61.72% frontend cycles idle ( +- 0.42% ) [33.32%]
2,193,266,208 stalled-cycles-backend # 26.37% backend cycles idle ( +- 5.51% ) [33.33%]
9,494,670,537 instructions # 1.14 insns per cycle
# 0.54 stalled cycles per insn ( +- 0.13% ) [41.68%]
2,108,522,738 branches # 601.451 M/sec ( +- 0.09% ) [41.68%]
158,746 branch-misses # 0.01% of all branches ( +- 1.60% ) [41.71%]
3,168,102,115 L1-dcache-loads
# 903.693 M/sec ( +- 0.11% ) [41.70%]
1,048,710,998 L1-dcache-misses
# 33.10% of all L1-dcache hits ( +- 0.11% ) [41.72%]
1,047,699,685 LLC-load
# 298.854 M/sec ( +- 0.03% ) [33.38%]
2,287 LLC-misses
# 0.00% of all LL-cache hits ( +- 8.27% ) [33.37%]
3,166,187,367 dTLB-loads
# 903.147 M/sec ( +- 0.02% ) [33.35%]
4,266,538 dTLB-misses
# 0.13% of all dTLB cache hits ( +- 0.03% ) [33.33%]
3.513339813 seconds time elapsed ( +- 0.26% )
vhzp:
Performance counter stats for 'taskset -c 0 ./test_memcmp2' (5 runs):
27313.891128 task-clock # 0.998 CPUs utilized ( +- 0.24% )
62 context-switches # 0.002 K/sec ( +- 0.61% )
4,384 page-faults # 0.160 K/sec ( +- 0.01% )
64,747,374,606 cycles # 2.370 GHz ( +- 0.24% ) [33.33%]
61,341,580,278 stalled-cycles-frontend # 94.74% frontend cycles idle ( +- 0.26% ) [33.33%]
56,702,237,511 stalled-cycles-backend # 87.57% backend cycles idle ( +- 0.07% ) [33.33%]
10,033,724,846 instructions # 0.15 insns per cycle
# 6.11 stalled cycles per insn ( +- 0.09% ) [41.65%]
2,190,424,932 branches # 80.195 M/sec ( +- 0.12% ) [41.66%]
1,028,630 branch-misses # 0.05% of all branches ( +- 1.50% ) [41.66%]
3,302,006,540 L1-dcache-loads
# 120.891 M/sec ( +- 0.11% ) [41.68%]
271,374,358 L1-dcache-misses
# 8.22% of all L1-dcache hits ( +- 0.04% ) [41.66%]
20,385,476 LLC-load
# 0.746 M/sec ( +- 1.64% ) [33.34%]
76,754 LLC-misses
# 0.38% of all LL-cache hits ( +- 2.35% ) [33.34%]
3,309,927,290 dTLB-loads
# 121.181 M/sec ( +- 0.03% ) [33.34%]
2,098,967,427 dTLB-misses
# 63.41% of all dTLB cache hits ( +- 0.03% ) [33.34%]
27.364448741 seconds time elapsed ( +- 0.24% )
===
I personally prefer implementation present in this patchset. It doesn't
touch arch-specific code.
This patch:
Huge zero page (hzp) is a non-movable huge page (2M on x86-64) filled with
zeros.
For now let's allocate the page on hugepage_init(). We'll switch to lazy
allocation later.
We are not going to map the huge zero page until we can handle it properly
on all code paths.
is_huge_zero_{pfn,pmd}() functions will be used by following patches to
check whether the pfn/pmd is huge zero page.
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@linux.intel.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-12-13 01:50:47 +04:00
}
2012-12-13 01:51:06 +04:00
static struct shrinker huge_zero_page_shrinker = {
2013-08-28 04:18:15 +04:00
. count_objects = shrink_huge_zero_page_count ,
. scan_objects = shrink_huge_zero_page_scan ,
2012-12-13 01:51:06 +04:00
. seeks = DEFAULT_SEEKS ,
} ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
# ifdef CONFIG_SYSFS
static ssize_t enabled_show ( struct kobject * kobj ,
struct kobj_attribute * attr , char * buf )
{
2020-12-15 06:14:46 +03:00
const char * output ;
mm: thp: set THP defrag by default to madvise and add a stall-free defrag option
THP defrag is enabled by default to direct reclaim/compact but not wake
kswapd in the event of a THP allocation failure. The problem is that
THP allocation requests potentially enter reclaim/compaction. This
potentially incurs a severe stall that is not guaranteed to be offset by
reduced TLB misses. While there has been considerable effort to reduce
the impact of reclaim/compaction, it is still a high cost and workloads
that should fit in memory fail to do so. Specifically, a simple
anon/file streaming workload will enter direct reclaim on NUMA at least
even though the working set size is 80% of RAM. It's been years and
it's time to throw in the towel.
First, this patch defines THP defrag as follows;
madvise: A failed allocation will direct reclaim/compact if the application requests it
never: Neither reclaim/compact nor wake kswapd
defer: A failed allocation will wake kswapd/kcompactd
always: A failed allocation will direct reclaim/compact (historical behaviour)
khugepaged defrag will enter direct/reclaim but not wake kswapd.
Next it sets the default defrag option to be "madvise" to only enter
direct reclaim/compaction for applications that specifically requested
it.
Lastly, it removes a check from the page allocator slowpath that is
related to __GFP_THISNODE to allow "defer" to work. The callers that
really cares are slub/slab and they are updated accordingly. The slab
one may be surprising because it also corrects a comment as kswapd was
never woken up by that path.
This means that a THP fault will no longer stall for most applications
by default and the ideal for most users that get THP if they are
immediately available. There are still options for users that prefer a
stall at startup of a new application by either restoring historical
behaviour with "always" or pick a half-way point with "defer" where
kswapd does some of the work in the background and wakes kcompactd if
necessary. THP defrag for khugepaged remains enabled and will enter
direct/reclaim but no wakeup kswapd or kcompactd.
After this patch a THP allocation failure will quickly fallback and rely
on khugepaged to recover the situation at some time in the future. In
some cases, this will reduce THP usage but the benefit of THP is hard to
measure and not a universal win where as a stall to reclaim/compaction
is definitely measurable and can be painful.
The first test for this is using "usemem" to read a large file and write
a large anonymous mapping (to avoid the zero page) multiple times. The
total size of the mappings is 80% of RAM and the benchmark simply
measures how long it takes to complete. It uses multiple threads to see
if that is a factor. On UMA, the performance is almost identical so is
not reported but on NUMA, we see this
usemem
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean System-1 102.86 ( 0.00%) 46.81 ( 54.50%)
Amean System-4 37.85 ( 0.00%) 34.02 ( 10.12%)
Amean System-7 48.12 ( 0.00%) 46.89 ( 2.56%)
Amean System-12 51.98 ( 0.00%) 56.96 ( -9.57%)
Amean System-21 80.16 ( 0.00%) 79.05 ( 1.39%)
Amean System-30 110.71 ( 0.00%) 107.17 ( 3.20%)
Amean System-48 127.98 ( 0.00%) 124.83 ( 2.46%)
Amean Elapsd-1 185.84 ( 0.00%) 105.51 ( 43.23%)
Amean Elapsd-4 26.19 ( 0.00%) 25.58 ( 2.33%)
Amean Elapsd-7 21.65 ( 0.00%) 21.62 ( 0.16%)
Amean Elapsd-12 18.58 ( 0.00%) 17.94 ( 3.43%)
Amean Elapsd-21 17.53 ( 0.00%) 16.60 ( 5.33%)
Amean Elapsd-30 17.45 ( 0.00%) 17.13 ( 1.84%)
Amean Elapsd-48 15.40 ( 0.00%) 15.27 ( 0.82%)
For a single thread, the benchmark completes 43.23% faster with this
patch applied with smaller benefits as the thread increases. Similar,
notice the large reduction in most cases in system CPU usage. The
overall CPU time is
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
User 10357.65 10438.33
System 3988.88 3543.94
Elapsed 2203.01 1634.41
Which is substantial. Now, the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 128458477 278352931
Major Faults 2174976 225
Swap Ins 16904701 0
Swap Outs 17359627 0
Allocation stalls 43611 0
DMA allocs 0 0
DMA32 allocs 19832646 19448017
Normal allocs 614488453 580941839
Movable allocs 0 0
Direct pages scanned 24163800 0
Kswapd pages scanned 0 0
Kswapd pages reclaimed 0 0
Direct pages reclaimed 20691346 0
Compaction stalls 42263 0
Compaction success 938 0
Compaction failures 41325 0
This patch eliminates almost all swapping and direct reclaim activity.
There is still overhead but it's from NUMA balancing which does not
identify that it's pointless trying to do anything with this workload.
I also tried the thpscale benchmark which forces a corner case where
compaction can be used heavily and measures the latency of whether base
or huge pages were used
thpscale Fault Latencies
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean fault-base-1 5288.84 ( 0.00%) 2817.12 ( 46.73%)
Amean fault-base-3 6365.53 ( 0.00%) 3499.11 ( 45.03%)
Amean fault-base-5 6526.19 ( 0.00%) 4363.06 ( 33.15%)
Amean fault-base-7 7142.25 ( 0.00%) 4858.08 ( 31.98%)
Amean fault-base-12 13827.64 ( 0.00%) 10292.11 ( 25.57%)
Amean fault-base-18 18235.07 ( 0.00%) 13788.84 ( 24.38%)
Amean fault-base-24 21597.80 ( 0.00%) 24388.03 (-12.92%)
Amean fault-base-30 26754.15 ( 0.00%) 19700.55 ( 26.36%)
Amean fault-base-32 26784.94 ( 0.00%) 19513.57 ( 27.15%)
Amean fault-huge-1 4223.96 ( 0.00%) 2178.57 ( 48.42%)
Amean fault-huge-3 2194.77 ( 0.00%) 2149.74 ( 2.05%)
Amean fault-huge-5 2569.60 ( 0.00%) 2346.95 ( 8.66%)
Amean fault-huge-7 3612.69 ( 0.00%) 2997.70 ( 17.02%)
Amean fault-huge-12 3301.75 ( 0.00%) 6727.02 (-103.74%)
Amean fault-huge-18 6696.47 ( 0.00%) 6685.72 ( 0.16%)
Amean fault-huge-24 8000.72 ( 0.00%) 9311.43 (-16.38%)
Amean fault-huge-30 13305.55 ( 0.00%) 9750.45 ( 26.72%)
Amean fault-huge-32 9981.71 ( 0.00%) 10316.06 ( -3.35%)
The average time to fault pages is substantially reduced in the majority
of caseds but with the obvious caveat that fewer THPs are actually used
in this adverse workload
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Percentage huge-1 0.71 ( 0.00%) 14.04 (1865.22%)
Percentage huge-3 10.77 ( 0.00%) 33.05 (206.85%)
Percentage huge-5 60.39 ( 0.00%) 38.51 (-36.23%)
Percentage huge-7 45.97 ( 0.00%) 34.57 (-24.79%)
Percentage huge-12 68.12 ( 0.00%) 40.07 (-41.17%)
Percentage huge-18 64.93 ( 0.00%) 47.82 (-26.35%)
Percentage huge-24 62.69 ( 0.00%) 44.23 (-29.44%)
Percentage huge-30 43.49 ( 0.00%) 55.38 ( 27.34%)
Percentage huge-32 50.72 ( 0.00%) 51.90 ( 2.35%)
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 37429143 47564000
Major Faults 1916 1558
Swap Ins 1466 1079
Swap Outs 2936863 149626
Allocation stalls 62510 3
DMA allocs 0 0
DMA32 allocs 6566458 6401314
Normal allocs 216361697 216538171
Movable allocs 0 0
Direct pages scanned 25977580 17998
Kswapd pages scanned 0 3638931
Kswapd pages reclaimed 0 207236
Direct pages reclaimed 8833714 88
Compaction stalls 103349 5
Compaction success 270 4
Compaction failures 103079 1
Note again that while this does swap as it's an aggressive workload, the
direct relcim activity and allocation stalls is substantially reduced.
There is some kswapd activity but ftrace showed that the kswapd activity
was due to normal wakeups from 4K pages being allocated.
Compaction-related stalls and activity are almost eliminated.
I also tried the stutter benchmark. For this, I do not have figures for
NUMA but it's something that does impact UMA so I'll report what is
available
stutter
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Min mmap 7.3571 ( 0.00%) 7.3438 ( 0.18%)
1st-qrtle mmap 7.5278 ( 0.00%) 17.9200 (-138.05%)
2nd-qrtle mmap 7.6818 ( 0.00%) 21.6055 (-181.25%)
3rd-qrtle mmap 11.0889 ( 0.00%) 21.8881 (-97.39%)
Max-90% mmap 27.8978 ( 0.00%) 22.1632 ( 20.56%)
Max-93% mmap 28.3202 ( 0.00%) 22.3044 ( 21.24%)
Max-95% mmap 28.5600 ( 0.00%) 22.4580 ( 21.37%)
Max-99% mmap 29.6032 ( 0.00%) 25.5216 ( 13.79%)
Max mmap 4109.7289 ( 0.00%) 4813.9832 (-17.14%)
Mean mmap 12.4474 ( 0.00%) 19.3027 (-55.07%)
This benchmark is trying to fault an anonymous mapping while there is a
heavy IO load -- a scenario that desktop users used to complain about
frequently. This shows a mix because the ideal case of mapping with THP
is not hit as often. However, note that 99% of the mappings complete
13.79% faster. The CPU usage here is particularly interesting
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
User 67.50 0.99
System 1327.88 91.30
Elapsed 2079.00 2128.98
And once again we look at the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 335241922 1314582827
Major Faults 715 819
Swap Ins 0 0
Swap Outs 0 0
Allocation stalls 532723 0
DMA allocs 0 0
DMA32 allocs 1822364341 1177950222
Normal allocs 1815640808 1517844854
Movable allocs 0 0
Direct pages scanned 21892772 0
Kswapd pages scanned 20015890 41879484
Kswapd pages reclaimed 19961986 41822072
Direct pages reclaimed 21892741 0
Compaction stalls 1065755 0
Compaction success 514 0
Compaction failures 1065241 0
Allocation stalls and all direct reclaim activity is eliminated as well
as compaction-related stalls.
THP gives impressive gains in some cases but only if they are quickly
available. We're not going to reach the point where they are completely
free so lets take the costs out of the fast paths finally and defer the
cost to kswapd, kcompactd and khugepaged where it belongs.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-18 00:19:23 +03:00
if ( test_bit ( TRANSPARENT_HUGEPAGE_FLAG , & transparent_hugepage_flags ) )
2020-12-15 06:14:46 +03:00
output = " [always] madvise never " ;
else if ( test_bit ( TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG ,
& transparent_hugepage_flags ) )
output = " always [madvise] never " ;
mm: thp: set THP defrag by default to madvise and add a stall-free defrag option
THP defrag is enabled by default to direct reclaim/compact but not wake
kswapd in the event of a THP allocation failure. The problem is that
THP allocation requests potentially enter reclaim/compaction. This
potentially incurs a severe stall that is not guaranteed to be offset by
reduced TLB misses. While there has been considerable effort to reduce
the impact of reclaim/compaction, it is still a high cost and workloads
that should fit in memory fail to do so. Specifically, a simple
anon/file streaming workload will enter direct reclaim on NUMA at least
even though the working set size is 80% of RAM. It's been years and
it's time to throw in the towel.
First, this patch defines THP defrag as follows;
madvise: A failed allocation will direct reclaim/compact if the application requests it
never: Neither reclaim/compact nor wake kswapd
defer: A failed allocation will wake kswapd/kcompactd
always: A failed allocation will direct reclaim/compact (historical behaviour)
khugepaged defrag will enter direct/reclaim but not wake kswapd.
Next it sets the default defrag option to be "madvise" to only enter
direct reclaim/compaction for applications that specifically requested
it.
Lastly, it removes a check from the page allocator slowpath that is
related to __GFP_THISNODE to allow "defer" to work. The callers that
really cares are slub/slab and they are updated accordingly. The slab
one may be surprising because it also corrects a comment as kswapd was
never woken up by that path.
This means that a THP fault will no longer stall for most applications
by default and the ideal for most users that get THP if they are
immediately available. There are still options for users that prefer a
stall at startup of a new application by either restoring historical
behaviour with "always" or pick a half-way point with "defer" where
kswapd does some of the work in the background and wakes kcompactd if
necessary. THP defrag for khugepaged remains enabled and will enter
direct/reclaim but no wakeup kswapd or kcompactd.
After this patch a THP allocation failure will quickly fallback and rely
on khugepaged to recover the situation at some time in the future. In
some cases, this will reduce THP usage but the benefit of THP is hard to
measure and not a universal win where as a stall to reclaim/compaction
is definitely measurable and can be painful.
The first test for this is using "usemem" to read a large file and write
a large anonymous mapping (to avoid the zero page) multiple times. The
total size of the mappings is 80% of RAM and the benchmark simply
measures how long it takes to complete. It uses multiple threads to see
if that is a factor. On UMA, the performance is almost identical so is
not reported but on NUMA, we see this
usemem
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean System-1 102.86 ( 0.00%) 46.81 ( 54.50%)
Amean System-4 37.85 ( 0.00%) 34.02 ( 10.12%)
Amean System-7 48.12 ( 0.00%) 46.89 ( 2.56%)
Amean System-12 51.98 ( 0.00%) 56.96 ( -9.57%)
Amean System-21 80.16 ( 0.00%) 79.05 ( 1.39%)
Amean System-30 110.71 ( 0.00%) 107.17 ( 3.20%)
Amean System-48 127.98 ( 0.00%) 124.83 ( 2.46%)
Amean Elapsd-1 185.84 ( 0.00%) 105.51 ( 43.23%)
Amean Elapsd-4 26.19 ( 0.00%) 25.58 ( 2.33%)
Amean Elapsd-7 21.65 ( 0.00%) 21.62 ( 0.16%)
Amean Elapsd-12 18.58 ( 0.00%) 17.94 ( 3.43%)
Amean Elapsd-21 17.53 ( 0.00%) 16.60 ( 5.33%)
Amean Elapsd-30 17.45 ( 0.00%) 17.13 ( 1.84%)
Amean Elapsd-48 15.40 ( 0.00%) 15.27 ( 0.82%)
For a single thread, the benchmark completes 43.23% faster with this
patch applied with smaller benefits as the thread increases. Similar,
notice the large reduction in most cases in system CPU usage. The
overall CPU time is
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
User 10357.65 10438.33
System 3988.88 3543.94
Elapsed 2203.01 1634.41
Which is substantial. Now, the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 128458477 278352931
Major Faults 2174976 225
Swap Ins 16904701 0
Swap Outs 17359627 0
Allocation stalls 43611 0
DMA allocs 0 0
DMA32 allocs 19832646 19448017
Normal allocs 614488453 580941839
Movable allocs 0 0
Direct pages scanned 24163800 0
Kswapd pages scanned 0 0
Kswapd pages reclaimed 0 0
Direct pages reclaimed 20691346 0
Compaction stalls 42263 0
Compaction success 938 0
Compaction failures 41325 0
This patch eliminates almost all swapping and direct reclaim activity.
There is still overhead but it's from NUMA balancing which does not
identify that it's pointless trying to do anything with this workload.
I also tried the thpscale benchmark which forces a corner case where
compaction can be used heavily and measures the latency of whether base
or huge pages were used
thpscale Fault Latencies
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean fault-base-1 5288.84 ( 0.00%) 2817.12 ( 46.73%)
Amean fault-base-3 6365.53 ( 0.00%) 3499.11 ( 45.03%)
Amean fault-base-5 6526.19 ( 0.00%) 4363.06 ( 33.15%)
Amean fault-base-7 7142.25 ( 0.00%) 4858.08 ( 31.98%)
Amean fault-base-12 13827.64 ( 0.00%) 10292.11 ( 25.57%)
Amean fault-base-18 18235.07 ( 0.00%) 13788.84 ( 24.38%)
Amean fault-base-24 21597.80 ( 0.00%) 24388.03 (-12.92%)
Amean fault-base-30 26754.15 ( 0.00%) 19700.55 ( 26.36%)
Amean fault-base-32 26784.94 ( 0.00%) 19513.57 ( 27.15%)
Amean fault-huge-1 4223.96 ( 0.00%) 2178.57 ( 48.42%)
Amean fault-huge-3 2194.77 ( 0.00%) 2149.74 ( 2.05%)
Amean fault-huge-5 2569.60 ( 0.00%) 2346.95 ( 8.66%)
Amean fault-huge-7 3612.69 ( 0.00%) 2997.70 ( 17.02%)
Amean fault-huge-12 3301.75 ( 0.00%) 6727.02 (-103.74%)
Amean fault-huge-18 6696.47 ( 0.00%) 6685.72 ( 0.16%)
Amean fault-huge-24 8000.72 ( 0.00%) 9311.43 (-16.38%)
Amean fault-huge-30 13305.55 ( 0.00%) 9750.45 ( 26.72%)
Amean fault-huge-32 9981.71 ( 0.00%) 10316.06 ( -3.35%)
The average time to fault pages is substantially reduced in the majority
of caseds but with the obvious caveat that fewer THPs are actually used
in this adverse workload
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Percentage huge-1 0.71 ( 0.00%) 14.04 (1865.22%)
Percentage huge-3 10.77 ( 0.00%) 33.05 (206.85%)
Percentage huge-5 60.39 ( 0.00%) 38.51 (-36.23%)
Percentage huge-7 45.97 ( 0.00%) 34.57 (-24.79%)
Percentage huge-12 68.12 ( 0.00%) 40.07 (-41.17%)
Percentage huge-18 64.93 ( 0.00%) 47.82 (-26.35%)
Percentage huge-24 62.69 ( 0.00%) 44.23 (-29.44%)
Percentage huge-30 43.49 ( 0.00%) 55.38 ( 27.34%)
Percentage huge-32 50.72 ( 0.00%) 51.90 ( 2.35%)
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 37429143 47564000
Major Faults 1916 1558
Swap Ins 1466 1079
Swap Outs 2936863 149626
Allocation stalls 62510 3
DMA allocs 0 0
DMA32 allocs 6566458 6401314
Normal allocs 216361697 216538171
Movable allocs 0 0
Direct pages scanned 25977580 17998
Kswapd pages scanned 0 3638931
Kswapd pages reclaimed 0 207236
Direct pages reclaimed 8833714 88
Compaction stalls 103349 5
Compaction success 270 4
Compaction failures 103079 1
Note again that while this does swap as it's an aggressive workload, the
direct relcim activity and allocation stalls is substantially reduced.
There is some kswapd activity but ftrace showed that the kswapd activity
was due to normal wakeups from 4K pages being allocated.
Compaction-related stalls and activity are almost eliminated.
I also tried the stutter benchmark. For this, I do not have figures for
NUMA but it's something that does impact UMA so I'll report what is
available
stutter
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Min mmap 7.3571 ( 0.00%) 7.3438 ( 0.18%)
1st-qrtle mmap 7.5278 ( 0.00%) 17.9200 (-138.05%)
2nd-qrtle mmap 7.6818 ( 0.00%) 21.6055 (-181.25%)
3rd-qrtle mmap 11.0889 ( 0.00%) 21.8881 (-97.39%)
Max-90% mmap 27.8978 ( 0.00%) 22.1632 ( 20.56%)
Max-93% mmap 28.3202 ( 0.00%) 22.3044 ( 21.24%)
Max-95% mmap 28.5600 ( 0.00%) 22.4580 ( 21.37%)
Max-99% mmap 29.6032 ( 0.00%) 25.5216 ( 13.79%)
Max mmap 4109.7289 ( 0.00%) 4813.9832 (-17.14%)
Mean mmap 12.4474 ( 0.00%) 19.3027 (-55.07%)
This benchmark is trying to fault an anonymous mapping while there is a
heavy IO load -- a scenario that desktop users used to complain about
frequently. This shows a mix because the ideal case of mapping with THP
is not hit as often. However, note that 99% of the mappings complete
13.79% faster. The CPU usage here is particularly interesting
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
User 67.50 0.99
System 1327.88 91.30
Elapsed 2079.00 2128.98
And once again we look at the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 335241922 1314582827
Major Faults 715 819
Swap Ins 0 0
Swap Outs 0 0
Allocation stalls 532723 0
DMA allocs 0 0
DMA32 allocs 1822364341 1177950222
Normal allocs 1815640808 1517844854
Movable allocs 0 0
Direct pages scanned 21892772 0
Kswapd pages scanned 20015890 41879484
Kswapd pages reclaimed 19961986 41822072
Direct pages reclaimed 21892741 0
Compaction stalls 1065755 0
Compaction success 514 0
Compaction failures 1065241 0
Allocation stalls and all direct reclaim activity is eliminated as well
as compaction-related stalls.
THP gives impressive gains in some cases but only if they are quickly
available. We're not going to reach the point where they are completely
free so lets take the costs out of the fast paths finally and defer the
cost to kswapd, kcompactd and khugepaged where it belongs.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-18 00:19:23 +03:00
else
2020-12-15 06:14:46 +03:00
output = " always madvise [never] " ;
return sysfs_emit ( buf , " %s \n " , output ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
}
mm: thp: set THP defrag by default to madvise and add a stall-free defrag option
THP defrag is enabled by default to direct reclaim/compact but not wake
kswapd in the event of a THP allocation failure. The problem is that
THP allocation requests potentially enter reclaim/compaction. This
potentially incurs a severe stall that is not guaranteed to be offset by
reduced TLB misses. While there has been considerable effort to reduce
the impact of reclaim/compaction, it is still a high cost and workloads
that should fit in memory fail to do so. Specifically, a simple
anon/file streaming workload will enter direct reclaim on NUMA at least
even though the working set size is 80% of RAM. It's been years and
it's time to throw in the towel.
First, this patch defines THP defrag as follows;
madvise: A failed allocation will direct reclaim/compact if the application requests it
never: Neither reclaim/compact nor wake kswapd
defer: A failed allocation will wake kswapd/kcompactd
always: A failed allocation will direct reclaim/compact (historical behaviour)
khugepaged defrag will enter direct/reclaim but not wake kswapd.
Next it sets the default defrag option to be "madvise" to only enter
direct reclaim/compaction for applications that specifically requested
it.
Lastly, it removes a check from the page allocator slowpath that is
related to __GFP_THISNODE to allow "defer" to work. The callers that
really cares are slub/slab and they are updated accordingly. The slab
one may be surprising because it also corrects a comment as kswapd was
never woken up by that path.
This means that a THP fault will no longer stall for most applications
by default and the ideal for most users that get THP if they are
immediately available. There are still options for users that prefer a
stall at startup of a new application by either restoring historical
behaviour with "always" or pick a half-way point with "defer" where
kswapd does some of the work in the background and wakes kcompactd if
necessary. THP defrag for khugepaged remains enabled and will enter
direct/reclaim but no wakeup kswapd or kcompactd.
After this patch a THP allocation failure will quickly fallback and rely
on khugepaged to recover the situation at some time in the future. In
some cases, this will reduce THP usage but the benefit of THP is hard to
measure and not a universal win where as a stall to reclaim/compaction
is definitely measurable and can be painful.
The first test for this is using "usemem" to read a large file and write
a large anonymous mapping (to avoid the zero page) multiple times. The
total size of the mappings is 80% of RAM and the benchmark simply
measures how long it takes to complete. It uses multiple threads to see
if that is a factor. On UMA, the performance is almost identical so is
not reported but on NUMA, we see this
usemem
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean System-1 102.86 ( 0.00%) 46.81 ( 54.50%)
Amean System-4 37.85 ( 0.00%) 34.02 ( 10.12%)
Amean System-7 48.12 ( 0.00%) 46.89 ( 2.56%)
Amean System-12 51.98 ( 0.00%) 56.96 ( -9.57%)
Amean System-21 80.16 ( 0.00%) 79.05 ( 1.39%)
Amean System-30 110.71 ( 0.00%) 107.17 ( 3.20%)
Amean System-48 127.98 ( 0.00%) 124.83 ( 2.46%)
Amean Elapsd-1 185.84 ( 0.00%) 105.51 ( 43.23%)
Amean Elapsd-4 26.19 ( 0.00%) 25.58 ( 2.33%)
Amean Elapsd-7 21.65 ( 0.00%) 21.62 ( 0.16%)
Amean Elapsd-12 18.58 ( 0.00%) 17.94 ( 3.43%)
Amean Elapsd-21 17.53 ( 0.00%) 16.60 ( 5.33%)
Amean Elapsd-30 17.45 ( 0.00%) 17.13 ( 1.84%)
Amean Elapsd-48 15.40 ( 0.00%) 15.27 ( 0.82%)
For a single thread, the benchmark completes 43.23% faster with this
patch applied with smaller benefits as the thread increases. Similar,
notice the large reduction in most cases in system CPU usage. The
overall CPU time is
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
User 10357.65 10438.33
System 3988.88 3543.94
Elapsed 2203.01 1634.41
Which is substantial. Now, the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 128458477 278352931
Major Faults 2174976 225
Swap Ins 16904701 0
Swap Outs 17359627 0
Allocation stalls 43611 0
DMA allocs 0 0
DMA32 allocs 19832646 19448017
Normal allocs 614488453 580941839
Movable allocs 0 0
Direct pages scanned 24163800 0
Kswapd pages scanned 0 0
Kswapd pages reclaimed 0 0
Direct pages reclaimed 20691346 0
Compaction stalls 42263 0
Compaction success 938 0
Compaction failures 41325 0
This patch eliminates almost all swapping and direct reclaim activity.
There is still overhead but it's from NUMA balancing which does not
identify that it's pointless trying to do anything with this workload.
I also tried the thpscale benchmark which forces a corner case where
compaction can be used heavily and measures the latency of whether base
or huge pages were used
thpscale Fault Latencies
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean fault-base-1 5288.84 ( 0.00%) 2817.12 ( 46.73%)
Amean fault-base-3 6365.53 ( 0.00%) 3499.11 ( 45.03%)
Amean fault-base-5 6526.19 ( 0.00%) 4363.06 ( 33.15%)
Amean fault-base-7 7142.25 ( 0.00%) 4858.08 ( 31.98%)
Amean fault-base-12 13827.64 ( 0.00%) 10292.11 ( 25.57%)
Amean fault-base-18 18235.07 ( 0.00%) 13788.84 ( 24.38%)
Amean fault-base-24 21597.80 ( 0.00%) 24388.03 (-12.92%)
Amean fault-base-30 26754.15 ( 0.00%) 19700.55 ( 26.36%)
Amean fault-base-32 26784.94 ( 0.00%) 19513.57 ( 27.15%)
Amean fault-huge-1 4223.96 ( 0.00%) 2178.57 ( 48.42%)
Amean fault-huge-3 2194.77 ( 0.00%) 2149.74 ( 2.05%)
Amean fault-huge-5 2569.60 ( 0.00%) 2346.95 ( 8.66%)
Amean fault-huge-7 3612.69 ( 0.00%) 2997.70 ( 17.02%)
Amean fault-huge-12 3301.75 ( 0.00%) 6727.02 (-103.74%)
Amean fault-huge-18 6696.47 ( 0.00%) 6685.72 ( 0.16%)
Amean fault-huge-24 8000.72 ( 0.00%) 9311.43 (-16.38%)
Amean fault-huge-30 13305.55 ( 0.00%) 9750.45 ( 26.72%)
Amean fault-huge-32 9981.71 ( 0.00%) 10316.06 ( -3.35%)
The average time to fault pages is substantially reduced in the majority
of caseds but with the obvious caveat that fewer THPs are actually used
in this adverse workload
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Percentage huge-1 0.71 ( 0.00%) 14.04 (1865.22%)
Percentage huge-3 10.77 ( 0.00%) 33.05 (206.85%)
Percentage huge-5 60.39 ( 0.00%) 38.51 (-36.23%)
Percentage huge-7 45.97 ( 0.00%) 34.57 (-24.79%)
Percentage huge-12 68.12 ( 0.00%) 40.07 (-41.17%)
Percentage huge-18 64.93 ( 0.00%) 47.82 (-26.35%)
Percentage huge-24 62.69 ( 0.00%) 44.23 (-29.44%)
Percentage huge-30 43.49 ( 0.00%) 55.38 ( 27.34%)
Percentage huge-32 50.72 ( 0.00%) 51.90 ( 2.35%)
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 37429143 47564000
Major Faults 1916 1558
Swap Ins 1466 1079
Swap Outs 2936863 149626
Allocation stalls 62510 3
DMA allocs 0 0
DMA32 allocs 6566458 6401314
Normal allocs 216361697 216538171
Movable allocs 0 0
Direct pages scanned 25977580 17998
Kswapd pages scanned 0 3638931
Kswapd pages reclaimed 0 207236
Direct pages reclaimed 8833714 88
Compaction stalls 103349 5
Compaction success 270 4
Compaction failures 103079 1
Note again that while this does swap as it's an aggressive workload, the
direct relcim activity and allocation stalls is substantially reduced.
There is some kswapd activity but ftrace showed that the kswapd activity
was due to normal wakeups from 4K pages being allocated.
Compaction-related stalls and activity are almost eliminated.
I also tried the stutter benchmark. For this, I do not have figures for
NUMA but it's something that does impact UMA so I'll report what is
available
stutter
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Min mmap 7.3571 ( 0.00%) 7.3438 ( 0.18%)
1st-qrtle mmap 7.5278 ( 0.00%) 17.9200 (-138.05%)
2nd-qrtle mmap 7.6818 ( 0.00%) 21.6055 (-181.25%)
3rd-qrtle mmap 11.0889 ( 0.00%) 21.8881 (-97.39%)
Max-90% mmap 27.8978 ( 0.00%) 22.1632 ( 20.56%)
Max-93% mmap 28.3202 ( 0.00%) 22.3044 ( 21.24%)
Max-95% mmap 28.5600 ( 0.00%) 22.4580 ( 21.37%)
Max-99% mmap 29.6032 ( 0.00%) 25.5216 ( 13.79%)
Max mmap 4109.7289 ( 0.00%) 4813.9832 (-17.14%)
Mean mmap 12.4474 ( 0.00%) 19.3027 (-55.07%)
This benchmark is trying to fault an anonymous mapping while there is a
heavy IO load -- a scenario that desktop users used to complain about
frequently. This shows a mix because the ideal case of mapping with THP
is not hit as often. However, note that 99% of the mappings complete
13.79% faster. The CPU usage here is particularly interesting
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
User 67.50 0.99
System 1327.88 91.30
Elapsed 2079.00 2128.98
And once again we look at the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 335241922 1314582827
Major Faults 715 819
Swap Ins 0 0
Swap Outs 0 0
Allocation stalls 532723 0
DMA allocs 0 0
DMA32 allocs 1822364341 1177950222
Normal allocs 1815640808 1517844854
Movable allocs 0 0
Direct pages scanned 21892772 0
Kswapd pages scanned 20015890 41879484
Kswapd pages reclaimed 19961986 41822072
Direct pages reclaimed 21892741 0
Compaction stalls 1065755 0
Compaction success 514 0
Compaction failures 1065241 0
Allocation stalls and all direct reclaim activity is eliminated as well
as compaction-related stalls.
THP gives impressive gains in some cases but only if they are quickly
available. We're not going to reach the point where they are completely
free so lets take the costs out of the fast paths finally and defer the
cost to kswapd, kcompactd and khugepaged where it belongs.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-18 00:19:23 +03:00
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
static ssize_t enabled_store ( struct kobject * kobj ,
struct kobj_attribute * attr ,
const char * buf , size_t count )
{
mm, thp: add new defer+madvise defrag option
There is no thp defrag option that currently allows MADV_HUGEPAGE
regions to do direct compaction and reclaim while all other thp
allocations simply trigger kswapd and kcompactd in the background and
fail immediately.
The "defer" setting simply triggers background reclaim and compaction
for all regions, regardless of MADV_HUGEPAGE, which makes it unusable
for our userspace where MADV_HUGEPAGE is being used to indicate the
application is willing to wait for work for thp memory to be available.
The "madvise" setting will do direct compaction and reclaim for these
MADV_HUGEPAGE regions, but does not trigger kswapd and kcompactd in the
background for anybody else.
For reasonable usage, there needs to be a mesh between the two options.
This patch introduces a fifth mode, "defer+madvise", that will do direct
reclaim and compaction for MADV_HUGEPAGE regions and trigger background
reclaim and compaction for everybody else so that hugepages may be
available in the near future.
A proposal to allow direct reclaim and compaction for MADV_HUGEPAGE
regions as part of the "defer" mode, making it a very powerful setting
and avoids breaking userspace, was offered:
http://marc.info/?t=148236612700003
This additional mode is a compromise.
A second proposal to allow both "defer" and "madvise" to be selected at
the same time was also offered:
http://marc.info/?t=148357345300001.
This is possible, but there was a concern that it might break existing
userspaces the parse the output of the defrag mode, so the fifth option
was introduced instead.
This patch also cleans up the helper function for storing to "enabled"
and "defrag" since the former supports three modes while the latter
supports five and triple_flag_store() was getting unnecessarily messy.
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1701101614330.41805@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 02:45:49 +03:00
ssize_t ret = count ;
2011-01-14 02:46:58 +03:00
2020-01-31 09:14:48 +03:00
if ( sysfs_streq ( buf , " always " ) ) {
mm, thp: add new defer+madvise defrag option
There is no thp defrag option that currently allows MADV_HUGEPAGE
regions to do direct compaction and reclaim while all other thp
allocations simply trigger kswapd and kcompactd in the background and
fail immediately.
The "defer" setting simply triggers background reclaim and compaction
for all regions, regardless of MADV_HUGEPAGE, which makes it unusable
for our userspace where MADV_HUGEPAGE is being used to indicate the
application is willing to wait for work for thp memory to be available.
The "madvise" setting will do direct compaction and reclaim for these
MADV_HUGEPAGE regions, but does not trigger kswapd and kcompactd in the
background for anybody else.
For reasonable usage, there needs to be a mesh between the two options.
This patch introduces a fifth mode, "defer+madvise", that will do direct
reclaim and compaction for MADV_HUGEPAGE regions and trigger background
reclaim and compaction for everybody else so that hugepages may be
available in the near future.
A proposal to allow direct reclaim and compaction for MADV_HUGEPAGE
regions as part of the "defer" mode, making it a very powerful setting
and avoids breaking userspace, was offered:
http://marc.info/?t=148236612700003
This additional mode is a compromise.
A second proposal to allow both "defer" and "madvise" to be selected at
the same time was also offered:
http://marc.info/?t=148357345300001.
This is possible, but there was a concern that it might break existing
userspaces the parse the output of the defrag mode, so the fifth option
was introduced instead.
This patch also cleans up the helper function for storing to "enabled"
and "defrag" since the former supports three modes while the latter
supports five and triple_flag_store() was getting unnecessarily messy.
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1701101614330.41805@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 02:45:49 +03:00
clear_bit ( TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG , & transparent_hugepage_flags ) ;
set_bit ( TRANSPARENT_HUGEPAGE_FLAG , & transparent_hugepage_flags ) ;
2020-01-31 09:14:48 +03:00
} else if ( sysfs_streq ( buf , " madvise " ) ) {
mm, thp: add new defer+madvise defrag option
There is no thp defrag option that currently allows MADV_HUGEPAGE
regions to do direct compaction and reclaim while all other thp
allocations simply trigger kswapd and kcompactd in the background and
fail immediately.
The "defer" setting simply triggers background reclaim and compaction
for all regions, regardless of MADV_HUGEPAGE, which makes it unusable
for our userspace where MADV_HUGEPAGE is being used to indicate the
application is willing to wait for work for thp memory to be available.
The "madvise" setting will do direct compaction and reclaim for these
MADV_HUGEPAGE regions, but does not trigger kswapd and kcompactd in the
background for anybody else.
For reasonable usage, there needs to be a mesh between the two options.
This patch introduces a fifth mode, "defer+madvise", that will do direct
reclaim and compaction for MADV_HUGEPAGE regions and trigger background
reclaim and compaction for everybody else so that hugepages may be
available in the near future.
A proposal to allow direct reclaim and compaction for MADV_HUGEPAGE
regions as part of the "defer" mode, making it a very powerful setting
and avoids breaking userspace, was offered:
http://marc.info/?t=148236612700003
This additional mode is a compromise.
A second proposal to allow both "defer" and "madvise" to be selected at
the same time was also offered:
http://marc.info/?t=148357345300001.
This is possible, but there was a concern that it might break existing
userspaces the parse the output of the defrag mode, so the fifth option
was introduced instead.
This patch also cleans up the helper function for storing to "enabled"
and "defrag" since the former supports three modes while the latter
supports five and triple_flag_store() was getting unnecessarily messy.
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1701101614330.41805@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 02:45:49 +03:00
clear_bit ( TRANSPARENT_HUGEPAGE_FLAG , & transparent_hugepage_flags ) ;
set_bit ( TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG , & transparent_hugepage_flags ) ;
2020-01-31 09:14:48 +03:00
} else if ( sysfs_streq ( buf , " never " ) ) {
mm, thp: add new defer+madvise defrag option
There is no thp defrag option that currently allows MADV_HUGEPAGE
regions to do direct compaction and reclaim while all other thp
allocations simply trigger kswapd and kcompactd in the background and
fail immediately.
The "defer" setting simply triggers background reclaim and compaction
for all regions, regardless of MADV_HUGEPAGE, which makes it unusable
for our userspace where MADV_HUGEPAGE is being used to indicate the
application is willing to wait for work for thp memory to be available.
The "madvise" setting will do direct compaction and reclaim for these
MADV_HUGEPAGE regions, but does not trigger kswapd and kcompactd in the
background for anybody else.
For reasonable usage, there needs to be a mesh between the two options.
This patch introduces a fifth mode, "defer+madvise", that will do direct
reclaim and compaction for MADV_HUGEPAGE regions and trigger background
reclaim and compaction for everybody else so that hugepages may be
available in the near future.
A proposal to allow direct reclaim and compaction for MADV_HUGEPAGE
regions as part of the "defer" mode, making it a very powerful setting
and avoids breaking userspace, was offered:
http://marc.info/?t=148236612700003
This additional mode is a compromise.
A second proposal to allow both "defer" and "madvise" to be selected at
the same time was also offered:
http://marc.info/?t=148357345300001.
This is possible, but there was a concern that it might break existing
userspaces the parse the output of the defrag mode, so the fifth option
was introduced instead.
This patch also cleans up the helper function for storing to "enabled"
and "defrag" since the former supports three modes while the latter
supports five and triple_flag_store() was getting unnecessarily messy.
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1701101614330.41805@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 02:45:49 +03:00
clear_bit ( TRANSPARENT_HUGEPAGE_FLAG , & transparent_hugepage_flags ) ;
clear_bit ( TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG , & transparent_hugepage_flags ) ;
} else
ret = - EINVAL ;
2011-01-14 02:46:58 +03:00
if ( ret > 0 ) {
2016-07-27 01:26:24 +03:00
int err = start_stop_khugepaged ( ) ;
2011-01-14 02:46:58 +03:00
if ( err )
ret = err ;
}
return ret ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
}
2022-07-04 16:21:53 +03:00
static struct kobj_attribute enabled_attr = __ATTR_RW ( enabled ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
2016-07-27 01:26:24 +03:00
ssize_t single_hugepage_flag_show ( struct kobject * kobj ,
2020-12-15 06:14:46 +03:00
struct kobj_attribute * attr , char * buf ,
enum transparent_hugepage_flag flag )
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
{
2020-12-15 06:14:46 +03:00
return sysfs_emit ( buf , " %d \n " ,
! ! test_bit ( flag , & transparent_hugepage_flags ) ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
}
2011-04-15 02:22:21 +04:00
2016-07-27 01:26:24 +03:00
ssize_t single_hugepage_flag_store ( struct kobject * kobj ,
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
struct kobj_attribute * attr ,
const char * buf , size_t count ,
enum transparent_hugepage_flag flag )
{
2011-04-15 02:22:21 +04:00
unsigned long value ;
int ret ;
ret = kstrtoul ( buf , 10 , & value ) ;
if ( ret < 0 )
return ret ;
if ( value > 1 )
return - EINVAL ;
if ( value )
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
set_bit ( flag , & transparent_hugepage_flags ) ;
2011-04-15 02:22:21 +04:00
else
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
clear_bit ( flag , & transparent_hugepage_flags ) ;
return count ;
}
static ssize_t defrag_show ( struct kobject * kobj ,
struct kobj_attribute * attr , char * buf )
{
2020-12-15 06:14:46 +03:00
const char * output ;
if ( test_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG ,
& transparent_hugepage_flags ) )
output = " [always] defer defer+madvise madvise never " ;
else if ( test_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG ,
& transparent_hugepage_flags ) )
output = " always [defer] defer+madvise madvise never " ;
else if ( test_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG ,
& transparent_hugepage_flags ) )
output = " always defer [defer+madvise] madvise never " ;
else if ( test_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG ,
& transparent_hugepage_flags ) )
output = " always defer defer+madvise [madvise] never " ;
else
output = " always defer defer+madvise madvise [never] " ;
return sysfs_emit ( buf , " %s \n " , output ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
}
mm, thp: add new defer+madvise defrag option
There is no thp defrag option that currently allows MADV_HUGEPAGE
regions to do direct compaction and reclaim while all other thp
allocations simply trigger kswapd and kcompactd in the background and
fail immediately.
The "defer" setting simply triggers background reclaim and compaction
for all regions, regardless of MADV_HUGEPAGE, which makes it unusable
for our userspace where MADV_HUGEPAGE is being used to indicate the
application is willing to wait for work for thp memory to be available.
The "madvise" setting will do direct compaction and reclaim for these
MADV_HUGEPAGE regions, but does not trigger kswapd and kcompactd in the
background for anybody else.
For reasonable usage, there needs to be a mesh between the two options.
This patch introduces a fifth mode, "defer+madvise", that will do direct
reclaim and compaction for MADV_HUGEPAGE regions and trigger background
reclaim and compaction for everybody else so that hugepages may be
available in the near future.
A proposal to allow direct reclaim and compaction for MADV_HUGEPAGE
regions as part of the "defer" mode, making it a very powerful setting
and avoids breaking userspace, was offered:
http://marc.info/?t=148236612700003
This additional mode is a compromise.
A second proposal to allow both "defer" and "madvise" to be selected at
the same time was also offered:
http://marc.info/?t=148357345300001.
This is possible, but there was a concern that it might break existing
userspaces the parse the output of the defrag mode, so the fifth option
was introduced instead.
This patch also cleans up the helper function for storing to "enabled"
and "defrag" since the former supports three modes while the latter
supports five and triple_flag_store() was getting unnecessarily messy.
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1701101614330.41805@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 02:45:49 +03:00
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
static ssize_t defrag_store ( struct kobject * kobj ,
struct kobj_attribute * attr ,
const char * buf , size_t count )
{
2020-01-31 09:14:48 +03:00
if ( sysfs_streq ( buf , " always " ) ) {
mm, thp: add new defer+madvise defrag option
There is no thp defrag option that currently allows MADV_HUGEPAGE
regions to do direct compaction and reclaim while all other thp
allocations simply trigger kswapd and kcompactd in the background and
fail immediately.
The "defer" setting simply triggers background reclaim and compaction
for all regions, regardless of MADV_HUGEPAGE, which makes it unusable
for our userspace where MADV_HUGEPAGE is being used to indicate the
application is willing to wait for work for thp memory to be available.
The "madvise" setting will do direct compaction and reclaim for these
MADV_HUGEPAGE regions, but does not trigger kswapd and kcompactd in the
background for anybody else.
For reasonable usage, there needs to be a mesh between the two options.
This patch introduces a fifth mode, "defer+madvise", that will do direct
reclaim and compaction for MADV_HUGEPAGE regions and trigger background
reclaim and compaction for everybody else so that hugepages may be
available in the near future.
A proposal to allow direct reclaim and compaction for MADV_HUGEPAGE
regions as part of the "defer" mode, making it a very powerful setting
and avoids breaking userspace, was offered:
http://marc.info/?t=148236612700003
This additional mode is a compromise.
A second proposal to allow both "defer" and "madvise" to be selected at
the same time was also offered:
http://marc.info/?t=148357345300001.
This is possible, but there was a concern that it might break existing
userspaces the parse the output of the defrag mode, so the fifth option
was introduced instead.
This patch also cleans up the helper function for storing to "enabled"
and "defrag" since the former supports three modes while the latter
supports five and triple_flag_store() was getting unnecessarily messy.
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1701101614330.41805@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 02:45:49 +03:00
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG , & transparent_hugepage_flags ) ;
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG , & transparent_hugepage_flags ) ;
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG , & transparent_hugepage_flags ) ;
set_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG , & transparent_hugepage_flags ) ;
2020-01-31 09:14:48 +03:00
} else if ( sysfs_streq ( buf , " defer+madvise " ) ) {
mm, thp: add new defer+madvise defrag option
There is no thp defrag option that currently allows MADV_HUGEPAGE
regions to do direct compaction and reclaim while all other thp
allocations simply trigger kswapd and kcompactd in the background and
fail immediately.
The "defer" setting simply triggers background reclaim and compaction
for all regions, regardless of MADV_HUGEPAGE, which makes it unusable
for our userspace where MADV_HUGEPAGE is being used to indicate the
application is willing to wait for work for thp memory to be available.
The "madvise" setting will do direct compaction and reclaim for these
MADV_HUGEPAGE regions, but does not trigger kswapd and kcompactd in the
background for anybody else.
For reasonable usage, there needs to be a mesh between the two options.
This patch introduces a fifth mode, "defer+madvise", that will do direct
reclaim and compaction for MADV_HUGEPAGE regions and trigger background
reclaim and compaction for everybody else so that hugepages may be
available in the near future.
A proposal to allow direct reclaim and compaction for MADV_HUGEPAGE
regions as part of the "defer" mode, making it a very powerful setting
and avoids breaking userspace, was offered:
http://marc.info/?t=148236612700003
This additional mode is a compromise.
A second proposal to allow both "defer" and "madvise" to be selected at
the same time was also offered:
http://marc.info/?t=148357345300001.
This is possible, but there was a concern that it might break existing
userspaces the parse the output of the defrag mode, so the fifth option
was introduced instead.
This patch also cleans up the helper function for storing to "enabled"
and "defrag" since the former supports three modes while the latter
supports five and triple_flag_store() was getting unnecessarily messy.
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1701101614330.41805@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 02:45:49 +03:00
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG , & transparent_hugepage_flags ) ;
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG , & transparent_hugepage_flags ) ;
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG , & transparent_hugepage_flags ) ;
set_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG , & transparent_hugepage_flags ) ;
2020-01-31 09:14:48 +03:00
} else if ( sysfs_streq ( buf , " defer " ) ) {
2017-04-08 02:04:54 +03:00
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG , & transparent_hugepage_flags ) ;
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG , & transparent_hugepage_flags ) ;
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG , & transparent_hugepage_flags ) ;
set_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG , & transparent_hugepage_flags ) ;
2020-01-31 09:14:48 +03:00
} else if ( sysfs_streq ( buf , " madvise " ) ) {
mm, thp: add new defer+madvise defrag option
There is no thp defrag option that currently allows MADV_HUGEPAGE
regions to do direct compaction and reclaim while all other thp
allocations simply trigger kswapd and kcompactd in the background and
fail immediately.
The "defer" setting simply triggers background reclaim and compaction
for all regions, regardless of MADV_HUGEPAGE, which makes it unusable
for our userspace where MADV_HUGEPAGE is being used to indicate the
application is willing to wait for work for thp memory to be available.
The "madvise" setting will do direct compaction and reclaim for these
MADV_HUGEPAGE regions, but does not trigger kswapd and kcompactd in the
background for anybody else.
For reasonable usage, there needs to be a mesh between the two options.
This patch introduces a fifth mode, "defer+madvise", that will do direct
reclaim and compaction for MADV_HUGEPAGE regions and trigger background
reclaim and compaction for everybody else so that hugepages may be
available in the near future.
A proposal to allow direct reclaim and compaction for MADV_HUGEPAGE
regions as part of the "defer" mode, making it a very powerful setting
and avoids breaking userspace, was offered:
http://marc.info/?t=148236612700003
This additional mode is a compromise.
A second proposal to allow both "defer" and "madvise" to be selected at
the same time was also offered:
http://marc.info/?t=148357345300001.
This is possible, but there was a concern that it might break existing
userspaces the parse the output of the defrag mode, so the fifth option
was introduced instead.
This patch also cleans up the helper function for storing to "enabled"
and "defrag" since the former supports three modes while the latter
supports five and triple_flag_store() was getting unnecessarily messy.
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1701101614330.41805@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 02:45:49 +03:00
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG , & transparent_hugepage_flags ) ;
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG , & transparent_hugepage_flags ) ;
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG , & transparent_hugepage_flags ) ;
set_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG , & transparent_hugepage_flags ) ;
2020-01-31 09:14:48 +03:00
} else if ( sysfs_streq ( buf , " never " ) ) {
mm, thp: add new defer+madvise defrag option
There is no thp defrag option that currently allows MADV_HUGEPAGE
regions to do direct compaction and reclaim while all other thp
allocations simply trigger kswapd and kcompactd in the background and
fail immediately.
The "defer" setting simply triggers background reclaim and compaction
for all regions, regardless of MADV_HUGEPAGE, which makes it unusable
for our userspace where MADV_HUGEPAGE is being used to indicate the
application is willing to wait for work for thp memory to be available.
The "madvise" setting will do direct compaction and reclaim for these
MADV_HUGEPAGE regions, but does not trigger kswapd and kcompactd in the
background for anybody else.
For reasonable usage, there needs to be a mesh between the two options.
This patch introduces a fifth mode, "defer+madvise", that will do direct
reclaim and compaction for MADV_HUGEPAGE regions and trigger background
reclaim and compaction for everybody else so that hugepages may be
available in the near future.
A proposal to allow direct reclaim and compaction for MADV_HUGEPAGE
regions as part of the "defer" mode, making it a very powerful setting
and avoids breaking userspace, was offered:
http://marc.info/?t=148236612700003
This additional mode is a compromise.
A second proposal to allow both "defer" and "madvise" to be selected at
the same time was also offered:
http://marc.info/?t=148357345300001.
This is possible, but there was a concern that it might break existing
userspaces the parse the output of the defrag mode, so the fifth option
was introduced instead.
This patch also cleans up the helper function for storing to "enabled"
and "defrag" since the former supports three modes while the latter
supports five and triple_flag_store() was getting unnecessarily messy.
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1701101614330.41805@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 02:45:49 +03:00
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG , & transparent_hugepage_flags ) ;
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG , & transparent_hugepage_flags ) ;
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG , & transparent_hugepage_flags ) ;
clear_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG , & transparent_hugepage_flags ) ;
} else
return - EINVAL ;
return count ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
}
2022-07-04 16:21:53 +03:00
static struct kobj_attribute defrag_attr = __ATTR_RW ( defrag ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
2012-12-13 01:51:12 +04:00
static ssize_t use_zero_page_show ( struct kobject * kobj ,
2020-12-15 06:14:42 +03:00
struct kobj_attribute * attr , char * buf )
2012-12-13 01:51:12 +04:00
{
2016-07-27 01:26:24 +03:00
return single_hugepage_flag_show ( kobj , attr , buf ,
2020-12-15 06:14:42 +03:00
TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG ) ;
2012-12-13 01:51:12 +04:00
}
static ssize_t use_zero_page_store ( struct kobject * kobj ,
struct kobj_attribute * attr , const char * buf , size_t count )
{
2016-07-27 01:26:24 +03:00
return single_hugepage_flag_store ( kobj , attr , buf , count ,
2012-12-13 01:51:12 +04:00
TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG ) ;
}
2022-07-04 16:21:53 +03:00
static struct kobj_attribute use_zero_page_attr = __ATTR_RW ( use_zero_page ) ;
2016-12-13 03:44:50 +03:00
static ssize_t hpage_pmd_size_show ( struct kobject * kobj ,
2020-12-15 06:14:42 +03:00
struct kobj_attribute * attr , char * buf )
2016-12-13 03:44:50 +03:00
{
2020-12-15 06:14:42 +03:00
return sysfs_emit ( buf , " %lu \n " , HPAGE_PMD_SIZE ) ;
2016-12-13 03:44:50 +03:00
}
static struct kobj_attribute hpage_pmd_size_attr =
__ATTR_RO ( hpage_pmd_size ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
static struct attribute * hugepage_attr [ ] = {
& enabled_attr . attr ,
& defrag_attr . attr ,
2012-12-13 01:51:12 +04:00
& use_zero_page_attr . attr ,
2016-12-13 03:44:50 +03:00
& hpage_pmd_size_attr . attr ,
2020-04-07 06:04:35 +03:00
# ifdef CONFIG_SHMEM
shmem: prepare huge= mount option and sysfs knob
This patch adds new mount option "huge=". It can have following values:
- "always":
Attempt to allocate huge pages every time we need a new page;
- "never":
Do not allocate huge pages;
- "within_size":
Only allocate huge page if it will be fully within i_size.
Also respect fadvise()/madvise() hints;
- "advise:
Only allocate huge pages if requested with fadvise()/madvise();
Default is "never" for now.
"mount -o remount,huge= /mountpoint" works fine after mount: remounting
huge=never will not attempt to break up huge pages at all, just stop
more from being allocated.
No new config option: put this under CONFIG_TRANSPARENT_HUGEPAGE, which
is the appropriate option to protect those who don't want the new bloat,
and with which we shall share some pmd code.
Prohibit the option when !CONFIG_TRANSPARENT_HUGEPAGE, just as mpol is
invalid without CONFIG_NUMA (was hidden in mpol_parse_str(): make it
explicit).
Allow enabling THP only if the machine has_transparent_hugepage().
But what about Shmem with no user-visible mount? SysV SHM, memfds,
shared anonymous mmaps (of /dev/zero or MAP_ANONYMOUS), GPU drivers' DRM
objects, Ashmem. Though unlikely to suit all usages, provide sysfs knob
/sys/kernel/mm/transparent_hugepage/shmem_enabled to experiment with
huge on those.
And allow shmem_enabled two further values:
- "deny":
For use in emergencies, to force the huge option off from
all mounts;
- "force":
Force the huge option on for all - very useful for testing;
Based on patch by Hugh Dickins.
Link: http://lkml.kernel.org/r/1466021202-61880-28-git-send-email-kirill.shutemov@linux.intel.com
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-27 01:26:13 +03:00
& shmem_enabled_attr . attr ,
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
# endif
NULL ,
} ;
2017-09-07 02:22:03 +03:00
static const struct attribute_group hugepage_attr_group = {
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
. attrs = hugepage_attr ,
2011-01-14 02:46:58 +03:00
} ;
2012-01-13 05:19:11 +04:00
static int __init hugepage_init_sysfs ( struct kobject * * hugepage_kobj )
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
{
int err ;
2012-01-13 05:19:11 +04:00
* hugepage_kobj = kobject_create_and_add ( " transparent_hugepage " , mm_kobj ) ;
if ( unlikely ( ! * hugepage_kobj ) ) {
2014-06-05 03:06:58 +04:00
pr_err ( " failed to create transparent hugepage kobject \n " ) ;
2012-01-13 05:19:11 +04:00
return - ENOMEM ;
2011-01-14 02:46:58 +03:00
}
2012-01-13 05:19:11 +04:00
err = sysfs_create_group ( * hugepage_kobj , & hugepage_attr_group ) ;
2011-01-14 02:46:58 +03:00
if ( err ) {
2014-06-05 03:06:58 +04:00
pr_err ( " failed to register transparent hugepage group \n " ) ;
2012-01-13 05:19:11 +04:00
goto delete_obj ;
2011-01-14 02:46:58 +03:00
}
2012-01-13 05:19:11 +04:00
err = sysfs_create_group ( * hugepage_kobj , & khugepaged_attr_group ) ;
2011-01-14 02:46:58 +03:00
if ( err ) {
2014-06-05 03:06:58 +04:00
pr_err ( " failed to register transparent hugepage group \n " ) ;
2012-01-13 05:19:11 +04:00
goto remove_hp_group ;
2011-01-14 02:46:58 +03:00
}
2012-01-13 05:19:11 +04:00
return 0 ;
remove_hp_group :
sysfs_remove_group ( * hugepage_kobj , & hugepage_attr_group ) ;
delete_obj :
kobject_put ( * hugepage_kobj ) ;
return err ;
}
static void __init hugepage_exit_sysfs ( struct kobject * hugepage_kobj )
{
sysfs_remove_group ( hugepage_kobj , & khugepaged_attr_group ) ;
sysfs_remove_group ( hugepage_kobj , & hugepage_attr_group ) ;
kobject_put ( hugepage_kobj ) ;
}
# else
static inline int hugepage_init_sysfs ( struct kobject * * hugepage_kobj )
{
return 0 ;
}
static inline void hugepage_exit_sysfs ( struct kobject * hugepage_kobj )
{
}
# endif /* CONFIG_SYSFS */
static int __init hugepage_init ( void )
{
int err ;
struct kobject * hugepage_kobj ;
if ( ! has_transparent_hugepage ( ) ) {
2021-02-24 23:07:32 +03:00
/*
* Hardware doesn ' t support hugepages , hence disable
* DAX PMD support .
*/
transparent_hugepage_flags = 1 < < TRANSPARENT_HUGEPAGE_NEVER_DAX ;
2012-01-13 05:19:11 +04:00
return - EINVAL ;
}
2016-03-01 07:15:14 +03:00
/*
* hugepages can ' t be allocated by the buddy allocator
*/
MAYBE_BUILD_BUG_ON ( HPAGE_PMD_ORDER > = MAX_ORDER ) ;
/*
* we use page - > mapping and page - > index in second tail page
* as list_head : assuming THP order > = 2
*/
MAYBE_BUILD_BUG_ON ( HPAGE_PMD_ORDER < 2 ) ;
2012-01-13 05:19:11 +04:00
err = hugepage_init_sysfs ( & hugepage_kobj ) ;
if ( err )
2015-04-16 02:14:20 +03:00
goto err_sysfs ;
2011-01-14 02:46:58 +03:00
2016-07-27 01:26:24 +03:00
err = khugepaged_init ( ) ;
2011-01-14 02:46:58 +03:00
if ( err )
2015-04-16 02:14:20 +03:00
goto err_slab ;
2011-01-14 02:46:58 +03:00
2022-06-01 06:22:24 +03:00
err = register_shrinker ( & huge_zero_page_shrinker , " thp-zero " ) ;
2015-04-16 02:14:20 +03:00
if ( err )
goto err_hzp_shrinker ;
2022-06-01 06:22:24 +03:00
err = register_shrinker ( & deferred_split_shrinker , " thp-deferred_split " ) ;
2016-01-16 03:54:17 +03:00
if ( err )
goto err_split_shrinker ;
2012-12-13 01:51:06 +04:00
2011-01-14 02:47:12 +03:00
/*
* By default disable transparent hugepages on smaller systems ,
* where the extra memory used could hurt more than TLB overhead
* is likely to save . The admin can still enable it through / sys .
*/
2018-12-28 11:34:29 +03:00
if ( totalram_pages ( ) < ( 512 < < ( 20 - PAGE_SHIFT ) ) ) {
2011-01-14 02:47:12 +03:00
transparent_hugepage_flags = 0 ;
2015-04-16 02:14:56 +03:00
return 0 ;
}
2011-01-14 02:47:12 +03:00
2015-04-16 02:14:56 +03:00
err = start_stop_khugepaged ( ) ;
2015-04-16 02:14:20 +03:00
if ( err )
goto err_khugepaged ;
2011-01-14 02:46:58 +03:00
2012-01-13 05:19:11 +04:00
return 0 ;
2015-04-16 02:14:20 +03:00
err_khugepaged :
2016-01-16 03:54:17 +03:00
unregister_shrinker ( & deferred_split_shrinker ) ;
err_split_shrinker :
2015-04-16 02:14:20 +03:00
unregister_shrinker ( & huge_zero_page_shrinker ) ;
err_hzp_shrinker :
2016-07-27 01:26:24 +03:00
khugepaged_destroy ( ) ;
2015-04-16 02:14:20 +03:00
err_slab :
2012-01-13 05:19:11 +04:00
hugepage_exit_sysfs ( hugepage_kobj ) ;
2015-04-16 02:14:20 +03:00
err_sysfs :
2011-01-14 02:46:58 +03:00
return err ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
}
2014-01-24 03:53:30 +04:00
subsys_initcall ( hugepage_init ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
static int __init setup_transparent_hugepage ( char * str )
{
int ret = 0 ;
if ( ! str )
goto out ;
if ( ! strcmp ( str , " always " ) ) {
set_bit ( TRANSPARENT_HUGEPAGE_FLAG ,
& transparent_hugepage_flags ) ;
clear_bit ( TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG ,
& transparent_hugepage_flags ) ;
ret = 1 ;
} else if ( ! strcmp ( str , " madvise " ) ) {
clear_bit ( TRANSPARENT_HUGEPAGE_FLAG ,
& transparent_hugepage_flags ) ;
set_bit ( TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG ,
& transparent_hugepage_flags ) ;
ret = 1 ;
} else if ( ! strcmp ( str , " never " ) ) {
clear_bit ( TRANSPARENT_HUGEPAGE_FLAG ,
& transparent_hugepage_flags ) ;
clear_bit ( TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG ,
& transparent_hugepage_flags ) ;
ret = 1 ;
}
out :
if ( ! ret )
2014-06-05 03:06:58 +04:00
pr_warn ( " transparent_hugepage= cannot parse, ignored \n " ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
return ret ;
}
__setup ( " transparent_hugepage= " , setup_transparent_hugepage ) ;
2017-11-29 20:01:01 +03:00
pmd_t maybe_pmd_mkwrite ( pmd_t pmd , struct vm_area_struct * vma )
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
{
2017-11-29 20:01:01 +03:00
if ( likely ( vma - > vm_flags & VM_WRITE ) )
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
pmd = pmd_mkwrite ( pmd ) ;
return pmd ;
}
2019-09-24 01:38:15 +03:00
# ifdef CONFIG_MEMCG
static inline struct deferred_split * get_deferred_split_queue ( struct page * page )
2016-01-16 03:54:17 +03:00
{
mm: memcontrol: Use helpers to read page's memcg data
Patch series "mm: allow mapping accounted kernel pages to userspace", v6.
Currently a non-slab kernel page which has been charged to a memory cgroup
can't be mapped to userspace. The underlying reason is simple: PageKmemcg
flag is defined as a page type (like buddy, offline, etc), so it takes a
bit from a page->mapped counter. Pages with a type set can't be mapped to
userspace.
But in general the kmemcg flag has nothing to do with mapping to
userspace. It only means that the page has been accounted by the page
allocator, so it has to be properly uncharged on release.
Some bpf maps are mapping the vmalloc-based memory to userspace, and their
memory can't be accounted because of this implementation detail.
This patchset removes this limitation by moving the PageKmemcg flag into
one of the free bits of the page->mem_cgroup pointer. Also it formalizes
accesses to the page->mem_cgroup and page->obj_cgroups using new helpers,
adds several checks and removes a couple of obsolete functions. As the
result the code became more robust with fewer open-coded bit tricks.
This patch (of 4):
Currently there are many open-coded reads of the page->mem_cgroup pointer,
as well as a couple of read helpers, which are barely used.
It creates an obstacle on a way to reuse some bits of the pointer for
storing additional bits of information. In fact, we already do this for
slab pages, where the last bit indicates that a pointer has an attached
vector of objcg pointers instead of a regular memcg pointer.
This commits uses 2 existing helpers and introduces a new helper to
converts all read sides to calls of these helpers:
struct mem_cgroup *page_memcg(struct page *page);
struct mem_cgroup *page_memcg_rcu(struct page *page);
struct mem_cgroup *page_memcg_check(struct page *page);
page_memcg_check() is intended to be used in cases when the page can be a
slab page and have a memcg pointer pointing at objcg vector. It does
check the lowest bit, and if set, returns NULL. page_memcg() contains a
VM_BUG_ON_PAGE() check for the page not being a slab page.
To make sure nobody uses a direct access, struct page's
mem_cgroup/obj_cgroups is converted to unsigned long memcg_data.
Signed-off-by: Roman Gushchin <guro@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Link: https://lkml.kernel.org/r/20201027001657.3398190-1-guro@fb.com
Link: https://lkml.kernel.org/r/20201027001657.3398190-2-guro@fb.com
Link: https://lore.kernel.org/bpf/20201201215900.3569844-2-guro@fb.com
2020-12-02 00:58:27 +03:00
struct mem_cgroup * memcg = page_memcg ( compound_head ( page ) ) ;
2019-09-24 01:38:15 +03:00
struct pglist_data * pgdat = NODE_DATA ( page_to_nid ( page ) ) ;
if ( memcg )
return & memcg - > deferred_split_queue ;
else
return & pgdat - > deferred_split_queue ;
2016-01-16 03:54:17 +03:00
}
2019-09-24 01:38:15 +03:00
# else
static inline struct deferred_split * get_deferred_split_queue ( struct page * page )
{
struct pglist_data * pgdat = NODE_DATA ( page_to_nid ( page ) ) ;
return & pgdat - > deferred_split_queue ;
}
# endif
2016-01-16 03:54:17 +03:00
void prep_transhuge_page ( struct page * page )
{
/*
2022-07-04 16:21:59 +03:00
* we use page - > mapping and page - > index in second tail page
2016-01-16 03:54:17 +03:00
* as list_head : assuming THP order > = 2
*/
INIT_LIST_HEAD ( page_deferred_list ( page ) ) ;
set_compound_page_dtor ( page , TRANSHUGE_PAGE_DTOR ) ;
}
2022-03-25 04:13:27 +03:00
static inline bool is_transparent_hugepage ( struct page * page )
2020-01-08 23:24:36 +03:00
{
if ( ! PageCompound ( page ) )
2020-06-05 02:49:46 +03:00
return false ;
2020-01-08 23:24:36 +03:00
page = compound_head ( page ) ;
return is_huge_zero_page ( page ) | |
page [ 1 ] . compound_dtor = = TRANSHUGE_PAGE_DTOR ;
}
2020-01-14 03:29:10 +03:00
static unsigned long __thp_get_unmapped_area ( struct file * filp ,
unsigned long addr , unsigned long len ,
2016-10-08 02:59:56 +03:00
loff_t off , unsigned long flags , unsigned long size )
{
loff_t off_end = off + len ;
loff_t off_align = round_up ( off , size ) ;
2020-01-14 03:29:10 +03:00
unsigned long len_pad , ret ;
2016-10-08 02:59:56 +03:00
if ( off_end < = off_align | | ( off_end - off_align ) < size )
return 0 ;
len_pad = len + size ;
if ( len_pad < len | | ( off + len_pad ) < off )
return 0 ;
2020-01-14 03:29:10 +03:00
ret = current - > mm - > get_unmapped_area ( filp , addr , len_pad ,
2016-10-08 02:59:56 +03:00
off > > PAGE_SHIFT , flags ) ;
2020-01-14 03:29:10 +03:00
/*
* The failure might be due to length padding . The caller will retry
* without the padding .
*/
if ( IS_ERR_VALUE ( ret ) )
2016-10-08 02:59:56 +03:00
return 0 ;
2020-01-14 03:29:10 +03:00
/*
* Do not try to align to THP boundary if allocation at the address
* hint succeeds .
*/
if ( ret = = addr )
return addr ;
ret + = ( off - ret ) & ( size - 1 ) ;
return ret ;
2016-10-08 02:59:56 +03:00
}
unsigned long thp_get_unmapped_area ( struct file * filp , unsigned long addr ,
unsigned long len , unsigned long pgoff , unsigned long flags )
{
2020-01-14 03:29:10 +03:00
unsigned long ret ;
2016-10-08 02:59:56 +03:00
loff_t off = ( loff_t ) pgoff < < PAGE_SHIFT ;
2020-01-14 03:29:10 +03:00
ret = __thp_get_unmapped_area ( filp , addr , len , off , flags , PMD_SIZE ) ;
if ( ret )
return ret ;
2019-09-22 15:43:15 +03:00
2016-10-08 02:59:56 +03:00
return current - > mm - > get_unmapped_area ( filp , addr , len , pgoff , flags ) ;
}
EXPORT_SYMBOL_GPL ( thp_get_unmapped_area ) ;
2018-08-24 03:01:36 +03:00
static vm_fault_t __do_huge_pmd_anonymous_page ( struct vm_fault * vmf ,
struct page * page , gfp_t gfp )
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
{
2016-12-15 02:06:58 +03:00
struct vm_area_struct * vma = vmf - > vma ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
pgtable_t pgtable ;
2016-12-15 02:06:58 +03:00
unsigned long haddr = vmf - > address & HPAGE_PMD_MASK ;
2018-08-24 03:01:36 +03:00
vm_fault_t ret = 0 ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
2014-01-24 03:52:54 +04:00
VM_BUG_ON_PAGE ( ! PageCompound ( page ) , page ) ;
mm: memcontrol: rewrite charge API
These patches rework memcg charge lifetime to integrate more naturally
with the lifetime of user pages. This drastically simplifies the code and
reduces charging and uncharging overhead. The most expensive part of
charging and uncharging is the page_cgroup bit spinlock, which is removed
entirely after this series.
Here are the top-10 profile entries of a stress test that reads a 128G
sparse file on a freshly booted box, without even a dedicated cgroup (i.e.
executing in the root memcg). Before:
15.36% cat [kernel.kallsyms] [k] copy_user_generic_string
13.31% cat [kernel.kallsyms] [k] memset
11.48% cat [kernel.kallsyms] [k] do_mpage_readpage
4.23% cat [kernel.kallsyms] [k] get_page_from_freelist
2.38% cat [kernel.kallsyms] [k] put_page
2.32% cat [kernel.kallsyms] [k] __mem_cgroup_commit_charge
2.18% kswapd0 [kernel.kallsyms] [k] __mem_cgroup_uncharge_common
1.92% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.86% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.62% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
After:
15.67% cat [kernel.kallsyms] [k] copy_user_generic_string
13.48% cat [kernel.kallsyms] [k] memset
11.42% cat [kernel.kallsyms] [k] do_mpage_readpage
3.98% cat [kernel.kallsyms] [k] get_page_from_freelist
2.46% cat [kernel.kallsyms] [k] put_page
2.13% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.88% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.67% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
1.39% kswapd0 [kernel.kallsyms] [k] free_pcppages_bulk
1.30% cat [kernel.kallsyms] [k] kfree
As you can see, the memcg footprint has shrunk quite a bit.
text data bss dec hex filename
37970 9892 400 48262 bc86 mm/memcontrol.o.old
35239 9892 400 45531 b1db mm/memcontrol.o
This patch (of 4):
The memcg charge API charges pages before they are rmapped - i.e. have an
actual "type" - and so every callsite needs its own set of charge and
uncharge functions to know what type is being operated on. Worse,
uncharge has to happen from a context that is still type-specific, rather
than at the end of the page's lifetime with exclusive access, and so
requires a lot of synchronization.
Rewrite the charge API to provide a generic set of try_charge(),
commit_charge() and cancel_charge() transaction operations, much like
what's currently done for swap-in:
mem_cgroup_try_charge() attempts to reserve a charge, reclaiming
pages from the memcg if necessary.
mem_cgroup_commit_charge() commits the page to the charge once it
has a valid page->mapping and PageAnon() reliably tells the type.
mem_cgroup_cancel_charge() aborts the transaction.
This reduces the charge API and enables subsequent patches to
drastically simplify uncharging.
As pages need to be committed after rmap is established but before they
are added to the LRU, page_add_new_anon_rmap() must stop doing LRU
additions again. Revive lru_cache_add_active_or_unevictable().
[hughd@google.com: fix shmem_unuse]
[hughd@google.com: Add comments on the private use of -EAGAIN]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-09 01:19:20 +04:00
2021-06-25 16:27:04 +03:00
if ( mem_cgroup_charge ( page_folio ( page ) , vma - > vm_mm , gfp ) ) {
2015-09-05 01:46:20 +03:00
put_page ( page ) ;
count_vm_event ( THP_FAULT_FALLBACK ) ;
2020-04-07 06:04:28 +03:00
count_vm_event ( THP_FAULT_FALLBACK_CHARGE ) ;
2015-09-05 01:46:20 +03:00
return VM_FAULT_FALLBACK ;
}
2020-06-04 02:02:04 +03:00
cgroup_throttle_swaprate ( page , gfp ) ;
mm: memcontrol: rewrite charge API
These patches rework memcg charge lifetime to integrate more naturally
with the lifetime of user pages. This drastically simplifies the code and
reduces charging and uncharging overhead. The most expensive part of
charging and uncharging is the page_cgroup bit spinlock, which is removed
entirely after this series.
Here are the top-10 profile entries of a stress test that reads a 128G
sparse file on a freshly booted box, without even a dedicated cgroup (i.e.
executing in the root memcg). Before:
15.36% cat [kernel.kallsyms] [k] copy_user_generic_string
13.31% cat [kernel.kallsyms] [k] memset
11.48% cat [kernel.kallsyms] [k] do_mpage_readpage
4.23% cat [kernel.kallsyms] [k] get_page_from_freelist
2.38% cat [kernel.kallsyms] [k] put_page
2.32% cat [kernel.kallsyms] [k] __mem_cgroup_commit_charge
2.18% kswapd0 [kernel.kallsyms] [k] __mem_cgroup_uncharge_common
1.92% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.86% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.62% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
After:
15.67% cat [kernel.kallsyms] [k] copy_user_generic_string
13.48% cat [kernel.kallsyms] [k] memset
11.42% cat [kernel.kallsyms] [k] do_mpage_readpage
3.98% cat [kernel.kallsyms] [k] get_page_from_freelist
2.46% cat [kernel.kallsyms] [k] put_page
2.13% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.88% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.67% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
1.39% kswapd0 [kernel.kallsyms] [k] free_pcppages_bulk
1.30% cat [kernel.kallsyms] [k] kfree
As you can see, the memcg footprint has shrunk quite a bit.
text data bss dec hex filename
37970 9892 400 48262 bc86 mm/memcontrol.o.old
35239 9892 400 45531 b1db mm/memcontrol.o
This patch (of 4):
The memcg charge API charges pages before they are rmapped - i.e. have an
actual "type" - and so every callsite needs its own set of charge and
uncharge functions to know what type is being operated on. Worse,
uncharge has to happen from a context that is still type-specific, rather
than at the end of the page's lifetime with exclusive access, and so
requires a lot of synchronization.
Rewrite the charge API to provide a generic set of try_charge(),
commit_charge() and cancel_charge() transaction operations, much like
what's currently done for swap-in:
mem_cgroup_try_charge() attempts to reserve a charge, reclaiming
pages from the memcg if necessary.
mem_cgroup_commit_charge() commits the page to the charge once it
has a valid page->mapping and PageAnon() reliably tells the type.
mem_cgroup_cancel_charge() aborts the transaction.
This reduces the charge API and enables subsequent patches to
drastically simplify uncharging.
As pages need to be committed after rmap is established but before they
are added to the LRU, page_add_new_anon_rmap() must stop doing LRU
additions again. Revive lru_cache_add_active_or_unevictable().
[hughd@google.com: fix shmem_unuse]
[hughd@google.com: Add comments on the private use of -EAGAIN]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-09 01:19:20 +04:00
mm: treewide: remove unused address argument from pte_alloc functions
Patch series "Add support for fast mremap".
This series speeds up the mremap(2) syscall by copying page tables at
the PMD level even for non-THP systems. There is concern that the extra
'address' argument that mremap passes to pte_alloc may do something
subtle architecture related in the future that may make the scheme not
work. Also we find that there is no point in passing the 'address' to
pte_alloc since its unused. This patch therefore removes this argument
tree-wide resulting in a nice negative diff as well. Also ensuring
along the way that the enabled architectures do not do anything funky
with the 'address' argument that goes unnoticed by the optimization.
Build and boot tested on x86-64. Build tested on arm64. The config
enablement patch for arm64 will be posted in the future after more
testing.
The changes were obtained by applying the following Coccinelle script.
(thanks Julia for answering all Coccinelle questions!).
Following fix ups were done manually:
* Removal of address argument from pte_fragment_alloc
* Removal of pte_alloc_one_fast definitions from m68k and microblaze.
// Options: --include-headers --no-includes
// Note: I split the 'identifier fn' line, so if you are manually
// running it, please unsplit it so it runs for you.
virtual patch
@pte_alloc_func_def depends on patch exists@
identifier E2;
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
type T2;
@@
fn(...
- , T2 E2
)
{ ... }
@pte_alloc_func_proto_noarg depends on patch exists@
type T1, T2, T3, T4;
identifier fn =~ "^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
@@
(
- T3 fn(T1, T2);
+ T3 fn(T1);
|
- T3 fn(T1, T2, T4);
+ T3 fn(T1, T2);
)
@pte_alloc_func_proto depends on patch exists@
identifier E1, E2, E4;
type T1, T2, T3, T4;
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
@@
(
- T3 fn(T1 E1, T2 E2);
+ T3 fn(T1 E1);
|
- T3 fn(T1 E1, T2 E2, T4 E4);
+ T3 fn(T1 E1, T2 E2);
)
@pte_alloc_func_call depends on patch exists@
expression E2;
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
@@
fn(...
-, E2
)
@pte_alloc_macro depends on patch exists@
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
identifier a, b, c;
expression e;
position p;
@@
(
- #define fn(a, b, c) e
+ #define fn(a, b) e
|
- #define fn(a, b) e
+ #define fn(a) e
)
Link: http://lkml.kernel.org/r/20181108181201.88826-2-joelaf@google.com
Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Acked-by: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Julia Lawall <Julia.Lawall@lip6.fr>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: William Kucharski <william.kucharski@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-04 02:28:34 +03:00
pgtable = pte_alloc_one ( vma - > vm_mm ) ;
mm: memcontrol: rewrite charge API
These patches rework memcg charge lifetime to integrate more naturally
with the lifetime of user pages. This drastically simplifies the code and
reduces charging and uncharging overhead. The most expensive part of
charging and uncharging is the page_cgroup bit spinlock, which is removed
entirely after this series.
Here are the top-10 profile entries of a stress test that reads a 128G
sparse file on a freshly booted box, without even a dedicated cgroup (i.e.
executing in the root memcg). Before:
15.36% cat [kernel.kallsyms] [k] copy_user_generic_string
13.31% cat [kernel.kallsyms] [k] memset
11.48% cat [kernel.kallsyms] [k] do_mpage_readpage
4.23% cat [kernel.kallsyms] [k] get_page_from_freelist
2.38% cat [kernel.kallsyms] [k] put_page
2.32% cat [kernel.kallsyms] [k] __mem_cgroup_commit_charge
2.18% kswapd0 [kernel.kallsyms] [k] __mem_cgroup_uncharge_common
1.92% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.86% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.62% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
After:
15.67% cat [kernel.kallsyms] [k] copy_user_generic_string
13.48% cat [kernel.kallsyms] [k] memset
11.42% cat [kernel.kallsyms] [k] do_mpage_readpage
3.98% cat [kernel.kallsyms] [k] get_page_from_freelist
2.46% cat [kernel.kallsyms] [k] put_page
2.13% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.88% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.67% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
1.39% kswapd0 [kernel.kallsyms] [k] free_pcppages_bulk
1.30% cat [kernel.kallsyms] [k] kfree
As you can see, the memcg footprint has shrunk quite a bit.
text data bss dec hex filename
37970 9892 400 48262 bc86 mm/memcontrol.o.old
35239 9892 400 45531 b1db mm/memcontrol.o
This patch (of 4):
The memcg charge API charges pages before they are rmapped - i.e. have an
actual "type" - and so every callsite needs its own set of charge and
uncharge functions to know what type is being operated on. Worse,
uncharge has to happen from a context that is still type-specific, rather
than at the end of the page's lifetime with exclusive access, and so
requires a lot of synchronization.
Rewrite the charge API to provide a generic set of try_charge(),
commit_charge() and cancel_charge() transaction operations, much like
what's currently done for swap-in:
mem_cgroup_try_charge() attempts to reserve a charge, reclaiming
pages from the memcg if necessary.
mem_cgroup_commit_charge() commits the page to the charge once it
has a valid page->mapping and PageAnon() reliably tells the type.
mem_cgroup_cancel_charge() aborts the transaction.
This reduces the charge API and enables subsequent patches to
drastically simplify uncharging.
As pages need to be committed after rmap is established but before they
are added to the LRU, page_add_new_anon_rmap() must stop doing LRU
additions again. Revive lru_cache_add_active_or_unevictable().
[hughd@google.com: fix shmem_unuse]
[hughd@google.com: Add comments on the private use of -EAGAIN]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-09 01:19:20 +04:00
if ( unlikely ( ! pgtable ) ) {
2017-08-19 01:16:15 +03:00
ret = VM_FAULT_OOM ;
goto release ;
mm: memcontrol: rewrite charge API
These patches rework memcg charge lifetime to integrate more naturally
with the lifetime of user pages. This drastically simplifies the code and
reduces charging and uncharging overhead. The most expensive part of
charging and uncharging is the page_cgroup bit spinlock, which is removed
entirely after this series.
Here are the top-10 profile entries of a stress test that reads a 128G
sparse file on a freshly booted box, without even a dedicated cgroup (i.e.
executing in the root memcg). Before:
15.36% cat [kernel.kallsyms] [k] copy_user_generic_string
13.31% cat [kernel.kallsyms] [k] memset
11.48% cat [kernel.kallsyms] [k] do_mpage_readpage
4.23% cat [kernel.kallsyms] [k] get_page_from_freelist
2.38% cat [kernel.kallsyms] [k] put_page
2.32% cat [kernel.kallsyms] [k] __mem_cgroup_commit_charge
2.18% kswapd0 [kernel.kallsyms] [k] __mem_cgroup_uncharge_common
1.92% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.86% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.62% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
After:
15.67% cat [kernel.kallsyms] [k] copy_user_generic_string
13.48% cat [kernel.kallsyms] [k] memset
11.42% cat [kernel.kallsyms] [k] do_mpage_readpage
3.98% cat [kernel.kallsyms] [k] get_page_from_freelist
2.46% cat [kernel.kallsyms] [k] put_page
2.13% kswapd0 [kernel.kallsyms] [k] shrink_page_list
1.88% cat [kernel.kallsyms] [k] __radix_tree_lookup
1.67% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn
1.39% kswapd0 [kernel.kallsyms] [k] free_pcppages_bulk
1.30% cat [kernel.kallsyms] [k] kfree
As you can see, the memcg footprint has shrunk quite a bit.
text data bss dec hex filename
37970 9892 400 48262 bc86 mm/memcontrol.o.old
35239 9892 400 45531 b1db mm/memcontrol.o
This patch (of 4):
The memcg charge API charges pages before they are rmapped - i.e. have an
actual "type" - and so every callsite needs its own set of charge and
uncharge functions to know what type is being operated on. Worse,
uncharge has to happen from a context that is still type-specific, rather
than at the end of the page's lifetime with exclusive access, and so
requires a lot of synchronization.
Rewrite the charge API to provide a generic set of try_charge(),
commit_charge() and cancel_charge() transaction operations, much like
what's currently done for swap-in:
mem_cgroup_try_charge() attempts to reserve a charge, reclaiming
pages from the memcg if necessary.
mem_cgroup_commit_charge() commits the page to the charge once it
has a valid page->mapping and PageAnon() reliably tells the type.
mem_cgroup_cancel_charge() aborts the transaction.
This reduces the charge API and enables subsequent patches to
drastically simplify uncharging.
As pages need to be committed after rmap is established but before they
are added to the LRU, page_add_new_anon_rmap() must stop doing LRU
additions again. Revive lru_cache_add_active_or_unevictable().
[hughd@google.com: fix shmem_unuse]
[hughd@google.com: Add comments on the private use of -EAGAIN]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-09 01:19:20 +04:00
}
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
mm: hugetlb: clear target sub-page last when clearing huge page
Huge page helps to reduce TLB miss rate, but it has higher cache
footprint, sometimes this may cause some issue. For example, when
clearing huge page on x86_64 platform, the cache footprint is 2M. But
on a Xeon E5 v3 2699 CPU, there are 18 cores, 36 threads, and only 45M
LLC (last level cache). That is, in average, there are 2.5M LLC for
each core and 1.25M LLC for each thread.
If the cache pressure is heavy when clearing the huge page, and we clear
the huge page from the begin to the end, it is possible that the begin
of huge page is evicted from the cache after we finishing clearing the
end of the huge page. And it is possible for the application to access
the begin of the huge page after clearing the huge page.
To help the above situation, in this patch, when we clear a huge page,
the order to clear sub-pages is changed. In quite some situation, we
can get the address that the application will access after we clear the
huge page, for example, in a page fault handler. Instead of clearing
the huge page from begin to end, we will clear the sub-pages farthest
from the the sub-page to access firstly, and clear the sub-page to
access last. This will make the sub-page to access most cache-hot and
sub-pages around it more cache-hot too. If we cannot know the address
the application will access, the begin of the huge page is assumed to be
the the address the application will access.
With this patch, the throughput increases ~28.3% in vm-scalability
anon-w-seq test case with 72 processes on a 2 socket Xeon E5 v3 2699
system (36 cores, 72 threads). The test case creates 72 processes, each
process mmap a big anonymous memory area and writes to it from the begin
to the end. For each process, other processes could be seen as other
workload which generates heavy cache pressure. At the same time, the
cache miss rate reduced from ~33.4% to ~31.7%, the IPC (instruction per
cycle) increased from 0.56 to 0.74, and the time spent in user space is
reduced ~7.9%
Christopher Lameter suggests to clear bytes inside a sub-page from end
to begin too. But tests show no visible performance difference in the
tests. May because the size of page is small compared with the cache
size.
Thanks Andi Kleen to propose to use address to access to determine the
order of sub-pages to clear.
The hugetlbfs access address could be improved, will do that in another
patch.
[ying.huang@intel.com: improve readability of clear_huge_page()]
Link: http://lkml.kernel.org/r/20170830051842.1397-1-ying.huang@intel.com
Link: http://lkml.kernel.org/r/20170815014618.15842-1-ying.huang@intel.com
Suggested-by: Andi Kleen <andi.kleen@intel.com>
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Jan Kara <jack@suse.cz>
Reviewed-by: Michal Hocko <mhocko@suse.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Nadia Yvette Chambers <nyc@holomorphy.com>
Cc: Matthew Wilcox <mawilcox@microsoft.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Shaohua Li <shli@fb.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 02:25:04 +03:00
clear_huge_page ( page , vmf - > address , HPAGE_PMD_NR ) ;
2013-04-30 02:08:15 +04:00
/*
* The memory barrier inside __SetPageUptodate makes sure that
* clear_huge_page writes become visible before the set_pmd_at ( )
* write .
*/
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
__SetPageUptodate ( page ) ;
2016-12-15 02:06:58 +03:00
vmf - > ptl = pmd_lock ( vma - > vm_mm , vmf - > pmd ) ;
if ( unlikely ( ! pmd_none ( * vmf - > pmd ) ) ) {
2017-08-19 01:16:15 +03:00
goto unlock_release ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
} else {
pmd_t entry ;
2015-09-05 01:46:20 +03:00
2017-08-19 01:16:15 +03:00
ret = check_stable_address_space ( vma - > vm_mm ) ;
if ( ret )
goto unlock_release ;
2015-09-05 01:46:20 +03:00
/* Deliver the page fault to userland */
if ( userfaultfd_missing ( vma ) ) {
2016-12-15 02:06:58 +03:00
spin_unlock ( vmf - > ptl ) ;
2015-09-05 01:46:20 +03:00
put_page ( page ) ;
2016-07-27 01:25:20 +03:00
pte_free ( vma - > vm_mm , pgtable ) ;
2021-05-05 04:33:49 +03:00
ret = handle_userfault ( vmf , VM_UFFD_MISSING ) ;
VM_BUG_ON ( ret & VM_FAULT_FALLBACK ) ;
return ret ;
2015-09-05 01:46:20 +03:00
}
2013-09-13 02:14:01 +04:00
entry = mk_huge_pmd ( page , vma - > vm_page_prot ) ;
2017-11-29 20:01:01 +03:00
entry = maybe_pmd_mkwrite ( pmd_mkdirty ( entry ) , vma ) ;
2022-05-10 04:20:43 +03:00
page_add_new_anon_rmap ( page , vma , haddr ) ;
2020-08-12 04:30:40 +03:00
lru_cache_add_inactive_or_unevictable ( page , vma ) ;
2016-12-15 02:06:58 +03:00
pgtable_trans_huge_deposit ( vma - > vm_mm , vmf - > pmd , pgtable ) ;
set_pmd_at ( vma - > vm_mm , haddr , vmf - > pmd , entry ) ;
2021-02-24 23:06:42 +03:00
update_mmu_cache_pmd ( vma , vmf - > address , vmf - > pmd ) ;
2016-07-27 01:25:20 +03:00
add_mm_counter ( vma - > vm_mm , MM_ANONPAGES , HPAGE_PMD_NR ) ;
2017-11-16 04:35:37 +03:00
mm_inc_nr_ptes ( vma - > vm_mm ) ;
2016-12-15 02:06:58 +03:00
spin_unlock ( vmf - > ptl ) ;
2015-09-05 01:46:20 +03:00
count_vm_event ( THP_FAULT_ALLOC ) ;
2020-06-04 02:02:04 +03:00
count_memcg_event_mm ( vma - > vm_mm , THP_FAULT_ALLOC ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
}
2012-05-30 02:06:17 +04:00
return 0 ;
2017-08-19 01:16:15 +03:00
unlock_release :
spin_unlock ( vmf - > ptl ) ;
release :
if ( pgtable )
pte_free ( vma - > vm_mm , pgtable ) ;
put_page ( page ) ;
return ret ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
}
mm: thp: set THP defrag by default to madvise and add a stall-free defrag option
THP defrag is enabled by default to direct reclaim/compact but not wake
kswapd in the event of a THP allocation failure. The problem is that
THP allocation requests potentially enter reclaim/compaction. This
potentially incurs a severe stall that is not guaranteed to be offset by
reduced TLB misses. While there has been considerable effort to reduce
the impact of reclaim/compaction, it is still a high cost and workloads
that should fit in memory fail to do so. Specifically, a simple
anon/file streaming workload will enter direct reclaim on NUMA at least
even though the working set size is 80% of RAM. It's been years and
it's time to throw in the towel.
First, this patch defines THP defrag as follows;
madvise: A failed allocation will direct reclaim/compact if the application requests it
never: Neither reclaim/compact nor wake kswapd
defer: A failed allocation will wake kswapd/kcompactd
always: A failed allocation will direct reclaim/compact (historical behaviour)
khugepaged defrag will enter direct/reclaim but not wake kswapd.
Next it sets the default defrag option to be "madvise" to only enter
direct reclaim/compaction for applications that specifically requested
it.
Lastly, it removes a check from the page allocator slowpath that is
related to __GFP_THISNODE to allow "defer" to work. The callers that
really cares are slub/slab and they are updated accordingly. The slab
one may be surprising because it also corrects a comment as kswapd was
never woken up by that path.
This means that a THP fault will no longer stall for most applications
by default and the ideal for most users that get THP if they are
immediately available. There are still options for users that prefer a
stall at startup of a new application by either restoring historical
behaviour with "always" or pick a half-way point with "defer" where
kswapd does some of the work in the background and wakes kcompactd if
necessary. THP defrag for khugepaged remains enabled and will enter
direct/reclaim but no wakeup kswapd or kcompactd.
After this patch a THP allocation failure will quickly fallback and rely
on khugepaged to recover the situation at some time in the future. In
some cases, this will reduce THP usage but the benefit of THP is hard to
measure and not a universal win where as a stall to reclaim/compaction
is definitely measurable and can be painful.
The first test for this is using "usemem" to read a large file and write
a large anonymous mapping (to avoid the zero page) multiple times. The
total size of the mappings is 80% of RAM and the benchmark simply
measures how long it takes to complete. It uses multiple threads to see
if that is a factor. On UMA, the performance is almost identical so is
not reported but on NUMA, we see this
usemem
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean System-1 102.86 ( 0.00%) 46.81 ( 54.50%)
Amean System-4 37.85 ( 0.00%) 34.02 ( 10.12%)
Amean System-7 48.12 ( 0.00%) 46.89 ( 2.56%)
Amean System-12 51.98 ( 0.00%) 56.96 ( -9.57%)
Amean System-21 80.16 ( 0.00%) 79.05 ( 1.39%)
Amean System-30 110.71 ( 0.00%) 107.17 ( 3.20%)
Amean System-48 127.98 ( 0.00%) 124.83 ( 2.46%)
Amean Elapsd-1 185.84 ( 0.00%) 105.51 ( 43.23%)
Amean Elapsd-4 26.19 ( 0.00%) 25.58 ( 2.33%)
Amean Elapsd-7 21.65 ( 0.00%) 21.62 ( 0.16%)
Amean Elapsd-12 18.58 ( 0.00%) 17.94 ( 3.43%)
Amean Elapsd-21 17.53 ( 0.00%) 16.60 ( 5.33%)
Amean Elapsd-30 17.45 ( 0.00%) 17.13 ( 1.84%)
Amean Elapsd-48 15.40 ( 0.00%) 15.27 ( 0.82%)
For a single thread, the benchmark completes 43.23% faster with this
patch applied with smaller benefits as the thread increases. Similar,
notice the large reduction in most cases in system CPU usage. The
overall CPU time is
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
User 10357.65 10438.33
System 3988.88 3543.94
Elapsed 2203.01 1634.41
Which is substantial. Now, the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 128458477 278352931
Major Faults 2174976 225
Swap Ins 16904701 0
Swap Outs 17359627 0
Allocation stalls 43611 0
DMA allocs 0 0
DMA32 allocs 19832646 19448017
Normal allocs 614488453 580941839
Movable allocs 0 0
Direct pages scanned 24163800 0
Kswapd pages scanned 0 0
Kswapd pages reclaimed 0 0
Direct pages reclaimed 20691346 0
Compaction stalls 42263 0
Compaction success 938 0
Compaction failures 41325 0
This patch eliminates almost all swapping and direct reclaim activity.
There is still overhead but it's from NUMA balancing which does not
identify that it's pointless trying to do anything with this workload.
I also tried the thpscale benchmark which forces a corner case where
compaction can be used heavily and measures the latency of whether base
or huge pages were used
thpscale Fault Latencies
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean fault-base-1 5288.84 ( 0.00%) 2817.12 ( 46.73%)
Amean fault-base-3 6365.53 ( 0.00%) 3499.11 ( 45.03%)
Amean fault-base-5 6526.19 ( 0.00%) 4363.06 ( 33.15%)
Amean fault-base-7 7142.25 ( 0.00%) 4858.08 ( 31.98%)
Amean fault-base-12 13827.64 ( 0.00%) 10292.11 ( 25.57%)
Amean fault-base-18 18235.07 ( 0.00%) 13788.84 ( 24.38%)
Amean fault-base-24 21597.80 ( 0.00%) 24388.03 (-12.92%)
Amean fault-base-30 26754.15 ( 0.00%) 19700.55 ( 26.36%)
Amean fault-base-32 26784.94 ( 0.00%) 19513.57 ( 27.15%)
Amean fault-huge-1 4223.96 ( 0.00%) 2178.57 ( 48.42%)
Amean fault-huge-3 2194.77 ( 0.00%) 2149.74 ( 2.05%)
Amean fault-huge-5 2569.60 ( 0.00%) 2346.95 ( 8.66%)
Amean fault-huge-7 3612.69 ( 0.00%) 2997.70 ( 17.02%)
Amean fault-huge-12 3301.75 ( 0.00%) 6727.02 (-103.74%)
Amean fault-huge-18 6696.47 ( 0.00%) 6685.72 ( 0.16%)
Amean fault-huge-24 8000.72 ( 0.00%) 9311.43 (-16.38%)
Amean fault-huge-30 13305.55 ( 0.00%) 9750.45 ( 26.72%)
Amean fault-huge-32 9981.71 ( 0.00%) 10316.06 ( -3.35%)
The average time to fault pages is substantially reduced in the majority
of caseds but with the obvious caveat that fewer THPs are actually used
in this adverse workload
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Percentage huge-1 0.71 ( 0.00%) 14.04 (1865.22%)
Percentage huge-3 10.77 ( 0.00%) 33.05 (206.85%)
Percentage huge-5 60.39 ( 0.00%) 38.51 (-36.23%)
Percentage huge-7 45.97 ( 0.00%) 34.57 (-24.79%)
Percentage huge-12 68.12 ( 0.00%) 40.07 (-41.17%)
Percentage huge-18 64.93 ( 0.00%) 47.82 (-26.35%)
Percentage huge-24 62.69 ( 0.00%) 44.23 (-29.44%)
Percentage huge-30 43.49 ( 0.00%) 55.38 ( 27.34%)
Percentage huge-32 50.72 ( 0.00%) 51.90 ( 2.35%)
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 37429143 47564000
Major Faults 1916 1558
Swap Ins 1466 1079
Swap Outs 2936863 149626
Allocation stalls 62510 3
DMA allocs 0 0
DMA32 allocs 6566458 6401314
Normal allocs 216361697 216538171
Movable allocs 0 0
Direct pages scanned 25977580 17998
Kswapd pages scanned 0 3638931
Kswapd pages reclaimed 0 207236
Direct pages reclaimed 8833714 88
Compaction stalls 103349 5
Compaction success 270 4
Compaction failures 103079 1
Note again that while this does swap as it's an aggressive workload, the
direct relcim activity and allocation stalls is substantially reduced.
There is some kswapd activity but ftrace showed that the kswapd activity
was due to normal wakeups from 4K pages being allocated.
Compaction-related stalls and activity are almost eliminated.
I also tried the stutter benchmark. For this, I do not have figures for
NUMA but it's something that does impact UMA so I'll report what is
available
stutter
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Min mmap 7.3571 ( 0.00%) 7.3438 ( 0.18%)
1st-qrtle mmap 7.5278 ( 0.00%) 17.9200 (-138.05%)
2nd-qrtle mmap 7.6818 ( 0.00%) 21.6055 (-181.25%)
3rd-qrtle mmap 11.0889 ( 0.00%) 21.8881 (-97.39%)
Max-90% mmap 27.8978 ( 0.00%) 22.1632 ( 20.56%)
Max-93% mmap 28.3202 ( 0.00%) 22.3044 ( 21.24%)
Max-95% mmap 28.5600 ( 0.00%) 22.4580 ( 21.37%)
Max-99% mmap 29.6032 ( 0.00%) 25.5216 ( 13.79%)
Max mmap 4109.7289 ( 0.00%) 4813.9832 (-17.14%)
Mean mmap 12.4474 ( 0.00%) 19.3027 (-55.07%)
This benchmark is trying to fault an anonymous mapping while there is a
heavy IO load -- a scenario that desktop users used to complain about
frequently. This shows a mix because the ideal case of mapping with THP
is not hit as often. However, note that 99% of the mappings complete
13.79% faster. The CPU usage here is particularly interesting
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
User 67.50 0.99
System 1327.88 91.30
Elapsed 2079.00 2128.98
And once again we look at the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 335241922 1314582827
Major Faults 715 819
Swap Ins 0 0
Swap Outs 0 0
Allocation stalls 532723 0
DMA allocs 0 0
DMA32 allocs 1822364341 1177950222
Normal allocs 1815640808 1517844854
Movable allocs 0 0
Direct pages scanned 21892772 0
Kswapd pages scanned 20015890 41879484
Kswapd pages reclaimed 19961986 41822072
Direct pages reclaimed 21892741 0
Compaction stalls 1065755 0
Compaction success 514 0
Compaction failures 1065241 0
Allocation stalls and all direct reclaim activity is eliminated as well
as compaction-related stalls.
THP gives impressive gains in some cases but only if they are quickly
available. We're not going to reach the point where they are completely
free so lets take the costs out of the fast paths finally and defer the
cost to kswapd, kcompactd and khugepaged where it belongs.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-18 00:19:23 +03:00
/*
mm, thp: add new defer+madvise defrag option
There is no thp defrag option that currently allows MADV_HUGEPAGE
regions to do direct compaction and reclaim while all other thp
allocations simply trigger kswapd and kcompactd in the background and
fail immediately.
The "defer" setting simply triggers background reclaim and compaction
for all regions, regardless of MADV_HUGEPAGE, which makes it unusable
for our userspace where MADV_HUGEPAGE is being used to indicate the
application is willing to wait for work for thp memory to be available.
The "madvise" setting will do direct compaction and reclaim for these
MADV_HUGEPAGE regions, but does not trigger kswapd and kcompactd in the
background for anybody else.
For reasonable usage, there needs to be a mesh between the two options.
This patch introduces a fifth mode, "defer+madvise", that will do direct
reclaim and compaction for MADV_HUGEPAGE regions and trigger background
reclaim and compaction for everybody else so that hugepages may be
available in the near future.
A proposal to allow direct reclaim and compaction for MADV_HUGEPAGE
regions as part of the "defer" mode, making it a very powerful setting
and avoids breaking userspace, was offered:
http://marc.info/?t=148236612700003
This additional mode is a compromise.
A second proposal to allow both "defer" and "madvise" to be selected at
the same time was also offered:
http://marc.info/?t=148357345300001.
This is possible, but there was a concern that it might break existing
userspaces the parse the output of the defrag mode, so the fifth option
was introduced instead.
This patch also cleans up the helper function for storing to "enabled"
and "defrag" since the former supports three modes while the latter
supports five and triple_flag_store() was getting unnecessarily messy.
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1701101614330.41805@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 02:45:49 +03:00
* always : directly stall for all thp allocations
* defer : wake kswapd and fail if not immediately available
* defer + madvise : wake kswapd and directly stall for MADV_HUGEPAGE , otherwise
* fail if not immediately available
* madvise : directly stall for MADV_HUGEPAGE , otherwise fail if not immediately
* available
* never : never stall for any thp allocation
mm: thp: set THP defrag by default to madvise and add a stall-free defrag option
THP defrag is enabled by default to direct reclaim/compact but not wake
kswapd in the event of a THP allocation failure. The problem is that
THP allocation requests potentially enter reclaim/compaction. This
potentially incurs a severe stall that is not guaranteed to be offset by
reduced TLB misses. While there has been considerable effort to reduce
the impact of reclaim/compaction, it is still a high cost and workloads
that should fit in memory fail to do so. Specifically, a simple
anon/file streaming workload will enter direct reclaim on NUMA at least
even though the working set size is 80% of RAM. It's been years and
it's time to throw in the towel.
First, this patch defines THP defrag as follows;
madvise: A failed allocation will direct reclaim/compact if the application requests it
never: Neither reclaim/compact nor wake kswapd
defer: A failed allocation will wake kswapd/kcompactd
always: A failed allocation will direct reclaim/compact (historical behaviour)
khugepaged defrag will enter direct/reclaim but not wake kswapd.
Next it sets the default defrag option to be "madvise" to only enter
direct reclaim/compaction for applications that specifically requested
it.
Lastly, it removes a check from the page allocator slowpath that is
related to __GFP_THISNODE to allow "defer" to work. The callers that
really cares are slub/slab and they are updated accordingly. The slab
one may be surprising because it also corrects a comment as kswapd was
never woken up by that path.
This means that a THP fault will no longer stall for most applications
by default and the ideal for most users that get THP if they are
immediately available. There are still options for users that prefer a
stall at startup of a new application by either restoring historical
behaviour with "always" or pick a half-way point with "defer" where
kswapd does some of the work in the background and wakes kcompactd if
necessary. THP defrag for khugepaged remains enabled and will enter
direct/reclaim but no wakeup kswapd or kcompactd.
After this patch a THP allocation failure will quickly fallback and rely
on khugepaged to recover the situation at some time in the future. In
some cases, this will reduce THP usage but the benefit of THP is hard to
measure and not a universal win where as a stall to reclaim/compaction
is definitely measurable and can be painful.
The first test for this is using "usemem" to read a large file and write
a large anonymous mapping (to avoid the zero page) multiple times. The
total size of the mappings is 80% of RAM and the benchmark simply
measures how long it takes to complete. It uses multiple threads to see
if that is a factor. On UMA, the performance is almost identical so is
not reported but on NUMA, we see this
usemem
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean System-1 102.86 ( 0.00%) 46.81 ( 54.50%)
Amean System-4 37.85 ( 0.00%) 34.02 ( 10.12%)
Amean System-7 48.12 ( 0.00%) 46.89 ( 2.56%)
Amean System-12 51.98 ( 0.00%) 56.96 ( -9.57%)
Amean System-21 80.16 ( 0.00%) 79.05 ( 1.39%)
Amean System-30 110.71 ( 0.00%) 107.17 ( 3.20%)
Amean System-48 127.98 ( 0.00%) 124.83 ( 2.46%)
Amean Elapsd-1 185.84 ( 0.00%) 105.51 ( 43.23%)
Amean Elapsd-4 26.19 ( 0.00%) 25.58 ( 2.33%)
Amean Elapsd-7 21.65 ( 0.00%) 21.62 ( 0.16%)
Amean Elapsd-12 18.58 ( 0.00%) 17.94 ( 3.43%)
Amean Elapsd-21 17.53 ( 0.00%) 16.60 ( 5.33%)
Amean Elapsd-30 17.45 ( 0.00%) 17.13 ( 1.84%)
Amean Elapsd-48 15.40 ( 0.00%) 15.27 ( 0.82%)
For a single thread, the benchmark completes 43.23% faster with this
patch applied with smaller benefits as the thread increases. Similar,
notice the large reduction in most cases in system CPU usage. The
overall CPU time is
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
User 10357.65 10438.33
System 3988.88 3543.94
Elapsed 2203.01 1634.41
Which is substantial. Now, the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 128458477 278352931
Major Faults 2174976 225
Swap Ins 16904701 0
Swap Outs 17359627 0
Allocation stalls 43611 0
DMA allocs 0 0
DMA32 allocs 19832646 19448017
Normal allocs 614488453 580941839
Movable allocs 0 0
Direct pages scanned 24163800 0
Kswapd pages scanned 0 0
Kswapd pages reclaimed 0 0
Direct pages reclaimed 20691346 0
Compaction stalls 42263 0
Compaction success 938 0
Compaction failures 41325 0
This patch eliminates almost all swapping and direct reclaim activity.
There is still overhead but it's from NUMA balancing which does not
identify that it's pointless trying to do anything with this workload.
I also tried the thpscale benchmark which forces a corner case where
compaction can be used heavily and measures the latency of whether base
or huge pages were used
thpscale Fault Latencies
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean fault-base-1 5288.84 ( 0.00%) 2817.12 ( 46.73%)
Amean fault-base-3 6365.53 ( 0.00%) 3499.11 ( 45.03%)
Amean fault-base-5 6526.19 ( 0.00%) 4363.06 ( 33.15%)
Amean fault-base-7 7142.25 ( 0.00%) 4858.08 ( 31.98%)
Amean fault-base-12 13827.64 ( 0.00%) 10292.11 ( 25.57%)
Amean fault-base-18 18235.07 ( 0.00%) 13788.84 ( 24.38%)
Amean fault-base-24 21597.80 ( 0.00%) 24388.03 (-12.92%)
Amean fault-base-30 26754.15 ( 0.00%) 19700.55 ( 26.36%)
Amean fault-base-32 26784.94 ( 0.00%) 19513.57 ( 27.15%)
Amean fault-huge-1 4223.96 ( 0.00%) 2178.57 ( 48.42%)
Amean fault-huge-3 2194.77 ( 0.00%) 2149.74 ( 2.05%)
Amean fault-huge-5 2569.60 ( 0.00%) 2346.95 ( 8.66%)
Amean fault-huge-7 3612.69 ( 0.00%) 2997.70 ( 17.02%)
Amean fault-huge-12 3301.75 ( 0.00%) 6727.02 (-103.74%)
Amean fault-huge-18 6696.47 ( 0.00%) 6685.72 ( 0.16%)
Amean fault-huge-24 8000.72 ( 0.00%) 9311.43 (-16.38%)
Amean fault-huge-30 13305.55 ( 0.00%) 9750.45 ( 26.72%)
Amean fault-huge-32 9981.71 ( 0.00%) 10316.06 ( -3.35%)
The average time to fault pages is substantially reduced in the majority
of caseds but with the obvious caveat that fewer THPs are actually used
in this adverse workload
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Percentage huge-1 0.71 ( 0.00%) 14.04 (1865.22%)
Percentage huge-3 10.77 ( 0.00%) 33.05 (206.85%)
Percentage huge-5 60.39 ( 0.00%) 38.51 (-36.23%)
Percentage huge-7 45.97 ( 0.00%) 34.57 (-24.79%)
Percentage huge-12 68.12 ( 0.00%) 40.07 (-41.17%)
Percentage huge-18 64.93 ( 0.00%) 47.82 (-26.35%)
Percentage huge-24 62.69 ( 0.00%) 44.23 (-29.44%)
Percentage huge-30 43.49 ( 0.00%) 55.38 ( 27.34%)
Percentage huge-32 50.72 ( 0.00%) 51.90 ( 2.35%)
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 37429143 47564000
Major Faults 1916 1558
Swap Ins 1466 1079
Swap Outs 2936863 149626
Allocation stalls 62510 3
DMA allocs 0 0
DMA32 allocs 6566458 6401314
Normal allocs 216361697 216538171
Movable allocs 0 0
Direct pages scanned 25977580 17998
Kswapd pages scanned 0 3638931
Kswapd pages reclaimed 0 207236
Direct pages reclaimed 8833714 88
Compaction stalls 103349 5
Compaction success 270 4
Compaction failures 103079 1
Note again that while this does swap as it's an aggressive workload, the
direct relcim activity and allocation stalls is substantially reduced.
There is some kswapd activity but ftrace showed that the kswapd activity
was due to normal wakeups from 4K pages being allocated.
Compaction-related stalls and activity are almost eliminated.
I also tried the stutter benchmark. For this, I do not have figures for
NUMA but it's something that does impact UMA so I'll report what is
available
stutter
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Min mmap 7.3571 ( 0.00%) 7.3438 ( 0.18%)
1st-qrtle mmap 7.5278 ( 0.00%) 17.9200 (-138.05%)
2nd-qrtle mmap 7.6818 ( 0.00%) 21.6055 (-181.25%)
3rd-qrtle mmap 11.0889 ( 0.00%) 21.8881 (-97.39%)
Max-90% mmap 27.8978 ( 0.00%) 22.1632 ( 20.56%)
Max-93% mmap 28.3202 ( 0.00%) 22.3044 ( 21.24%)
Max-95% mmap 28.5600 ( 0.00%) 22.4580 ( 21.37%)
Max-99% mmap 29.6032 ( 0.00%) 25.5216 ( 13.79%)
Max mmap 4109.7289 ( 0.00%) 4813.9832 (-17.14%)
Mean mmap 12.4474 ( 0.00%) 19.3027 (-55.07%)
This benchmark is trying to fault an anonymous mapping while there is a
heavy IO load -- a scenario that desktop users used to complain about
frequently. This shows a mix because the ideal case of mapping with THP
is not hit as often. However, note that 99% of the mappings complete
13.79% faster. The CPU usage here is particularly interesting
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
User 67.50 0.99
System 1327.88 91.30
Elapsed 2079.00 2128.98
And once again we look at the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 335241922 1314582827
Major Faults 715 819
Swap Ins 0 0
Swap Outs 0 0
Allocation stalls 532723 0
DMA allocs 0 0
DMA32 allocs 1822364341 1177950222
Normal allocs 1815640808 1517844854
Movable allocs 0 0
Direct pages scanned 21892772 0
Kswapd pages scanned 20015890 41879484
Kswapd pages reclaimed 19961986 41822072
Direct pages reclaimed 21892741 0
Compaction stalls 1065755 0
Compaction success 514 0
Compaction failures 1065241 0
Allocation stalls and all direct reclaim activity is eliminated as well
as compaction-related stalls.
THP gives impressive gains in some cases but only if they are quickly
available. We're not going to reach the point where they are completely
free so lets take the costs out of the fast paths finally and defer the
cost to kswapd, kcompactd and khugepaged where it belongs.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-18 00:19:23 +03:00
*/
mm,thp,shmem: limit shmem THP alloc gfp_mask
Patch series "mm,thp,shm: limit shmem THP alloc gfp_mask", v6.
The allocation flags of anonymous transparent huge pages can be controlled
through the files in /sys/kernel/mm/transparent_hugepage/defrag, which can
help the system from getting bogged down in the page reclaim and
compaction code when many THPs are getting allocated simultaneously.
However, the gfp_mask for shmem THP allocations were not limited by those
configuration settings, and some workloads ended up with all CPUs stuck on
the LRU lock in the page reclaim code, trying to allocate dozens of THPs
simultaneously.
This patch applies the same configurated limitation of THPs to shmem
hugepage allocations, to prevent that from happening.
This way a THP defrag setting of "never" or "defer+madvise" will result in
quick allocation failures without direct reclaim when no 2MB free pages
are available.
With this patch applied, THP allocations for tmpfs will be a little more
aggressive than today for files mmapped with MADV_HUGEPAGE, and a little
less aggressive for files that are not mmapped or mapped without that
flag.
This patch (of 4):
The allocation flags of anonymous transparent huge pages can be controlled
through the files in /sys/kernel/mm/transparent_hugepage/defrag, which can
help the system from getting bogged down in the page reclaim and
compaction code when many THPs are getting allocated simultaneously.
However, the gfp_mask for shmem THP allocations were not limited by those
configuration settings, and some workloads ended up with all CPUs stuck on
the LRU lock in the page reclaim code, trying to allocate dozens of THPs
simultaneously.
This patch applies the same configurated limitation of THPs to shmem
hugepage allocations, to prevent that from happening.
Controlling the gfp_mask of THP allocations through the knobs in sysfs
allows users to determine the balance between how aggressively the system
tries to allocate THPs at fault time, and how much the application may end
up stalling attempting those allocations.
This way a THP defrag setting of "never" or "defer+madvise" will result in
quick allocation failures without direct reclaim when no 2MB free pages
are available.
With this patch applied, THP allocations for tmpfs will be a little more
aggressive than today for files mmapped with MADV_HUGEPAGE, and a little
less aggressive for files that are not mmapped or mapped without that
flag.
Link: https://lkml.kernel.org/r/20201124194925.623931-1-riel@surriel.com
Link: https://lkml.kernel.org/r/20201124194925.623931-2-riel@surriel.com
Signed-off-by: Rik van Riel <riel@surriel.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Xu Yu <xuyu@linux.alibaba.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 04:16:18 +03:00
gfp_t vma_thp_gfp_mask ( struct vm_area_struct * vma )
mm: thp: set THP defrag by default to madvise and add a stall-free defrag option
THP defrag is enabled by default to direct reclaim/compact but not wake
kswapd in the event of a THP allocation failure. The problem is that
THP allocation requests potentially enter reclaim/compaction. This
potentially incurs a severe stall that is not guaranteed to be offset by
reduced TLB misses. While there has been considerable effort to reduce
the impact of reclaim/compaction, it is still a high cost and workloads
that should fit in memory fail to do so. Specifically, a simple
anon/file streaming workload will enter direct reclaim on NUMA at least
even though the working set size is 80% of RAM. It's been years and
it's time to throw in the towel.
First, this patch defines THP defrag as follows;
madvise: A failed allocation will direct reclaim/compact if the application requests it
never: Neither reclaim/compact nor wake kswapd
defer: A failed allocation will wake kswapd/kcompactd
always: A failed allocation will direct reclaim/compact (historical behaviour)
khugepaged defrag will enter direct/reclaim but not wake kswapd.
Next it sets the default defrag option to be "madvise" to only enter
direct reclaim/compaction for applications that specifically requested
it.
Lastly, it removes a check from the page allocator slowpath that is
related to __GFP_THISNODE to allow "defer" to work. The callers that
really cares are slub/slab and they are updated accordingly. The slab
one may be surprising because it also corrects a comment as kswapd was
never woken up by that path.
This means that a THP fault will no longer stall for most applications
by default and the ideal for most users that get THP if they are
immediately available. There are still options for users that prefer a
stall at startup of a new application by either restoring historical
behaviour with "always" or pick a half-way point with "defer" where
kswapd does some of the work in the background and wakes kcompactd if
necessary. THP defrag for khugepaged remains enabled and will enter
direct/reclaim but no wakeup kswapd or kcompactd.
After this patch a THP allocation failure will quickly fallback and rely
on khugepaged to recover the situation at some time in the future. In
some cases, this will reduce THP usage but the benefit of THP is hard to
measure and not a universal win where as a stall to reclaim/compaction
is definitely measurable and can be painful.
The first test for this is using "usemem" to read a large file and write
a large anonymous mapping (to avoid the zero page) multiple times. The
total size of the mappings is 80% of RAM and the benchmark simply
measures how long it takes to complete. It uses multiple threads to see
if that is a factor. On UMA, the performance is almost identical so is
not reported but on NUMA, we see this
usemem
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean System-1 102.86 ( 0.00%) 46.81 ( 54.50%)
Amean System-4 37.85 ( 0.00%) 34.02 ( 10.12%)
Amean System-7 48.12 ( 0.00%) 46.89 ( 2.56%)
Amean System-12 51.98 ( 0.00%) 56.96 ( -9.57%)
Amean System-21 80.16 ( 0.00%) 79.05 ( 1.39%)
Amean System-30 110.71 ( 0.00%) 107.17 ( 3.20%)
Amean System-48 127.98 ( 0.00%) 124.83 ( 2.46%)
Amean Elapsd-1 185.84 ( 0.00%) 105.51 ( 43.23%)
Amean Elapsd-4 26.19 ( 0.00%) 25.58 ( 2.33%)
Amean Elapsd-7 21.65 ( 0.00%) 21.62 ( 0.16%)
Amean Elapsd-12 18.58 ( 0.00%) 17.94 ( 3.43%)
Amean Elapsd-21 17.53 ( 0.00%) 16.60 ( 5.33%)
Amean Elapsd-30 17.45 ( 0.00%) 17.13 ( 1.84%)
Amean Elapsd-48 15.40 ( 0.00%) 15.27 ( 0.82%)
For a single thread, the benchmark completes 43.23% faster with this
patch applied with smaller benefits as the thread increases. Similar,
notice the large reduction in most cases in system CPU usage. The
overall CPU time is
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
User 10357.65 10438.33
System 3988.88 3543.94
Elapsed 2203.01 1634.41
Which is substantial. Now, the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 128458477 278352931
Major Faults 2174976 225
Swap Ins 16904701 0
Swap Outs 17359627 0
Allocation stalls 43611 0
DMA allocs 0 0
DMA32 allocs 19832646 19448017
Normal allocs 614488453 580941839
Movable allocs 0 0
Direct pages scanned 24163800 0
Kswapd pages scanned 0 0
Kswapd pages reclaimed 0 0
Direct pages reclaimed 20691346 0
Compaction stalls 42263 0
Compaction success 938 0
Compaction failures 41325 0
This patch eliminates almost all swapping and direct reclaim activity.
There is still overhead but it's from NUMA balancing which does not
identify that it's pointless trying to do anything with this workload.
I also tried the thpscale benchmark which forces a corner case where
compaction can be used heavily and measures the latency of whether base
or huge pages were used
thpscale Fault Latencies
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean fault-base-1 5288.84 ( 0.00%) 2817.12 ( 46.73%)
Amean fault-base-3 6365.53 ( 0.00%) 3499.11 ( 45.03%)
Amean fault-base-5 6526.19 ( 0.00%) 4363.06 ( 33.15%)
Amean fault-base-7 7142.25 ( 0.00%) 4858.08 ( 31.98%)
Amean fault-base-12 13827.64 ( 0.00%) 10292.11 ( 25.57%)
Amean fault-base-18 18235.07 ( 0.00%) 13788.84 ( 24.38%)
Amean fault-base-24 21597.80 ( 0.00%) 24388.03 (-12.92%)
Amean fault-base-30 26754.15 ( 0.00%) 19700.55 ( 26.36%)
Amean fault-base-32 26784.94 ( 0.00%) 19513.57 ( 27.15%)
Amean fault-huge-1 4223.96 ( 0.00%) 2178.57 ( 48.42%)
Amean fault-huge-3 2194.77 ( 0.00%) 2149.74 ( 2.05%)
Amean fault-huge-5 2569.60 ( 0.00%) 2346.95 ( 8.66%)
Amean fault-huge-7 3612.69 ( 0.00%) 2997.70 ( 17.02%)
Amean fault-huge-12 3301.75 ( 0.00%) 6727.02 (-103.74%)
Amean fault-huge-18 6696.47 ( 0.00%) 6685.72 ( 0.16%)
Amean fault-huge-24 8000.72 ( 0.00%) 9311.43 (-16.38%)
Amean fault-huge-30 13305.55 ( 0.00%) 9750.45 ( 26.72%)
Amean fault-huge-32 9981.71 ( 0.00%) 10316.06 ( -3.35%)
The average time to fault pages is substantially reduced in the majority
of caseds but with the obvious caveat that fewer THPs are actually used
in this adverse workload
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Percentage huge-1 0.71 ( 0.00%) 14.04 (1865.22%)
Percentage huge-3 10.77 ( 0.00%) 33.05 (206.85%)
Percentage huge-5 60.39 ( 0.00%) 38.51 (-36.23%)
Percentage huge-7 45.97 ( 0.00%) 34.57 (-24.79%)
Percentage huge-12 68.12 ( 0.00%) 40.07 (-41.17%)
Percentage huge-18 64.93 ( 0.00%) 47.82 (-26.35%)
Percentage huge-24 62.69 ( 0.00%) 44.23 (-29.44%)
Percentage huge-30 43.49 ( 0.00%) 55.38 ( 27.34%)
Percentage huge-32 50.72 ( 0.00%) 51.90 ( 2.35%)
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 37429143 47564000
Major Faults 1916 1558
Swap Ins 1466 1079
Swap Outs 2936863 149626
Allocation stalls 62510 3
DMA allocs 0 0
DMA32 allocs 6566458 6401314
Normal allocs 216361697 216538171
Movable allocs 0 0
Direct pages scanned 25977580 17998
Kswapd pages scanned 0 3638931
Kswapd pages reclaimed 0 207236
Direct pages reclaimed 8833714 88
Compaction stalls 103349 5
Compaction success 270 4
Compaction failures 103079 1
Note again that while this does swap as it's an aggressive workload, the
direct relcim activity and allocation stalls is substantially reduced.
There is some kswapd activity but ftrace showed that the kswapd activity
was due to normal wakeups from 4K pages being allocated.
Compaction-related stalls and activity are almost eliminated.
I also tried the stutter benchmark. For this, I do not have figures for
NUMA but it's something that does impact UMA so I'll report what is
available
stutter
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Min mmap 7.3571 ( 0.00%) 7.3438 ( 0.18%)
1st-qrtle mmap 7.5278 ( 0.00%) 17.9200 (-138.05%)
2nd-qrtle mmap 7.6818 ( 0.00%) 21.6055 (-181.25%)
3rd-qrtle mmap 11.0889 ( 0.00%) 21.8881 (-97.39%)
Max-90% mmap 27.8978 ( 0.00%) 22.1632 ( 20.56%)
Max-93% mmap 28.3202 ( 0.00%) 22.3044 ( 21.24%)
Max-95% mmap 28.5600 ( 0.00%) 22.4580 ( 21.37%)
Max-99% mmap 29.6032 ( 0.00%) 25.5216 ( 13.79%)
Max mmap 4109.7289 ( 0.00%) 4813.9832 (-17.14%)
Mean mmap 12.4474 ( 0.00%) 19.3027 (-55.07%)
This benchmark is trying to fault an anonymous mapping while there is a
heavy IO load -- a scenario that desktop users used to complain about
frequently. This shows a mix because the ideal case of mapping with THP
is not hit as often. However, note that 99% of the mappings complete
13.79% faster. The CPU usage here is particularly interesting
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
User 67.50 0.99
System 1327.88 91.30
Elapsed 2079.00 2128.98
And once again we look at the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 335241922 1314582827
Major Faults 715 819
Swap Ins 0 0
Swap Outs 0 0
Allocation stalls 532723 0
DMA allocs 0 0
DMA32 allocs 1822364341 1177950222
Normal allocs 1815640808 1517844854
Movable allocs 0 0
Direct pages scanned 21892772 0
Kswapd pages scanned 20015890 41879484
Kswapd pages reclaimed 19961986 41822072
Direct pages reclaimed 21892741 0
Compaction stalls 1065755 0
Compaction success 514 0
Compaction failures 1065241 0
Allocation stalls and all direct reclaim activity is eliminated as well
as compaction-related stalls.
THP gives impressive gains in some cases but only if they are quickly
available. We're not going to reach the point where they are completely
free so lets take the costs out of the fast paths finally and defer the
cost to kswapd, kcompactd and khugepaged where it belongs.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-18 00:19:23 +03:00
{
mm,thp,shmem: limit shmem THP alloc gfp_mask
Patch series "mm,thp,shm: limit shmem THP alloc gfp_mask", v6.
The allocation flags of anonymous transparent huge pages can be controlled
through the files in /sys/kernel/mm/transparent_hugepage/defrag, which can
help the system from getting bogged down in the page reclaim and
compaction code when many THPs are getting allocated simultaneously.
However, the gfp_mask for shmem THP allocations were not limited by those
configuration settings, and some workloads ended up with all CPUs stuck on
the LRU lock in the page reclaim code, trying to allocate dozens of THPs
simultaneously.
This patch applies the same configurated limitation of THPs to shmem
hugepage allocations, to prevent that from happening.
This way a THP defrag setting of "never" or "defer+madvise" will result in
quick allocation failures without direct reclaim when no 2MB free pages
are available.
With this patch applied, THP allocations for tmpfs will be a little more
aggressive than today for files mmapped with MADV_HUGEPAGE, and a little
less aggressive for files that are not mmapped or mapped without that
flag.
This patch (of 4):
The allocation flags of anonymous transparent huge pages can be controlled
through the files in /sys/kernel/mm/transparent_hugepage/defrag, which can
help the system from getting bogged down in the page reclaim and
compaction code when many THPs are getting allocated simultaneously.
However, the gfp_mask for shmem THP allocations were not limited by those
configuration settings, and some workloads ended up with all CPUs stuck on
the LRU lock in the page reclaim code, trying to allocate dozens of THPs
simultaneously.
This patch applies the same configurated limitation of THPs to shmem
hugepage allocations, to prevent that from happening.
Controlling the gfp_mask of THP allocations through the knobs in sysfs
allows users to determine the balance between how aggressively the system
tries to allocate THPs at fault time, and how much the application may end
up stalling attempting those allocations.
This way a THP defrag setting of "never" or "defer+madvise" will result in
quick allocation failures without direct reclaim when no 2MB free pages
are available.
With this patch applied, THP allocations for tmpfs will be a little more
aggressive than today for files mmapped with MADV_HUGEPAGE, and a little
less aggressive for files that are not mmapped or mapped without that
flag.
Link: https://lkml.kernel.org/r/20201124194925.623931-1-riel@surriel.com
Link: https://lkml.kernel.org/r/20201124194925.623931-2-riel@surriel.com
Signed-off-by: Rik van Riel <riel@surriel.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Xu Yu <xuyu@linux.alibaba.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 04:16:18 +03:00
const bool vma_madvised = vma & & ( vma - > vm_flags & VM_HUGEPAGE ) ;
2018-12-06 02:45:54 +03:00
Revert "Revert "mm, thp: restore node-local hugepage allocations""
This reverts commit a8282608c88e08b1782141026eab61204c1e533f.
The commit references the original intended semantic for MADV_HUGEPAGE
which has subsequently taken on three unique purposes:
- enables or disables thp for a range of memory depending on the system's
config (is thp "enabled" set to "always" or "madvise"),
- determines the synchronous compaction behavior for thp allocations at
fault (is thp "defrag" set to "always", "defer+madvise", or "madvise"),
and
- reverts a previous MADV_NOHUGEPAGE (there is no madvise mode to only
clear previous hugepage advice).
These are the three purposes that currently exist in 5.2 and over the
past several years that userspace has been written around. Adding a
NUMA locality preference adds a fourth dimension to an already conflated
advice mode.
Based on the semantic that MADV_HUGEPAGE has provided over the past
several years, there exist workloads that use the tunable based on these
principles: specifically that the allocation should attempt to
defragment a local node before falling back. It is agreed that remote
hugepages typically (but not always) have a better access latency than
remote native pages, although on Naples this is at parity for
intersocket.
The revert commit that this patch reverts allows hugepage allocation to
immediately allocate remotely when local memory is fragmented. This is
contrary to the semantic of MADV_HUGEPAGE over the past several years:
that is, memory compaction should be attempted locally before falling
back.
The performance degradation of remote hugepages over local hugepages on
Rome, for example, is 53.5% increased access latency. For this reason,
the goal is to revert back to the 5.2 and previous behavior that would
attempt local defragmentation before falling back. With the patch that
is reverted by this patch, we see performance degradations at the tail
because the allocator happily allocates the remote hugepage rather than
even attempting to make a local hugepage available.
zone_reclaim_mode is not a solution to this problem since it does not
only impact hugepage allocations but rather changes the memory
allocation strategy for *all* page allocations.
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Stefan Priebe - Profihost AG <s.priebe@profihost.ag>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-04 22:54:18 +03:00
/* Always do synchronous compaction */
Revert "mm, thp: restore node-local hugepage allocations"
This reverts commit 2f0799a0ffc033b ("mm, thp: restore node-local
hugepage allocations").
commit 2f0799a0ffc033b was rightfully applied to avoid the risk of a
severe regression that was reported by the kernel test robot at the end
of the merge window. Now we understood the regression was a false
positive and was caused by a significant increase in fairness during a
swap trashing benchmark. So it's safe to re-apply the fix and continue
improving the code from there. The benchmark that reported the
regression is very useful, but it provides a meaningful result only when
there is no significant alteration in fairness during the workload. The
removal of __GFP_THISNODE increased fairness.
__GFP_THISNODE cannot be used in the generic page faults path for new
memory allocations under the MPOL_DEFAULT mempolicy, or the allocation
behavior significantly deviates from what the MPOL_DEFAULT semantics are
supposed to be for THP and 4k allocations alike.
Setting THP defrag to "always" or using MADV_HUGEPAGE (with THP defrag
set to "madvise") has never meant to provide an implicit MPOL_BIND on
the "current" node the task is running on, causing swap storms and
providing a much more aggressive behavior than even zone_reclaim_node =
3.
Any workload who could have benefited from __GFP_THISNODE has now to
enable zone_reclaim_mode=1||2||3. __GFP_THISNODE implicitly provided
the zone_reclaim_mode behavior, but it only did so if THP was enabled:
if THP was disabled, there would have been no chance to get any 4k page
from the current node if the current node was full of pagecache, which
further shows how this __GFP_THISNODE was misplaced in MADV_HUGEPAGE.
MADV_HUGEPAGE has never been intended to provide any zone_reclaim_mode
semantics, in fact the two are orthogonal, zone_reclaim_mode = 1|2|3
must work exactly the same with MADV_HUGEPAGE set or not.
The performance characteristic of memory depends on the hardware
details. The numbers below are obtained on Naples/EPYC architecture and
the N/A projection extends them to show what we should aim for in the
future as a good THP NUMA locality default. The benchmark used
exercises random memory seeks (note: the cost of the page faults is not
part of the measurement).
D0 THP | D0 4k | D1 THP | D1 4k | D2 THP | D2 4k | D3 THP | D3 4k | ...
0% | +43% | +45% | +106% | +131% | +224% | N/A | N/A
D0 means distance zero (i.e. local memory), D1 means distance one (i.e.
intra socket memory), D2 means distance two (i.e. inter socket memory),
etc...
For the guest physical memory allocated by qemu and for guest mode
kernel the performance characteristic of RAM is more complex and an
ideal default could be:
D0 THP | D1 THP | D0 4k | D2 THP | D1 4k | D3 THP | D2 4k | D3 4k | ...
0% | +58% | +101% | N/A | +222% | N/A | N/A | N/A
NOTE: the N/A are projections and haven't been measured yet, the
measurement in this case is done on a 1950x with only two NUMA nodes.
The THP case here means THP was used both in the host and in the guest.
After applying this commit the THP NUMA locality order that we'll get
out of MADV_HUGEPAGE is this:
D0 THP | D1 THP | D2 THP | D3 THP | ... | D0 4k | D1 4k | D2 4k | D3 4k | ...
Before this commit it was:
D0 THP | D0 4k | D1 4k | D2 4k | D3 4k | ...
Even if we ignore the breakage of large workloads that can't fit in a
single node that the __GFP_THISNODE implicit "current node" mbind
caused, the THP NUMA locality order provided by __GFP_THISNODE was still
not the one we shall aim for in the long term (i.e. the first one at
the top).
After this commit is applied, we can introduce a new allocator multi
order API and to replace those two alloc_pages_vmas calls in the page
fault path, with a single multi order call:
unsigned int order = (1 << HPAGE_PMD_ORDER) | (1 << 0);
page = alloc_pages_multi_order(..., &order);
if (!page)
goto out;
if (!(order & (1 << 0))) {
VM_WARN_ON(order != 1 << HPAGE_PMD_ORDER);
/* THP fault */
} else {
VM_WARN_ON(order != 1 << 0);
/* 4k fallback */
}
The page allocator logic has to be altered so that when it fails on any
zone with order 9, it has to try again with a order 0 before falling
back to the next zone in the zonelist.
After that we need to do more measurements and evaluate if adding an
opt-in feature for guest mode is worth it, to swap "DN 4k | DN+1 THP"
with "DN+1 THP | DN 4k" at every NUMA distance crossing.
Link: http://lkml.kernel.org/r/20190503223146.2312-3-aarcange@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Stefan Priebe - Profihost AG <s.priebe@profihost.ag>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-08-14 01:37:53 +03:00
if ( test_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG , & transparent_hugepage_flags ) )
return GFP_TRANSHUGE | ( vma_madvised ? 0 : __GFP_NORETRY ) ;
Revert "Revert "mm, thp: restore node-local hugepage allocations""
This reverts commit a8282608c88e08b1782141026eab61204c1e533f.
The commit references the original intended semantic for MADV_HUGEPAGE
which has subsequently taken on three unique purposes:
- enables or disables thp for a range of memory depending on the system's
config (is thp "enabled" set to "always" or "madvise"),
- determines the synchronous compaction behavior for thp allocations at
fault (is thp "defrag" set to "always", "defer+madvise", or "madvise"),
and
- reverts a previous MADV_NOHUGEPAGE (there is no madvise mode to only
clear previous hugepage advice).
These are the three purposes that currently exist in 5.2 and over the
past several years that userspace has been written around. Adding a
NUMA locality preference adds a fourth dimension to an already conflated
advice mode.
Based on the semantic that MADV_HUGEPAGE has provided over the past
several years, there exist workloads that use the tunable based on these
principles: specifically that the allocation should attempt to
defragment a local node before falling back. It is agreed that remote
hugepages typically (but not always) have a better access latency than
remote native pages, although on Naples this is at parity for
intersocket.
The revert commit that this patch reverts allows hugepage allocation to
immediately allocate remotely when local memory is fragmented. This is
contrary to the semantic of MADV_HUGEPAGE over the past several years:
that is, memory compaction should be attempted locally before falling
back.
The performance degradation of remote hugepages over local hugepages on
Rome, for example, is 53.5% increased access latency. For this reason,
the goal is to revert back to the 5.2 and previous behavior that would
attempt local defragmentation before falling back. With the patch that
is reverted by this patch, we see performance degradations at the tail
because the allocator happily allocates the remote hugepage rather than
even attempting to make a local hugepage available.
zone_reclaim_mode is not a solution to this problem since it does not
only impact hugepage allocations but rather changes the memory
allocation strategy for *all* page allocations.
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Stefan Priebe - Profihost AG <s.priebe@profihost.ag>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-04 22:54:18 +03:00
/* Kick kcompactd and fail quickly */
mm, thp: add new defer+madvise defrag option
There is no thp defrag option that currently allows MADV_HUGEPAGE
regions to do direct compaction and reclaim while all other thp
allocations simply trigger kswapd and kcompactd in the background and
fail immediately.
The "defer" setting simply triggers background reclaim and compaction
for all regions, regardless of MADV_HUGEPAGE, which makes it unusable
for our userspace where MADV_HUGEPAGE is being used to indicate the
application is willing to wait for work for thp memory to be available.
The "madvise" setting will do direct compaction and reclaim for these
MADV_HUGEPAGE regions, but does not trigger kswapd and kcompactd in the
background for anybody else.
For reasonable usage, there needs to be a mesh between the two options.
This patch introduces a fifth mode, "defer+madvise", that will do direct
reclaim and compaction for MADV_HUGEPAGE regions and trigger background
reclaim and compaction for everybody else so that hugepages may be
available in the near future.
A proposal to allow direct reclaim and compaction for MADV_HUGEPAGE
regions as part of the "defer" mode, making it a very powerful setting
and avoids breaking userspace, was offered:
http://marc.info/?t=148236612700003
This additional mode is a compromise.
A second proposal to allow both "defer" and "madvise" to be selected at
the same time was also offered:
http://marc.info/?t=148357345300001.
This is possible, but there was a concern that it might break existing
userspaces the parse the output of the defrag mode, so the fifth option
was introduced instead.
This patch also cleans up the helper function for storing to "enabled"
and "defrag" since the former supports three modes while the latter
supports five and triple_flag_store() was getting unnecessarily messy.
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1701101614330.41805@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 02:45:49 +03:00
if ( test_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG , & transparent_hugepage_flags ) )
2019-09-04 22:54:20 +03:00
return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM ;
Revert "Revert "mm, thp: restore node-local hugepage allocations""
This reverts commit a8282608c88e08b1782141026eab61204c1e533f.
The commit references the original intended semantic for MADV_HUGEPAGE
which has subsequently taken on three unique purposes:
- enables or disables thp for a range of memory depending on the system's
config (is thp "enabled" set to "always" or "madvise"),
- determines the synchronous compaction behavior for thp allocations at
fault (is thp "defrag" set to "always", "defer+madvise", or "madvise"),
and
- reverts a previous MADV_NOHUGEPAGE (there is no madvise mode to only
clear previous hugepage advice).
These are the three purposes that currently exist in 5.2 and over the
past several years that userspace has been written around. Adding a
NUMA locality preference adds a fourth dimension to an already conflated
advice mode.
Based on the semantic that MADV_HUGEPAGE has provided over the past
several years, there exist workloads that use the tunable based on these
principles: specifically that the allocation should attempt to
defragment a local node before falling back. It is agreed that remote
hugepages typically (but not always) have a better access latency than
remote native pages, although on Naples this is at parity for
intersocket.
The revert commit that this patch reverts allows hugepage allocation to
immediately allocate remotely when local memory is fragmented. This is
contrary to the semantic of MADV_HUGEPAGE over the past several years:
that is, memory compaction should be attempted locally before falling
back.
The performance degradation of remote hugepages over local hugepages on
Rome, for example, is 53.5% increased access latency. For this reason,
the goal is to revert back to the 5.2 and previous behavior that would
attempt local defragmentation before falling back. With the patch that
is reverted by this patch, we see performance degradations at the tail
because the allocator happily allocates the remote hugepage rather than
even attempting to make a local hugepage available.
zone_reclaim_mode is not a solution to this problem since it does not
only impact hugepage allocations but rather changes the memory
allocation strategy for *all* page allocations.
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Stefan Priebe - Profihost AG <s.priebe@profihost.ag>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-04 22:54:18 +03:00
/* Synchronous compaction if madvised, otherwise kick kcompactd */
mm, thp: add new defer+madvise defrag option
There is no thp defrag option that currently allows MADV_HUGEPAGE
regions to do direct compaction and reclaim while all other thp
allocations simply trigger kswapd and kcompactd in the background and
fail immediately.
The "defer" setting simply triggers background reclaim and compaction
for all regions, regardless of MADV_HUGEPAGE, which makes it unusable
for our userspace where MADV_HUGEPAGE is being used to indicate the
application is willing to wait for work for thp memory to be available.
The "madvise" setting will do direct compaction and reclaim for these
MADV_HUGEPAGE regions, but does not trigger kswapd and kcompactd in the
background for anybody else.
For reasonable usage, there needs to be a mesh between the two options.
This patch introduces a fifth mode, "defer+madvise", that will do direct
reclaim and compaction for MADV_HUGEPAGE regions and trigger background
reclaim and compaction for everybody else so that hugepages may be
available in the near future.
A proposal to allow direct reclaim and compaction for MADV_HUGEPAGE
regions as part of the "defer" mode, making it a very powerful setting
and avoids breaking userspace, was offered:
http://marc.info/?t=148236612700003
This additional mode is a compromise.
A second proposal to allow both "defer" and "madvise" to be selected at
the same time was also offered:
http://marc.info/?t=148357345300001.
This is possible, but there was a concern that it might break existing
userspaces the parse the output of the defrag mode, so the fifth option
was introduced instead.
This patch also cleans up the helper function for storing to "enabled"
and "defrag" since the former supports three modes while the latter
supports five and triple_flag_store() was getting unnecessarily messy.
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1701101614330.41805@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 02:45:49 +03:00
if ( test_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG , & transparent_hugepage_flags ) )
2019-09-04 22:54:20 +03:00
return GFP_TRANSHUGE_LIGHT |
( vma_madvised ? __GFP_DIRECT_RECLAIM :
__GFP_KSWAPD_RECLAIM ) ;
Revert "Revert "mm, thp: restore node-local hugepage allocations""
This reverts commit a8282608c88e08b1782141026eab61204c1e533f.
The commit references the original intended semantic for MADV_HUGEPAGE
which has subsequently taken on three unique purposes:
- enables or disables thp for a range of memory depending on the system's
config (is thp "enabled" set to "always" or "madvise"),
- determines the synchronous compaction behavior for thp allocations at
fault (is thp "defrag" set to "always", "defer+madvise", or "madvise"),
and
- reverts a previous MADV_NOHUGEPAGE (there is no madvise mode to only
clear previous hugepage advice).
These are the three purposes that currently exist in 5.2 and over the
past several years that userspace has been written around. Adding a
NUMA locality preference adds a fourth dimension to an already conflated
advice mode.
Based on the semantic that MADV_HUGEPAGE has provided over the past
several years, there exist workloads that use the tunable based on these
principles: specifically that the allocation should attempt to
defragment a local node before falling back. It is agreed that remote
hugepages typically (but not always) have a better access latency than
remote native pages, although on Naples this is at parity for
intersocket.
The revert commit that this patch reverts allows hugepage allocation to
immediately allocate remotely when local memory is fragmented. This is
contrary to the semantic of MADV_HUGEPAGE over the past several years:
that is, memory compaction should be attempted locally before falling
back.
The performance degradation of remote hugepages over local hugepages on
Rome, for example, is 53.5% increased access latency. For this reason,
the goal is to revert back to the 5.2 and previous behavior that would
attempt local defragmentation before falling back. With the patch that
is reverted by this patch, we see performance degradations at the tail
because the allocator happily allocates the remote hugepage rather than
even attempting to make a local hugepage available.
zone_reclaim_mode is not a solution to this problem since it does not
only impact hugepage allocations but rather changes the memory
allocation strategy for *all* page allocations.
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Stefan Priebe - Profihost AG <s.priebe@profihost.ag>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-04 22:54:18 +03:00
/* Only do synchronous compaction if madvised */
mm, thp: add new defer+madvise defrag option
There is no thp defrag option that currently allows MADV_HUGEPAGE
regions to do direct compaction and reclaim while all other thp
allocations simply trigger kswapd and kcompactd in the background and
fail immediately.
The "defer" setting simply triggers background reclaim and compaction
for all regions, regardless of MADV_HUGEPAGE, which makes it unusable
for our userspace where MADV_HUGEPAGE is being used to indicate the
application is willing to wait for work for thp memory to be available.
The "madvise" setting will do direct compaction and reclaim for these
MADV_HUGEPAGE regions, but does not trigger kswapd and kcompactd in the
background for anybody else.
For reasonable usage, there needs to be a mesh between the two options.
This patch introduces a fifth mode, "defer+madvise", that will do direct
reclaim and compaction for MADV_HUGEPAGE regions and trigger background
reclaim and compaction for everybody else so that hugepages may be
available in the near future.
A proposal to allow direct reclaim and compaction for MADV_HUGEPAGE
regions as part of the "defer" mode, making it a very powerful setting
and avoids breaking userspace, was offered:
http://marc.info/?t=148236612700003
This additional mode is a compromise.
A second proposal to allow both "defer" and "madvise" to be selected at
the same time was also offered:
http://marc.info/?t=148357345300001.
This is possible, but there was a concern that it might break existing
userspaces the parse the output of the defrag mode, so the fifth option
was introduced instead.
This patch also cleans up the helper function for storing to "enabled"
and "defrag" since the former supports three modes while the latter
supports five and triple_flag_store() was getting unnecessarily messy.
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1701101614330.41805@chino.kir.corp.google.com
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 02:45:49 +03:00
if ( test_bit ( TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG , & transparent_hugepage_flags ) )
2019-09-04 22:54:20 +03:00
return GFP_TRANSHUGE_LIGHT |
( vma_madvised ? __GFP_DIRECT_RECLAIM : 0 ) ;
Revert "Revert "mm, thp: restore node-local hugepage allocations""
This reverts commit a8282608c88e08b1782141026eab61204c1e533f.
The commit references the original intended semantic for MADV_HUGEPAGE
which has subsequently taken on three unique purposes:
- enables or disables thp for a range of memory depending on the system's
config (is thp "enabled" set to "always" or "madvise"),
- determines the synchronous compaction behavior for thp allocations at
fault (is thp "defrag" set to "always", "defer+madvise", or "madvise"),
and
- reverts a previous MADV_NOHUGEPAGE (there is no madvise mode to only
clear previous hugepage advice).
These are the three purposes that currently exist in 5.2 and over the
past several years that userspace has been written around. Adding a
NUMA locality preference adds a fourth dimension to an already conflated
advice mode.
Based on the semantic that MADV_HUGEPAGE has provided over the past
several years, there exist workloads that use the tunable based on these
principles: specifically that the allocation should attempt to
defragment a local node before falling back. It is agreed that remote
hugepages typically (but not always) have a better access latency than
remote native pages, although on Naples this is at parity for
intersocket.
The revert commit that this patch reverts allows hugepage allocation to
immediately allocate remotely when local memory is fragmented. This is
contrary to the semantic of MADV_HUGEPAGE over the past several years:
that is, memory compaction should be attempted locally before falling
back.
The performance degradation of remote hugepages over local hugepages on
Rome, for example, is 53.5% increased access latency. For this reason,
the goal is to revert back to the 5.2 and previous behavior that would
attempt local defragmentation before falling back. With the patch that
is reverted by this patch, we see performance degradations at the tail
because the allocator happily allocates the remote hugepage rather than
even attempting to make a local hugepage available.
zone_reclaim_mode is not a solution to this problem since it does not
only impact hugepage allocations but rather changes the memory
allocation strategy for *all* page allocations.
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Stefan Priebe - Profihost AG <s.priebe@profihost.ag>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-04 22:54:18 +03:00
2019-09-04 22:54:20 +03:00
return GFP_TRANSHUGE_LIGHT ;
mm: thp: set THP defrag by default to madvise and add a stall-free defrag option
THP defrag is enabled by default to direct reclaim/compact but not wake
kswapd in the event of a THP allocation failure. The problem is that
THP allocation requests potentially enter reclaim/compaction. This
potentially incurs a severe stall that is not guaranteed to be offset by
reduced TLB misses. While there has been considerable effort to reduce
the impact of reclaim/compaction, it is still a high cost and workloads
that should fit in memory fail to do so. Specifically, a simple
anon/file streaming workload will enter direct reclaim on NUMA at least
even though the working set size is 80% of RAM. It's been years and
it's time to throw in the towel.
First, this patch defines THP defrag as follows;
madvise: A failed allocation will direct reclaim/compact if the application requests it
never: Neither reclaim/compact nor wake kswapd
defer: A failed allocation will wake kswapd/kcompactd
always: A failed allocation will direct reclaim/compact (historical behaviour)
khugepaged defrag will enter direct/reclaim but not wake kswapd.
Next it sets the default defrag option to be "madvise" to only enter
direct reclaim/compaction for applications that specifically requested
it.
Lastly, it removes a check from the page allocator slowpath that is
related to __GFP_THISNODE to allow "defer" to work. The callers that
really cares are slub/slab and they are updated accordingly. The slab
one may be surprising because it also corrects a comment as kswapd was
never woken up by that path.
This means that a THP fault will no longer stall for most applications
by default and the ideal for most users that get THP if they are
immediately available. There are still options for users that prefer a
stall at startup of a new application by either restoring historical
behaviour with "always" or pick a half-way point with "defer" where
kswapd does some of the work in the background and wakes kcompactd if
necessary. THP defrag for khugepaged remains enabled and will enter
direct/reclaim but no wakeup kswapd or kcompactd.
After this patch a THP allocation failure will quickly fallback and rely
on khugepaged to recover the situation at some time in the future. In
some cases, this will reduce THP usage but the benefit of THP is hard to
measure and not a universal win where as a stall to reclaim/compaction
is definitely measurable and can be painful.
The first test for this is using "usemem" to read a large file and write
a large anonymous mapping (to avoid the zero page) multiple times. The
total size of the mappings is 80% of RAM and the benchmark simply
measures how long it takes to complete. It uses multiple threads to see
if that is a factor. On UMA, the performance is almost identical so is
not reported but on NUMA, we see this
usemem
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean System-1 102.86 ( 0.00%) 46.81 ( 54.50%)
Amean System-4 37.85 ( 0.00%) 34.02 ( 10.12%)
Amean System-7 48.12 ( 0.00%) 46.89 ( 2.56%)
Amean System-12 51.98 ( 0.00%) 56.96 ( -9.57%)
Amean System-21 80.16 ( 0.00%) 79.05 ( 1.39%)
Amean System-30 110.71 ( 0.00%) 107.17 ( 3.20%)
Amean System-48 127.98 ( 0.00%) 124.83 ( 2.46%)
Amean Elapsd-1 185.84 ( 0.00%) 105.51 ( 43.23%)
Amean Elapsd-4 26.19 ( 0.00%) 25.58 ( 2.33%)
Amean Elapsd-7 21.65 ( 0.00%) 21.62 ( 0.16%)
Amean Elapsd-12 18.58 ( 0.00%) 17.94 ( 3.43%)
Amean Elapsd-21 17.53 ( 0.00%) 16.60 ( 5.33%)
Amean Elapsd-30 17.45 ( 0.00%) 17.13 ( 1.84%)
Amean Elapsd-48 15.40 ( 0.00%) 15.27 ( 0.82%)
For a single thread, the benchmark completes 43.23% faster with this
patch applied with smaller benefits as the thread increases. Similar,
notice the large reduction in most cases in system CPU usage. The
overall CPU time is
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
User 10357.65 10438.33
System 3988.88 3543.94
Elapsed 2203.01 1634.41
Which is substantial. Now, the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 128458477 278352931
Major Faults 2174976 225
Swap Ins 16904701 0
Swap Outs 17359627 0
Allocation stalls 43611 0
DMA allocs 0 0
DMA32 allocs 19832646 19448017
Normal allocs 614488453 580941839
Movable allocs 0 0
Direct pages scanned 24163800 0
Kswapd pages scanned 0 0
Kswapd pages reclaimed 0 0
Direct pages reclaimed 20691346 0
Compaction stalls 42263 0
Compaction success 938 0
Compaction failures 41325 0
This patch eliminates almost all swapping and direct reclaim activity.
There is still overhead but it's from NUMA balancing which does not
identify that it's pointless trying to do anything with this workload.
I also tried the thpscale benchmark which forces a corner case where
compaction can be used heavily and measures the latency of whether base
or huge pages were used
thpscale Fault Latencies
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Amean fault-base-1 5288.84 ( 0.00%) 2817.12 ( 46.73%)
Amean fault-base-3 6365.53 ( 0.00%) 3499.11 ( 45.03%)
Amean fault-base-5 6526.19 ( 0.00%) 4363.06 ( 33.15%)
Amean fault-base-7 7142.25 ( 0.00%) 4858.08 ( 31.98%)
Amean fault-base-12 13827.64 ( 0.00%) 10292.11 ( 25.57%)
Amean fault-base-18 18235.07 ( 0.00%) 13788.84 ( 24.38%)
Amean fault-base-24 21597.80 ( 0.00%) 24388.03 (-12.92%)
Amean fault-base-30 26754.15 ( 0.00%) 19700.55 ( 26.36%)
Amean fault-base-32 26784.94 ( 0.00%) 19513.57 ( 27.15%)
Amean fault-huge-1 4223.96 ( 0.00%) 2178.57 ( 48.42%)
Amean fault-huge-3 2194.77 ( 0.00%) 2149.74 ( 2.05%)
Amean fault-huge-5 2569.60 ( 0.00%) 2346.95 ( 8.66%)
Amean fault-huge-7 3612.69 ( 0.00%) 2997.70 ( 17.02%)
Amean fault-huge-12 3301.75 ( 0.00%) 6727.02 (-103.74%)
Amean fault-huge-18 6696.47 ( 0.00%) 6685.72 ( 0.16%)
Amean fault-huge-24 8000.72 ( 0.00%) 9311.43 (-16.38%)
Amean fault-huge-30 13305.55 ( 0.00%) 9750.45 ( 26.72%)
Amean fault-huge-32 9981.71 ( 0.00%) 10316.06 ( -3.35%)
The average time to fault pages is substantially reduced in the majority
of caseds but with the obvious caveat that fewer THPs are actually used
in this adverse workload
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Percentage huge-1 0.71 ( 0.00%) 14.04 (1865.22%)
Percentage huge-3 10.77 ( 0.00%) 33.05 (206.85%)
Percentage huge-5 60.39 ( 0.00%) 38.51 (-36.23%)
Percentage huge-7 45.97 ( 0.00%) 34.57 (-24.79%)
Percentage huge-12 68.12 ( 0.00%) 40.07 (-41.17%)
Percentage huge-18 64.93 ( 0.00%) 47.82 (-26.35%)
Percentage huge-24 62.69 ( 0.00%) 44.23 (-29.44%)
Percentage huge-30 43.49 ( 0.00%) 55.38 ( 27.34%)
Percentage huge-32 50.72 ( 0.00%) 51.90 ( 2.35%)
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 37429143 47564000
Major Faults 1916 1558
Swap Ins 1466 1079
Swap Outs 2936863 149626
Allocation stalls 62510 3
DMA allocs 0 0
DMA32 allocs 6566458 6401314
Normal allocs 216361697 216538171
Movable allocs 0 0
Direct pages scanned 25977580 17998
Kswapd pages scanned 0 3638931
Kswapd pages reclaimed 0 207236
Direct pages reclaimed 8833714 88
Compaction stalls 103349 5
Compaction success 270 4
Compaction failures 103079 1
Note again that while this does swap as it's an aggressive workload, the
direct relcim activity and allocation stalls is substantially reduced.
There is some kswapd activity but ftrace showed that the kswapd activity
was due to normal wakeups from 4K pages being allocated.
Compaction-related stalls and activity are almost eliminated.
I also tried the stutter benchmark. For this, I do not have figures for
NUMA but it's something that does impact UMA so I'll report what is
available
stutter
4.4.0 4.4.0
kcompactd-v1r1 nodefrag-v1r3
Min mmap 7.3571 ( 0.00%) 7.3438 ( 0.18%)
1st-qrtle mmap 7.5278 ( 0.00%) 17.9200 (-138.05%)
2nd-qrtle mmap 7.6818 ( 0.00%) 21.6055 (-181.25%)
3rd-qrtle mmap 11.0889 ( 0.00%) 21.8881 (-97.39%)
Max-90% mmap 27.8978 ( 0.00%) 22.1632 ( 20.56%)
Max-93% mmap 28.3202 ( 0.00%) 22.3044 ( 21.24%)
Max-95% mmap 28.5600 ( 0.00%) 22.4580 ( 21.37%)
Max-99% mmap 29.6032 ( 0.00%) 25.5216 ( 13.79%)
Max mmap 4109.7289 ( 0.00%) 4813.9832 (-17.14%)
Mean mmap 12.4474 ( 0.00%) 19.3027 (-55.07%)
This benchmark is trying to fault an anonymous mapping while there is a
heavy IO load -- a scenario that desktop users used to complain about
frequently. This shows a mix because the ideal case of mapping with THP
is not hit as often. However, note that 99% of the mappings complete
13.79% faster. The CPU usage here is particularly interesting
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
User 67.50 0.99
System 1327.88 91.30
Elapsed 2079.00 2128.98
And once again we look at the reclaim figures
4.4.0 4.4.0
kcompactd-v1r1nodefrag-v1r3
Minor Faults 335241922 1314582827
Major Faults 715 819
Swap Ins 0 0
Swap Outs 0 0
Allocation stalls 532723 0
DMA allocs 0 0
DMA32 allocs 1822364341 1177950222
Normal allocs 1815640808 1517844854
Movable allocs 0 0
Direct pages scanned 21892772 0
Kswapd pages scanned 20015890 41879484
Kswapd pages reclaimed 19961986 41822072
Direct pages reclaimed 21892741 0
Compaction stalls 1065755 0
Compaction success 514 0
Compaction failures 1065241 0
Allocation stalls and all direct reclaim activity is eliminated as well
as compaction-related stalls.
THP gives impressive gains in some cases but only if they are quickly
available. We're not going to reach the point where they are completely
free so lets take the costs out of the fast paths finally and defer the
cost to kswapd, kcompactd and khugepaged where it belongs.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-18 00:19:23 +03:00
}
2013-11-15 02:31:04 +04:00
/* Caller must hold page table lock. */
2021-02-24 23:07:29 +03:00
static void set_huge_zero_page ( pgtable_t pgtable , struct mm_struct * mm ,
2012-12-13 01:51:06 +04:00
struct vm_area_struct * vma , unsigned long haddr , pmd_t * pmd ,
2013-04-30 02:08:44 +04:00
struct page * zero_page )
2012-12-13 01:50:51 +04:00
{
pmd_t entry ;
2015-09-09 00:58:43 +03:00
if ( ! pmd_none ( * pmd ) )
2021-02-24 23:07:29 +03:00
return ;
2013-04-30 02:08:44 +04:00
entry = mk_pmd ( zero_page , vma - > vm_page_prot ) ;
2012-12-13 01:50:51 +04:00
entry = pmd_mkhuge ( entry ) ;
2022-08-18 11:27:48 +03:00
pgtable_trans_huge_deposit ( mm , pmd , pgtable ) ;
2012-12-13 01:50:51 +04:00
set_pmd_at ( mm , haddr , pmd , entry ) ;
2017-11-16 04:35:37 +03:00
mm_inc_nr_ptes ( mm ) ;
2012-12-13 01:50:51 +04:00
}
2018-08-24 03:01:36 +03:00
vm_fault_t do_huge_pmd_anonymous_page ( struct vm_fault * vmf )
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
{
2016-12-15 02:06:58 +03:00
struct vm_area_struct * vma = vmf - > vma ;
2015-02-12 02:27:12 +03:00
gfp_t gfp ;
2022-05-13 06:23:01 +03:00
struct folio * folio ;
2016-12-15 02:06:58 +03:00
unsigned long haddr = vmf - > address & HPAGE_PMD_MASK ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
2019-07-19 01:57:24 +03:00
if ( ! transhuge_vma_suitable ( vma , haddr ) )
2013-09-13 02:14:05 +04:00
return VM_FAULT_FALLBACK ;
2013-09-13 02:14:03 +04:00
if ( unlikely ( anon_vma_prepare ( vma ) ) )
return VM_FAULT_OOM ;
2022-06-16 20:48:35 +03:00
khugepaged_enter_vma ( vma , vma - > vm_flags ) ;
2022-05-20 00:08:49 +03:00
2016-12-15 02:06:58 +03:00
if ( ! ( vmf - > flags & FAULT_FLAG_WRITE ) & &
2016-07-27 01:25:20 +03:00
! mm_forbids_zeropage ( vma - > vm_mm ) & &
2013-09-13 02:14:03 +04:00
transparent_hugepage_use_zero_page ( ) ) {
pgtable_t pgtable ;
struct page * zero_page ;
2018-08-24 03:01:36 +03:00
vm_fault_t ret ;
mm: treewide: remove unused address argument from pte_alloc functions
Patch series "Add support for fast mremap".
This series speeds up the mremap(2) syscall by copying page tables at
the PMD level even for non-THP systems. There is concern that the extra
'address' argument that mremap passes to pte_alloc may do something
subtle architecture related in the future that may make the scheme not
work. Also we find that there is no point in passing the 'address' to
pte_alloc since its unused. This patch therefore removes this argument
tree-wide resulting in a nice negative diff as well. Also ensuring
along the way that the enabled architectures do not do anything funky
with the 'address' argument that goes unnoticed by the optimization.
Build and boot tested on x86-64. Build tested on arm64. The config
enablement patch for arm64 will be posted in the future after more
testing.
The changes were obtained by applying the following Coccinelle script.
(thanks Julia for answering all Coccinelle questions!).
Following fix ups were done manually:
* Removal of address argument from pte_fragment_alloc
* Removal of pte_alloc_one_fast definitions from m68k and microblaze.
// Options: --include-headers --no-includes
// Note: I split the 'identifier fn' line, so if you are manually
// running it, please unsplit it so it runs for you.
virtual patch
@pte_alloc_func_def depends on patch exists@
identifier E2;
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
type T2;
@@
fn(...
- , T2 E2
)
{ ... }
@pte_alloc_func_proto_noarg depends on patch exists@
type T1, T2, T3, T4;
identifier fn =~ "^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
@@
(
- T3 fn(T1, T2);
+ T3 fn(T1);
|
- T3 fn(T1, T2, T4);
+ T3 fn(T1, T2);
)
@pte_alloc_func_proto depends on patch exists@
identifier E1, E2, E4;
type T1, T2, T3, T4;
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
@@
(
- T3 fn(T1 E1, T2 E2);
+ T3 fn(T1 E1);
|
- T3 fn(T1 E1, T2 E2, T4 E4);
+ T3 fn(T1 E1, T2 E2);
)
@pte_alloc_func_call depends on patch exists@
expression E2;
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
@@
fn(...
-, E2
)
@pte_alloc_macro depends on patch exists@
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
identifier a, b, c;
expression e;
position p;
@@
(
- #define fn(a, b, c) e
+ #define fn(a, b) e
|
- #define fn(a, b) e
+ #define fn(a) e
)
Link: http://lkml.kernel.org/r/20181108181201.88826-2-joelaf@google.com
Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Acked-by: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Julia Lawall <Julia.Lawall@lip6.fr>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: William Kucharski <william.kucharski@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-04 02:28:34 +03:00
pgtable = pte_alloc_one ( vma - > vm_mm ) ;
2013-09-13 02:14:03 +04:00
if ( unlikely ( ! pgtable ) )
2011-01-14 02:46:58 +03:00
return VM_FAULT_OOM ;
2016-10-08 03:00:08 +03:00
zero_page = mm_get_huge_zero_page ( vma - > vm_mm ) ;
2013-09-13 02:14:03 +04:00
if ( unlikely ( ! zero_page ) ) {
2016-07-27 01:25:20 +03:00
pte_free ( vma - > vm_mm , pgtable ) ;
2011-04-15 02:22:06 +04:00
count_vm_event ( THP_FAULT_FALLBACK ) ;
2013-09-13 02:14:05 +04:00
return VM_FAULT_FALLBACK ;
2011-01-14 02:46:57 +03:00
}
2016-12-15 02:06:58 +03:00
vmf - > ptl = pmd_lock ( vma - > vm_mm , vmf - > pmd ) ;
2015-09-05 01:46:20 +03:00
ret = 0 ;
2016-12-15 02:06:58 +03:00
if ( pmd_none ( * vmf - > pmd ) ) {
2017-08-19 01:16:15 +03:00
ret = check_stable_address_space ( vma - > vm_mm ) ;
if ( ret ) {
spin_unlock ( vmf - > ptl ) ;
mm/userfaultfd: do not access vma->vm_mm after calling handle_userfault()
Alexander reported a syzkaller / KASAN finding on s390, see below for
complete output.
In do_huge_pmd_anonymous_page(), the pre-allocated pagetable will be
freed in some cases. In the case of userfaultfd_missing(), this will
happen after calling handle_userfault(), which might have released the
mmap_lock. Therefore, the following pte_free(vma->vm_mm, pgtable) will
access an unstable vma->vm_mm, which could have been freed or re-used
already.
For all architectures other than s390 this will go w/o any negative
impact, because pte_free() simply frees the page and ignores the
passed-in mm. The implementation for SPARC32 would also access
mm->page_table_lock for pte_free(), but there is no THP support in
SPARC32, so the buggy code path will not be used there.
For s390, the mm->context.pgtable_list is being used to maintain the 2K
pagetable fragments, and operating on an already freed or even re-used
mm could result in various more or less subtle bugs due to list /
pagetable corruption.
Fix this by calling pte_free() before handle_userfault(), similar to how
it is already done in __do_huge_pmd_anonymous_page() for the WRITE /
non-huge_zero_page case.
Commit 6b251fc96cf2c ("userfaultfd: call handle_userfault() for
userfaultfd_missing() faults") actually introduced both, the
do_huge_pmd_anonymous_page() and also __do_huge_pmd_anonymous_page()
changes wrt to calling handle_userfault(), but only in the latter case
it put the pte_free() before calling handle_userfault().
BUG: KASAN: use-after-free in do_huge_pmd_anonymous_page+0xcda/0xd90 mm/huge_memory.c:744
Read of size 8 at addr 00000000962d6988 by task syz-executor.0/9334
CPU: 1 PID: 9334 Comm: syz-executor.0 Not tainted 5.10.0-rc1-syzkaller-07083-g4c9720875573 #0
Hardware name: IBM 3906 M04 701 (KVM/Linux)
Call Trace:
do_huge_pmd_anonymous_page+0xcda/0xd90 mm/huge_memory.c:744
create_huge_pmd mm/memory.c:4256 [inline]
__handle_mm_fault+0xe6e/0x1068 mm/memory.c:4480
handle_mm_fault+0x288/0x748 mm/memory.c:4607
do_exception+0x394/0xae0 arch/s390/mm/fault.c:479
do_dat_exception+0x34/0x80 arch/s390/mm/fault.c:567
pgm_check_handler+0x1da/0x22c arch/s390/kernel/entry.S:706
copy_from_user_mvcos arch/s390/lib/uaccess.c:111 [inline]
raw_copy_from_user+0x3a/0x88 arch/s390/lib/uaccess.c:174
_copy_from_user+0x48/0xa8 lib/usercopy.c:16
copy_from_user include/linux/uaccess.h:192 [inline]
__do_sys_sigaltstack kernel/signal.c:4064 [inline]
__s390x_sys_sigaltstack+0xc8/0x240 kernel/signal.c:4060
system_call+0xe0/0x28c arch/s390/kernel/entry.S:415
Allocated by task 9334:
slab_alloc_node mm/slub.c:2891 [inline]
slab_alloc mm/slub.c:2899 [inline]
kmem_cache_alloc+0x118/0x348 mm/slub.c:2904
vm_area_dup+0x9c/0x2b8 kernel/fork.c:356
__split_vma+0xba/0x560 mm/mmap.c:2742
split_vma+0xca/0x108 mm/mmap.c:2800
mlock_fixup+0x4ae/0x600 mm/mlock.c:550
apply_vma_lock_flags+0x2c6/0x398 mm/mlock.c:619
do_mlock+0x1aa/0x718 mm/mlock.c:711
__do_sys_mlock2 mm/mlock.c:738 [inline]
__s390x_sys_mlock2+0x86/0xa8 mm/mlock.c:728
system_call+0xe0/0x28c arch/s390/kernel/entry.S:415
Freed by task 9333:
slab_free mm/slub.c:3142 [inline]
kmem_cache_free+0x7c/0x4b8 mm/slub.c:3158
__vma_adjust+0x7b2/0x2508 mm/mmap.c:960
vma_merge+0x87e/0xce0 mm/mmap.c:1209
userfaultfd_release+0x412/0x6b8 fs/userfaultfd.c:868
__fput+0x22c/0x7a8 fs/file_table.c:281
task_work_run+0x200/0x320 kernel/task_work.c:151
tracehook_notify_resume include/linux/tracehook.h:188 [inline]
do_notify_resume+0x100/0x148 arch/s390/kernel/signal.c:538
system_call+0xe6/0x28c arch/s390/kernel/entry.S:416
The buggy address belongs to the object at 00000000962d6948 which belongs to the cache vm_area_struct of size 200
The buggy address is located 64 bytes inside of 200-byte region [00000000962d6948, 00000000962d6a10)
The buggy address belongs to the page: page:00000000313a09fe refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x962d6 flags: 0x3ffff00000000200(slab)
raw: 3ffff00000000200 000040000257e080 0000000c0000000c 000000008020ba00
raw: 0000000000000000 000f001e00000000 ffffffff00000001 0000000096959501
page dumped because: kasan: bad access detected
page->mem_cgroup:0000000096959501
Memory state around the buggy address:
00000000962d6880: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
00000000962d6900: 00 fc fc fc fc fc fc fc fc fa fb fb fb fb fb fb
>00000000962d6980: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb
^
00000000962d6a00: fb fb fc fc fc fc fc fc fc fc 00 00 00 00 00 00
00000000962d6a80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
==================================================================
Fixes: 6b251fc96cf2c ("userfaultfd: call handle_userfault() for userfaultfd_missing() faults")
Reported-by: Alexander Egorenkov <egorenar@linux.ibm.com>
Signed-off-by: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: <stable@vger.kernel.org> [4.3+]
Link: https://lkml.kernel.org/r/20201110190329.11920-1-gerald.schaefer@linux.ibm.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-11-22 09:17:15 +03:00
pte_free ( vma - > vm_mm , pgtable ) ;
2017-08-19 01:16:15 +03:00
} else if ( userfaultfd_missing ( vma ) ) {
2016-12-15 02:06:58 +03:00
spin_unlock ( vmf - > ptl ) ;
mm/userfaultfd: do not access vma->vm_mm after calling handle_userfault()
Alexander reported a syzkaller / KASAN finding on s390, see below for
complete output.
In do_huge_pmd_anonymous_page(), the pre-allocated pagetable will be
freed in some cases. In the case of userfaultfd_missing(), this will
happen after calling handle_userfault(), which might have released the
mmap_lock. Therefore, the following pte_free(vma->vm_mm, pgtable) will
access an unstable vma->vm_mm, which could have been freed or re-used
already.
For all architectures other than s390 this will go w/o any negative
impact, because pte_free() simply frees the page and ignores the
passed-in mm. The implementation for SPARC32 would also access
mm->page_table_lock for pte_free(), but there is no THP support in
SPARC32, so the buggy code path will not be used there.
For s390, the mm->context.pgtable_list is being used to maintain the 2K
pagetable fragments, and operating on an already freed or even re-used
mm could result in various more or less subtle bugs due to list /
pagetable corruption.
Fix this by calling pte_free() before handle_userfault(), similar to how
it is already done in __do_huge_pmd_anonymous_page() for the WRITE /
non-huge_zero_page case.
Commit 6b251fc96cf2c ("userfaultfd: call handle_userfault() for
userfaultfd_missing() faults") actually introduced both, the
do_huge_pmd_anonymous_page() and also __do_huge_pmd_anonymous_page()
changes wrt to calling handle_userfault(), but only in the latter case
it put the pte_free() before calling handle_userfault().
BUG: KASAN: use-after-free in do_huge_pmd_anonymous_page+0xcda/0xd90 mm/huge_memory.c:744
Read of size 8 at addr 00000000962d6988 by task syz-executor.0/9334
CPU: 1 PID: 9334 Comm: syz-executor.0 Not tainted 5.10.0-rc1-syzkaller-07083-g4c9720875573 #0
Hardware name: IBM 3906 M04 701 (KVM/Linux)
Call Trace:
do_huge_pmd_anonymous_page+0xcda/0xd90 mm/huge_memory.c:744
create_huge_pmd mm/memory.c:4256 [inline]
__handle_mm_fault+0xe6e/0x1068 mm/memory.c:4480
handle_mm_fault+0x288/0x748 mm/memory.c:4607
do_exception+0x394/0xae0 arch/s390/mm/fault.c:479
do_dat_exception+0x34/0x80 arch/s390/mm/fault.c:567
pgm_check_handler+0x1da/0x22c arch/s390/kernel/entry.S:706
copy_from_user_mvcos arch/s390/lib/uaccess.c:111 [inline]
raw_copy_from_user+0x3a/0x88 arch/s390/lib/uaccess.c:174
_copy_from_user+0x48/0xa8 lib/usercopy.c:16
copy_from_user include/linux/uaccess.h:192 [inline]
__do_sys_sigaltstack kernel/signal.c:4064 [inline]
__s390x_sys_sigaltstack+0xc8/0x240 kernel/signal.c:4060
system_call+0xe0/0x28c arch/s390/kernel/entry.S:415
Allocated by task 9334:
slab_alloc_node mm/slub.c:2891 [inline]
slab_alloc mm/slub.c:2899 [inline]
kmem_cache_alloc+0x118/0x348 mm/slub.c:2904
vm_area_dup+0x9c/0x2b8 kernel/fork.c:356
__split_vma+0xba/0x560 mm/mmap.c:2742
split_vma+0xca/0x108 mm/mmap.c:2800
mlock_fixup+0x4ae/0x600 mm/mlock.c:550
apply_vma_lock_flags+0x2c6/0x398 mm/mlock.c:619
do_mlock+0x1aa/0x718 mm/mlock.c:711
__do_sys_mlock2 mm/mlock.c:738 [inline]
__s390x_sys_mlock2+0x86/0xa8 mm/mlock.c:728
system_call+0xe0/0x28c arch/s390/kernel/entry.S:415
Freed by task 9333:
slab_free mm/slub.c:3142 [inline]
kmem_cache_free+0x7c/0x4b8 mm/slub.c:3158
__vma_adjust+0x7b2/0x2508 mm/mmap.c:960
vma_merge+0x87e/0xce0 mm/mmap.c:1209
userfaultfd_release+0x412/0x6b8 fs/userfaultfd.c:868
__fput+0x22c/0x7a8 fs/file_table.c:281
task_work_run+0x200/0x320 kernel/task_work.c:151
tracehook_notify_resume include/linux/tracehook.h:188 [inline]
do_notify_resume+0x100/0x148 arch/s390/kernel/signal.c:538
system_call+0xe6/0x28c arch/s390/kernel/entry.S:416
The buggy address belongs to the object at 00000000962d6948 which belongs to the cache vm_area_struct of size 200
The buggy address is located 64 bytes inside of 200-byte region [00000000962d6948, 00000000962d6a10)
The buggy address belongs to the page: page:00000000313a09fe refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x962d6 flags: 0x3ffff00000000200(slab)
raw: 3ffff00000000200 000040000257e080 0000000c0000000c 000000008020ba00
raw: 0000000000000000 000f001e00000000 ffffffff00000001 0000000096959501
page dumped because: kasan: bad access detected
page->mem_cgroup:0000000096959501
Memory state around the buggy address:
00000000962d6880: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
00000000962d6900: 00 fc fc fc fc fc fc fc fc fa fb fb fb fb fb fb
>00000000962d6980: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb
^
00000000962d6a00: fb fb fc fc fc fc fc fc fc fc 00 00 00 00 00 00
00000000962d6a80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
==================================================================
Fixes: 6b251fc96cf2c ("userfaultfd: call handle_userfault() for userfaultfd_missing() faults")
Reported-by: Alexander Egorenkov <egorenar@linux.ibm.com>
Signed-off-by: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: <stable@vger.kernel.org> [4.3+]
Link: https://lkml.kernel.org/r/20201110190329.11920-1-gerald.schaefer@linux.ibm.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-11-22 09:17:15 +03:00
pte_free ( vma - > vm_mm , pgtable ) ;
2016-12-15 02:06:58 +03:00
ret = handle_userfault ( vmf , VM_UFFD_MISSING ) ;
2015-09-05 01:46:20 +03:00
VM_BUG_ON ( ret & VM_FAULT_FALLBACK ) ;
} else {
2016-07-27 01:25:20 +03:00
set_huge_zero_page ( pgtable , vma - > vm_mm , vma ,
2016-12-15 02:06:58 +03:00
haddr , vmf - > pmd , zero_page ) ;
2021-02-24 23:06:42 +03:00
update_mmu_cache_pmd ( vma , vmf - > address , vmf - > pmd ) ;
2016-12-15 02:06:58 +03:00
spin_unlock ( vmf - > ptl ) ;
2015-09-05 01:46:20 +03:00
}
mm/userfaultfd: do not access vma->vm_mm after calling handle_userfault()
Alexander reported a syzkaller / KASAN finding on s390, see below for
complete output.
In do_huge_pmd_anonymous_page(), the pre-allocated pagetable will be
freed in some cases. In the case of userfaultfd_missing(), this will
happen after calling handle_userfault(), which might have released the
mmap_lock. Therefore, the following pte_free(vma->vm_mm, pgtable) will
access an unstable vma->vm_mm, which could have been freed or re-used
already.
For all architectures other than s390 this will go w/o any negative
impact, because pte_free() simply frees the page and ignores the
passed-in mm. The implementation for SPARC32 would also access
mm->page_table_lock for pte_free(), but there is no THP support in
SPARC32, so the buggy code path will not be used there.
For s390, the mm->context.pgtable_list is being used to maintain the 2K
pagetable fragments, and operating on an already freed or even re-used
mm could result in various more or less subtle bugs due to list /
pagetable corruption.
Fix this by calling pte_free() before handle_userfault(), similar to how
it is already done in __do_huge_pmd_anonymous_page() for the WRITE /
non-huge_zero_page case.
Commit 6b251fc96cf2c ("userfaultfd: call handle_userfault() for
userfaultfd_missing() faults") actually introduced both, the
do_huge_pmd_anonymous_page() and also __do_huge_pmd_anonymous_page()
changes wrt to calling handle_userfault(), but only in the latter case
it put the pte_free() before calling handle_userfault().
BUG: KASAN: use-after-free in do_huge_pmd_anonymous_page+0xcda/0xd90 mm/huge_memory.c:744
Read of size 8 at addr 00000000962d6988 by task syz-executor.0/9334
CPU: 1 PID: 9334 Comm: syz-executor.0 Not tainted 5.10.0-rc1-syzkaller-07083-g4c9720875573 #0
Hardware name: IBM 3906 M04 701 (KVM/Linux)
Call Trace:
do_huge_pmd_anonymous_page+0xcda/0xd90 mm/huge_memory.c:744
create_huge_pmd mm/memory.c:4256 [inline]
__handle_mm_fault+0xe6e/0x1068 mm/memory.c:4480
handle_mm_fault+0x288/0x748 mm/memory.c:4607
do_exception+0x394/0xae0 arch/s390/mm/fault.c:479
do_dat_exception+0x34/0x80 arch/s390/mm/fault.c:567
pgm_check_handler+0x1da/0x22c arch/s390/kernel/entry.S:706
copy_from_user_mvcos arch/s390/lib/uaccess.c:111 [inline]
raw_copy_from_user+0x3a/0x88 arch/s390/lib/uaccess.c:174
_copy_from_user+0x48/0xa8 lib/usercopy.c:16
copy_from_user include/linux/uaccess.h:192 [inline]
__do_sys_sigaltstack kernel/signal.c:4064 [inline]
__s390x_sys_sigaltstack+0xc8/0x240 kernel/signal.c:4060
system_call+0xe0/0x28c arch/s390/kernel/entry.S:415
Allocated by task 9334:
slab_alloc_node mm/slub.c:2891 [inline]
slab_alloc mm/slub.c:2899 [inline]
kmem_cache_alloc+0x118/0x348 mm/slub.c:2904
vm_area_dup+0x9c/0x2b8 kernel/fork.c:356
__split_vma+0xba/0x560 mm/mmap.c:2742
split_vma+0xca/0x108 mm/mmap.c:2800
mlock_fixup+0x4ae/0x600 mm/mlock.c:550
apply_vma_lock_flags+0x2c6/0x398 mm/mlock.c:619
do_mlock+0x1aa/0x718 mm/mlock.c:711
__do_sys_mlock2 mm/mlock.c:738 [inline]
__s390x_sys_mlock2+0x86/0xa8 mm/mlock.c:728
system_call+0xe0/0x28c arch/s390/kernel/entry.S:415
Freed by task 9333:
slab_free mm/slub.c:3142 [inline]
kmem_cache_free+0x7c/0x4b8 mm/slub.c:3158
__vma_adjust+0x7b2/0x2508 mm/mmap.c:960
vma_merge+0x87e/0xce0 mm/mmap.c:1209
userfaultfd_release+0x412/0x6b8 fs/userfaultfd.c:868
__fput+0x22c/0x7a8 fs/file_table.c:281
task_work_run+0x200/0x320 kernel/task_work.c:151
tracehook_notify_resume include/linux/tracehook.h:188 [inline]
do_notify_resume+0x100/0x148 arch/s390/kernel/signal.c:538
system_call+0xe6/0x28c arch/s390/kernel/entry.S:416
The buggy address belongs to the object at 00000000962d6948 which belongs to the cache vm_area_struct of size 200
The buggy address is located 64 bytes inside of 200-byte region [00000000962d6948, 00000000962d6a10)
The buggy address belongs to the page: page:00000000313a09fe refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x962d6 flags: 0x3ffff00000000200(slab)
raw: 3ffff00000000200 000040000257e080 0000000c0000000c 000000008020ba00
raw: 0000000000000000 000f001e00000000 ffffffff00000001 0000000096959501
page dumped because: kasan: bad access detected
page->mem_cgroup:0000000096959501
Memory state around the buggy address:
00000000962d6880: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
00000000962d6900: 00 fc fc fc fc fc fc fc fc fa fb fb fb fb fb fb
>00000000962d6980: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb
^
00000000962d6a00: fb fb fc fc fc fc fc fc fc fc 00 00 00 00 00 00
00000000962d6a80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
==================================================================
Fixes: 6b251fc96cf2c ("userfaultfd: call handle_userfault() for userfaultfd_missing() faults")
Reported-by: Alexander Egorenkov <egorenar@linux.ibm.com>
Signed-off-by: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: <stable@vger.kernel.org> [4.3+]
Link: https://lkml.kernel.org/r/20201110190329.11920-1-gerald.schaefer@linux.ibm.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-11-22 09:17:15 +03:00
} else {
2016-12-15 02:06:58 +03:00
spin_unlock ( vmf - > ptl ) ;
2016-07-27 01:25:20 +03:00
pte_free ( vma - > vm_mm , pgtable ) ;
mm/userfaultfd: do not access vma->vm_mm after calling handle_userfault()
Alexander reported a syzkaller / KASAN finding on s390, see below for
complete output.
In do_huge_pmd_anonymous_page(), the pre-allocated pagetable will be
freed in some cases. In the case of userfaultfd_missing(), this will
happen after calling handle_userfault(), which might have released the
mmap_lock. Therefore, the following pte_free(vma->vm_mm, pgtable) will
access an unstable vma->vm_mm, which could have been freed or re-used
already.
For all architectures other than s390 this will go w/o any negative
impact, because pte_free() simply frees the page and ignores the
passed-in mm. The implementation for SPARC32 would also access
mm->page_table_lock for pte_free(), but there is no THP support in
SPARC32, so the buggy code path will not be used there.
For s390, the mm->context.pgtable_list is being used to maintain the 2K
pagetable fragments, and operating on an already freed or even re-used
mm could result in various more or less subtle bugs due to list /
pagetable corruption.
Fix this by calling pte_free() before handle_userfault(), similar to how
it is already done in __do_huge_pmd_anonymous_page() for the WRITE /
non-huge_zero_page case.
Commit 6b251fc96cf2c ("userfaultfd: call handle_userfault() for
userfaultfd_missing() faults") actually introduced both, the
do_huge_pmd_anonymous_page() and also __do_huge_pmd_anonymous_page()
changes wrt to calling handle_userfault(), but only in the latter case
it put the pte_free() before calling handle_userfault().
BUG: KASAN: use-after-free in do_huge_pmd_anonymous_page+0xcda/0xd90 mm/huge_memory.c:744
Read of size 8 at addr 00000000962d6988 by task syz-executor.0/9334
CPU: 1 PID: 9334 Comm: syz-executor.0 Not tainted 5.10.0-rc1-syzkaller-07083-g4c9720875573 #0
Hardware name: IBM 3906 M04 701 (KVM/Linux)
Call Trace:
do_huge_pmd_anonymous_page+0xcda/0xd90 mm/huge_memory.c:744
create_huge_pmd mm/memory.c:4256 [inline]
__handle_mm_fault+0xe6e/0x1068 mm/memory.c:4480
handle_mm_fault+0x288/0x748 mm/memory.c:4607
do_exception+0x394/0xae0 arch/s390/mm/fault.c:479
do_dat_exception+0x34/0x80 arch/s390/mm/fault.c:567
pgm_check_handler+0x1da/0x22c arch/s390/kernel/entry.S:706
copy_from_user_mvcos arch/s390/lib/uaccess.c:111 [inline]
raw_copy_from_user+0x3a/0x88 arch/s390/lib/uaccess.c:174
_copy_from_user+0x48/0xa8 lib/usercopy.c:16
copy_from_user include/linux/uaccess.h:192 [inline]
__do_sys_sigaltstack kernel/signal.c:4064 [inline]
__s390x_sys_sigaltstack+0xc8/0x240 kernel/signal.c:4060
system_call+0xe0/0x28c arch/s390/kernel/entry.S:415
Allocated by task 9334:
slab_alloc_node mm/slub.c:2891 [inline]
slab_alloc mm/slub.c:2899 [inline]
kmem_cache_alloc+0x118/0x348 mm/slub.c:2904
vm_area_dup+0x9c/0x2b8 kernel/fork.c:356
__split_vma+0xba/0x560 mm/mmap.c:2742
split_vma+0xca/0x108 mm/mmap.c:2800
mlock_fixup+0x4ae/0x600 mm/mlock.c:550
apply_vma_lock_flags+0x2c6/0x398 mm/mlock.c:619
do_mlock+0x1aa/0x718 mm/mlock.c:711
__do_sys_mlock2 mm/mlock.c:738 [inline]
__s390x_sys_mlock2+0x86/0xa8 mm/mlock.c:728
system_call+0xe0/0x28c arch/s390/kernel/entry.S:415
Freed by task 9333:
slab_free mm/slub.c:3142 [inline]
kmem_cache_free+0x7c/0x4b8 mm/slub.c:3158
__vma_adjust+0x7b2/0x2508 mm/mmap.c:960
vma_merge+0x87e/0xce0 mm/mmap.c:1209
userfaultfd_release+0x412/0x6b8 fs/userfaultfd.c:868
__fput+0x22c/0x7a8 fs/file_table.c:281
task_work_run+0x200/0x320 kernel/task_work.c:151
tracehook_notify_resume include/linux/tracehook.h:188 [inline]
do_notify_resume+0x100/0x148 arch/s390/kernel/signal.c:538
system_call+0xe6/0x28c arch/s390/kernel/entry.S:416
The buggy address belongs to the object at 00000000962d6948 which belongs to the cache vm_area_struct of size 200
The buggy address is located 64 bytes inside of 200-byte region [00000000962d6948, 00000000962d6a10)
The buggy address belongs to the page: page:00000000313a09fe refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x962d6 flags: 0x3ffff00000000200(slab)
raw: 3ffff00000000200 000040000257e080 0000000c0000000c 000000008020ba00
raw: 0000000000000000 000f001e00000000 ffffffff00000001 0000000096959501
page dumped because: kasan: bad access detected
page->mem_cgroup:0000000096959501
Memory state around the buggy address:
00000000962d6880: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
00000000962d6900: 00 fc fc fc fc fc fc fc fc fa fb fb fb fb fb fb
>00000000962d6980: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb
^
00000000962d6a00: fb fb fc fc fc fc fc fc fc fc 00 00 00 00 00 00
00000000962d6a80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
==================================================================
Fixes: 6b251fc96cf2c ("userfaultfd: call handle_userfault() for userfaultfd_missing() faults")
Reported-by: Alexander Egorenkov <egorenar@linux.ibm.com>
Signed-off-by: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: <stable@vger.kernel.org> [4.3+]
Link: https://lkml.kernel.org/r/20201110190329.11920-1-gerald.schaefer@linux.ibm.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-11-22 09:17:15 +03:00
}
2015-09-05 01:46:20 +03:00
return ret ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
}
mm,thp,shmem: limit shmem THP alloc gfp_mask
Patch series "mm,thp,shm: limit shmem THP alloc gfp_mask", v6.
The allocation flags of anonymous transparent huge pages can be controlled
through the files in /sys/kernel/mm/transparent_hugepage/defrag, which can
help the system from getting bogged down in the page reclaim and
compaction code when many THPs are getting allocated simultaneously.
However, the gfp_mask for shmem THP allocations were not limited by those
configuration settings, and some workloads ended up with all CPUs stuck on
the LRU lock in the page reclaim code, trying to allocate dozens of THPs
simultaneously.
This patch applies the same configurated limitation of THPs to shmem
hugepage allocations, to prevent that from happening.
This way a THP defrag setting of "never" or "defer+madvise" will result in
quick allocation failures without direct reclaim when no 2MB free pages
are available.
With this patch applied, THP allocations for tmpfs will be a little more
aggressive than today for files mmapped with MADV_HUGEPAGE, and a little
less aggressive for files that are not mmapped or mapped without that
flag.
This patch (of 4):
The allocation flags of anonymous transparent huge pages can be controlled
through the files in /sys/kernel/mm/transparent_hugepage/defrag, which can
help the system from getting bogged down in the page reclaim and
compaction code when many THPs are getting allocated simultaneously.
However, the gfp_mask for shmem THP allocations were not limited by those
configuration settings, and some workloads ended up with all CPUs stuck on
the LRU lock in the page reclaim code, trying to allocate dozens of THPs
simultaneously.
This patch applies the same configurated limitation of THPs to shmem
hugepage allocations, to prevent that from happening.
Controlling the gfp_mask of THP allocations through the knobs in sysfs
allows users to determine the balance between how aggressively the system
tries to allocate THPs at fault time, and how much the application may end
up stalling attempting those allocations.
This way a THP defrag setting of "never" or "defer+madvise" will result in
quick allocation failures without direct reclaim when no 2MB free pages
are available.
With this patch applied, THP allocations for tmpfs will be a little more
aggressive than today for files mmapped with MADV_HUGEPAGE, and a little
less aggressive for files that are not mmapped or mapped without that
flag.
Link: https://lkml.kernel.org/r/20201124194925.623931-1-riel@surriel.com
Link: https://lkml.kernel.org/r/20201124194925.623931-2-riel@surriel.com
Signed-off-by: Rik van Riel <riel@surriel.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Xu Yu <xuyu@linux.alibaba.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 04:16:18 +03:00
gfp = vma_thp_gfp_mask ( vma ) ;
2022-05-13 06:23:01 +03:00
folio = vma_alloc_folio ( gfp , HPAGE_PMD_ORDER , vma , haddr , true ) ;
if ( unlikely ( ! folio ) ) {
2013-09-13 02:14:03 +04:00
count_vm_event ( THP_FAULT_FALLBACK ) ;
2013-09-13 02:14:05 +04:00
return VM_FAULT_FALLBACK ;
2013-09-13 02:14:03 +04:00
}
2022-05-13 06:23:01 +03:00
return __do_huge_pmd_anonymous_page ( vmf , & folio - > page , gfp ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
}
2015-09-09 00:59:14 +03:00
static void insert_pfn_pmd ( struct vm_area_struct * vma , unsigned long addr ,
2017-05-09 01:59:43 +03:00
pmd_t * pmd , pfn_t pfn , pgprot_t prot , bool write ,
pgtable_t pgtable )
2015-09-09 00:58:54 +03:00
{
struct mm_struct * mm = vma - > vm_mm ;
pmd_t entry ;
spinlock_t * ptl ;
ptl = pmd_lock ( mm , pmd ) ;
2019-04-06 04:39:10 +03:00
if ( ! pmd_none ( * pmd ) ) {
if ( write ) {
if ( pmd_pfn ( * pmd ) ! = pfn_t_to_pfn ( pfn ) ) {
WARN_ON_ONCE ( ! is_huge_zero_pmd ( * pmd ) ) ;
goto out_unlock ;
}
entry = pmd_mkyoung ( * pmd ) ;
entry = maybe_pmd_mkwrite ( pmd_mkdirty ( entry ) , vma ) ;
if ( pmdp_set_access_flags ( vma , addr , pmd , entry , 1 ) )
update_mmu_cache_pmd ( vma , addr , pmd ) ;
}
goto out_unlock ;
}
2016-01-16 03:56:43 +03:00
entry = pmd_mkhuge ( pfn_t_pmd ( pfn , prot ) ) ;
if ( pfn_t_devmap ( pfn ) )
entry = pmd_mkdevmap ( entry ) ;
2016-01-16 03:56:02 +03:00
if ( write ) {
2017-11-29 20:01:01 +03:00
entry = pmd_mkyoung ( pmd_mkdirty ( entry ) ) ;
entry = maybe_pmd_mkwrite ( entry , vma ) ;
2015-09-09 00:58:54 +03:00
}
2017-05-09 01:59:43 +03:00
if ( pgtable ) {
pgtable_trans_huge_deposit ( mm , pmd , pgtable ) ;
2017-11-16 04:35:37 +03:00
mm_inc_nr_ptes ( mm ) ;
2019-04-06 04:39:10 +03:00
pgtable = NULL ;
2017-05-09 01:59:43 +03:00
}
2016-01-16 03:56:02 +03:00
set_pmd_at ( mm , addr , pmd , entry ) ;
update_mmu_cache_pmd ( vma , addr , pmd ) ;
2019-04-06 04:39:10 +03:00
out_unlock :
2015-09-09 00:58:54 +03:00
spin_unlock ( ptl ) ;
2019-04-06 04:39:10 +03:00
if ( pgtable )
pte_free ( mm , pgtable ) ;
2015-09-09 00:58:54 +03:00
}
2020-03-24 20:48:09 +03:00
/**
* vmf_insert_pfn_pmd_prot - insert a pmd size pfn
* @ vmf : Structure describing the fault
* @ pfn : pfn to insert
* @ pgprot : page protection to use
* @ write : whether it ' s a write fault
*
* Insert a pmd size pfn . See vmf_insert_pfn ( ) for additional info and
* also consult the vmf_insert_mixed_prot ( ) documentation when
* @ pgprot ! = @ vmf - > vma - > vm_page_prot .
*
* Return : vm_fault_t value .
*/
vm_fault_t vmf_insert_pfn_pmd_prot ( struct vm_fault * vmf , pfn_t pfn ,
pgprot_t pgprot , bool write )
2015-09-09 00:58:54 +03:00
{
2019-05-14 03:15:33 +03:00
unsigned long addr = vmf - > address & PMD_MASK ;
struct vm_area_struct * vma = vmf - > vma ;
2017-05-09 01:59:43 +03:00
pgtable_t pgtable = NULL ;
2019-05-14 03:15:33 +03:00
2015-09-09 00:58:54 +03:00
/*
* If we had pmd_special , we could avoid all these restrictions ,
* but we need to be consistent with PTEs and architectures that
* can ' t support a ' special ' bit .
*/
2018-08-18 01:43:40 +03:00
BUG_ON ( ! ( vma - > vm_flags & ( VM_PFNMAP | VM_MIXEDMAP ) ) & &
! pfn_t_devmap ( pfn ) ) ;
2015-09-09 00:58:54 +03:00
BUG_ON ( ( vma - > vm_flags & ( VM_PFNMAP | VM_MIXEDMAP ) ) = =
( VM_PFNMAP | VM_MIXEDMAP ) ) ;
BUG_ON ( ( vma - > vm_flags & VM_PFNMAP ) & & is_cow_mapping ( vma - > vm_flags ) ) ;
if ( addr < vma - > vm_start | | addr > = vma - > vm_end )
return VM_FAULT_SIGBUS ;
2016-10-26 20:43:43 +03:00
2017-05-09 01:59:43 +03:00
if ( arch_needs_pgtable_deposit ( ) ) {
mm: treewide: remove unused address argument from pte_alloc functions
Patch series "Add support for fast mremap".
This series speeds up the mremap(2) syscall by copying page tables at
the PMD level even for non-THP systems. There is concern that the extra
'address' argument that mremap passes to pte_alloc may do something
subtle architecture related in the future that may make the scheme not
work. Also we find that there is no point in passing the 'address' to
pte_alloc since its unused. This patch therefore removes this argument
tree-wide resulting in a nice negative diff as well. Also ensuring
along the way that the enabled architectures do not do anything funky
with the 'address' argument that goes unnoticed by the optimization.
Build and boot tested on x86-64. Build tested on arm64. The config
enablement patch for arm64 will be posted in the future after more
testing.
The changes were obtained by applying the following Coccinelle script.
(thanks Julia for answering all Coccinelle questions!).
Following fix ups were done manually:
* Removal of address argument from pte_fragment_alloc
* Removal of pte_alloc_one_fast definitions from m68k and microblaze.
// Options: --include-headers --no-includes
// Note: I split the 'identifier fn' line, so if you are manually
// running it, please unsplit it so it runs for you.
virtual patch
@pte_alloc_func_def depends on patch exists@
identifier E2;
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
type T2;
@@
fn(...
- , T2 E2
)
{ ... }
@pte_alloc_func_proto_noarg depends on patch exists@
type T1, T2, T3, T4;
identifier fn =~ "^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
@@
(
- T3 fn(T1, T2);
+ T3 fn(T1);
|
- T3 fn(T1, T2, T4);
+ T3 fn(T1, T2);
)
@pte_alloc_func_proto depends on patch exists@
identifier E1, E2, E4;
type T1, T2, T3, T4;
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
@@
(
- T3 fn(T1 E1, T2 E2);
+ T3 fn(T1 E1);
|
- T3 fn(T1 E1, T2 E2, T4 E4);
+ T3 fn(T1 E1, T2 E2);
)
@pte_alloc_func_call depends on patch exists@
expression E2;
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
@@
fn(...
-, E2
)
@pte_alloc_macro depends on patch exists@
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
identifier a, b, c;
expression e;
position p;
@@
(
- #define fn(a, b, c) e
+ #define fn(a, b) e
|
- #define fn(a, b) e
+ #define fn(a) e
)
Link: http://lkml.kernel.org/r/20181108181201.88826-2-joelaf@google.com
Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Acked-by: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Julia Lawall <Julia.Lawall@lip6.fr>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: William Kucharski <william.kucharski@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-04 02:28:34 +03:00
pgtable = pte_alloc_one ( vma - > vm_mm ) ;
2017-05-09 01:59:43 +03:00
if ( ! pgtable )
return VM_FAULT_OOM ;
}
2016-10-26 20:43:43 +03:00
track_pfn_insert ( vma , & pgprot , pfn ) ;
2019-05-14 03:15:33 +03:00
insert_pfn_pmd ( vma , addr , vmf - > pmd , pfn , pgprot , write , pgtable ) ;
2015-09-09 00:59:14 +03:00
return VM_FAULT_NOPAGE ;
2015-09-09 00:58:54 +03:00
}
2020-03-24 20:48:09 +03:00
EXPORT_SYMBOL_GPL ( vmf_insert_pfn_pmd_prot ) ;
2015-09-09 00:58:54 +03:00
2017-02-25 01:57:02 +03:00
# ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2017-11-29 20:01:01 +03:00
static pud_t maybe_pud_mkwrite ( pud_t pud , struct vm_area_struct * vma )
2017-02-25 01:57:02 +03:00
{
2017-11-29 20:01:01 +03:00
if ( likely ( vma - > vm_flags & VM_WRITE ) )
2017-02-25 01:57:02 +03:00
pud = pud_mkwrite ( pud ) ;
return pud ;
}
static void insert_pfn_pud ( struct vm_area_struct * vma , unsigned long addr ,
pud_t * pud , pfn_t pfn , pgprot_t prot , bool write )
{
struct mm_struct * mm = vma - > vm_mm ;
pud_t entry ;
spinlock_t * ptl ;
ptl = pud_lock ( mm , pud ) ;
2019-04-06 04:39:10 +03:00
if ( ! pud_none ( * pud ) ) {
if ( write ) {
if ( pud_pfn ( * pud ) ! = pfn_t_to_pfn ( pfn ) ) {
WARN_ON_ONCE ( ! is_huge_zero_pud ( * pud ) ) ;
goto out_unlock ;
}
entry = pud_mkyoung ( * pud ) ;
entry = maybe_pud_mkwrite ( pud_mkdirty ( entry ) , vma ) ;
if ( pudp_set_access_flags ( vma , addr , pud , entry , 1 ) )
update_mmu_cache_pud ( vma , addr , pud ) ;
}
goto out_unlock ;
}
2017-02-25 01:57:02 +03:00
entry = pud_mkhuge ( pfn_t_pud ( pfn , prot ) ) ;
if ( pfn_t_devmap ( pfn ) )
entry = pud_mkdevmap ( entry ) ;
if ( write ) {
2017-11-29 20:01:01 +03:00
entry = pud_mkyoung ( pud_mkdirty ( entry ) ) ;
entry = maybe_pud_mkwrite ( entry , vma ) ;
2017-02-25 01:57:02 +03:00
}
set_pud_at ( mm , addr , pud , entry ) ;
update_mmu_cache_pud ( vma , addr , pud ) ;
2019-04-06 04:39:10 +03:00
out_unlock :
2017-02-25 01:57:02 +03:00
spin_unlock ( ptl ) ;
}
2020-03-24 20:48:09 +03:00
/**
* vmf_insert_pfn_pud_prot - insert a pud size pfn
* @ vmf : Structure describing the fault
* @ pfn : pfn to insert
* @ pgprot : page protection to use
* @ write : whether it ' s a write fault
*
* Insert a pud size pfn . See vmf_insert_pfn ( ) for additional info and
* also consult the vmf_insert_mixed_prot ( ) documentation when
* @ pgprot ! = @ vmf - > vma - > vm_page_prot .
*
* Return : vm_fault_t value .
*/
vm_fault_t vmf_insert_pfn_pud_prot ( struct vm_fault * vmf , pfn_t pfn ,
pgprot_t pgprot , bool write )
2017-02-25 01:57:02 +03:00
{
2019-05-14 03:15:33 +03:00
unsigned long addr = vmf - > address & PUD_MASK ;
struct vm_area_struct * vma = vmf - > vma ;
2017-02-25 01:57:02 +03:00
/*
* If we had pud_special , we could avoid all these restrictions ,
* but we need to be consistent with PTEs and architectures that
* can ' t support a ' special ' bit .
*/
2018-09-05 01:46:16 +03:00
BUG_ON ( ! ( vma - > vm_flags & ( VM_PFNMAP | VM_MIXEDMAP ) ) & &
! pfn_t_devmap ( pfn ) ) ;
2017-02-25 01:57:02 +03:00
BUG_ON ( ( vma - > vm_flags & ( VM_PFNMAP | VM_MIXEDMAP ) ) = =
( VM_PFNMAP | VM_MIXEDMAP ) ) ;
BUG_ON ( ( vma - > vm_flags & VM_PFNMAP ) & & is_cow_mapping ( vma - > vm_flags ) ) ;
if ( addr < vma - > vm_start | | addr > = vma - > vm_end )
return VM_FAULT_SIGBUS ;
track_pfn_insert ( vma , & pgprot , pfn ) ;
2019-05-14 03:15:33 +03:00
insert_pfn_pud ( vma , addr , vmf - > pud , pfn , pgprot , write ) ;
2017-02-25 01:57:02 +03:00
return VM_FAULT_NOPAGE ;
}
2020-03-24 20:48:09 +03:00
EXPORT_SYMBOL_GPL ( vmf_insert_pfn_pud_prot ) ;
2017-02-25 01:57:02 +03:00
# endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2016-01-16 03:56:55 +03:00
static void touch_pmd ( struct vm_area_struct * vma , unsigned long addr ,
2022-07-04 16:21:50 +03:00
pmd_t * pmd , bool write )
2016-01-16 03:56:55 +03:00
{
pmd_t _pmd ;
2017-11-27 06:21:25 +03:00
_pmd = pmd_mkyoung ( * pmd ) ;
2022-07-04 16:21:50 +03:00
if ( write )
2017-11-27 06:21:25 +03:00
_pmd = pmd_mkdirty ( _pmd ) ;
2016-01-16 03:56:55 +03:00
if ( pmdp_set_access_flags ( vma , addr & HPAGE_PMD_MASK ,
2022-07-04 16:21:50 +03:00
pmd , _pmd , write ) )
2016-01-16 03:56:55 +03:00
update_mmu_cache_pmd ( vma , addr , pmd ) ;
}
struct page * follow_devmap_pmd ( struct vm_area_struct * vma , unsigned long addr ,
2018-10-27 01:10:28 +03:00
pmd_t * pmd , int flags , struct dev_pagemap * * pgmap )
2016-01-16 03:56:55 +03:00
{
unsigned long pfn = pmd_pfn ( * pmd ) ;
struct mm_struct * mm = vma - > vm_mm ;
struct page * page ;
2022-10-21 20:41:08 +03:00
int ret ;
2016-01-16 03:56:55 +03:00
assert_spin_locked ( pmd_lockptr ( mm , pmd ) ) ;
mm/gup: track FOLL_PIN pages
Add tracking of pages that were pinned via FOLL_PIN. This tracking is
implemented via overloading of page->_refcount: pins are added by adding
GUP_PIN_COUNTING_BIAS (1024) to the refcount. This provides a fuzzy
indication of pinning, and it can have false positives (and that's OK).
Please see the pre-existing Documentation/core-api/pin_user_pages.rst for
details.
As mentioned in pin_user_pages.rst, callers who effectively set FOLL_PIN
(typically via pin_user_pages*()) are required to ultimately free such
pages via unpin_user_page().
Please also note the limitation, discussed in pin_user_pages.rst under the
"TODO: for 1GB and larger huge pages" section. (That limitation will be
removed in a following patch.)
The effect of a FOLL_PIN flag is similar to that of FOLL_GET, and may be
thought of as "FOLL_GET for DIO and/or RDMA use".
Pages that have been pinned via FOLL_PIN are identifiable via a new
function call:
bool page_maybe_dma_pinned(struct page *page);
What to do in response to encountering such a page, is left to later
patchsets. There is discussion about this in [1], [2], [3], and [4].
This also changes a BUG_ON(), to a WARN_ON(), in follow_page_mask().
[1] Some slow progress on get_user_pages() (Apr 2, 2019):
https://lwn.net/Articles/784574/
[2] DMA and get_user_pages() (LPC: Dec 12, 2018):
https://lwn.net/Articles/774411/
[3] The trouble with get_user_pages() (Apr 30, 2018):
https://lwn.net/Articles/753027/
[4] LWN kernel index: get_user_pages():
https://lwn.net/Kernel/Index/#Memory_management-get_user_pages
[jhubbard@nvidia.com: add kerneldoc]
Link: http://lkml.kernel.org/r/20200307021157.235726-1-jhubbard@nvidia.com
[imbrenda@linux.ibm.com: if pin fails, we need to unpin, a simple put_page will not be enough]
Link: http://lkml.kernel.org/r/20200306132537.783769-2-imbrenda@linux.ibm.com
[akpm@linux-foundation.org: fix put_compound_head defined but not used]
Suggested-by: Jan Kara <jack@suse.cz>
Suggested-by: Jérôme Glisse <jglisse@redhat.com>
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Ira Weiny <ira.weiny@intel.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Link: http://lkml.kernel.org/r/20200211001536.1027652-7-jhubbard@nvidia.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:05:29 +03:00
/* FOLL_GET and FOLL_PIN are mutually exclusive. */
if ( WARN_ON_ONCE ( ( flags & ( FOLL_PIN | FOLL_GET ) ) = =
( FOLL_PIN | FOLL_GET ) ) )
return NULL ;
Revert "mm: replace p??_write with pte_access_permitted in fault + gup paths"
This reverts commits 5c9d2d5c269c, c7da82b894e9, and e7fe7b5cae90.
We'll probably need to revisit this, but basically we should not
complicate the get_user_pages_fast() case, and checking the actual page
table protection key bits will require more care anyway, since the
protection keys depend on the exact state of the VM in question.
Particularly when doing a "remote" page lookup (ie in somebody elses VM,
not your own), you need to be much more careful than this was. Dave
Hansen says:
"So, the underlying bug here is that we now a get_user_pages_remote()
and then go ahead and do the p*_access_permitted() checks against the
current PKRU. This was introduced recently with the addition of the
new p??_access_permitted() calls.
We have checks in the VMA path for the "remote" gups and we avoid
consulting PKRU for them. This got missed in the pkeys selftests
because I did a ptrace read, but not a *write*. I also didn't
explicitly test it against something where a COW needed to be done"
It's also not entirely clear that it makes sense to check the protection
key bits at this level at all. But one possible eventual solution is to
make the get_user_pages_fast() case just abort if it sees protection key
bits set, which makes us fall back to the regular get_user_pages() case,
which then has a vma and can do the check there if we want to.
We'll see.
Somewhat related to this all: what we _do_ want to do some day is to
check the PAGE_USER bit - it should obviously always be set for user
pages, but it would be a good check to have back. Because we have no
generic way to test for it, we lost it as part of moving over from the
architecture-specific x86 GUP implementation to the generic one in
commit e585513b76f7 ("x86/mm/gup: Switch GUP to the generic
get_user_page_fast() implementation").
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "Jérôme Glisse" <jglisse@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-12-16 05:53:22 +03:00
if ( flags & FOLL_WRITE & & ! pmd_write ( * pmd ) )
2016-01-16 03:56:55 +03:00
return NULL ;
if ( pmd_present ( * pmd ) & & pmd_devmap ( * pmd ) )
/* pass */ ;
else
return NULL ;
if ( flags & FOLL_TOUCH )
2022-07-04 16:21:50 +03:00
touch_pmd ( vma , addr , pmd , flags & FOLL_WRITE ) ;
2016-01-16 03:56:55 +03:00
/*
* device mapped pages can only be returned if the
* caller will manage the page reference count .
*/
mm/gup: track FOLL_PIN pages
Add tracking of pages that were pinned via FOLL_PIN. This tracking is
implemented via overloading of page->_refcount: pins are added by adding
GUP_PIN_COUNTING_BIAS (1024) to the refcount. This provides a fuzzy
indication of pinning, and it can have false positives (and that's OK).
Please see the pre-existing Documentation/core-api/pin_user_pages.rst for
details.
As mentioned in pin_user_pages.rst, callers who effectively set FOLL_PIN
(typically via pin_user_pages*()) are required to ultimately free such
pages via unpin_user_page().
Please also note the limitation, discussed in pin_user_pages.rst under the
"TODO: for 1GB and larger huge pages" section. (That limitation will be
removed in a following patch.)
The effect of a FOLL_PIN flag is similar to that of FOLL_GET, and may be
thought of as "FOLL_GET for DIO and/or RDMA use".
Pages that have been pinned via FOLL_PIN are identifiable via a new
function call:
bool page_maybe_dma_pinned(struct page *page);
What to do in response to encountering such a page, is left to later
patchsets. There is discussion about this in [1], [2], [3], and [4].
This also changes a BUG_ON(), to a WARN_ON(), in follow_page_mask().
[1] Some slow progress on get_user_pages() (Apr 2, 2019):
https://lwn.net/Articles/784574/
[2] DMA and get_user_pages() (LPC: Dec 12, 2018):
https://lwn.net/Articles/774411/
[3] The trouble with get_user_pages() (Apr 30, 2018):
https://lwn.net/Articles/753027/
[4] LWN kernel index: get_user_pages():
https://lwn.net/Kernel/Index/#Memory_management-get_user_pages
[jhubbard@nvidia.com: add kerneldoc]
Link: http://lkml.kernel.org/r/20200307021157.235726-1-jhubbard@nvidia.com
[imbrenda@linux.ibm.com: if pin fails, we need to unpin, a simple put_page will not be enough]
Link: http://lkml.kernel.org/r/20200306132537.783769-2-imbrenda@linux.ibm.com
[akpm@linux-foundation.org: fix put_compound_head defined but not used]
Suggested-by: Jan Kara <jack@suse.cz>
Suggested-by: Jérôme Glisse <jglisse@redhat.com>
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Ira Weiny <ira.weiny@intel.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Link: http://lkml.kernel.org/r/20200211001536.1027652-7-jhubbard@nvidia.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:05:29 +03:00
if ( ! ( flags & ( FOLL_GET | FOLL_PIN ) ) )
2016-01-16 03:56:55 +03:00
return ERR_PTR ( - EEXIST ) ;
pfn + = ( addr & ~ PMD_MASK ) > > PAGE_SHIFT ;
2018-10-27 01:10:28 +03:00
* pgmap = get_dev_pagemap ( pfn , * pgmap ) ;
if ( ! * pgmap )
2016-01-16 03:56:55 +03:00
return ERR_PTR ( - EFAULT ) ;
page = pfn_to_page ( pfn ) ;
2022-10-21 20:41:08 +03:00
ret = try_grab_page ( page , flags ) ;
if ( ret )
page = ERR_PTR ( ret ) ;
2016-01-16 03:56:55 +03:00
return page ;
}
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
int copy_huge_pmd ( struct mm_struct * dst_mm , struct mm_struct * src_mm ,
pmd_t * dst_pmd , pmd_t * src_pmd , unsigned long addr ,
mm/userfaultfd: fix uffd-wp special cases for fork()
We tried to do something similar in b569a1760782 ("userfaultfd: wp: drop
_PAGE_UFFD_WP properly when fork") previously, but it's not doing it all
right.. A few fixes around the code path:
1. We were referencing VM_UFFD_WP vm_flags on the _old_ vma rather
than the new vma. That's overlooked in b569a1760782, so it won't work
as expected. Thanks to the recent rework on fork code
(7a4830c380f3a8b3), we can easily get the new vma now, so switch the
checks to that.
2. Dropping the uffd-wp bit in copy_huge_pmd() could be wrong if the
huge pmd is a migration huge pmd. When it happens, instead of using
pmd_uffd_wp(), we should use pmd_swp_uffd_wp(). The fix is simply to
handle them separately.
3. Forget to carry over uffd-wp bit for a write migration huge pmd
entry. This also happens in copy_huge_pmd(), where we converted a
write huge migration entry into a read one.
4. In copy_nonpresent_pte(), drop uffd-wp if necessary for swap ptes.
5. In copy_present_page() when COW is enforced when fork(), we also
need to pass over the uffd-wp bit if VM_UFFD_WP is armed on the new
vma, and when the pte to be copied has uffd-wp bit set.
Remove the comment in copy_present_pte() about this. It won't help a huge
lot to only comment there, but comment everywhere would be an overkill.
Let's assume the commit messages would help.
[peterx@redhat.com: fix a few thp pmd missing uffd-wp bit]
Link: https://lkml.kernel.org/r/20210428225030.9708-4-peterx@redhat.com
Link: https://lkml.kernel.org/r/20210428225030.9708-3-peterx@redhat.com
Fixes: b569a1760782f ("userfaultfd: wp: drop _PAGE_UFFD_WP properly when fork")
Signed-off-by: Peter Xu <peterx@redhat.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Brian Geffon <bgeffon@google.com>
Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Joe Perches <joe@perches.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Lokesh Gidra <lokeshgidra@google.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Oliver Upton <oupton@google.com>
Cc: Shaohua Li <shli@fb.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Wang Qing <wangqing@vivo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-01 04:49:02 +03:00
struct vm_area_struct * dst_vma , struct vm_area_struct * src_vma )
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
{
2013-11-15 02:31:04 +04:00
spinlock_t * dst_ptl , * src_ptl ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
struct page * src_page ;
pmd_t pmd ;
2016-02-03 03:57:57 +03:00
pgtable_t pgtable = NULL ;
2016-07-27 01:25:42 +03:00
int ret = - ENOMEM ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
2016-07-27 01:25:42 +03:00
/* Skip if can be re-fill on fault */
mm/userfaultfd: fix uffd-wp special cases for fork()
We tried to do something similar in b569a1760782 ("userfaultfd: wp: drop
_PAGE_UFFD_WP properly when fork") previously, but it's not doing it all
right.. A few fixes around the code path:
1. We were referencing VM_UFFD_WP vm_flags on the _old_ vma rather
than the new vma. That's overlooked in b569a1760782, so it won't work
as expected. Thanks to the recent rework on fork code
(7a4830c380f3a8b3), we can easily get the new vma now, so switch the
checks to that.
2. Dropping the uffd-wp bit in copy_huge_pmd() could be wrong if the
huge pmd is a migration huge pmd. When it happens, instead of using
pmd_uffd_wp(), we should use pmd_swp_uffd_wp(). The fix is simply to
handle them separately.
3. Forget to carry over uffd-wp bit for a write migration huge pmd
entry. This also happens in copy_huge_pmd(), where we converted a
write huge migration entry into a read one.
4. In copy_nonpresent_pte(), drop uffd-wp if necessary for swap ptes.
5. In copy_present_page() when COW is enforced when fork(), we also
need to pass over the uffd-wp bit if VM_UFFD_WP is armed on the new
vma, and when the pte to be copied has uffd-wp bit set.
Remove the comment in copy_present_pte() about this. It won't help a huge
lot to only comment there, but comment everywhere would be an overkill.
Let's assume the commit messages would help.
[peterx@redhat.com: fix a few thp pmd missing uffd-wp bit]
Link: https://lkml.kernel.org/r/20210428225030.9708-4-peterx@redhat.com
Link: https://lkml.kernel.org/r/20210428225030.9708-3-peterx@redhat.com
Fixes: b569a1760782f ("userfaultfd: wp: drop _PAGE_UFFD_WP properly when fork")
Signed-off-by: Peter Xu <peterx@redhat.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Brian Geffon <bgeffon@google.com>
Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Joe Perches <joe@perches.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Lokesh Gidra <lokeshgidra@google.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Oliver Upton <oupton@google.com>
Cc: Shaohua Li <shli@fb.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Wang Qing <wangqing@vivo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-01 04:49:02 +03:00
if ( ! vma_is_anonymous ( dst_vma ) )
2016-07-27 01:25:42 +03:00
return 0 ;
mm: treewide: remove unused address argument from pte_alloc functions
Patch series "Add support for fast mremap".
This series speeds up the mremap(2) syscall by copying page tables at
the PMD level even for non-THP systems. There is concern that the extra
'address' argument that mremap passes to pte_alloc may do something
subtle architecture related in the future that may make the scheme not
work. Also we find that there is no point in passing the 'address' to
pte_alloc since its unused. This patch therefore removes this argument
tree-wide resulting in a nice negative diff as well. Also ensuring
along the way that the enabled architectures do not do anything funky
with the 'address' argument that goes unnoticed by the optimization.
Build and boot tested on x86-64. Build tested on arm64. The config
enablement patch for arm64 will be posted in the future after more
testing.
The changes were obtained by applying the following Coccinelle script.
(thanks Julia for answering all Coccinelle questions!).
Following fix ups were done manually:
* Removal of address argument from pte_fragment_alloc
* Removal of pte_alloc_one_fast definitions from m68k and microblaze.
// Options: --include-headers --no-includes
// Note: I split the 'identifier fn' line, so if you are manually
// running it, please unsplit it so it runs for you.
virtual patch
@pte_alloc_func_def depends on patch exists@
identifier E2;
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
type T2;
@@
fn(...
- , T2 E2
)
{ ... }
@pte_alloc_func_proto_noarg depends on patch exists@
type T1, T2, T3, T4;
identifier fn =~ "^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
@@
(
- T3 fn(T1, T2);
+ T3 fn(T1);
|
- T3 fn(T1, T2, T4);
+ T3 fn(T1, T2);
)
@pte_alloc_func_proto depends on patch exists@
identifier E1, E2, E4;
type T1, T2, T3, T4;
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
@@
(
- T3 fn(T1 E1, T2 E2);
+ T3 fn(T1 E1);
|
- T3 fn(T1 E1, T2 E2, T4 E4);
+ T3 fn(T1 E1, T2 E2);
)
@pte_alloc_func_call depends on patch exists@
expression E2;
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
@@
fn(...
-, E2
)
@pte_alloc_macro depends on patch exists@
identifier fn =~
"^(__pte_alloc|pte_alloc_one|pte_alloc|__pte_alloc_kernel|pte_alloc_one_kernel)$";
identifier a, b, c;
expression e;
position p;
@@
(
- #define fn(a, b, c) e
+ #define fn(a, b) e
|
- #define fn(a, b) e
+ #define fn(a) e
)
Link: http://lkml.kernel.org/r/20181108181201.88826-2-joelaf@google.com
Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Acked-by: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Julia Lawall <Julia.Lawall@lip6.fr>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: William Kucharski <william.kucharski@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-04 02:28:34 +03:00
pgtable = pte_alloc_one ( dst_mm ) ;
2016-07-27 01:25:42 +03:00
if ( unlikely ( ! pgtable ) )
goto out ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
2013-11-15 02:31:04 +04:00
dst_ptl = pmd_lock ( dst_mm , dst_pmd ) ;
src_ptl = pmd_lockptr ( src_mm , src_pmd ) ;
spin_lock_nested ( src_ptl , SINGLE_DEPTH_NESTING ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
ret = - EAGAIN ;
pmd = * src_pmd ;
mm: thp: check pmd migration entry in common path
When THP migration is being used, memory management code needs to handle
pmd migration entries properly. This patch uses !pmd_present() or
is_swap_pmd() (depending on whether pmd_none() needs separate code or
not) to check pmd migration entries at the places where a pmd entry is
present.
Since pmd-related code uses split_huge_page(), split_huge_pmd(),
pmd_trans_huge(), pmd_trans_unstable(), or
pmd_none_or_trans_huge_or_clear_bad(), this patch:
1. adds pmd migration entry split code in split_huge_pmd(),
2. takes care of pmd migration entries whenever pmd_trans_huge() is present,
3. makes pmd_none_or_trans_huge_or_clear_bad() pmd migration entry aware.
Since split_huge_page() uses split_huge_pmd() and pmd_trans_unstable()
is equivalent to pmd_none_or_trans_huge_or_clear_bad(), we do not change
them.
Until this commit, a pmd entry should be:
1. pointing to a pte page,
2. is_swap_pmd(),
3. pmd_trans_huge(),
4. pmd_devmap(), or
5. pmd_none().
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-09 02:11:01 +03:00
# ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
if ( unlikely ( is_swap_pmd ( pmd ) ) ) {
swp_entry_t entry = pmd_to_swp_entry ( pmd ) ;
VM_BUG_ON ( ! is_pmd_migration_entry ( pmd ) ) ;
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
if ( ! is_readable_migration_entry ( entry ) ) {
2021-07-01 04:54:09 +03:00
entry = make_readable_migration_entry (
swp_offset ( entry ) ) ;
mm: thp: check pmd migration entry in common path
When THP migration is being used, memory management code needs to handle
pmd migration entries properly. This patch uses !pmd_present() or
is_swap_pmd() (depending on whether pmd_none() needs separate code or
not) to check pmd migration entries at the places where a pmd entry is
present.
Since pmd-related code uses split_huge_page(), split_huge_pmd(),
pmd_trans_huge(), pmd_trans_unstable(), or
pmd_none_or_trans_huge_or_clear_bad(), this patch:
1. adds pmd migration entry split code in split_huge_pmd(),
2. takes care of pmd migration entries whenever pmd_trans_huge() is present,
3. makes pmd_none_or_trans_huge_or_clear_bad() pmd migration entry aware.
Since split_huge_page() uses split_huge_pmd() and pmd_trans_unstable()
is equivalent to pmd_none_or_trans_huge_or_clear_bad(), we do not change
them.
Until this commit, a pmd entry should be:
1. pointing to a pte page,
2. is_swap_pmd(),
3. pmd_trans_huge(),
4. pmd_devmap(), or
5. pmd_none().
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-09 02:11:01 +03:00
pmd = swp_entry_to_pmd ( entry ) ;
2017-09-09 02:11:04 +03:00
if ( pmd_swp_soft_dirty ( * src_pmd ) )
pmd = pmd_swp_mksoft_dirty ( pmd ) ;
mm/userfaultfd: fix uffd-wp special cases for fork()
We tried to do something similar in b569a1760782 ("userfaultfd: wp: drop
_PAGE_UFFD_WP properly when fork") previously, but it's not doing it all
right.. A few fixes around the code path:
1. We were referencing VM_UFFD_WP vm_flags on the _old_ vma rather
than the new vma. That's overlooked in b569a1760782, so it won't work
as expected. Thanks to the recent rework on fork code
(7a4830c380f3a8b3), we can easily get the new vma now, so switch the
checks to that.
2. Dropping the uffd-wp bit in copy_huge_pmd() could be wrong if the
huge pmd is a migration huge pmd. When it happens, instead of using
pmd_uffd_wp(), we should use pmd_swp_uffd_wp(). The fix is simply to
handle them separately.
3. Forget to carry over uffd-wp bit for a write migration huge pmd
entry. This also happens in copy_huge_pmd(), where we converted a
write huge migration entry into a read one.
4. In copy_nonpresent_pte(), drop uffd-wp if necessary for swap ptes.
5. In copy_present_page() when COW is enforced when fork(), we also
need to pass over the uffd-wp bit if VM_UFFD_WP is armed on the new
vma, and when the pte to be copied has uffd-wp bit set.
Remove the comment in copy_present_pte() about this. It won't help a huge
lot to only comment there, but comment everywhere would be an overkill.
Let's assume the commit messages would help.
[peterx@redhat.com: fix a few thp pmd missing uffd-wp bit]
Link: https://lkml.kernel.org/r/20210428225030.9708-4-peterx@redhat.com
Link: https://lkml.kernel.org/r/20210428225030.9708-3-peterx@redhat.com
Fixes: b569a1760782f ("userfaultfd: wp: drop _PAGE_UFFD_WP properly when fork")
Signed-off-by: Peter Xu <peterx@redhat.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Brian Geffon <bgeffon@google.com>
Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Joe Perches <joe@perches.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Lokesh Gidra <lokeshgidra@google.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Oliver Upton <oupton@google.com>
Cc: Shaohua Li <shli@fb.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Wang Qing <wangqing@vivo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-01 04:49:02 +03:00
if ( pmd_swp_uffd_wp ( * src_pmd ) )
pmd = pmd_swp_mkuffd_wp ( pmd ) ;
mm: thp: check pmd migration entry in common path
When THP migration is being used, memory management code needs to handle
pmd migration entries properly. This patch uses !pmd_present() or
is_swap_pmd() (depending on whether pmd_none() needs separate code or
not) to check pmd migration entries at the places where a pmd entry is
present.
Since pmd-related code uses split_huge_page(), split_huge_pmd(),
pmd_trans_huge(), pmd_trans_unstable(), or
pmd_none_or_trans_huge_or_clear_bad(), this patch:
1. adds pmd migration entry split code in split_huge_pmd(),
2. takes care of pmd migration entries whenever pmd_trans_huge() is present,
3. makes pmd_none_or_trans_huge_or_clear_bad() pmd migration entry aware.
Since split_huge_page() uses split_huge_pmd() and pmd_trans_unstable()
is equivalent to pmd_none_or_trans_huge_or_clear_bad(), we do not change
them.
Until this commit, a pmd entry should be:
1. pointing to a pte page,
2. is_swap_pmd(),
3. pmd_trans_huge(),
4. pmd_devmap(), or
5. pmd_none().
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-09 02:11:01 +03:00
set_pmd_at ( src_mm , addr , src_pmd , pmd ) ;
}
2017-11-03 01:59:47 +03:00
add_mm_counter ( dst_mm , MM_ANONPAGES , HPAGE_PMD_NR ) ;
2017-11-16 04:35:40 +03:00
mm_inc_nr_ptes ( dst_mm ) ;
2017-11-03 01:59:47 +03:00
pgtable_trans_huge_deposit ( dst_mm , dst_pmd , pgtable ) ;
mm/userfaultfd: fix uffd-wp special cases for fork()
We tried to do something similar in b569a1760782 ("userfaultfd: wp: drop
_PAGE_UFFD_WP properly when fork") previously, but it's not doing it all
right.. A few fixes around the code path:
1. We were referencing VM_UFFD_WP vm_flags on the _old_ vma rather
than the new vma. That's overlooked in b569a1760782, so it won't work
as expected. Thanks to the recent rework on fork code
(7a4830c380f3a8b3), we can easily get the new vma now, so switch the
checks to that.
2. Dropping the uffd-wp bit in copy_huge_pmd() could be wrong if the
huge pmd is a migration huge pmd. When it happens, instead of using
pmd_uffd_wp(), we should use pmd_swp_uffd_wp(). The fix is simply to
handle them separately.
3. Forget to carry over uffd-wp bit for a write migration huge pmd
entry. This also happens in copy_huge_pmd(), where we converted a
write huge migration entry into a read one.
4. In copy_nonpresent_pte(), drop uffd-wp if necessary for swap ptes.
5. In copy_present_page() when COW is enforced when fork(), we also
need to pass over the uffd-wp bit if VM_UFFD_WP is armed on the new
vma, and when the pte to be copied has uffd-wp bit set.
Remove the comment in copy_present_pte() about this. It won't help a huge
lot to only comment there, but comment everywhere would be an overkill.
Let's assume the commit messages would help.
[peterx@redhat.com: fix a few thp pmd missing uffd-wp bit]
Link: https://lkml.kernel.org/r/20210428225030.9708-4-peterx@redhat.com
Link: https://lkml.kernel.org/r/20210428225030.9708-3-peterx@redhat.com
Fixes: b569a1760782f ("userfaultfd: wp: drop _PAGE_UFFD_WP properly when fork")
Signed-off-by: Peter Xu <peterx@redhat.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Brian Geffon <bgeffon@google.com>
Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Joe Perches <joe@perches.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Lokesh Gidra <lokeshgidra@google.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Oliver Upton <oupton@google.com>
Cc: Shaohua Li <shli@fb.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Wang Qing <wangqing@vivo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-01 04:49:02 +03:00
if ( ! userfaultfd_wp ( dst_vma ) )
pmd = pmd_swp_clear_uffd_wp ( pmd ) ;
mm: thp: check pmd migration entry in common path
When THP migration is being used, memory management code needs to handle
pmd migration entries properly. This patch uses !pmd_present() or
is_swap_pmd() (depending on whether pmd_none() needs separate code or
not) to check pmd migration entries at the places where a pmd entry is
present.
Since pmd-related code uses split_huge_page(), split_huge_pmd(),
pmd_trans_huge(), pmd_trans_unstable(), or
pmd_none_or_trans_huge_or_clear_bad(), this patch:
1. adds pmd migration entry split code in split_huge_pmd(),
2. takes care of pmd migration entries whenever pmd_trans_huge() is present,
3. makes pmd_none_or_trans_huge_or_clear_bad() pmd migration entry aware.
Since split_huge_page() uses split_huge_pmd() and pmd_trans_unstable()
is equivalent to pmd_none_or_trans_huge_or_clear_bad(), we do not change
them.
Until this commit, a pmd entry should be:
1. pointing to a pte page,
2. is_swap_pmd(),
3. pmd_trans_huge(),
4. pmd_devmap(), or
5. pmd_none().
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-09 02:11:01 +03:00
set_pmd_at ( dst_mm , addr , dst_pmd , pmd ) ;
ret = 0 ;
goto out_unlock ;
}
# endif
2016-07-27 01:25:42 +03:00
if ( unlikely ( ! pmd_trans_huge ( pmd ) ) ) {
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
pte_free ( dst_mm , pgtable ) ;
goto out_unlock ;
}
2012-12-13 01:50:51 +04:00
/*
2013-11-15 02:31:04 +04:00
* When page table lock is held , the huge zero pmd should not be
2012-12-13 01:50:51 +04:00
* under splitting since we don ' t split the page itself , only pmd to
* a page table .
*/
if ( is_huge_zero_pmd ( pmd ) ) {
2012-12-13 01:51:06 +04:00
/*
* get_huge_zero_page ( ) will never allocate a new page here ,
* since we already have a zero page to copy . It just takes a
* reference .
*/
mm/thp: simplify copying of huge zero page pmd when fork
Patch series "mm/uffd: Misc fix for uffd-wp and one more test".
This series tries to fix some corner case bugs for uffd-wp on either thp
or fork(). Then it introduced a new test with pagemap/pageout.
Patch layout:
Patch 1: cleanup for THP, it'll slightly simplify the follow up patches
Patch 2-4: misc fixes for uffd-wp here and there; please refer to each patch
Patch 5: add pagemap support for uffd-wp
Patch 6: add pagemap/pageout test for uffd-wp
The last test introduced can also verify some of the fixes in previous
patches, as the test will fail without the fixes. However it's not easy
to verify all the changes in patch 2-4, but hopefully they can still be
properly reviewed.
Note that if considering the ongoing uffd-wp shmem & hugetlbfs work, patch
5 will be incomplete as it's missing e.g. hugetlbfs part or the special
swap pte detection. However that's not needed in this series, and since
that series is still during review, this series does not depend on that
one (the last test only runs with anonymous memory, not file-backed). So
this series can be merged even before that series.
This patch (of 6):
Huge zero page is handled in a special path in copy_huge_pmd(), however it
should share most codes with a normal thp page. Trying to share more code
with it by removing the special path. The only leftover so far is the
huge zero page refcounting (mm_get_huge_zero_page()), because that's
separately done with a global counter.
This prepares for a future patch to modify the huge pmd to be installed,
so that we don't need to duplicate it explicitly into huge zero page case
too.
Link: https://lkml.kernel.org/r/20210428225030.9708-1-peterx@redhat.com
Link: https://lkml.kernel.org/r/20210428225030.9708-2-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Mike Kravetz <mike.kravetz@oracle.com>, peterx@redhat.com
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Brian Geffon <bgeffon@google.com>
Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com>
Cc: Joe Perches <joe@perches.com>
Cc: Lokesh Gidra <lokeshgidra@google.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Oliver Upton <oupton@google.com>
Cc: Shaohua Li <shli@fb.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Wang Qing <wangqing@vivo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-01 04:48:59 +03:00
mm_get_huge_zero_page ( dst_mm ) ;
goto out_zero_page ;
2012-12-13 01:50:51 +04:00
}
2013-12-19 05:08:42 +04:00
2016-07-27 01:25:42 +03:00
src_page = pmd_page ( pmd ) ;
VM_BUG_ON_PAGE ( ! PageHead ( src_page ) , src_page ) ;
2020-09-26 01:26:00 +03:00
2022-05-10 04:20:43 +03:00
get_page ( src_page ) ;
if ( unlikely ( page_try_dup_anon_rmap ( src_page , true , src_vma ) ) ) {
/* Page maybe pinned: split and retry the fault on PTEs. */
put_page ( src_page ) ;
2020-09-26 01:26:00 +03:00
pte_free ( dst_mm , pgtable ) ;
spin_unlock ( src_ptl ) ;
spin_unlock ( dst_ptl ) ;
mm/userfaultfd: fix uffd-wp special cases for fork()
We tried to do something similar in b569a1760782 ("userfaultfd: wp: drop
_PAGE_UFFD_WP properly when fork") previously, but it's not doing it all
right.. A few fixes around the code path:
1. We were referencing VM_UFFD_WP vm_flags on the _old_ vma rather
than the new vma. That's overlooked in b569a1760782, so it won't work
as expected. Thanks to the recent rework on fork code
(7a4830c380f3a8b3), we can easily get the new vma now, so switch the
checks to that.
2. Dropping the uffd-wp bit in copy_huge_pmd() could be wrong if the
huge pmd is a migration huge pmd. When it happens, instead of using
pmd_uffd_wp(), we should use pmd_swp_uffd_wp(). The fix is simply to
handle them separately.
3. Forget to carry over uffd-wp bit for a write migration huge pmd
entry. This also happens in copy_huge_pmd(), where we converted a
write huge migration entry into a read one.
4. In copy_nonpresent_pte(), drop uffd-wp if necessary for swap ptes.
5. In copy_present_page() when COW is enforced when fork(), we also
need to pass over the uffd-wp bit if VM_UFFD_WP is armed on the new
vma, and when the pte to be copied has uffd-wp bit set.
Remove the comment in copy_present_pte() about this. It won't help a huge
lot to only comment there, but comment everywhere would be an overkill.
Let's assume the commit messages would help.
[peterx@redhat.com: fix a few thp pmd missing uffd-wp bit]
Link: https://lkml.kernel.org/r/20210428225030.9708-4-peterx@redhat.com
Link: https://lkml.kernel.org/r/20210428225030.9708-3-peterx@redhat.com
Fixes: b569a1760782f ("userfaultfd: wp: drop _PAGE_UFFD_WP properly when fork")
Signed-off-by: Peter Xu <peterx@redhat.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Brian Geffon <bgeffon@google.com>
Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Joe Perches <joe@perches.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Lokesh Gidra <lokeshgidra@google.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Oliver Upton <oupton@google.com>
Cc: Shaohua Li <shli@fb.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Wang Qing <wangqing@vivo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-01 04:49:02 +03:00
__split_huge_pmd ( src_vma , src_pmd , addr , false , NULL ) ;
2020-09-26 01:26:00 +03:00
return - EAGAIN ;
}
2016-07-27 01:25:42 +03:00
add_mm_counter ( dst_mm , MM_ANONPAGES , HPAGE_PMD_NR ) ;
mm/thp: simplify copying of huge zero page pmd when fork
Patch series "mm/uffd: Misc fix for uffd-wp and one more test".
This series tries to fix some corner case bugs for uffd-wp on either thp
or fork(). Then it introduced a new test with pagemap/pageout.
Patch layout:
Patch 1: cleanup for THP, it'll slightly simplify the follow up patches
Patch 2-4: misc fixes for uffd-wp here and there; please refer to each patch
Patch 5: add pagemap support for uffd-wp
Patch 6: add pagemap/pageout test for uffd-wp
The last test introduced can also verify some of the fixes in previous
patches, as the test will fail without the fixes. However it's not easy
to verify all the changes in patch 2-4, but hopefully they can still be
properly reviewed.
Note that if considering the ongoing uffd-wp shmem & hugetlbfs work, patch
5 will be incomplete as it's missing e.g. hugetlbfs part or the special
swap pte detection. However that's not needed in this series, and since
that series is still during review, this series does not depend on that
one (the last test only runs with anonymous memory, not file-backed). So
this series can be merged even before that series.
This patch (of 6):
Huge zero page is handled in a special path in copy_huge_pmd(), however it
should share most codes with a normal thp page. Trying to share more code
with it by removing the special path. The only leftover so far is the
huge zero page refcounting (mm_get_huge_zero_page()), because that's
separately done with a global counter.
This prepares for a future patch to modify the huge pmd to be installed,
so that we don't need to duplicate it explicitly into huge zero page case
too.
Link: https://lkml.kernel.org/r/20210428225030.9708-1-peterx@redhat.com
Link: https://lkml.kernel.org/r/20210428225030.9708-2-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Mike Kravetz <mike.kravetz@oracle.com>, peterx@redhat.com
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Brian Geffon <bgeffon@google.com>
Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com>
Cc: Joe Perches <joe@perches.com>
Cc: Lokesh Gidra <lokeshgidra@google.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Oliver Upton <oupton@google.com>
Cc: Shaohua Li <shli@fb.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Wang Qing <wangqing@vivo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-01 04:48:59 +03:00
out_zero_page :
2017-11-16 04:35:37 +03:00
mm_inc_nr_ptes ( dst_mm ) ;
2016-07-27 01:25:42 +03:00
pgtable_trans_huge_deposit ( dst_mm , dst_pmd , pgtable ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
pmdp_set_wrprotect ( src_mm , addr , src_pmd ) ;
mm/userfaultfd: fix uffd-wp special cases for fork()
We tried to do something similar in b569a1760782 ("userfaultfd: wp: drop
_PAGE_UFFD_WP properly when fork") previously, but it's not doing it all
right.. A few fixes around the code path:
1. We were referencing VM_UFFD_WP vm_flags on the _old_ vma rather
than the new vma. That's overlooked in b569a1760782, so it won't work
as expected. Thanks to the recent rework on fork code
(7a4830c380f3a8b3), we can easily get the new vma now, so switch the
checks to that.
2. Dropping the uffd-wp bit in copy_huge_pmd() could be wrong if the
huge pmd is a migration huge pmd. When it happens, instead of using
pmd_uffd_wp(), we should use pmd_swp_uffd_wp(). The fix is simply to
handle them separately.
3. Forget to carry over uffd-wp bit for a write migration huge pmd
entry. This also happens in copy_huge_pmd(), where we converted a
write huge migration entry into a read one.
4. In copy_nonpresent_pte(), drop uffd-wp if necessary for swap ptes.
5. In copy_present_page() when COW is enforced when fork(), we also
need to pass over the uffd-wp bit if VM_UFFD_WP is armed on the new
vma, and when the pte to be copied has uffd-wp bit set.
Remove the comment in copy_present_pte() about this. It won't help a huge
lot to only comment there, but comment everywhere would be an overkill.
Let's assume the commit messages would help.
[peterx@redhat.com: fix a few thp pmd missing uffd-wp bit]
Link: https://lkml.kernel.org/r/20210428225030.9708-4-peterx@redhat.com
Link: https://lkml.kernel.org/r/20210428225030.9708-3-peterx@redhat.com
Fixes: b569a1760782f ("userfaultfd: wp: drop _PAGE_UFFD_WP properly when fork")
Signed-off-by: Peter Xu <peterx@redhat.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Brian Geffon <bgeffon@google.com>
Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Joe Perches <joe@perches.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Lokesh Gidra <lokeshgidra@google.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Oliver Upton <oupton@google.com>
Cc: Shaohua Li <shli@fb.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Wang Qing <wangqing@vivo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-01 04:49:02 +03:00
if ( ! userfaultfd_wp ( dst_vma ) )
pmd = pmd_clear_uffd_wp ( pmd ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
pmd = pmd_mkold ( pmd_wrprotect ( pmd ) ) ;
set_pmd_at ( dst_mm , addr , dst_pmd , pmd ) ;
ret = 0 ;
out_unlock :
2013-11-15 02:31:04 +04:00
spin_unlock ( src_ptl ) ;
spin_unlock ( dst_ptl ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
out :
return ret ;
}
2017-02-25 01:57:02 +03:00
# ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
static void touch_pud ( struct vm_area_struct * vma , unsigned long addr ,
2022-07-04 16:21:49 +03:00
pud_t * pud , bool write )
2017-02-25 01:57:02 +03:00
{
pud_t _pud ;
2017-11-27 06:21:25 +03:00
_pud = pud_mkyoung ( * pud ) ;
2022-07-04 16:21:49 +03:00
if ( write )
2017-11-27 06:21:25 +03:00
_pud = pud_mkdirty ( _pud ) ;
2017-02-25 01:57:02 +03:00
if ( pudp_set_access_flags ( vma , addr & HPAGE_PUD_MASK ,
2022-07-04 16:21:49 +03:00
pud , _pud , write ) )
2017-02-25 01:57:02 +03:00
update_mmu_cache_pud ( vma , addr , pud ) ;
}
struct page * follow_devmap_pud ( struct vm_area_struct * vma , unsigned long addr ,
2018-10-27 01:10:28 +03:00
pud_t * pud , int flags , struct dev_pagemap * * pgmap )
2017-02-25 01:57:02 +03:00
{
unsigned long pfn = pud_pfn ( * pud ) ;
struct mm_struct * mm = vma - > vm_mm ;
struct page * page ;
2022-10-21 20:41:08 +03:00
int ret ;
2017-02-25 01:57:02 +03:00
assert_spin_locked ( pud_lockptr ( mm , pud ) ) ;
Revert "mm: replace p??_write with pte_access_permitted in fault + gup paths"
This reverts commits 5c9d2d5c269c, c7da82b894e9, and e7fe7b5cae90.
We'll probably need to revisit this, but basically we should not
complicate the get_user_pages_fast() case, and checking the actual page
table protection key bits will require more care anyway, since the
protection keys depend on the exact state of the VM in question.
Particularly when doing a "remote" page lookup (ie in somebody elses VM,
not your own), you need to be much more careful than this was. Dave
Hansen says:
"So, the underlying bug here is that we now a get_user_pages_remote()
and then go ahead and do the p*_access_permitted() checks against the
current PKRU. This was introduced recently with the addition of the
new p??_access_permitted() calls.
We have checks in the VMA path for the "remote" gups and we avoid
consulting PKRU for them. This got missed in the pkeys selftests
because I did a ptrace read, but not a *write*. I also didn't
explicitly test it against something where a COW needed to be done"
It's also not entirely clear that it makes sense to check the protection
key bits at this level at all. But one possible eventual solution is to
make the get_user_pages_fast() case just abort if it sees protection key
bits set, which makes us fall back to the regular get_user_pages() case,
which then has a vma and can do the check there if we want to.
We'll see.
Somewhat related to this all: what we _do_ want to do some day is to
check the PAGE_USER bit - it should obviously always be set for user
pages, but it would be a good check to have back. Because we have no
generic way to test for it, we lost it as part of moving over from the
architecture-specific x86 GUP implementation to the generic one in
commit e585513b76f7 ("x86/mm/gup: Switch GUP to the generic
get_user_page_fast() implementation").
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "Jérôme Glisse" <jglisse@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-12-16 05:53:22 +03:00
if ( flags & FOLL_WRITE & & ! pud_write ( * pud ) )
2017-02-25 01:57:02 +03:00
return NULL ;
mm/gup: track FOLL_PIN pages
Add tracking of pages that were pinned via FOLL_PIN. This tracking is
implemented via overloading of page->_refcount: pins are added by adding
GUP_PIN_COUNTING_BIAS (1024) to the refcount. This provides a fuzzy
indication of pinning, and it can have false positives (and that's OK).
Please see the pre-existing Documentation/core-api/pin_user_pages.rst for
details.
As mentioned in pin_user_pages.rst, callers who effectively set FOLL_PIN
(typically via pin_user_pages*()) are required to ultimately free such
pages via unpin_user_page().
Please also note the limitation, discussed in pin_user_pages.rst under the
"TODO: for 1GB and larger huge pages" section. (That limitation will be
removed in a following patch.)
The effect of a FOLL_PIN flag is similar to that of FOLL_GET, and may be
thought of as "FOLL_GET for DIO and/or RDMA use".
Pages that have been pinned via FOLL_PIN are identifiable via a new
function call:
bool page_maybe_dma_pinned(struct page *page);
What to do in response to encountering such a page, is left to later
patchsets. There is discussion about this in [1], [2], [3], and [4].
This also changes a BUG_ON(), to a WARN_ON(), in follow_page_mask().
[1] Some slow progress on get_user_pages() (Apr 2, 2019):
https://lwn.net/Articles/784574/
[2] DMA and get_user_pages() (LPC: Dec 12, 2018):
https://lwn.net/Articles/774411/
[3] The trouble with get_user_pages() (Apr 30, 2018):
https://lwn.net/Articles/753027/
[4] LWN kernel index: get_user_pages():
https://lwn.net/Kernel/Index/#Memory_management-get_user_pages
[jhubbard@nvidia.com: add kerneldoc]
Link: http://lkml.kernel.org/r/20200307021157.235726-1-jhubbard@nvidia.com
[imbrenda@linux.ibm.com: if pin fails, we need to unpin, a simple put_page will not be enough]
Link: http://lkml.kernel.org/r/20200306132537.783769-2-imbrenda@linux.ibm.com
[akpm@linux-foundation.org: fix put_compound_head defined but not used]
Suggested-by: Jan Kara <jack@suse.cz>
Suggested-by: Jérôme Glisse <jglisse@redhat.com>
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Ira Weiny <ira.weiny@intel.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Link: http://lkml.kernel.org/r/20200211001536.1027652-7-jhubbard@nvidia.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:05:29 +03:00
/* FOLL_GET and FOLL_PIN are mutually exclusive. */
if ( WARN_ON_ONCE ( ( flags & ( FOLL_PIN | FOLL_GET ) ) = =
( FOLL_PIN | FOLL_GET ) ) )
return NULL ;
2017-02-25 01:57:02 +03:00
if ( pud_present ( * pud ) & & pud_devmap ( * pud ) )
/* pass */ ;
else
return NULL ;
if ( flags & FOLL_TOUCH )
2022-07-04 16:21:49 +03:00
touch_pud ( vma , addr , pud , flags & FOLL_WRITE ) ;
2017-02-25 01:57:02 +03:00
/*
* device mapped pages can only be returned if the
* caller will manage the page reference count .
mm/gup: track FOLL_PIN pages
Add tracking of pages that were pinned via FOLL_PIN. This tracking is
implemented via overloading of page->_refcount: pins are added by adding
GUP_PIN_COUNTING_BIAS (1024) to the refcount. This provides a fuzzy
indication of pinning, and it can have false positives (and that's OK).
Please see the pre-existing Documentation/core-api/pin_user_pages.rst for
details.
As mentioned in pin_user_pages.rst, callers who effectively set FOLL_PIN
(typically via pin_user_pages*()) are required to ultimately free such
pages via unpin_user_page().
Please also note the limitation, discussed in pin_user_pages.rst under the
"TODO: for 1GB and larger huge pages" section. (That limitation will be
removed in a following patch.)
The effect of a FOLL_PIN flag is similar to that of FOLL_GET, and may be
thought of as "FOLL_GET for DIO and/or RDMA use".
Pages that have been pinned via FOLL_PIN are identifiable via a new
function call:
bool page_maybe_dma_pinned(struct page *page);
What to do in response to encountering such a page, is left to later
patchsets. There is discussion about this in [1], [2], [3], and [4].
This also changes a BUG_ON(), to a WARN_ON(), in follow_page_mask().
[1] Some slow progress on get_user_pages() (Apr 2, 2019):
https://lwn.net/Articles/784574/
[2] DMA and get_user_pages() (LPC: Dec 12, 2018):
https://lwn.net/Articles/774411/
[3] The trouble with get_user_pages() (Apr 30, 2018):
https://lwn.net/Articles/753027/
[4] LWN kernel index: get_user_pages():
https://lwn.net/Kernel/Index/#Memory_management-get_user_pages
[jhubbard@nvidia.com: add kerneldoc]
Link: http://lkml.kernel.org/r/20200307021157.235726-1-jhubbard@nvidia.com
[imbrenda@linux.ibm.com: if pin fails, we need to unpin, a simple put_page will not be enough]
Link: http://lkml.kernel.org/r/20200306132537.783769-2-imbrenda@linux.ibm.com
[akpm@linux-foundation.org: fix put_compound_head defined but not used]
Suggested-by: Jan Kara <jack@suse.cz>
Suggested-by: Jérôme Glisse <jglisse@redhat.com>
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Ira Weiny <ira.weiny@intel.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Link: http://lkml.kernel.org/r/20200211001536.1027652-7-jhubbard@nvidia.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:05:29 +03:00
*
* At least one of FOLL_GET | FOLL_PIN must be set , so assert that here :
2017-02-25 01:57:02 +03:00
*/
mm/gup: track FOLL_PIN pages
Add tracking of pages that were pinned via FOLL_PIN. This tracking is
implemented via overloading of page->_refcount: pins are added by adding
GUP_PIN_COUNTING_BIAS (1024) to the refcount. This provides a fuzzy
indication of pinning, and it can have false positives (and that's OK).
Please see the pre-existing Documentation/core-api/pin_user_pages.rst for
details.
As mentioned in pin_user_pages.rst, callers who effectively set FOLL_PIN
(typically via pin_user_pages*()) are required to ultimately free such
pages via unpin_user_page().
Please also note the limitation, discussed in pin_user_pages.rst under the
"TODO: for 1GB and larger huge pages" section. (That limitation will be
removed in a following patch.)
The effect of a FOLL_PIN flag is similar to that of FOLL_GET, and may be
thought of as "FOLL_GET for DIO and/or RDMA use".
Pages that have been pinned via FOLL_PIN are identifiable via a new
function call:
bool page_maybe_dma_pinned(struct page *page);
What to do in response to encountering such a page, is left to later
patchsets. There is discussion about this in [1], [2], [3], and [4].
This also changes a BUG_ON(), to a WARN_ON(), in follow_page_mask().
[1] Some slow progress on get_user_pages() (Apr 2, 2019):
https://lwn.net/Articles/784574/
[2] DMA and get_user_pages() (LPC: Dec 12, 2018):
https://lwn.net/Articles/774411/
[3] The trouble with get_user_pages() (Apr 30, 2018):
https://lwn.net/Articles/753027/
[4] LWN kernel index: get_user_pages():
https://lwn.net/Kernel/Index/#Memory_management-get_user_pages
[jhubbard@nvidia.com: add kerneldoc]
Link: http://lkml.kernel.org/r/20200307021157.235726-1-jhubbard@nvidia.com
[imbrenda@linux.ibm.com: if pin fails, we need to unpin, a simple put_page will not be enough]
Link: http://lkml.kernel.org/r/20200306132537.783769-2-imbrenda@linux.ibm.com
[akpm@linux-foundation.org: fix put_compound_head defined but not used]
Suggested-by: Jan Kara <jack@suse.cz>
Suggested-by: Jérôme Glisse <jglisse@redhat.com>
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Ira Weiny <ira.weiny@intel.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Link: http://lkml.kernel.org/r/20200211001536.1027652-7-jhubbard@nvidia.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:05:29 +03:00
if ( ! ( flags & ( FOLL_GET | FOLL_PIN ) ) )
2017-02-25 01:57:02 +03:00
return ERR_PTR ( - EEXIST ) ;
pfn + = ( addr & ~ PUD_MASK ) > > PAGE_SHIFT ;
2018-10-27 01:10:28 +03:00
* pgmap = get_dev_pagemap ( pfn , * pgmap ) ;
if ( ! * pgmap )
2017-02-25 01:57:02 +03:00
return ERR_PTR ( - EFAULT ) ;
page = pfn_to_page ( pfn ) ;
2022-10-21 20:41:08 +03:00
ret = try_grab_page ( page , flags ) ;
if ( ret )
page = ERR_PTR ( ret ) ;
2017-02-25 01:57:02 +03:00
return page ;
}
int copy_huge_pud ( struct mm_struct * dst_mm , struct mm_struct * src_mm ,
pud_t * dst_pud , pud_t * src_pud , unsigned long addr ,
struct vm_area_struct * vma )
{
spinlock_t * dst_ptl , * src_ptl ;
pud_t pud ;
int ret ;
dst_ptl = pud_lock ( dst_mm , dst_pud ) ;
src_ptl = pud_lockptr ( src_mm , src_pud ) ;
spin_lock_nested ( src_ptl , SINGLE_DEPTH_NESTING ) ;
ret = - EAGAIN ;
pud = * src_pud ;
if ( unlikely ( ! pud_trans_huge ( pud ) & & ! pud_devmap ( pud ) ) )
goto out_unlock ;
/*
* When page table lock is held , the huge zero pud should not be
* under splitting since we don ' t split the page itself , only pud to
* a page table .
*/
if ( is_huge_zero_pud ( pud ) ) {
/* No huge zero pud yet */
}
2022-05-10 04:20:43 +03:00
/*
* TODO : once we support anonymous pages , use page_try_dup_anon_rmap ( )
* and split if duplicating fails .
*/
2017-02-25 01:57:02 +03:00
pudp_set_wrprotect ( src_mm , addr , src_pud ) ;
pud = pud_mkold ( pud_wrprotect ( pud ) ) ;
set_pud_at ( dst_mm , addr , dst_pud , pud ) ;
ret = 0 ;
out_unlock :
spin_unlock ( src_ptl ) ;
spin_unlock ( dst_ptl ) ;
return ret ;
}
void huge_pud_set_accessed ( struct vm_fault * vmf , pud_t orig_pud )
{
bool write = vmf - > flags & FAULT_FLAG_WRITE ;
vmf - > ptl = pud_lock ( vmf - > vma - > vm_mm , vmf - > pud ) ;
if ( unlikely ( ! pud_same ( * vmf - > pud , orig_pud ) ) )
goto unlock ;
2022-07-04 16:21:49 +03:00
touch_pud ( vmf - > vma , vmf - > address , vmf - > pud , write ) ;
2017-02-25 01:57:02 +03:00
unlock :
spin_unlock ( vmf - > ptl ) ;
}
# endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
mm: memory: add orig_pmd to struct vm_fault
Pach series "mm: thp: use generic THP migration for NUMA hinting fault", v3.
When the THP NUMA fault support was added THP migration was not supported
yet. So the ad hoc THP migration was implemented in NUMA fault handling.
Since v4.14 THP migration has been supported so it doesn't make too much
sense to still keep another THP migration implementation rather than using
the generic migration code. It is definitely a maintenance burden to keep
two THP migration implementation for different code paths and it is more
error prone. Using the generic THP migration implementation allows us
remove the duplicate code and some hacks needed by the old ad hoc
implementation.
A quick grep shows x86_64, PowerPC (book3s), ARM64 ans S390 support both
THP and NUMA balancing. The most of them support THP migration except for
S390. Zi Yan tried to add THP migration support for S390 before but it
was not accepted due to the design of S390 PMD. For the discussion,
please see: https://lkml.org/lkml/2018/4/27/953.
Per the discussion with Gerald Schaefer in v1 it is acceptible to skip
huge PMD for S390 for now.
I saw there were some hacks about gup from git history, but I didn't
figure out if they have been removed or not since I just found FOLL_NUMA
code in the current gup implementation and they seems useful.
Patch #1 ~ #2 are preparation patches.
Patch #3 is the real meat.
Patch #4 ~ #6 keep consistent counters and behaviors with before.
Patch #7 skips change huge PMD to prot_none if thp migration is not supported.
Test
----
Did some tests to measure the latency of do_huge_pmd_numa_page. The test
VM has 80 vcpus and 64G memory. The test would create 2 processes to
consume 128G memory together which would incur memory pressure to cause
THP splits. And it also creates 80 processes to hog cpu, and the memory
consumer processes are bound to different nodes periodically in order to
increase NUMA faults.
The below test script is used:
echo 3 > /proc/sys/vm/drop_caches
# Run stress-ng for 24 hours
./stress-ng/stress-ng --vm 2 --vm-bytes 64G --timeout 24h &
PID=$!
./stress-ng/stress-ng --cpu $NR_CPUS --timeout 24h &
# Wait for vm stressors forked
sleep 5
PID_1=`pgrep -P $PID | awk 'NR == 1'`
PID_2=`pgrep -P $PID | awk 'NR == 2'`
JOB1=`pgrep -P $PID_1`
JOB2=`pgrep -P $PID_2`
# Bind load jobs to different nodes periodically to force generate
# cross node memory access
while [ -d "/proc/$PID" ]
do
taskset -apc 8 $JOB1
taskset -apc 8 $JOB2
sleep 300
taskset -apc 58 $JOB1
taskset -apc 58 $JOB2
sleep 300
done
With the above test the histogram of latency of do_huge_pmd_numa_page is
as shown below. Since the number of do_huge_pmd_numa_page varies
drastically for each run (should be due to scheduler), so I converted the
raw number to percentage.
patched base
@us[stress-ng]:
[0] 3.57% 0.16%
[1] 55.68% 18.36%
[2, 4) 10.46% 40.44%
[4, 8) 7.26% 17.82%
[8, 16) 21.12% 13.41%
[16, 32) 1.06% 4.27%
[32, 64) 0.56% 4.07%
[64, 128) 0.16% 0.35%
[128, 256) < 0.1% < 0.1%
[256, 512) < 0.1% < 0.1%
[512, 1K) < 0.1% < 0.1%
[1K, 2K) < 0.1% < 0.1%
[2K, 4K) < 0.1% < 0.1%
[4K, 8K) < 0.1% < 0.1%
[8K, 16K) < 0.1% < 0.1%
[16K, 32K) < 0.1% < 0.1%
[32K, 64K) < 0.1% < 0.1%
Per the result, patched kernel is even slightly better than the base
kernel. I think this is because the lock contention against THP split is
less than base kernel due to the refactor.
To exclude the affect from THP split, I also did test w/o memory pressure.
No obvious regression is spotted. The below is the test result *w/o*
memory pressure.
patched base
@us[stress-ng]:
[0] 7.97% 18.4%
[1] 69.63% 58.24%
[2, 4) 4.18% 2.63%
[4, 8) 0.22% 0.17%
[8, 16) 1.03% 0.92%
[16, 32) 0.14% < 0.1%
[32, 64) < 0.1% < 0.1%
[64, 128) < 0.1% < 0.1%
[128, 256) < 0.1% < 0.1%
[256, 512) 0.45% 1.19%
[512, 1K) 15.45% 17.27%
[1K, 2K) < 0.1% < 0.1%
[2K, 4K) < 0.1% < 0.1%
[4K, 8K) < 0.1% < 0.1%
[8K, 16K) 0.86% 0.88%
[16K, 32K) < 0.1% 0.15%
[32K, 64K) < 0.1% < 0.1%
[64K, 128K) < 0.1% < 0.1%
[128K, 256K) < 0.1% < 0.1%
The series also survived a series of tests that exercise NUMA balancing
migrations by Mel.
This patch (of 7):
Add orig_pmd to struct vm_fault so the "orig_pmd" parameter used by huge
page fault could be removed, just like its PTE counterpart does.
Link: https://lkml.kernel.org/r/20210518200801.7413-1-shy828301@gmail.com
Link: https://lkml.kernel.org/r/20210518200801.7413-2-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-01 04:51:35 +03:00
void huge_pmd_set_accessed ( struct vm_fault * vmf )
2012-12-12 04:01:27 +04:00
{
2017-01-11 03:57:51 +03:00
bool write = vmf - > flags & FAULT_FLAG_WRITE ;
2012-12-12 04:01:27 +04:00
2016-12-15 02:06:58 +03:00
vmf - > ptl = pmd_lock ( vmf - > vma - > vm_mm , vmf - > pmd ) ;
2022-07-04 16:21:50 +03:00
if ( unlikely ( ! pmd_same ( * vmf - > pmd , vmf - > orig_pmd ) ) )
2012-12-12 04:01:27 +04:00
goto unlock ;
2022-07-04 16:21:50 +03:00
touch_pmd ( vmf - > vma , vmf - > address , vmf - > pmd , write ) ;
2012-12-12 04:01:27 +04:00
unlock :
2016-12-15 02:06:58 +03:00
spin_unlock ( vmf - > ptl ) ;
2012-12-12 04:01:27 +04:00
}
mm: memory: add orig_pmd to struct vm_fault
Pach series "mm: thp: use generic THP migration for NUMA hinting fault", v3.
When the THP NUMA fault support was added THP migration was not supported
yet. So the ad hoc THP migration was implemented in NUMA fault handling.
Since v4.14 THP migration has been supported so it doesn't make too much
sense to still keep another THP migration implementation rather than using
the generic migration code. It is definitely a maintenance burden to keep
two THP migration implementation for different code paths and it is more
error prone. Using the generic THP migration implementation allows us
remove the duplicate code and some hacks needed by the old ad hoc
implementation.
A quick grep shows x86_64, PowerPC (book3s), ARM64 ans S390 support both
THP and NUMA balancing. The most of them support THP migration except for
S390. Zi Yan tried to add THP migration support for S390 before but it
was not accepted due to the design of S390 PMD. For the discussion,
please see: https://lkml.org/lkml/2018/4/27/953.
Per the discussion with Gerald Schaefer in v1 it is acceptible to skip
huge PMD for S390 for now.
I saw there were some hacks about gup from git history, but I didn't
figure out if they have been removed or not since I just found FOLL_NUMA
code in the current gup implementation and they seems useful.
Patch #1 ~ #2 are preparation patches.
Patch #3 is the real meat.
Patch #4 ~ #6 keep consistent counters and behaviors with before.
Patch #7 skips change huge PMD to prot_none if thp migration is not supported.
Test
----
Did some tests to measure the latency of do_huge_pmd_numa_page. The test
VM has 80 vcpus and 64G memory. The test would create 2 processes to
consume 128G memory together which would incur memory pressure to cause
THP splits. And it also creates 80 processes to hog cpu, and the memory
consumer processes are bound to different nodes periodically in order to
increase NUMA faults.
The below test script is used:
echo 3 > /proc/sys/vm/drop_caches
# Run stress-ng for 24 hours
./stress-ng/stress-ng --vm 2 --vm-bytes 64G --timeout 24h &
PID=$!
./stress-ng/stress-ng --cpu $NR_CPUS --timeout 24h &
# Wait for vm stressors forked
sleep 5
PID_1=`pgrep -P $PID | awk 'NR == 1'`
PID_2=`pgrep -P $PID | awk 'NR == 2'`
JOB1=`pgrep -P $PID_1`
JOB2=`pgrep -P $PID_2`
# Bind load jobs to different nodes periodically to force generate
# cross node memory access
while [ -d "/proc/$PID" ]
do
taskset -apc 8 $JOB1
taskset -apc 8 $JOB2
sleep 300
taskset -apc 58 $JOB1
taskset -apc 58 $JOB2
sleep 300
done
With the above test the histogram of latency of do_huge_pmd_numa_page is
as shown below. Since the number of do_huge_pmd_numa_page varies
drastically for each run (should be due to scheduler), so I converted the
raw number to percentage.
patched base
@us[stress-ng]:
[0] 3.57% 0.16%
[1] 55.68% 18.36%
[2, 4) 10.46% 40.44%
[4, 8) 7.26% 17.82%
[8, 16) 21.12% 13.41%
[16, 32) 1.06% 4.27%
[32, 64) 0.56% 4.07%
[64, 128) 0.16% 0.35%
[128, 256) < 0.1% < 0.1%
[256, 512) < 0.1% < 0.1%
[512, 1K) < 0.1% < 0.1%
[1K, 2K) < 0.1% < 0.1%
[2K, 4K) < 0.1% < 0.1%
[4K, 8K) < 0.1% < 0.1%
[8K, 16K) < 0.1% < 0.1%
[16K, 32K) < 0.1% < 0.1%
[32K, 64K) < 0.1% < 0.1%
Per the result, patched kernel is even slightly better than the base
kernel. I think this is because the lock contention against THP split is
less than base kernel due to the refactor.
To exclude the affect from THP split, I also did test w/o memory pressure.
No obvious regression is spotted. The below is the test result *w/o*
memory pressure.
patched base
@us[stress-ng]:
[0] 7.97% 18.4%
[1] 69.63% 58.24%
[2, 4) 4.18% 2.63%
[4, 8) 0.22% 0.17%
[8, 16) 1.03% 0.92%
[16, 32) 0.14% < 0.1%
[32, 64) < 0.1% < 0.1%
[64, 128) < 0.1% < 0.1%
[128, 256) < 0.1% < 0.1%
[256, 512) 0.45% 1.19%
[512, 1K) 15.45% 17.27%
[1K, 2K) < 0.1% < 0.1%
[2K, 4K) < 0.1% < 0.1%
[4K, 8K) < 0.1% < 0.1%
[8K, 16K) 0.86% 0.88%
[16K, 32K) < 0.1% 0.15%
[32K, 64K) < 0.1% < 0.1%
[64K, 128K) < 0.1% < 0.1%
[128K, 256K) < 0.1% < 0.1%
The series also survived a series of tests that exercise NUMA balancing
migrations by Mel.
This patch (of 7):
Add orig_pmd to struct vm_fault so the "orig_pmd" parameter used by huge
page fault could be removed, just like its PTE counterpart does.
Link: https://lkml.kernel.org/r/20210518200801.7413-1-shy828301@gmail.com
Link: https://lkml.kernel.org/r/20210518200801.7413-2-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-01 04:51:35 +03:00
vm_fault_t do_huge_pmd_wp_page ( struct vm_fault * vmf )
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
{
mm: support GUP-triggered unsharing of anonymous pages
Whenever GUP currently ends up taking a R/O pin on an anonymous page that
might be shared -- mapped R/O and !PageAnonExclusive() -- any write fault
on the page table entry will end up replacing the mapped anonymous page
due to COW, resulting in the GUP pin no longer being consistent with the
page actually mapped into the page table.
The possible ways to deal with this situation are:
(1) Ignore and pin -- what we do right now.
(2) Fail to pin -- which would be rather surprising to callers and
could break user space.
(3) Trigger unsharing and pin the now exclusive page -- reliable R/O
pins.
We want to implement 3) because it provides the clearest semantics and
allows for checking in unpin_user_pages() and friends for possible BUGs:
when trying to unpin a page that's no longer exclusive, clearly something
went very wrong and might result in memory corruptions that might be hard
to debug. So we better have a nice way to spot such issues.
To implement 3), we need a way for GUP to trigger unsharing:
FAULT_FLAG_UNSHARE. FAULT_FLAG_UNSHARE is only applicable to R/O mapped
anonymous pages and resembles COW logic during a write fault. However, in
contrast to a write fault, GUP-triggered unsharing will, for example,
still maintain the write protection.
Let's implement FAULT_FLAG_UNSHARE by hooking into the existing write
fault handlers for all applicable anonymous page types: ordinary pages,
THP and hugetlb.
* If FAULT_FLAG_UNSHARE finds a R/O-mapped anonymous page that has been
marked exclusive in the meantime by someone else, there is nothing to do.
* If FAULT_FLAG_UNSHARE finds a R/O-mapped anonymous page that's not
marked exclusive, it will try detecting if the process is the exclusive
owner. If exclusive, it can be set exclusive similar to reuse logic
during write faults via page_move_anon_rmap() and there is nothing
else to do; otherwise, we either have to copy and map a fresh,
anonymous exclusive page R/O (ordinary pages, hugetlb), or split the
THP.
This commit is heavily based on patches by Andrea.
Link: https://lkml.kernel.org/r/20220428083441.37290-16-david@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Co-developed-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:45 +03:00
const bool unshare = vmf - > flags & FAULT_FLAG_UNSHARE ;
2016-12-15 02:06:58 +03:00
struct vm_area_struct * vma = vmf - > vma ;
2022-09-02 22:46:38 +03:00
struct folio * folio ;
2020-06-04 02:00:27 +03:00
struct page * page ;
2016-12-15 02:06:58 +03:00
unsigned long haddr = vmf - > address & HPAGE_PMD_MASK ;
mm: memory: add orig_pmd to struct vm_fault
Pach series "mm: thp: use generic THP migration for NUMA hinting fault", v3.
When the THP NUMA fault support was added THP migration was not supported
yet. So the ad hoc THP migration was implemented in NUMA fault handling.
Since v4.14 THP migration has been supported so it doesn't make too much
sense to still keep another THP migration implementation rather than using
the generic migration code. It is definitely a maintenance burden to keep
two THP migration implementation for different code paths and it is more
error prone. Using the generic THP migration implementation allows us
remove the duplicate code and some hacks needed by the old ad hoc
implementation.
A quick grep shows x86_64, PowerPC (book3s), ARM64 ans S390 support both
THP and NUMA balancing. The most of them support THP migration except for
S390. Zi Yan tried to add THP migration support for S390 before but it
was not accepted due to the design of S390 PMD. For the discussion,
please see: https://lkml.org/lkml/2018/4/27/953.
Per the discussion with Gerald Schaefer in v1 it is acceptible to skip
huge PMD for S390 for now.
I saw there were some hacks about gup from git history, but I didn't
figure out if they have been removed or not since I just found FOLL_NUMA
code in the current gup implementation and they seems useful.
Patch #1 ~ #2 are preparation patches.
Patch #3 is the real meat.
Patch #4 ~ #6 keep consistent counters and behaviors with before.
Patch #7 skips change huge PMD to prot_none if thp migration is not supported.
Test
----
Did some tests to measure the latency of do_huge_pmd_numa_page. The test
VM has 80 vcpus and 64G memory. The test would create 2 processes to
consume 128G memory together which would incur memory pressure to cause
THP splits. And it also creates 80 processes to hog cpu, and the memory
consumer processes are bound to different nodes periodically in order to
increase NUMA faults.
The below test script is used:
echo 3 > /proc/sys/vm/drop_caches
# Run stress-ng for 24 hours
./stress-ng/stress-ng --vm 2 --vm-bytes 64G --timeout 24h &
PID=$!
./stress-ng/stress-ng --cpu $NR_CPUS --timeout 24h &
# Wait for vm stressors forked
sleep 5
PID_1=`pgrep -P $PID | awk 'NR == 1'`
PID_2=`pgrep -P $PID | awk 'NR == 2'`
JOB1=`pgrep -P $PID_1`
JOB2=`pgrep -P $PID_2`
# Bind load jobs to different nodes periodically to force generate
# cross node memory access
while [ -d "/proc/$PID" ]
do
taskset -apc 8 $JOB1
taskset -apc 8 $JOB2
sleep 300
taskset -apc 58 $JOB1
taskset -apc 58 $JOB2
sleep 300
done
With the above test the histogram of latency of do_huge_pmd_numa_page is
as shown below. Since the number of do_huge_pmd_numa_page varies
drastically for each run (should be due to scheduler), so I converted the
raw number to percentage.
patched base
@us[stress-ng]:
[0] 3.57% 0.16%
[1] 55.68% 18.36%
[2, 4) 10.46% 40.44%
[4, 8) 7.26% 17.82%
[8, 16) 21.12% 13.41%
[16, 32) 1.06% 4.27%
[32, 64) 0.56% 4.07%
[64, 128) 0.16% 0.35%
[128, 256) < 0.1% < 0.1%
[256, 512) < 0.1% < 0.1%
[512, 1K) < 0.1% < 0.1%
[1K, 2K) < 0.1% < 0.1%
[2K, 4K) < 0.1% < 0.1%
[4K, 8K) < 0.1% < 0.1%
[8K, 16K) < 0.1% < 0.1%
[16K, 32K) < 0.1% < 0.1%
[32K, 64K) < 0.1% < 0.1%
Per the result, patched kernel is even slightly better than the base
kernel. I think this is because the lock contention against THP split is
less than base kernel due to the refactor.
To exclude the affect from THP split, I also did test w/o memory pressure.
No obvious regression is spotted. The below is the test result *w/o*
memory pressure.
patched base
@us[stress-ng]:
[0] 7.97% 18.4%
[1] 69.63% 58.24%
[2, 4) 4.18% 2.63%
[4, 8) 0.22% 0.17%
[8, 16) 1.03% 0.92%
[16, 32) 0.14% < 0.1%
[32, 64) < 0.1% < 0.1%
[64, 128) < 0.1% < 0.1%
[128, 256) < 0.1% < 0.1%
[256, 512) 0.45% 1.19%
[512, 1K) 15.45% 17.27%
[1K, 2K) < 0.1% < 0.1%
[2K, 4K) < 0.1% < 0.1%
[4K, 8K) < 0.1% < 0.1%
[8K, 16K) 0.86% 0.88%
[16K, 32K) < 0.1% 0.15%
[32K, 64K) < 0.1% < 0.1%
[64K, 128K) < 0.1% < 0.1%
[128K, 256K) < 0.1% < 0.1%
The series also survived a series of tests that exercise NUMA balancing
migrations by Mel.
This patch (of 7):
Add orig_pmd to struct vm_fault so the "orig_pmd" parameter used by huge
page fault could be removed, just like its PTE counterpart does.
Link: https://lkml.kernel.org/r/20210518200801.7413-1-shy828301@gmail.com
Link: https://lkml.kernel.org/r/20210518200801.7413-2-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-01 04:51:35 +03:00
pmd_t orig_pmd = vmf - > orig_pmd ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
2016-12-15 02:06:58 +03:00
vmf - > ptl = pmd_lockptr ( vma - > vm_mm , vmf - > pmd ) ;
2014-10-10 02:28:10 +04:00
VM_BUG_ON_VMA ( ! vma - > anon_vma , vma ) ;
2020-06-04 02:00:27 +03:00
2012-12-13 01:50:54 +04:00
if ( is_huge_zero_pmd ( orig_pmd ) )
2020-06-04 02:00:27 +03:00
goto fallback ;
2016-12-15 02:06:58 +03:00
spin_lock ( vmf - > ptl ) ;
2020-06-04 02:00:27 +03:00
if ( unlikely ( ! pmd_same ( * vmf - > pmd , orig_pmd ) ) ) {
spin_unlock ( vmf - > ptl ) ;
return 0 ;
}
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
page = pmd_page ( orig_pmd ) ;
2022-09-02 22:46:38 +03:00
folio = page_folio ( page ) ;
2021-05-05 04:34:02 +03:00
VM_BUG_ON_PAGE ( ! PageHead ( page ) , page ) ;
2020-06-04 02:00:27 +03:00
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
/* Early check when only holding the PT lock. */
if ( PageAnonExclusive ( page ) )
goto reuse ;
2022-09-02 22:46:38 +03:00
if ( ! folio_trylock ( folio ) ) {
folio_get ( folio ) ;
2017-09-07 02:22:19 +03:00
spin_unlock ( vmf - > ptl ) ;
2022-09-02 22:46:38 +03:00
folio_lock ( folio ) ;
2017-09-07 02:22:19 +03:00
spin_lock ( vmf - > ptl ) ;
if ( unlikely ( ! pmd_same ( * vmf - > pmd , orig_pmd ) ) ) {
2020-06-04 02:00:27 +03:00
spin_unlock ( vmf - > ptl ) ;
2022-09-02 22:46:38 +03:00
folio_unlock ( folio ) ;
folio_put ( folio ) ;
2020-06-04 02:00:27 +03:00
return 0 ;
2017-09-07 02:22:19 +03:00
}
2022-09-02 22:46:38 +03:00
folio_put ( folio ) ;
2017-09-07 02:22:19 +03:00
}
2020-06-04 02:00:27 +03:00
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
/* Recheck after temporarily dropping the PT lock. */
if ( PageAnonExclusive ( page ) ) {
2022-09-02 22:46:38 +03:00
folio_unlock ( folio ) ;
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
goto reuse ;
}
2020-06-04 02:00:27 +03:00
/*
2022-09-02 22:46:38 +03:00
* See do_wp_page ( ) : we can only reuse the folio exclusively if
* there are no additional references . Note that we always drain
* the LRU pagevecs immediately after adding a THP .
2020-06-04 02:00:27 +03:00
*/
2022-09-02 22:46:38 +03:00
if ( folio_ref_count ( folio ) >
1 + folio_test_swapcache ( folio ) * folio_nr_pages ( folio ) )
mm/huge_memory: streamline COW logic in do_huge_pmd_wp_page()
We currently have a different COW logic for anon THP than we have for
ordinary anon pages in do_wp_page(): the effect is that the issue reported
in CVE-2020-29374 is currently still possible for anon THP: an unintended
information leak from the parent to the child.
Let's apply the same logic (page_count() == 1), with similar optimizations
to remove additional references first as we really want to avoid
PTE-mapping the THP and copying individual pages best we can.
If we end up with a page that has page_count() != 1, we'll have to PTE-map
the THP and fallback to do_wp_page(), which will always copy the page.
Note that KSM does not apply to THP.
I. Interaction with the swapcache and writeback
While a THP is in the swapcache, the swapcache holds one reference on each
subpage of the THP. So with PageSwapCache() set, we expect as many
additional references as we have subpages. If we manage to remove the THP
from the swapcache, all these references will be gone.
Usually, a THP is not split when entered into the swapcache and stays a
compound page. However, try_to_unmap() will PTE-map the THP and use PTE
swap entries. There are no PMD swap entries for that purpose,
consequently, we always only swapin subpages into PTEs.
Removing a page from the swapcache can fail either when there are
remaining swap entries (in which case COW is the right thing to do) or if
the page is currently under writeback.
Having a locked, R/O PMD-mapped THP that is in the swapcache seems to be
possible only in corner cases, for example, if try_to_unmap() failed after
adding the page to the swapcache. However, it's comparatively easy to
handle.
As we have to fully unmap a THP before starting writeback, and swapin is
always done on the PTE level, we shouldn't find a R/O PMD-mapped THP in
the swapcache that is under writeback. This should at least leave
writeback out of the picture.
II. Interaction with GUP references
Having a R/O PMD-mapped THP with GUP references (i.e., R/O references)
will result in PTE-mapping the THP on a write fault. Similar to ordinary
anon pages, do_wp_page() will have to copy sub-pages and result in a
disconnect between the GUP references and the pages actually mapped into
the page tables. To improve the situation in the future, we'll need
additional handling to mark anonymous pages as definitely exclusive to a
single process, only allow GUP pins on exclusive anon pages, and disallow
sharing of exclusive anon pages with GUP pins e.g., during fork().
III. Interaction with references from LRU pagevecs
There is no need to try draining the (local) LRU pagevecs in case we would
stumble over a !PageLRU() page: folio_add_lru() and friends will always
flush the affected pagevec after adding a compound page to it immediately
-- pagevec_add_and_need_flush() always returns "true" for them. Note that
the LRU pagevecs will hold a reference on the compound page for a very
short time, between adding the page to the pagevec and draining it
immediately afterwards.
IV. Interaction with speculative/temporary references
Similar to ordinary anon pages, other speculative/temporary references on
the THP, for example, from the pagecache or page migration code, will
disallow exclusive reuse of the page. We'll have to PTE-map the THP.
Link: https://lkml.kernel.org/r/20220131162940.210846-6-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <nadav.amit@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-03-25 04:13:43 +03:00
goto unlock_fallback ;
2022-09-02 22:46:38 +03:00
if ( folio_test_swapcache ( folio ) )
folio_free_swap ( folio ) ;
if ( folio_ref_count ( folio ) = = 1 ) {
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
pmd_t entry ;
2022-05-10 04:20:43 +03:00
page_move_anon_rmap ( page , vma ) ;
2022-09-02 22:46:38 +03:00
folio_unlock ( folio ) ;
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
reuse :
mm: support GUP-triggered unsharing of anonymous pages
Whenever GUP currently ends up taking a R/O pin on an anonymous page that
might be shared -- mapped R/O and !PageAnonExclusive() -- any write fault
on the page table entry will end up replacing the mapped anonymous page
due to COW, resulting in the GUP pin no longer being consistent with the
page actually mapped into the page table.
The possible ways to deal with this situation are:
(1) Ignore and pin -- what we do right now.
(2) Fail to pin -- which would be rather surprising to callers and
could break user space.
(3) Trigger unsharing and pin the now exclusive page -- reliable R/O
pins.
We want to implement 3) because it provides the clearest semantics and
allows for checking in unpin_user_pages() and friends for possible BUGs:
when trying to unpin a page that's no longer exclusive, clearly something
went very wrong and might result in memory corruptions that might be hard
to debug. So we better have a nice way to spot such issues.
To implement 3), we need a way for GUP to trigger unsharing:
FAULT_FLAG_UNSHARE. FAULT_FLAG_UNSHARE is only applicable to R/O mapped
anonymous pages and resembles COW logic during a write fault. However, in
contrast to a write fault, GUP-triggered unsharing will, for example,
still maintain the write protection.
Let's implement FAULT_FLAG_UNSHARE by hooking into the existing write
fault handlers for all applicable anonymous page types: ordinary pages,
THP and hugetlb.
* If FAULT_FLAG_UNSHARE finds a R/O-mapped anonymous page that has been
marked exclusive in the meantime by someone else, there is nothing to do.
* If FAULT_FLAG_UNSHARE finds a R/O-mapped anonymous page that's not
marked exclusive, it will try detecting if the process is the exclusive
owner. If exclusive, it can be set exclusive similar to reuse logic
during write faults via page_move_anon_rmap() and there is nothing
else to do; otherwise, we either have to copy and map a fresh,
anonymous exclusive page R/O (ordinary pages, hugetlb), or split the
THP.
This commit is heavily based on patches by Andrea.
Link: https://lkml.kernel.org/r/20220428083441.37290-16-david@redhat.com
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Co-developed-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:45 +03:00
if ( unlikely ( unshare ) ) {
spin_unlock ( vmf - > ptl ) ;
return 0 ;
}
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
entry = pmd_mkyoung ( orig_pmd ) ;
2017-11-29 20:01:01 +03:00
entry = maybe_pmd_mkwrite ( pmd_mkdirty ( entry ) , vma ) ;
2020-06-04 02:00:27 +03:00
if ( pmdp_set_access_flags ( vma , haddr , vmf - > pmd , entry , 1 ) )
2016-12-15 02:06:58 +03:00
update_mmu_cache_pmd ( vma , vmf - > address , vmf - > pmd ) ;
spin_unlock ( vmf - > ptl ) ;
2022-10-21 13:11:35 +03:00
return 0 ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
}
2020-06-04 02:00:27 +03:00
mm/huge_memory: streamline COW logic in do_huge_pmd_wp_page()
We currently have a different COW logic for anon THP than we have for
ordinary anon pages in do_wp_page(): the effect is that the issue reported
in CVE-2020-29374 is currently still possible for anon THP: an unintended
information leak from the parent to the child.
Let's apply the same logic (page_count() == 1), with similar optimizations
to remove additional references first as we really want to avoid
PTE-mapping the THP and copying individual pages best we can.
If we end up with a page that has page_count() != 1, we'll have to PTE-map
the THP and fallback to do_wp_page(), which will always copy the page.
Note that KSM does not apply to THP.
I. Interaction with the swapcache and writeback
While a THP is in the swapcache, the swapcache holds one reference on each
subpage of the THP. So with PageSwapCache() set, we expect as many
additional references as we have subpages. If we manage to remove the THP
from the swapcache, all these references will be gone.
Usually, a THP is not split when entered into the swapcache and stays a
compound page. However, try_to_unmap() will PTE-map the THP and use PTE
swap entries. There are no PMD swap entries for that purpose,
consequently, we always only swapin subpages into PTEs.
Removing a page from the swapcache can fail either when there are
remaining swap entries (in which case COW is the right thing to do) or if
the page is currently under writeback.
Having a locked, R/O PMD-mapped THP that is in the swapcache seems to be
possible only in corner cases, for example, if try_to_unmap() failed after
adding the page to the swapcache. However, it's comparatively easy to
handle.
As we have to fully unmap a THP before starting writeback, and swapin is
always done on the PTE level, we shouldn't find a R/O PMD-mapped THP in
the swapcache that is under writeback. This should at least leave
writeback out of the picture.
II. Interaction with GUP references
Having a R/O PMD-mapped THP with GUP references (i.e., R/O references)
will result in PTE-mapping the THP on a write fault. Similar to ordinary
anon pages, do_wp_page() will have to copy sub-pages and result in a
disconnect between the GUP references and the pages actually mapped into
the page tables. To improve the situation in the future, we'll need
additional handling to mark anonymous pages as definitely exclusive to a
single process, only allow GUP pins on exclusive anon pages, and disallow
sharing of exclusive anon pages with GUP pins e.g., during fork().
III. Interaction with references from LRU pagevecs
There is no need to try draining the (local) LRU pagevecs in case we would
stumble over a !PageLRU() page: folio_add_lru() and friends will always
flush the affected pagevec after adding a compound page to it immediately
-- pagevec_add_and_need_flush() always returns "true" for them. Note that
the LRU pagevecs will hold a reference on the compound page for a very
short time, between adding the page to the pagevec and draining it
immediately afterwards.
IV. Interaction with speculative/temporary references
Similar to ordinary anon pages, other speculative/temporary references on
the THP, for example, from the pagecache or page migration code, will
disallow exclusive reuse of the page. We'll have to PTE-map the THP.
Link: https://lkml.kernel.org/r/20220131162940.210846-6-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <nadav.amit@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-03-25 04:13:43 +03:00
unlock_fallback :
2022-09-02 22:46:38 +03:00
folio_unlock ( folio ) ;
2016-12-15 02:06:58 +03:00
spin_unlock ( vmf - > ptl ) ;
2020-06-04 02:00:27 +03:00
fallback :
__split_huge_pmd ( vma , vmf - > pmd , vmf - > address , false , NULL ) ;
return VM_FAULT_FALLBACK ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
}
2022-11-08 20:46:48 +03:00
static inline bool can_change_pmd_writable ( struct vm_area_struct * vma ,
unsigned long addr , pmd_t pmd )
{
struct page * page ;
if ( WARN_ON_ONCE ( ! ( vma - > vm_flags & VM_WRITE ) ) )
return false ;
/* Don't touch entries that are not even readable (NUMA hinting). */
if ( pmd_protnone ( pmd ) )
return false ;
/* Do we need write faults for softdirty tracking? */
if ( vma_soft_dirty_enabled ( vma ) & & ! pmd_soft_dirty ( pmd ) )
return false ;
/* Do we need write faults for uffd-wp tracking? */
if ( userfaultfd_huge_pmd_wp ( vma , pmd ) )
return false ;
if ( ! ( vma - > vm_flags & VM_SHARED ) ) {
/* See can_change_pte_writable(). */
page = vm_normal_page_pmd ( vma , addr , pmd ) ;
return page & & PageAnon ( page ) & & PageAnonExclusive ( page ) ;
}
/* See can_change_pte_writable(). */
return pmd_dirty ( pmd ) ;
}
mm/gup: fix FOLL_FORCE COW security issue and remove FOLL_COW
Ever since the Dirty COW (CVE-2016-5195) security issue happened, we know
that FOLL_FORCE can be possibly dangerous, especially if there are races
that can be exploited by user space.
Right now, it would be sufficient to have some code that sets a PTE of a
R/O-mapped shared page dirty, in order for it to erroneously become
writable by FOLL_FORCE. The implications of setting a write-protected PTE
dirty might not be immediately obvious to everyone.
And in fact ever since commit 9ae0f87d009c ("mm/shmem: unconditionally set
pte dirty in mfill_atomic_install_pte"), we can use UFFDIO_CONTINUE to map
a shmem page R/O while marking the pte dirty. This can be used by
unprivileged user space to modify tmpfs/shmem file content even if the
user does not have write permissions to the file, and to bypass memfd
write sealing -- Dirty COW restricted to tmpfs/shmem (CVE-2022-2590).
To fix such security issues for good, the insight is that we really only
need that fancy retry logic (FOLL_COW) for COW mappings that are not
writable (!VM_WRITE). And in a COW mapping, we really only broke COW if
we have an exclusive anonymous page mapped. If we have something else
mapped, or the mapped anonymous page might be shared (!PageAnonExclusive),
we have to trigger a write fault to break COW. If we don't find an
exclusive anonymous page when we retry, we have to trigger COW breaking
once again because something intervened.
Let's move away from this mandatory-retry + dirty handling and rely on our
PageAnonExclusive() flag for making a similar decision, to use the same
COW logic as in other kernel parts here as well. In case we stumble over
a PTE in a COW mapping that does not map an exclusive anonymous page, COW
was not properly broken and we have to trigger a fake write-fault to break
COW.
Just like we do in can_change_pte_writable() added via commit 64fe24a3e05e
("mm/mprotect: try avoiding write faults for exclusive anonymous pages
when changing protection") and commit 76aefad628aa ("mm/mprotect: fix
soft-dirty check in can_change_pte_writable()"), take care of softdirty
and uffd-wp manually.
For example, a write() via /proc/self/mem to a uffd-wp-protected range has
to fail instead of silently granting write access and bypassing the
userspace fault handler. Note that FOLL_FORCE is not only used for debug
access, but also triggered by applications without debug intentions, for
example, when pinning pages via RDMA.
This fixes CVE-2022-2590. Note that only x86_64 and aarch64 are
affected, because only those support CONFIG_HAVE_ARCH_USERFAULTFD_MINOR.
Fortunately, FOLL_COW is no longer required to handle FOLL_FORCE. So
let's just get rid of it.
Thanks to Nadav Amit for pointing out that the pte_dirty() check in
FOLL_FORCE code is problematic and might be exploitable.
Note 1: We don't check for the PTE being dirty because it doesn't matter
for making a "was COWed" decision anymore, and whoever modifies the
page has to set the page dirty either way.
Note 2: Kernels before extended uffd-wp support and before
PageAnonExclusive (< 5.19) can simply revert the problematic
commit instead and be safe regarding UFFDIO_CONTINUE. A backport to
v5.19 requires minor adjustments due to lack of
vma_soft_dirty_enabled().
Link: https://lkml.kernel.org/r/20220809205640.70916-1-david@redhat.com
Fixes: 9ae0f87d009c ("mm/shmem: unconditionally set pte dirty in mfill_atomic_install_pte")
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Nadav Amit <nadav.amit@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: David Laight <David.Laight@ACULAB.COM>
Cc: <stable@vger.kernel.org> [5.16]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-09 23:56:40 +03:00
/* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
static inline bool can_follow_write_pmd ( pmd_t pmd , struct page * page ,
struct vm_area_struct * vma ,
unsigned int flags )
mm/huge_memory.c: respect FOLL_FORCE/FOLL_COW for thp
In commit 19be0eaffa3a ("mm: remove gup_flags FOLL_WRITE games from
__get_user_pages()"), the mm code was changed from unsetting FOLL_WRITE
after a COW was resolved to setting the (newly introduced) FOLL_COW
instead. Simultaneously, the check in gup.c was updated to still allow
writes with FOLL_FORCE set if FOLL_COW had also been set.
However, a similar check in huge_memory.c was forgotten. As a result,
remote memory writes to ro regions of memory backed by transparent huge
pages cause an infinite loop in the kernel (handle_mm_fault sets
FOLL_COW and returns 0 causing a retry, but follow_trans_huge_pmd bails
out immidiately because `(flags & FOLL_WRITE) && !pmd_write(*pmd)` is
true.
While in this state the process is stil SIGKILLable, but little else
works (e.g. no ptrace attach, no other signals). This is easily
reproduced with the following code (assuming thp are set to always):
#include <assert.h>
#include <fcntl.h>
#include <stdint.h>
#include <stdio.h>
#include <string.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <unistd.h>
#define TEST_SIZE 5 * 1024 * 1024
int main(void) {
int status;
pid_t child;
int fd = open("/proc/self/mem", O_RDWR);
void *addr = mmap(NULL, TEST_SIZE, PROT_READ,
MAP_ANONYMOUS | MAP_PRIVATE, 0, 0);
assert(addr != MAP_FAILED);
pid_t parent_pid = getpid();
if ((child = fork()) == 0) {
void *addr2 = mmap(NULL, TEST_SIZE, PROT_READ | PROT_WRITE,
MAP_ANONYMOUS | MAP_PRIVATE, 0, 0);
assert(addr2 != MAP_FAILED);
memset(addr2, 'a', TEST_SIZE);
pwrite(fd, addr2, TEST_SIZE, (uintptr_t)addr);
return 0;
}
assert(child == waitpid(child, &status, 0));
assert(WIFEXITED(status) && WEXITSTATUS(status) == 0);
return 0;
}
Fix this by updating follow_trans_huge_pmd in huge_memory.c analogously
to the update in gup.c in the original commit. The same pattern exists
in follow_devmap_pmd. However, we should not be able to reach that
check with FOLL_COW set, so add WARN_ONCE to make sure we notice if we
ever do.
[akpm@linux-foundation.org: coding-style fixes]
Link: http://lkml.kernel.org/r/20170106015025.GA38411@juliacomputing.com
Signed-off-by: Keno Fischer <keno@juliacomputing.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Willy Tarreau <w@1wt.eu>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Kees Cook <keescook@chromium.org>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-01-25 02:17:48 +03:00
{
mm/gup: fix FOLL_FORCE COW security issue and remove FOLL_COW
Ever since the Dirty COW (CVE-2016-5195) security issue happened, we know
that FOLL_FORCE can be possibly dangerous, especially if there are races
that can be exploited by user space.
Right now, it would be sufficient to have some code that sets a PTE of a
R/O-mapped shared page dirty, in order for it to erroneously become
writable by FOLL_FORCE. The implications of setting a write-protected PTE
dirty might not be immediately obvious to everyone.
And in fact ever since commit 9ae0f87d009c ("mm/shmem: unconditionally set
pte dirty in mfill_atomic_install_pte"), we can use UFFDIO_CONTINUE to map
a shmem page R/O while marking the pte dirty. This can be used by
unprivileged user space to modify tmpfs/shmem file content even if the
user does not have write permissions to the file, and to bypass memfd
write sealing -- Dirty COW restricted to tmpfs/shmem (CVE-2022-2590).
To fix such security issues for good, the insight is that we really only
need that fancy retry logic (FOLL_COW) for COW mappings that are not
writable (!VM_WRITE). And in a COW mapping, we really only broke COW if
we have an exclusive anonymous page mapped. If we have something else
mapped, or the mapped anonymous page might be shared (!PageAnonExclusive),
we have to trigger a write fault to break COW. If we don't find an
exclusive anonymous page when we retry, we have to trigger COW breaking
once again because something intervened.
Let's move away from this mandatory-retry + dirty handling and rely on our
PageAnonExclusive() flag for making a similar decision, to use the same
COW logic as in other kernel parts here as well. In case we stumble over
a PTE in a COW mapping that does not map an exclusive anonymous page, COW
was not properly broken and we have to trigger a fake write-fault to break
COW.
Just like we do in can_change_pte_writable() added via commit 64fe24a3e05e
("mm/mprotect: try avoiding write faults for exclusive anonymous pages
when changing protection") and commit 76aefad628aa ("mm/mprotect: fix
soft-dirty check in can_change_pte_writable()"), take care of softdirty
and uffd-wp manually.
For example, a write() via /proc/self/mem to a uffd-wp-protected range has
to fail instead of silently granting write access and bypassing the
userspace fault handler. Note that FOLL_FORCE is not only used for debug
access, but also triggered by applications without debug intentions, for
example, when pinning pages via RDMA.
This fixes CVE-2022-2590. Note that only x86_64 and aarch64 are
affected, because only those support CONFIG_HAVE_ARCH_USERFAULTFD_MINOR.
Fortunately, FOLL_COW is no longer required to handle FOLL_FORCE. So
let's just get rid of it.
Thanks to Nadav Amit for pointing out that the pte_dirty() check in
FOLL_FORCE code is problematic and might be exploitable.
Note 1: We don't check for the PTE being dirty because it doesn't matter
for making a "was COWed" decision anymore, and whoever modifies the
page has to set the page dirty either way.
Note 2: Kernels before extended uffd-wp support and before
PageAnonExclusive (< 5.19) can simply revert the problematic
commit instead and be safe regarding UFFDIO_CONTINUE. A backport to
v5.19 requires minor adjustments due to lack of
vma_soft_dirty_enabled().
Link: https://lkml.kernel.org/r/20220809205640.70916-1-david@redhat.com
Fixes: 9ae0f87d009c ("mm/shmem: unconditionally set pte dirty in mfill_atomic_install_pte")
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Nadav Amit <nadav.amit@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: David Laight <David.Laight@ACULAB.COM>
Cc: <stable@vger.kernel.org> [5.16]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-09 23:56:40 +03:00
/* If the pmd is writable, we can write to the page. */
if ( pmd_write ( pmd ) )
return true ;
/* Maybe FOLL_FORCE is set to override it? */
if ( ! ( flags & FOLL_FORCE ) )
return false ;
/* But FOLL_FORCE has no effect on shared mappings */
if ( vma - > vm_flags & ( VM_MAYSHARE | VM_SHARED ) )
return false ;
/* ... or read-only private ones */
if ( ! ( vma - > vm_flags & VM_MAYWRITE ) )
return false ;
/* ... or already writable ones that just need to take a write fault */
if ( vma - > vm_flags & VM_WRITE )
return false ;
/*
* See can_change_pte_writable ( ) : we broke COW and could map the page
* writable if we have an exclusive anonymous page . . .
*/
if ( ! page | | ! PageAnon ( page ) | | ! PageAnonExclusive ( page ) )
return false ;
/* ... and a write-fault isn't required for other reasons. */
if ( vma_soft_dirty_enabled ( vma ) & & ! pmd_soft_dirty ( pmd ) )
return false ;
return ! userfaultfd_huge_pmd_wp ( vma , pmd ) ;
mm/huge_memory.c: respect FOLL_FORCE/FOLL_COW for thp
In commit 19be0eaffa3a ("mm: remove gup_flags FOLL_WRITE games from
__get_user_pages()"), the mm code was changed from unsetting FOLL_WRITE
after a COW was resolved to setting the (newly introduced) FOLL_COW
instead. Simultaneously, the check in gup.c was updated to still allow
writes with FOLL_FORCE set if FOLL_COW had also been set.
However, a similar check in huge_memory.c was forgotten. As a result,
remote memory writes to ro regions of memory backed by transparent huge
pages cause an infinite loop in the kernel (handle_mm_fault sets
FOLL_COW and returns 0 causing a retry, but follow_trans_huge_pmd bails
out immidiately because `(flags & FOLL_WRITE) && !pmd_write(*pmd)` is
true.
While in this state the process is stil SIGKILLable, but little else
works (e.g. no ptrace attach, no other signals). This is easily
reproduced with the following code (assuming thp are set to always):
#include <assert.h>
#include <fcntl.h>
#include <stdint.h>
#include <stdio.h>
#include <string.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <unistd.h>
#define TEST_SIZE 5 * 1024 * 1024
int main(void) {
int status;
pid_t child;
int fd = open("/proc/self/mem", O_RDWR);
void *addr = mmap(NULL, TEST_SIZE, PROT_READ,
MAP_ANONYMOUS | MAP_PRIVATE, 0, 0);
assert(addr != MAP_FAILED);
pid_t parent_pid = getpid();
if ((child = fork()) == 0) {
void *addr2 = mmap(NULL, TEST_SIZE, PROT_READ | PROT_WRITE,
MAP_ANONYMOUS | MAP_PRIVATE, 0, 0);
assert(addr2 != MAP_FAILED);
memset(addr2, 'a', TEST_SIZE);
pwrite(fd, addr2, TEST_SIZE, (uintptr_t)addr);
return 0;
}
assert(child == waitpid(child, &status, 0));
assert(WIFEXITED(status) && WEXITSTATUS(status) == 0);
return 0;
}
Fix this by updating follow_trans_huge_pmd in huge_memory.c analogously
to the update in gup.c in the original commit. The same pattern exists
in follow_devmap_pmd. However, we should not be able to reach that
check with FOLL_COW set, so add WARN_ONCE to make sure we notice if we
ever do.
[akpm@linux-foundation.org: coding-style fixes]
Link: http://lkml.kernel.org/r/20170106015025.GA38411@juliacomputing.com
Signed-off-by: Keno Fischer <keno@juliacomputing.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Willy Tarreau <w@1wt.eu>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Kees Cook <keescook@chromium.org>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-01-25 02:17:48 +03:00
}
mm, thp: fix mapped pages avoiding unevictable list on mlock
When a transparent hugepage is mapped and it is included in an mlock()
range, follow_page() incorrectly avoids setting the page's mlock bit and
moving it to the unevictable lru.
This is evident if you try to mlock(), munlock(), and then mlock() a
range again. Currently:
#define MAP_SIZE (4 << 30) /* 4GB */
void *ptr = mmap(NULL, MAP_SIZE, PROT_READ | PROT_WRITE,
MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
mlock(ptr, MAP_SIZE);
$ grep -E "Unevictable|Inactive\(anon" /proc/meminfo
Inactive(anon): 6304 kB
Unevictable: 4213924 kB
munlock(ptr, MAP_SIZE);
Inactive(anon): 4186252 kB
Unevictable: 19652 kB
mlock(ptr, MAP_SIZE);
Inactive(anon): 4198556 kB
Unevictable: 21684 kB
Notice that less than 2MB was added to the unevictable list; this is
because these pages in the range are not transparent hugepages since the
4GB range was allocated with mmap() and has no specific alignment. If
posix_memalign() were used instead, unevictable would not have grown at
all on the second mlock().
The fix is to call mlock_vma_page() so that the mlock bit is set and the
page is added to the unevictable list. With this patch:
mlock(ptr, MAP_SIZE);
Inactive(anon): 4056 kB
Unevictable: 4213940 kB
munlock(ptr, MAP_SIZE);
Inactive(anon): 4198268 kB
Unevictable: 19636 kB
mlock(ptr, MAP_SIZE);
Inactive(anon): 4008 kB
Unevictable: 4213940 kB
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michel Lespinasse <walken@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-09 03:34:03 +04:00
struct page * follow_trans_huge_pmd ( struct vm_area_struct * vma ,
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
unsigned long addr ,
pmd_t * pmd ,
unsigned int flags )
{
mm, thp: fix mapped pages avoiding unevictable list on mlock
When a transparent hugepage is mapped and it is included in an mlock()
range, follow_page() incorrectly avoids setting the page's mlock bit and
moving it to the unevictable lru.
This is evident if you try to mlock(), munlock(), and then mlock() a
range again. Currently:
#define MAP_SIZE (4 << 30) /* 4GB */
void *ptr = mmap(NULL, MAP_SIZE, PROT_READ | PROT_WRITE,
MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
mlock(ptr, MAP_SIZE);
$ grep -E "Unevictable|Inactive\(anon" /proc/meminfo
Inactive(anon): 6304 kB
Unevictable: 4213924 kB
munlock(ptr, MAP_SIZE);
Inactive(anon): 4186252 kB
Unevictable: 19652 kB
mlock(ptr, MAP_SIZE);
Inactive(anon): 4198556 kB
Unevictable: 21684 kB
Notice that less than 2MB was added to the unevictable list; this is
because these pages in the range are not transparent hugepages since the
4GB range was allocated with mmap() and has no specific alignment. If
posix_memalign() were used instead, unevictable would not have grown at
all on the second mlock().
The fix is to call mlock_vma_page() so that the mlock bit is set and the
page is added to the unevictable list. With this patch:
mlock(ptr, MAP_SIZE);
Inactive(anon): 4056 kB
Unevictable: 4213940 kB
munlock(ptr, MAP_SIZE);
Inactive(anon): 4198268 kB
Unevictable: 19636 kB
mlock(ptr, MAP_SIZE);
Inactive(anon): 4008 kB
Unevictable: 4213940 kB
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michel Lespinasse <walken@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-09 03:34:03 +04:00
struct mm_struct * mm = vma - > vm_mm ;
mm/gup: fix FOLL_FORCE COW security issue and remove FOLL_COW
Ever since the Dirty COW (CVE-2016-5195) security issue happened, we know
that FOLL_FORCE can be possibly dangerous, especially if there are races
that can be exploited by user space.
Right now, it would be sufficient to have some code that sets a PTE of a
R/O-mapped shared page dirty, in order for it to erroneously become
writable by FOLL_FORCE. The implications of setting a write-protected PTE
dirty might not be immediately obvious to everyone.
And in fact ever since commit 9ae0f87d009c ("mm/shmem: unconditionally set
pte dirty in mfill_atomic_install_pte"), we can use UFFDIO_CONTINUE to map
a shmem page R/O while marking the pte dirty. This can be used by
unprivileged user space to modify tmpfs/shmem file content even if the
user does not have write permissions to the file, and to bypass memfd
write sealing -- Dirty COW restricted to tmpfs/shmem (CVE-2022-2590).
To fix such security issues for good, the insight is that we really only
need that fancy retry logic (FOLL_COW) for COW mappings that are not
writable (!VM_WRITE). And in a COW mapping, we really only broke COW if
we have an exclusive anonymous page mapped. If we have something else
mapped, or the mapped anonymous page might be shared (!PageAnonExclusive),
we have to trigger a write fault to break COW. If we don't find an
exclusive anonymous page when we retry, we have to trigger COW breaking
once again because something intervened.
Let's move away from this mandatory-retry + dirty handling and rely on our
PageAnonExclusive() flag for making a similar decision, to use the same
COW logic as in other kernel parts here as well. In case we stumble over
a PTE in a COW mapping that does not map an exclusive anonymous page, COW
was not properly broken and we have to trigger a fake write-fault to break
COW.
Just like we do in can_change_pte_writable() added via commit 64fe24a3e05e
("mm/mprotect: try avoiding write faults for exclusive anonymous pages
when changing protection") and commit 76aefad628aa ("mm/mprotect: fix
soft-dirty check in can_change_pte_writable()"), take care of softdirty
and uffd-wp manually.
For example, a write() via /proc/self/mem to a uffd-wp-protected range has
to fail instead of silently granting write access and bypassing the
userspace fault handler. Note that FOLL_FORCE is not only used for debug
access, but also triggered by applications without debug intentions, for
example, when pinning pages via RDMA.
This fixes CVE-2022-2590. Note that only x86_64 and aarch64 are
affected, because only those support CONFIG_HAVE_ARCH_USERFAULTFD_MINOR.
Fortunately, FOLL_COW is no longer required to handle FOLL_FORCE. So
let's just get rid of it.
Thanks to Nadav Amit for pointing out that the pte_dirty() check in
FOLL_FORCE code is problematic and might be exploitable.
Note 1: We don't check for the PTE being dirty because it doesn't matter
for making a "was COWed" decision anymore, and whoever modifies the
page has to set the page dirty either way.
Note 2: Kernels before extended uffd-wp support and before
PageAnonExclusive (< 5.19) can simply revert the problematic
commit instead and be safe regarding UFFDIO_CONTINUE. A backport to
v5.19 requires minor adjustments due to lack of
vma_soft_dirty_enabled().
Link: https://lkml.kernel.org/r/20220809205640.70916-1-david@redhat.com
Fixes: 9ae0f87d009c ("mm/shmem: unconditionally set pte dirty in mfill_atomic_install_pte")
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Nadav Amit <nadav.amit@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: David Laight <David.Laight@ACULAB.COM>
Cc: <stable@vger.kernel.org> [5.16]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-09 23:56:40 +03:00
struct page * page ;
2022-10-21 20:41:08 +03:00
int ret ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
2013-11-15 02:31:04 +04:00
assert_spin_locked ( pmd_lockptr ( mm , pmd ) ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
mm/gup: fix FOLL_FORCE COW security issue and remove FOLL_COW
Ever since the Dirty COW (CVE-2016-5195) security issue happened, we know
that FOLL_FORCE can be possibly dangerous, especially if there are races
that can be exploited by user space.
Right now, it would be sufficient to have some code that sets a PTE of a
R/O-mapped shared page dirty, in order for it to erroneously become
writable by FOLL_FORCE. The implications of setting a write-protected PTE
dirty might not be immediately obvious to everyone.
And in fact ever since commit 9ae0f87d009c ("mm/shmem: unconditionally set
pte dirty in mfill_atomic_install_pte"), we can use UFFDIO_CONTINUE to map
a shmem page R/O while marking the pte dirty. This can be used by
unprivileged user space to modify tmpfs/shmem file content even if the
user does not have write permissions to the file, and to bypass memfd
write sealing -- Dirty COW restricted to tmpfs/shmem (CVE-2022-2590).
To fix such security issues for good, the insight is that we really only
need that fancy retry logic (FOLL_COW) for COW mappings that are not
writable (!VM_WRITE). And in a COW mapping, we really only broke COW if
we have an exclusive anonymous page mapped. If we have something else
mapped, or the mapped anonymous page might be shared (!PageAnonExclusive),
we have to trigger a write fault to break COW. If we don't find an
exclusive anonymous page when we retry, we have to trigger COW breaking
once again because something intervened.
Let's move away from this mandatory-retry + dirty handling and rely on our
PageAnonExclusive() flag for making a similar decision, to use the same
COW logic as in other kernel parts here as well. In case we stumble over
a PTE in a COW mapping that does not map an exclusive anonymous page, COW
was not properly broken and we have to trigger a fake write-fault to break
COW.
Just like we do in can_change_pte_writable() added via commit 64fe24a3e05e
("mm/mprotect: try avoiding write faults for exclusive anonymous pages
when changing protection") and commit 76aefad628aa ("mm/mprotect: fix
soft-dirty check in can_change_pte_writable()"), take care of softdirty
and uffd-wp manually.
For example, a write() via /proc/self/mem to a uffd-wp-protected range has
to fail instead of silently granting write access and bypassing the
userspace fault handler. Note that FOLL_FORCE is not only used for debug
access, but also triggered by applications without debug intentions, for
example, when pinning pages via RDMA.
This fixes CVE-2022-2590. Note that only x86_64 and aarch64 are
affected, because only those support CONFIG_HAVE_ARCH_USERFAULTFD_MINOR.
Fortunately, FOLL_COW is no longer required to handle FOLL_FORCE. So
let's just get rid of it.
Thanks to Nadav Amit for pointing out that the pte_dirty() check in
FOLL_FORCE code is problematic and might be exploitable.
Note 1: We don't check for the PTE being dirty because it doesn't matter
for making a "was COWed" decision anymore, and whoever modifies the
page has to set the page dirty either way.
Note 2: Kernels before extended uffd-wp support and before
PageAnonExclusive (< 5.19) can simply revert the problematic
commit instead and be safe regarding UFFDIO_CONTINUE. A backport to
v5.19 requires minor adjustments due to lack of
vma_soft_dirty_enabled().
Link: https://lkml.kernel.org/r/20220809205640.70916-1-david@redhat.com
Fixes: 9ae0f87d009c ("mm/shmem: unconditionally set pte dirty in mfill_atomic_install_pte")
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Nadav Amit <nadav.amit@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: David Laight <David.Laight@ACULAB.COM>
Cc: <stable@vger.kernel.org> [5.16]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-09 23:56:40 +03:00
page = pmd_page ( * pmd ) ;
VM_BUG_ON_PAGE ( ! PageHead ( page ) & & ! is_zone_device_page ( page ) , page ) ;
if ( ( flags & FOLL_WRITE ) & &
! can_follow_write_pmd ( * pmd , page , vma , flags ) )
return NULL ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
2013-02-05 02:28:42 +04:00
/* Avoid dumping huge zero page */
if ( ( flags & FOLL_DUMP ) & & is_huge_zero_pmd ( * pmd ) )
return ERR_PTR ( - EFAULT ) ;
2013-12-19 05:08:32 +04:00
/* Full NUMA hinting faults to serialise migration in fault paths */
2022-08-25 19:46:57 +03:00
if ( pmd_protnone ( * pmd ) & & ! gup_can_follow_protnone ( flags ) )
mm/gup: fix FOLL_FORCE COW security issue and remove FOLL_COW
Ever since the Dirty COW (CVE-2016-5195) security issue happened, we know
that FOLL_FORCE can be possibly dangerous, especially if there are races
that can be exploited by user space.
Right now, it would be sufficient to have some code that sets a PTE of a
R/O-mapped shared page dirty, in order for it to erroneously become
writable by FOLL_FORCE. The implications of setting a write-protected PTE
dirty might not be immediately obvious to everyone.
And in fact ever since commit 9ae0f87d009c ("mm/shmem: unconditionally set
pte dirty in mfill_atomic_install_pte"), we can use UFFDIO_CONTINUE to map
a shmem page R/O while marking the pte dirty. This can be used by
unprivileged user space to modify tmpfs/shmem file content even if the
user does not have write permissions to the file, and to bypass memfd
write sealing -- Dirty COW restricted to tmpfs/shmem (CVE-2022-2590).
To fix such security issues for good, the insight is that we really only
need that fancy retry logic (FOLL_COW) for COW mappings that are not
writable (!VM_WRITE). And in a COW mapping, we really only broke COW if
we have an exclusive anonymous page mapped. If we have something else
mapped, or the mapped anonymous page might be shared (!PageAnonExclusive),
we have to trigger a write fault to break COW. If we don't find an
exclusive anonymous page when we retry, we have to trigger COW breaking
once again because something intervened.
Let's move away from this mandatory-retry + dirty handling and rely on our
PageAnonExclusive() flag for making a similar decision, to use the same
COW logic as in other kernel parts here as well. In case we stumble over
a PTE in a COW mapping that does not map an exclusive anonymous page, COW
was not properly broken and we have to trigger a fake write-fault to break
COW.
Just like we do in can_change_pte_writable() added via commit 64fe24a3e05e
("mm/mprotect: try avoiding write faults for exclusive anonymous pages
when changing protection") and commit 76aefad628aa ("mm/mprotect: fix
soft-dirty check in can_change_pte_writable()"), take care of softdirty
and uffd-wp manually.
For example, a write() via /proc/self/mem to a uffd-wp-protected range has
to fail instead of silently granting write access and bypassing the
userspace fault handler. Note that FOLL_FORCE is not only used for debug
access, but also triggered by applications without debug intentions, for
example, when pinning pages via RDMA.
This fixes CVE-2022-2590. Note that only x86_64 and aarch64 are
affected, because only those support CONFIG_HAVE_ARCH_USERFAULTFD_MINOR.
Fortunately, FOLL_COW is no longer required to handle FOLL_FORCE. So
let's just get rid of it.
Thanks to Nadav Amit for pointing out that the pte_dirty() check in
FOLL_FORCE code is problematic and might be exploitable.
Note 1: We don't check for the PTE being dirty because it doesn't matter
for making a "was COWed" decision anymore, and whoever modifies the
page has to set the page dirty either way.
Note 2: Kernels before extended uffd-wp support and before
PageAnonExclusive (< 5.19) can simply revert the problematic
commit instead and be safe regarding UFFDIO_CONTINUE. A backport to
v5.19 requires minor adjustments due to lack of
vma_soft_dirty_enabled().
Link: https://lkml.kernel.org/r/20220809205640.70916-1-david@redhat.com
Fixes: 9ae0f87d009c ("mm/shmem: unconditionally set pte dirty in mfill_atomic_install_pte")
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Nadav Amit <nadav.amit@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: David Laight <David.Laight@ACULAB.COM>
Cc: <stable@vger.kernel.org> [5.16]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-09 23:56:40 +03:00
return NULL ;
mm/gup: track FOLL_PIN pages
Add tracking of pages that were pinned via FOLL_PIN. This tracking is
implemented via overloading of page->_refcount: pins are added by adding
GUP_PIN_COUNTING_BIAS (1024) to the refcount. This provides a fuzzy
indication of pinning, and it can have false positives (and that's OK).
Please see the pre-existing Documentation/core-api/pin_user_pages.rst for
details.
As mentioned in pin_user_pages.rst, callers who effectively set FOLL_PIN
(typically via pin_user_pages*()) are required to ultimately free such
pages via unpin_user_page().
Please also note the limitation, discussed in pin_user_pages.rst under the
"TODO: for 1GB and larger huge pages" section. (That limitation will be
removed in a following patch.)
The effect of a FOLL_PIN flag is similar to that of FOLL_GET, and may be
thought of as "FOLL_GET for DIO and/or RDMA use".
Pages that have been pinned via FOLL_PIN are identifiable via a new
function call:
bool page_maybe_dma_pinned(struct page *page);
What to do in response to encountering such a page, is left to later
patchsets. There is discussion about this in [1], [2], [3], and [4].
This also changes a BUG_ON(), to a WARN_ON(), in follow_page_mask().
[1] Some slow progress on get_user_pages() (Apr 2, 2019):
https://lwn.net/Articles/784574/
[2] DMA and get_user_pages() (LPC: Dec 12, 2018):
https://lwn.net/Articles/774411/
[3] The trouble with get_user_pages() (Apr 30, 2018):
https://lwn.net/Articles/753027/
[4] LWN kernel index: get_user_pages():
https://lwn.net/Kernel/Index/#Memory_management-get_user_pages
[jhubbard@nvidia.com: add kerneldoc]
Link: http://lkml.kernel.org/r/20200307021157.235726-1-jhubbard@nvidia.com
[imbrenda@linux.ibm.com: if pin fails, we need to unpin, a simple put_page will not be enough]
Link: http://lkml.kernel.org/r/20200306132537.783769-2-imbrenda@linux.ibm.com
[akpm@linux-foundation.org: fix put_compound_head defined but not used]
Suggested-by: Jan Kara <jack@suse.cz>
Suggested-by: Jérôme Glisse <jglisse@redhat.com>
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Ira Weiny <ira.weiny@intel.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Link: http://lkml.kernel.org/r/20200211001536.1027652-7-jhubbard@nvidia.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:05:29 +03:00
mm/gup: reliable R/O long-term pinning in COW mappings
We already support reliable R/O pinning of anonymous memory. However,
assume we end up pinning (R/O long-term) a pagecache page or the shared
zeropage inside a writable private ("COW") mapping. The next write access
will trigger a write-fault and replace the pinned page by an exclusive
anonymous page in the process page tables to break COW: the pinned page no
longer corresponds to the page mapped into the process' page table.
Now that FAULT_FLAG_UNSHARE can break COW on anything mapped into a
COW mapping, let's properly break COW first before R/O long-term
pinning something that's not an exclusive anon page inside a COW
mapping. FAULT_FLAG_UNSHARE will break COW and map an exclusive anon page
instead that can get pinned safely.
With this change, we can stop using FOLL_FORCE|FOLL_WRITE for reliable
R/O long-term pinning in COW mappings.
With this change, the new R/O long-term pinning tests for non-anonymous
memory succeed:
# [RUN] R/O longterm GUP pin ... with shared zeropage
ok 151 Longterm R/O pin is reliable
# [RUN] R/O longterm GUP pin ... with memfd
ok 152 Longterm R/O pin is reliable
# [RUN] R/O longterm GUP pin ... with tmpfile
ok 153 Longterm R/O pin is reliable
# [RUN] R/O longterm GUP pin ... with huge zeropage
ok 154 Longterm R/O pin is reliable
# [RUN] R/O longterm GUP pin ... with memfd hugetlb (2048 kB)
ok 155 Longterm R/O pin is reliable
# [RUN] R/O longterm GUP pin ... with memfd hugetlb (1048576 kB)
ok 156 Longterm R/O pin is reliable
# [RUN] R/O longterm GUP-fast pin ... with shared zeropage
ok 157 Longterm R/O pin is reliable
# [RUN] R/O longterm GUP-fast pin ... with memfd
ok 158 Longterm R/O pin is reliable
# [RUN] R/O longterm GUP-fast pin ... with tmpfile
ok 159 Longterm R/O pin is reliable
# [RUN] R/O longterm GUP-fast pin ... with huge zeropage
ok 160 Longterm R/O pin is reliable
# [RUN] R/O longterm GUP-fast pin ... with memfd hugetlb (2048 kB)
ok 161 Longterm R/O pin is reliable
# [RUN] R/O longterm GUP-fast pin ... with memfd hugetlb (1048576 kB)
ok 162 Longterm R/O pin is reliable
Note 1: We don't care about short-term R/O-pinning, because they have
snapshot semantics: they are not supposed to observe modifications that
happen after pinning.
As one example, assume we start direct I/O to read from a page and store
page content into a file: modifications to page content after starting
direct I/O are not guaranteed to end up in the file. So even if we'd pin
the shared zeropage, the end result would be as expected -- getting zeroes
stored to the file.
Note 2: For shared mappings we'll now always fallback to the slow path to
lookup the VMA when R/O long-term pining. While that's the necessary price
we have to pay right now, it's actually not that bad in practice: most
FOLL_LONGTERM users already specify FOLL_WRITE, for example, along with
FOLL_FORCE because they tried dealing with COW mappings correctly ...
Note 3: For users that use FOLL_LONGTERM right now without FOLL_WRITE,
such as VFIO, we'd now no longer pin the shared zeropage. Instead, we'd
populate exclusive anon pages that we can pin. There was a concern that
this could affect the memlock limit of existing setups.
For example, a VM running with VFIO could run into the memlock limit and
fail to run. However, we essentially had the same behavior already in
commit 17839856fd58 ("gup: document and work around "COW can break either
way" issue") which got merged into some enterprise distros, and there were
not any such complaints. So most probably, we're fine.
Link: https://lkml.kernel.org/r/20221116102659.70287-10-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: John Hubbard <jhubbard@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-16 13:26:48 +03:00
if ( ! pmd_write ( * pmd ) & & gup_must_unshare ( vma , flags , page ) )
2022-05-10 04:20:45 +03:00
return ERR_PTR ( - EMLINK ) ;
mm/gup: sanity-check with CONFIG_DEBUG_VM that anonymous pages are exclusive when (un)pinning
Let's verify when (un)pinning anonymous pages that we always deal with
exclusive anonymous pages, which guarantees that we'll have a reliable
PIN, meaning that we cannot end up with the GUP pin being inconsistent
with he pages mapped into the page tables due to a COW triggered by a
write fault.
When pinning pages, after conditionally triggering GUP unsharing of
possibly shared anonymous pages, we should always only see exclusive
anonymous pages. Note that anonymous pages that are mapped writable must
be marked exclusive, otherwise we'd have a BUG.
When pinning during ordinary GUP, simply add a check after our conditional
GUP-triggered unsharing checks. As we know exactly how the page is
mapped, we know exactly in which page we have to check for
PageAnonExclusive().
When pinning via GUP-fast we have to be careful, because we can race with
fork(): verify only after we made sure via the seqcount that we didn't
race with concurrent fork() that we didn't end up pinning a possibly
shared anonymous page.
Similarly, when unpinning, verify that the pages are still marked as
exclusive: otherwise something turned the pages possibly shared, which can
result in random memory corruptions, which we really want to catch.
With only the pinned pages at hand and not the actual page table entries
we have to be a bit careful: hugetlb pages are always mapped via a single
logical page table entry referencing the head page and PG_anon_exclusive
of the head page applies. Anon THP are a bit more complicated, because we
might have obtained the page reference either via a PMD or a PTE --
depending on the mapping type we either have to check PageAnonExclusive of
the head page (PMD-mapped THP) or the tail page (PTE-mapped THP) applies:
as we don't know and to make our life easier, check that either is set.
Take care to not verify in case we're unpinning during GUP-fast because we
detected concurrent fork(): we might stumble over an anonymous page that
is now shared.
Link: https://lkml.kernel.org/r/20220428083441.37290-18-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:45 +03:00
VM_BUG_ON_PAGE ( ( flags & FOLL_PIN ) & & PageAnon ( page ) & &
! PageAnonExclusive ( page ) , page ) ;
2022-10-21 20:41:08 +03:00
ret = try_grab_page ( page , flags ) ;
if ( ret )
return ERR_PTR ( ret ) ;
mm/gup: track FOLL_PIN pages
Add tracking of pages that were pinned via FOLL_PIN. This tracking is
implemented via overloading of page->_refcount: pins are added by adding
GUP_PIN_COUNTING_BIAS (1024) to the refcount. This provides a fuzzy
indication of pinning, and it can have false positives (and that's OK).
Please see the pre-existing Documentation/core-api/pin_user_pages.rst for
details.
As mentioned in pin_user_pages.rst, callers who effectively set FOLL_PIN
(typically via pin_user_pages*()) are required to ultimately free such
pages via unpin_user_page().
Please also note the limitation, discussed in pin_user_pages.rst under the
"TODO: for 1GB and larger huge pages" section. (That limitation will be
removed in a following patch.)
The effect of a FOLL_PIN flag is similar to that of FOLL_GET, and may be
thought of as "FOLL_GET for DIO and/or RDMA use".
Pages that have been pinned via FOLL_PIN are identifiable via a new
function call:
bool page_maybe_dma_pinned(struct page *page);
What to do in response to encountering such a page, is left to later
patchsets. There is discussion about this in [1], [2], [3], and [4].
This also changes a BUG_ON(), to a WARN_ON(), in follow_page_mask().
[1] Some slow progress on get_user_pages() (Apr 2, 2019):
https://lwn.net/Articles/784574/
[2] DMA and get_user_pages() (LPC: Dec 12, 2018):
https://lwn.net/Articles/774411/
[3] The trouble with get_user_pages() (Apr 30, 2018):
https://lwn.net/Articles/753027/
[4] LWN kernel index: get_user_pages():
https://lwn.net/Kernel/Index/#Memory_management-get_user_pages
[jhubbard@nvidia.com: add kerneldoc]
Link: http://lkml.kernel.org/r/20200307021157.235726-1-jhubbard@nvidia.com
[imbrenda@linux.ibm.com: if pin fails, we need to unpin, a simple put_page will not be enough]
Link: http://lkml.kernel.org/r/20200306132537.783769-2-imbrenda@linux.ibm.com
[akpm@linux-foundation.org: fix put_compound_head defined but not used]
Suggested-by: Jan Kara <jack@suse.cz>
Suggested-by: Jérôme Glisse <jglisse@redhat.com>
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Ira Weiny <ira.weiny@intel.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Link: http://lkml.kernel.org/r/20200211001536.1027652-7-jhubbard@nvidia.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:05:29 +03:00
2016-01-16 03:56:55 +03:00
if ( flags & FOLL_TOUCH )
2022-07-04 16:21:50 +03:00
touch_pmd ( vma , addr , pmd , flags & FOLL_WRITE ) ;
mm/gup: track FOLL_PIN pages
Add tracking of pages that were pinned via FOLL_PIN. This tracking is
implemented via overloading of page->_refcount: pins are added by adding
GUP_PIN_COUNTING_BIAS (1024) to the refcount. This provides a fuzzy
indication of pinning, and it can have false positives (and that's OK).
Please see the pre-existing Documentation/core-api/pin_user_pages.rst for
details.
As mentioned in pin_user_pages.rst, callers who effectively set FOLL_PIN
(typically via pin_user_pages*()) are required to ultimately free such
pages via unpin_user_page().
Please also note the limitation, discussed in pin_user_pages.rst under the
"TODO: for 1GB and larger huge pages" section. (That limitation will be
removed in a following patch.)
The effect of a FOLL_PIN flag is similar to that of FOLL_GET, and may be
thought of as "FOLL_GET for DIO and/or RDMA use".
Pages that have been pinned via FOLL_PIN are identifiable via a new
function call:
bool page_maybe_dma_pinned(struct page *page);
What to do in response to encountering such a page, is left to later
patchsets. There is discussion about this in [1], [2], [3], and [4].
This also changes a BUG_ON(), to a WARN_ON(), in follow_page_mask().
[1] Some slow progress on get_user_pages() (Apr 2, 2019):
https://lwn.net/Articles/784574/
[2] DMA and get_user_pages() (LPC: Dec 12, 2018):
https://lwn.net/Articles/774411/
[3] The trouble with get_user_pages() (Apr 30, 2018):
https://lwn.net/Articles/753027/
[4] LWN kernel index: get_user_pages():
https://lwn.net/Kernel/Index/#Memory_management-get_user_pages
[jhubbard@nvidia.com: add kerneldoc]
Link: http://lkml.kernel.org/r/20200307021157.235726-1-jhubbard@nvidia.com
[imbrenda@linux.ibm.com: if pin fails, we need to unpin, a simple put_page will not be enough]
Link: http://lkml.kernel.org/r/20200306132537.783769-2-imbrenda@linux.ibm.com
[akpm@linux-foundation.org: fix put_compound_head defined but not used]
Suggested-by: Jan Kara <jack@suse.cz>
Suggested-by: Jérôme Glisse <jglisse@redhat.com>
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Ira Weiny <ira.weiny@intel.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Link: http://lkml.kernel.org/r/20200211001536.1027652-7-jhubbard@nvidia.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 07:05:29 +03:00
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
page + = ( addr & ~ HPAGE_PMD_MASK ) > > PAGE_SHIFT ;
2016-09-03 20:38:03 +03:00
VM_BUG_ON_PAGE ( ! PageCompound ( page ) & & ! is_zone_device_page ( page ) , page ) ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
return page ;
}
2012-10-25 16:16:31 +04:00
/* NUMA hinting page fault entry point for trans huge pmds */
mm: memory: add orig_pmd to struct vm_fault
Pach series "mm: thp: use generic THP migration for NUMA hinting fault", v3.
When the THP NUMA fault support was added THP migration was not supported
yet. So the ad hoc THP migration was implemented in NUMA fault handling.
Since v4.14 THP migration has been supported so it doesn't make too much
sense to still keep another THP migration implementation rather than using
the generic migration code. It is definitely a maintenance burden to keep
two THP migration implementation for different code paths and it is more
error prone. Using the generic THP migration implementation allows us
remove the duplicate code and some hacks needed by the old ad hoc
implementation.
A quick grep shows x86_64, PowerPC (book3s), ARM64 ans S390 support both
THP and NUMA balancing. The most of them support THP migration except for
S390. Zi Yan tried to add THP migration support for S390 before but it
was not accepted due to the design of S390 PMD. For the discussion,
please see: https://lkml.org/lkml/2018/4/27/953.
Per the discussion with Gerald Schaefer in v1 it is acceptible to skip
huge PMD for S390 for now.
I saw there were some hacks about gup from git history, but I didn't
figure out if they have been removed or not since I just found FOLL_NUMA
code in the current gup implementation and they seems useful.
Patch #1 ~ #2 are preparation patches.
Patch #3 is the real meat.
Patch #4 ~ #6 keep consistent counters and behaviors with before.
Patch #7 skips change huge PMD to prot_none if thp migration is not supported.
Test
----
Did some tests to measure the latency of do_huge_pmd_numa_page. The test
VM has 80 vcpus and 64G memory. The test would create 2 processes to
consume 128G memory together which would incur memory pressure to cause
THP splits. And it also creates 80 processes to hog cpu, and the memory
consumer processes are bound to different nodes periodically in order to
increase NUMA faults.
The below test script is used:
echo 3 > /proc/sys/vm/drop_caches
# Run stress-ng for 24 hours
./stress-ng/stress-ng --vm 2 --vm-bytes 64G --timeout 24h &
PID=$!
./stress-ng/stress-ng --cpu $NR_CPUS --timeout 24h &
# Wait for vm stressors forked
sleep 5
PID_1=`pgrep -P $PID | awk 'NR == 1'`
PID_2=`pgrep -P $PID | awk 'NR == 2'`
JOB1=`pgrep -P $PID_1`
JOB2=`pgrep -P $PID_2`
# Bind load jobs to different nodes periodically to force generate
# cross node memory access
while [ -d "/proc/$PID" ]
do
taskset -apc 8 $JOB1
taskset -apc 8 $JOB2
sleep 300
taskset -apc 58 $JOB1
taskset -apc 58 $JOB2
sleep 300
done
With the above test the histogram of latency of do_huge_pmd_numa_page is
as shown below. Since the number of do_huge_pmd_numa_page varies
drastically for each run (should be due to scheduler), so I converted the
raw number to percentage.
patched base
@us[stress-ng]:
[0] 3.57% 0.16%
[1] 55.68% 18.36%
[2, 4) 10.46% 40.44%
[4, 8) 7.26% 17.82%
[8, 16) 21.12% 13.41%
[16, 32) 1.06% 4.27%
[32, 64) 0.56% 4.07%
[64, 128) 0.16% 0.35%
[128, 256) < 0.1% < 0.1%
[256, 512) < 0.1% < 0.1%
[512, 1K) < 0.1% < 0.1%
[1K, 2K) < 0.1% < 0.1%
[2K, 4K) < 0.1% < 0.1%
[4K, 8K) < 0.1% < 0.1%
[8K, 16K) < 0.1% < 0.1%
[16K, 32K) < 0.1% < 0.1%
[32K, 64K) < 0.1% < 0.1%
Per the result, patched kernel is even slightly better than the base
kernel. I think this is because the lock contention against THP split is
less than base kernel due to the refactor.
To exclude the affect from THP split, I also did test w/o memory pressure.
No obvious regression is spotted. The below is the test result *w/o*
memory pressure.
patched base
@us[stress-ng]:
[0] 7.97% 18.4%
[1] 69.63% 58.24%
[2, 4) 4.18% 2.63%
[4, 8) 0.22% 0.17%
[8, 16) 1.03% 0.92%
[16, 32) 0.14% < 0.1%
[32, 64) < 0.1% < 0.1%
[64, 128) < 0.1% < 0.1%
[128, 256) < 0.1% < 0.1%
[256, 512) 0.45% 1.19%
[512, 1K) 15.45% 17.27%
[1K, 2K) < 0.1% < 0.1%
[2K, 4K) < 0.1% < 0.1%
[4K, 8K) < 0.1% < 0.1%
[8K, 16K) 0.86% 0.88%
[16K, 32K) < 0.1% 0.15%
[32K, 64K) < 0.1% < 0.1%
[64K, 128K) < 0.1% < 0.1%
[128K, 256K) < 0.1% < 0.1%
The series also survived a series of tests that exercise NUMA balancing
migrations by Mel.
This patch (of 7):
Add orig_pmd to struct vm_fault so the "orig_pmd" parameter used by huge
page fault could be removed, just like its PTE counterpart does.
Link: https://lkml.kernel.org/r/20210518200801.7413-1-shy828301@gmail.com
Link: https://lkml.kernel.org/r/20210518200801.7413-2-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-01 04:51:35 +03:00
vm_fault_t do_huge_pmd_numa_page ( struct vm_fault * vmf )
2012-10-25 16:16:31 +04:00
{
2016-12-15 02:06:58 +03:00
struct vm_area_struct * vma = vmf - > vma ;
2021-07-01 04:51:42 +03:00
pmd_t oldpmd = vmf - > orig_pmd ;
pmd_t pmd ;
2012-11-19 16:35:47 +04:00
struct page * page ;
2016-12-15 02:06:58 +03:00
unsigned long haddr = vmf - > address & HPAGE_PMD_MASK ;
2021-07-01 04:51:42 +03:00
int page_nid = NUMA_NO_NODE ;
memory tiering: hot page selection with hint page fault latency
Patch series "memory tiering: hot page selection", v4.
To optimize page placement in a memory tiering system with NUMA balancing,
the hot pages in the slow memory nodes need to be identified.
Essentially, the original NUMA balancing implementation selects the mostly
recently accessed (MRU) pages to promote. But this isn't a perfect
algorithm to identify the hot pages. Because the pages with quite low
access frequency may be accessed eventually given the NUMA balancing page
table scanning period could be quite long (e.g. 60 seconds). So in this
patchset, we implement a new hot page identification algorithm based on
the latency between NUMA balancing page table scanning and hint page
fault. Which is a kind of mostly frequently accessed (MFU) algorithm.
In NUMA balancing memory tiering mode, if there are hot pages in slow
memory node and cold pages in fast memory node, we need to promote/demote
hot/cold pages between the fast and cold memory nodes.
A choice is to promote/demote as fast as possible. But the CPU cycles and
memory bandwidth consumed by the high promoting/demoting throughput will
hurt the latency of some workload because of accessing inflating and slow
memory bandwidth contention.
A way to resolve this issue is to restrict the max promoting/demoting
throughput. It will take longer to finish the promoting/demoting. But
the workload latency will be better. This is implemented in this patchset
as the page promotion rate limit mechanism.
The promotion hot threshold is workload and system configuration
dependent. So in this patchset, a method to adjust the hot threshold
automatically is implemented. The basic idea is to control the number of
the candidate promotion pages to match the promotion rate limit.
We used the pmbench memory accessing benchmark tested the patchset on a
2-socket server system with DRAM and PMEM installed. The test results are
as follows,
pmbench score promote rate
(accesses/s) MB/s
------------- ------------
base 146887704.1 725.6
hot selection 165695601.2 544.0
rate limit 162814569.8 165.2
auto adjustment 170495294.0 136.9
From the results above,
With hot page selection patch [1/3], the pmbench score increases about
12.8%, and promote rate (overhead) decreases about 25.0%, compared with
base kernel.
With rate limit patch [2/3], pmbench score decreases about 1.7%, and
promote rate decreases about 69.6%, compared with hot page selection
patch.
With threshold auto adjustment patch [3/3], pmbench score increases about
4.7%, and promote rate decrease about 17.1%, compared with rate limit
patch.
Baolin helped to test the patchset with MySQL on a machine which contains
1 DRAM node (30G) and 1 PMEM node (126G).
sysbench /usr/share/sysbench/oltp_read_write.lua \
......
--tables=200 \
--table-size=1000000 \
--report-interval=10 \
--threads=16 \
--time=120
The tps can be improved about 5%.
This patch (of 3):
To optimize page placement in a memory tiering system with NUMA balancing,
the hot pages in the slow memory node need to be identified. Essentially,
the original NUMA balancing implementation selects the mostly recently
accessed (MRU) pages to promote. But this isn't a perfect algorithm to
identify the hot pages. Because the pages with quite low access frequency
may be accessed eventually given the NUMA balancing page table scanning
period could be quite long (e.g. 60 seconds). The most frequently
accessed (MFU) algorithm is better.
So, in this patch we implemented a better hot page selection algorithm.
Which is based on NUMA balancing page table scanning and hint page fault
as follows,
- When the page tables of the processes are scanned to change PTE/PMD
to be PROT_NONE, the current time is recorded in struct page as scan
time.
- When the page is accessed, hint page fault will occur. The scan
time is gotten from the struct page. And The hint page fault
latency is defined as
hint page fault time - scan time
The shorter the hint page fault latency of a page is, the higher the
probability of their access frequency to be higher. So the hint page
fault latency is a better estimation of the page hot/cold.
It's hard to find some extra space in struct page to hold the scan time.
Fortunately, we can reuse some bits used by the original NUMA balancing.
NUMA balancing uses some bits in struct page to store the page accessing
CPU and PID (referring to page_cpupid_xchg_last()). Which is used by the
multi-stage node selection algorithm to avoid to migrate pages shared
accessed by the NUMA nodes back and forth. But for pages in the slow
memory node, even if they are shared accessed by multiple NUMA nodes, as
long as the pages are hot, they need to be promoted to the fast memory
node. So the accessing CPU and PID information are unnecessary for the
slow memory pages. We can reuse these bits in struct page to record the
scan time. For the fast memory pages, these bits are used as before.
For the hot threshold, the default value is 1 second, which works well in
our performance test. All pages with hint page fault latency < hot
threshold will be considered hot.
It's hard for users to determine the hot threshold. So we don't provide a
kernel ABI to set it, just provide a debugfs interface for advanced users
to experiment. We will continue to work on a hot threshold automatic
adjustment mechanism.
The downside of the above method is that the response time to the workload
hot spot changing may be much longer. For example,
- A previous cold memory area becomes hot
- The hint page fault will be triggered. But the hint page fault
latency isn't shorter than the hot threshold. So the pages will
not be promoted.
- When the memory area is scanned again, maybe after a scan period,
the hint page fault latency measured will be shorter than the hot
threshold and the pages will be promoted.
To mitigate this, if there are enough free space in the fast memory node,
the hot threshold will not be used, all pages will be promoted upon the
hint page fault for fast response.
Thanks Zhong Jiang reported and tested the fix for a bug when disabling
memory tiering mode dynamically.
Link: https://lkml.kernel.org/r/20220713083954.34196-1-ying.huang@intel.com
Link: https://lkml.kernel.org/r/20220713083954.34196-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Wei Xu <weixugc@google.com>
Cc: osalvador <osalvador@suse.de>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Zhong Jiang <zhongjiang-ali@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-07-13 11:39:51 +03:00
int target_nid , last_cpupid = ( - 1 & LAST_CPUPID_MASK ) ;
2022-11-08 20:46:50 +03:00
bool migrated = false , writable = false ;
2013-10-07 14:29:24 +04:00
int flags = 0 ;
2012-10-25 16:16:31 +04:00
2016-12-15 02:06:58 +03:00
vmf - > ptl = pmd_lock ( vma - > vm_mm , vmf - > pmd ) ;
2021-07-01 04:51:42 +03:00
if ( unlikely ( ! pmd_same ( oldpmd , * vmf - > pmd ) ) ) {
2016-12-15 02:06:58 +03:00
spin_unlock ( vmf - > ptl ) ;
2013-12-19 05:08:42 +04:00
goto out ;
}
2021-07-01 04:51:42 +03:00
pmd = pmd_modify ( oldpmd , vma - > vm_page_prot ) ;
2022-11-08 20:46:50 +03:00
/*
* Detect now whether the PMD could be writable ; this information
* is only valid while holding the PT lock .
*/
writable = pmd_write ( pmd ) ;
if ( ! writable & & vma_wants_manual_pte_write_upgrade ( vma ) & &
can_change_pmd_writable ( vma , vmf - > address , pmd ) )
writable = true ;
2021-07-01 04:51:42 +03:00
page = vm_normal_page_pmd ( vma , haddr , pmd ) ;
if ( ! page )
goto out_map ;
/* See similar comment in do_numa_page for explanation */
2022-11-08 20:46:50 +03:00
if ( ! writable )
2021-07-01 04:51:42 +03:00
flags | = TNF_NO_GROUP ;
page_nid = page_to_nid ( page ) ;
memory tiering: hot page selection with hint page fault latency
Patch series "memory tiering: hot page selection", v4.
To optimize page placement in a memory tiering system with NUMA balancing,
the hot pages in the slow memory nodes need to be identified.
Essentially, the original NUMA balancing implementation selects the mostly
recently accessed (MRU) pages to promote. But this isn't a perfect
algorithm to identify the hot pages. Because the pages with quite low
access frequency may be accessed eventually given the NUMA balancing page
table scanning period could be quite long (e.g. 60 seconds). So in this
patchset, we implement a new hot page identification algorithm based on
the latency between NUMA balancing page table scanning and hint page
fault. Which is a kind of mostly frequently accessed (MFU) algorithm.
In NUMA balancing memory tiering mode, if there are hot pages in slow
memory node and cold pages in fast memory node, we need to promote/demote
hot/cold pages between the fast and cold memory nodes.
A choice is to promote/demote as fast as possible. But the CPU cycles and
memory bandwidth consumed by the high promoting/demoting throughput will
hurt the latency of some workload because of accessing inflating and slow
memory bandwidth contention.
A way to resolve this issue is to restrict the max promoting/demoting
throughput. It will take longer to finish the promoting/demoting. But
the workload latency will be better. This is implemented in this patchset
as the page promotion rate limit mechanism.
The promotion hot threshold is workload and system configuration
dependent. So in this patchset, a method to adjust the hot threshold
automatically is implemented. The basic idea is to control the number of
the candidate promotion pages to match the promotion rate limit.
We used the pmbench memory accessing benchmark tested the patchset on a
2-socket server system with DRAM and PMEM installed. The test results are
as follows,
pmbench score promote rate
(accesses/s) MB/s
------------- ------------
base 146887704.1 725.6
hot selection 165695601.2 544.0
rate limit 162814569.8 165.2
auto adjustment 170495294.0 136.9
From the results above,
With hot page selection patch [1/3], the pmbench score increases about
12.8%, and promote rate (overhead) decreases about 25.0%, compared with
base kernel.
With rate limit patch [2/3], pmbench score decreases about 1.7%, and
promote rate decreases about 69.6%, compared with hot page selection
patch.
With threshold auto adjustment patch [3/3], pmbench score increases about
4.7%, and promote rate decrease about 17.1%, compared with rate limit
patch.
Baolin helped to test the patchset with MySQL on a machine which contains
1 DRAM node (30G) and 1 PMEM node (126G).
sysbench /usr/share/sysbench/oltp_read_write.lua \
......
--tables=200 \
--table-size=1000000 \
--report-interval=10 \
--threads=16 \
--time=120
The tps can be improved about 5%.
This patch (of 3):
To optimize page placement in a memory tiering system with NUMA balancing,
the hot pages in the slow memory node need to be identified. Essentially,
the original NUMA balancing implementation selects the mostly recently
accessed (MRU) pages to promote. But this isn't a perfect algorithm to
identify the hot pages. Because the pages with quite low access frequency
may be accessed eventually given the NUMA balancing page table scanning
period could be quite long (e.g. 60 seconds). The most frequently
accessed (MFU) algorithm is better.
So, in this patch we implemented a better hot page selection algorithm.
Which is based on NUMA balancing page table scanning and hint page fault
as follows,
- When the page tables of the processes are scanned to change PTE/PMD
to be PROT_NONE, the current time is recorded in struct page as scan
time.
- When the page is accessed, hint page fault will occur. The scan
time is gotten from the struct page. And The hint page fault
latency is defined as
hint page fault time - scan time
The shorter the hint page fault latency of a page is, the higher the
probability of their access frequency to be higher. So the hint page
fault latency is a better estimation of the page hot/cold.
It's hard to find some extra space in struct page to hold the scan time.
Fortunately, we can reuse some bits used by the original NUMA balancing.
NUMA balancing uses some bits in struct page to store the page accessing
CPU and PID (referring to page_cpupid_xchg_last()). Which is used by the
multi-stage node selection algorithm to avoid to migrate pages shared
accessed by the NUMA nodes back and forth. But for pages in the slow
memory node, even if they are shared accessed by multiple NUMA nodes, as
long as the pages are hot, they need to be promoted to the fast memory
node. So the accessing CPU and PID information are unnecessary for the
slow memory pages. We can reuse these bits in struct page to record the
scan time. For the fast memory pages, these bits are used as before.
For the hot threshold, the default value is 1 second, which works well in
our performance test. All pages with hint page fault latency < hot
threshold will be considered hot.
It's hard for users to determine the hot threshold. So we don't provide a
kernel ABI to set it, just provide a debugfs interface for advanced users
to experiment. We will continue to work on a hot threshold automatic
adjustment mechanism.
The downside of the above method is that the response time to the workload
hot spot changing may be much longer. For example,
- A previous cold memory area becomes hot
- The hint page fault will be triggered. But the hint page fault
latency isn't shorter than the hot threshold. So the pages will
not be promoted.
- When the memory area is scanned again, maybe after a scan period,
the hint page fault latency measured will be shorter than the hot
threshold and the pages will be promoted.
To mitigate this, if there are enough free space in the fast memory node,
the hot threshold will not be used, all pages will be promoted upon the
hint page fault for fast response.
Thanks Zhong Jiang reported and tested the fix for a bug when disabling
memory tiering mode dynamically.
Link: https://lkml.kernel.org/r/20220713083954.34196-1-ying.huang@intel.com
Link: https://lkml.kernel.org/r/20220713083954.34196-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Wei Xu <weixugc@google.com>
Cc: osalvador <osalvador@suse.de>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Zhong Jiang <zhongjiang-ali@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-07-13 11:39:51 +03:00
/*
* For memory tiering mode , cpupid of slow memory page is used
* to record page access time . So use default value .
*/
if ( node_is_toptier ( page_nid ) )
last_cpupid = page_cpupid_last ( page ) ;
2021-07-01 04:51:42 +03:00
target_nid = numa_migrate_prep ( page , vma , haddr , page_nid ,
& flags ) ;
if ( target_nid = = NUMA_NO_NODE ) {
put_page ( page ) ;
goto out_map ;
}
2016-12-15 02:06:58 +03:00
spin_unlock ( vmf - > ptl ) ;
2022-11-08 20:46:50 +03:00
writable = false ;
mm, locking: Rework {set,clear,mm}_tlb_flush_pending()
Commit:
af2c1401e6f9 ("mm: numa: guarantee that tlb_flush_pending updates are visible before page table updates")
added smp_mb__before_spinlock() to set_tlb_flush_pending(). I think we
can solve the same problem without this barrier.
If instead we mandate that mm_tlb_flush_pending() is used while
holding the PTL we're guaranteed to observe prior
set_tlb_flush_pending() instances.
For this to work we need to rework migrate_misplaced_transhuge_page()
a little and move the test up into do_huge_pmd_numa_page().
NOTE: this relies on flush_tlb_range() to guarantee:
(1) it ensures that prior page table updates are visible to the
page table walker and
(2) it ensures that subsequent memory accesses are only made
visible after the invalidation has completed
This is required for architectures that implement TRANSPARENT_HUGEPAGE
(arc, arm, arm64, mips, powerpc, s390, sparc, x86) or otherwise use
mm_tlb_flush_pending() in their page-table operations (arm, arm64,
x86).
This appears true for:
- arm (DSB ISB before and after),
- arm64 (DSB ISHST before, and DSB ISH after),
- powerpc (PTESYNC before and after),
- s390 and x86 TLB invalidate are serializing instructions
But I failed to understand the situation for:
- arc, mips, sparc
Now SPARC64 is a wee bit special in that flush_tlb_range() is a no-op
and it flushes the TLBs using arch_{enter,leave}_lazy_mmu_mode()
inside the PTL. It still needs to guarantee the PTL unlock happens
_after_ the invalidate completes.
Vineet, Ralf and Dave could you guys please have a look?
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Will Deacon <will.deacon@arm.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Russell King <linux@armlinux.org.uk>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vineet Gupta <vgupta@synopsys.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-07 19:05:07 +03:00
2021-07-01 04:51:42 +03:00
migrated = migrate_misplaced_page ( page , vma , target_nid ) ;
2013-10-07 14:29:24 +04:00
if ( migrated ) {
flags | = TNF_MIGRATED ;
mm: numa: Sanitize task_numa_fault() callsites
There are three callers of task_numa_fault():
- do_huge_pmd_numa_page():
Accounts against the current node, not the node where the
page resides, unless we migrated, in which case it accounts
against the node we migrated to.
- do_numa_page():
Accounts against the current node, not the node where the
page resides, unless we migrated, in which case it accounts
against the node we migrated to.
- do_pmd_numa_page():
Accounts not at all when the page isn't migrated, otherwise
accounts against the node we migrated towards.
This seems wrong to me; all three sites should have the same
sementaics, furthermore we should accounts against where the page
really is, we already know where the task is.
So modify all three sites to always account; we did after all receive
the fault; and always account to where the page is after migration,
regardless of success.
They all still differ on when they clear the PTE/PMD; ideally that
would get sorted too.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1381141781-10992-8-git-send-email-mgorman@suse.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 14:28:45 +04:00
page_nid = target_nid ;
2021-07-01 04:51:42 +03:00
} else {
2015-03-26 01:55:42 +03:00
flags | = TNF_MIGRATE_FAIL ;
2021-07-01 04:51:42 +03:00
vmf - > ptl = pmd_lock ( vma - > vm_mm , vmf - > pmd ) ;
if ( unlikely ( ! pmd_same ( oldpmd , * vmf - > pmd ) ) ) {
spin_unlock ( vmf - > ptl ) ;
goto out ;
}
goto out_map ;
}
2013-10-07 14:28:44 +04:00
out :
2019-03-06 02:42:58 +03:00
if ( page_nid ! = NUMA_NO_NODE )
2016-12-15 02:06:58 +03:00
task_numa_fault ( last_cpupid , page_nid , HPAGE_PMD_NR ,
2017-02-25 01:59:56 +03:00
flags ) ;
mm: numa: Sanitize task_numa_fault() callsites
There are three callers of task_numa_fault():
- do_huge_pmd_numa_page():
Accounts against the current node, not the node where the
page resides, unless we migrated, in which case it accounts
against the node we migrated to.
- do_numa_page():
Accounts against the current node, not the node where the
page resides, unless we migrated, in which case it accounts
against the node we migrated to.
- do_pmd_numa_page():
Accounts not at all when the page isn't migrated, otherwise
accounts against the node we migrated towards.
This seems wrong to me; all three sites should have the same
sementaics, furthermore we should accounts against where the page
really is, we already know where the task is.
So modify all three sites to always account; we did after all receive
the fault; and always account to where the page is after migration,
regardless of success.
They all still differ on when they clear the PTE/PMD; ideally that
would get sorted too.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1381141781-10992-8-git-send-email-mgorman@suse.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-10-07 14:28:45 +04:00
2012-10-25 16:16:31 +04:00
return 0 ;
2021-07-01 04:51:42 +03:00
out_map :
/* Restore the PMD */
pmd = pmd_modify ( oldpmd , vma - > vm_page_prot ) ;
pmd = pmd_mkyoung ( pmd ) ;
2022-11-08 20:46:50 +03:00
if ( writable )
2021-07-01 04:51:42 +03:00
pmd = pmd_mkwrite ( pmd ) ;
set_pmd_at ( vma - > vm_mm , haddr , vmf - > pmd , pmd ) ;
update_mmu_cache_pmd ( vma , vmf - > address , vmf - > pmd ) ;
spin_unlock ( vmf - > ptl ) ;
goto out ;
2012-10-25 16:16:31 +04:00
}
2016-07-29 01:48:03 +03:00
/*
* Return true if we do MADV_FREE successfully on entire pmd page .
* Otherwise , return false .
*/
bool madvise_free_huge_pmd ( struct mmu_gather * tlb , struct vm_area_struct * vma ,
mm/huge_memory.c: don't split THP page when MADV_FREE syscall is called
We don't need to split THP page when MADV_FREE syscall is called if
[start, len] is aligned with THP size. The split could be done when VM
decide to free it in reclaim path if memory pressure is heavy. With
that, we could avoid unnecessary THP split.
For the feature, this patch changes pte dirtness marking logic of THP.
Now, it marks every ptes of pages dirty unconditionally in splitting,
which makes MADV_FREE void. So, instead, this patch propagates pmd
dirtiness to all pages via PG_dirty and restores pte dirtiness from
PG_dirty. With this, if pmd is clean(ie, MADV_FREEed) when split
happens(e,g, shrink_page_list), all of pages are clean too so we could
discard them.
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Shaohua Li <shli@kernel.org>
Cc: <yalin.wang2010@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Daniel Micay <danielmicay@gmail.com>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Helge Deller <deller@gmx.de>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: Jason Evans <je@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttil <mika.penttila@nextfour.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: Roland Dreier <roland@kernel.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Shaohua Li <shli@kernel.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-16 03:55:42 +03:00
pmd_t * pmd , unsigned long addr , unsigned long next )
{
spinlock_t * ptl ;
pmd_t orig_pmd ;
struct page * page ;
struct mm_struct * mm = tlb - > mm ;
2016-07-29 01:48:03 +03:00
bool ret = false ;
mm/huge_memory.c: don't split THP page when MADV_FREE syscall is called
We don't need to split THP page when MADV_FREE syscall is called if
[start, len] is aligned with THP size. The split could be done when VM
decide to free it in reclaim path if memory pressure is heavy. With
that, we could avoid unnecessary THP split.
For the feature, this patch changes pte dirtness marking logic of THP.
Now, it marks every ptes of pages dirty unconditionally in splitting,
which makes MADV_FREE void. So, instead, this patch propagates pmd
dirtiness to all pages via PG_dirty and restores pte dirtiness from
PG_dirty. With this, if pmd is clean(ie, MADV_FREEed) when split
happens(e,g, shrink_page_list), all of pages are clean too so we could
discard them.
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Shaohua Li <shli@kernel.org>
Cc: <yalin.wang2010@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Daniel Micay <danielmicay@gmail.com>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Helge Deller <deller@gmx.de>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: Jason Evans <je@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttil <mika.penttila@nextfour.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: Roland Dreier <roland@kernel.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Shaohua Li <shli@kernel.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-16 03:55:42 +03:00
2018-08-31 15:46:08 +03:00
tlb_change_page_size ( tlb , HPAGE_PMD_SIZE ) ;
2016-12-13 03:42:40 +03:00
2016-01-22 03:40:25 +03:00
ptl = pmd_trans_huge_lock ( pmd , vma ) ;
if ( ! ptl )
2016-01-18 05:33:15 +03:00
goto out_unlocked ;
mm/huge_memory.c: don't split THP page when MADV_FREE syscall is called
We don't need to split THP page when MADV_FREE syscall is called if
[start, len] is aligned with THP size. The split could be done when VM
decide to free it in reclaim path if memory pressure is heavy. With
that, we could avoid unnecessary THP split.
For the feature, this patch changes pte dirtness marking logic of THP.
Now, it marks every ptes of pages dirty unconditionally in splitting,
which makes MADV_FREE void. So, instead, this patch propagates pmd
dirtiness to all pages via PG_dirty and restores pte dirtiness from
PG_dirty. With this, if pmd is clean(ie, MADV_FREEed) when split
happens(e,g, shrink_page_list), all of pages are clean too so we could
discard them.
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Shaohua Li <shli@kernel.org>
Cc: <yalin.wang2010@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Daniel Micay <danielmicay@gmail.com>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Helge Deller <deller@gmx.de>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: Jason Evans <je@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttil <mika.penttila@nextfour.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: Roland Dreier <roland@kernel.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Shaohua Li <shli@kernel.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-16 03:55:42 +03:00
orig_pmd = * pmd ;
2016-07-29 01:48:03 +03:00
if ( is_huge_zero_pmd ( orig_pmd ) )
mm/huge_memory.c: don't split THP page when MADV_FREE syscall is called
We don't need to split THP page when MADV_FREE syscall is called if
[start, len] is aligned with THP size. The split could be done when VM
decide to free it in reclaim path if memory pressure is heavy. With
that, we could avoid unnecessary THP split.
For the feature, this patch changes pte dirtness marking logic of THP.
Now, it marks every ptes of pages dirty unconditionally in splitting,
which makes MADV_FREE void. So, instead, this patch propagates pmd
dirtiness to all pages via PG_dirty and restores pte dirtiness from
PG_dirty. With this, if pmd is clean(ie, MADV_FREEed) when split
happens(e,g, shrink_page_list), all of pages are clean too so we could
discard them.
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Shaohua Li <shli@kernel.org>
Cc: <yalin.wang2010@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Daniel Micay <danielmicay@gmail.com>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Helge Deller <deller@gmx.de>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: Jason Evans <je@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttil <mika.penttila@nextfour.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: Roland Dreier <roland@kernel.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Shaohua Li <shli@kernel.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-16 03:55:42 +03:00
goto out ;
mm: thp: check pmd migration entry in common path
When THP migration is being used, memory management code needs to handle
pmd migration entries properly. This patch uses !pmd_present() or
is_swap_pmd() (depending on whether pmd_none() needs separate code or
not) to check pmd migration entries at the places where a pmd entry is
present.
Since pmd-related code uses split_huge_page(), split_huge_pmd(),
pmd_trans_huge(), pmd_trans_unstable(), or
pmd_none_or_trans_huge_or_clear_bad(), this patch:
1. adds pmd migration entry split code in split_huge_pmd(),
2. takes care of pmd migration entries whenever pmd_trans_huge() is present,
3. makes pmd_none_or_trans_huge_or_clear_bad() pmd migration entry aware.
Since split_huge_page() uses split_huge_pmd() and pmd_trans_unstable()
is equivalent to pmd_none_or_trans_huge_or_clear_bad(), we do not change
them.
Until this commit, a pmd entry should be:
1. pointing to a pte page,
2. is_swap_pmd(),
3. pmd_trans_huge(),
4. pmd_devmap(), or
5. pmd_none().
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-09 02:11:01 +03:00
if ( unlikely ( ! pmd_present ( orig_pmd ) ) ) {
VM_BUG_ON ( thp_migration_supported ( ) & &
! is_pmd_migration_entry ( orig_pmd ) ) ;
goto out ;
}
mm/huge_memory.c: don't split THP page when MADV_FREE syscall is called
We don't need to split THP page when MADV_FREE syscall is called if
[start, len] is aligned with THP size. The split could be done when VM
decide to free it in reclaim path if memory pressure is heavy. With
that, we could avoid unnecessary THP split.
For the feature, this patch changes pte dirtness marking logic of THP.
Now, it marks every ptes of pages dirty unconditionally in splitting,
which makes MADV_FREE void. So, instead, this patch propagates pmd
dirtiness to all pages via PG_dirty and restores pte dirtiness from
PG_dirty. With this, if pmd is clean(ie, MADV_FREEed) when split
happens(e,g, shrink_page_list), all of pages are clean too so we could
discard them.
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Shaohua Li <shli@kernel.org>
Cc: <yalin.wang2010@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Daniel Micay <danielmicay@gmail.com>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Helge Deller <deller@gmx.de>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: Jason Evans <je@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttil <mika.penttila@nextfour.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: Roland Dreier <roland@kernel.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Shaohua Li <shli@kernel.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-16 03:55:42 +03:00
page = pmd_page ( orig_pmd ) ;
/*
* If other processes are mapping this page , we couldn ' t discard
* the page unless they all do MADV_FREE so let ' s skip the page .
*/
2021-07-01 04:47:57 +03:00
if ( total_mapcount ( page ) ! = 1 )
mm/huge_memory.c: don't split THP page when MADV_FREE syscall is called
We don't need to split THP page when MADV_FREE syscall is called if
[start, len] is aligned with THP size. The split could be done when VM
decide to free it in reclaim path if memory pressure is heavy. With
that, we could avoid unnecessary THP split.
For the feature, this patch changes pte dirtness marking logic of THP.
Now, it marks every ptes of pages dirty unconditionally in splitting,
which makes MADV_FREE void. So, instead, this patch propagates pmd
dirtiness to all pages via PG_dirty and restores pte dirtiness from
PG_dirty. With this, if pmd is clean(ie, MADV_FREEed) when split
happens(e,g, shrink_page_list), all of pages are clean too so we could
discard them.
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Shaohua Li <shli@kernel.org>
Cc: <yalin.wang2010@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Daniel Micay <danielmicay@gmail.com>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Helge Deller <deller@gmx.de>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: Jason Evans <je@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttil <mika.penttila@nextfour.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: Roland Dreier <roland@kernel.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Shaohua Li <shli@kernel.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-16 03:55:42 +03:00
goto out ;
if ( ! trylock_page ( page ) )
goto out ;
/*
* If user want to discard part - pages of THP , split it so MADV_FREE
* will deactivate only them .
*/
if ( next - addr ! = HPAGE_PMD_SIZE ) {
get_page ( page ) ;
spin_unlock ( ptl ) ;
2016-07-14 22:07:12 +03:00
split_huge_page ( page ) ;
mm/huge_memory.c: don't split THP page when MADV_FREE syscall is called
We don't need to split THP page when MADV_FREE syscall is called if
[start, len] is aligned with THP size. The split could be done when VM
decide to free it in reclaim path if memory pressure is heavy. With
that, we could avoid unnecessary THP split.
For the feature, this patch changes pte dirtness marking logic of THP.
Now, it marks every ptes of pages dirty unconditionally in splitting,
which makes MADV_FREE void. So, instead, this patch propagates pmd
dirtiness to all pages via PG_dirty and restores pte dirtiness from
PG_dirty. With this, if pmd is clean(ie, MADV_FREEed) when split
happens(e,g, shrink_page_list), all of pages are clean too so we could
discard them.
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Shaohua Li <shli@kernel.org>
Cc: <yalin.wang2010@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Daniel Micay <danielmicay@gmail.com>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Helge Deller <deller@gmx.de>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: Jason Evans <je@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttil <mika.penttila@nextfour.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: Roland Dreier <roland@kernel.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Shaohua Li <shli@kernel.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-16 03:55:42 +03:00
unlock_page ( page ) ;
2017-07-07 01:35:28 +03:00
put_page ( page ) ;
mm/huge_memory.c: don't split THP page when MADV_FREE syscall is called
We don't need to split THP page when MADV_FREE syscall is called if
[start, len] is aligned with THP size. The split could be done when VM
decide to free it in reclaim path if memory pressure is heavy. With
that, we could avoid unnecessary THP split.
For the feature, this patch changes pte dirtness marking logic of THP.
Now, it marks every ptes of pages dirty unconditionally in splitting,
which makes MADV_FREE void. So, instead, this patch propagates pmd
dirtiness to all pages via PG_dirty and restores pte dirtiness from
PG_dirty. With this, if pmd is clean(ie, MADV_FREEed) when split
happens(e,g, shrink_page_list), all of pages are clean too so we could
discard them.
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Shaohua Li <shli@kernel.org>
Cc: <yalin.wang2010@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Daniel Micay <danielmicay@gmail.com>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Helge Deller <deller@gmx.de>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: Jason Evans <je@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttil <mika.penttila@nextfour.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: Roland Dreier <roland@kernel.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Shaohua Li <shli@kernel.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-16 03:55:42 +03:00
goto out_unlocked ;
}
if ( PageDirty ( page ) )
ClearPageDirty ( page ) ;
unlock_page ( page ) ;
if ( pmd_young ( orig_pmd ) | | pmd_dirty ( orig_pmd ) ) {
2017-04-14 00:56:26 +03:00
pmdp_invalidate ( vma , addr , pmd ) ;
mm/huge_memory.c: don't split THP page when MADV_FREE syscall is called
We don't need to split THP page when MADV_FREE syscall is called if
[start, len] is aligned with THP size. The split could be done when VM
decide to free it in reclaim path if memory pressure is heavy. With
that, we could avoid unnecessary THP split.
For the feature, this patch changes pte dirtness marking logic of THP.
Now, it marks every ptes of pages dirty unconditionally in splitting,
which makes MADV_FREE void. So, instead, this patch propagates pmd
dirtiness to all pages via PG_dirty and restores pte dirtiness from
PG_dirty. With this, if pmd is clean(ie, MADV_FREEed) when split
happens(e,g, shrink_page_list), all of pages are clean too so we could
discard them.
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Shaohua Li <shli@kernel.org>
Cc: <yalin.wang2010@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Daniel Micay <danielmicay@gmail.com>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Helge Deller <deller@gmx.de>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: Jason Evans <je@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttil <mika.penttila@nextfour.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: Roland Dreier <roland@kernel.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Shaohua Li <shli@kernel.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-16 03:55:42 +03:00
orig_pmd = pmd_mkold ( orig_pmd ) ;
orig_pmd = pmd_mkclean ( orig_pmd ) ;
set_pmd_at ( mm , addr , pmd , orig_pmd ) ;
tlb_remove_pmd_tlb_entry ( tlb , pmd , addr ) ;
}
2017-05-04 00:52:32 +03:00
mark_page_lazyfree ( page ) ;
2016-07-29 01:48:03 +03:00
ret = true ;
mm/huge_memory.c: don't split THP page when MADV_FREE syscall is called
We don't need to split THP page when MADV_FREE syscall is called if
[start, len] is aligned with THP size. The split could be done when VM
decide to free it in reclaim path if memory pressure is heavy. With
that, we could avoid unnecessary THP split.
For the feature, this patch changes pte dirtness marking logic of THP.
Now, it marks every ptes of pages dirty unconditionally in splitting,
which makes MADV_FREE void. So, instead, this patch propagates pmd
dirtiness to all pages via PG_dirty and restores pte dirtiness from
PG_dirty. With this, if pmd is clean(ie, MADV_FREEed) when split
happens(e,g, shrink_page_list), all of pages are clean too so we could
discard them.
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Shaohua Li <shli@kernel.org>
Cc: <yalin.wang2010@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Daniel Micay <danielmicay@gmail.com>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Helge Deller <deller@gmx.de>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: Jason Evans <je@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttil <mika.penttila@nextfour.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: Roland Dreier <roland@kernel.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Shaohua Li <shli@kernel.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-16 03:55:42 +03:00
out :
spin_unlock ( ptl ) ;
out_unlocked :
return ret ;
}
2016-12-13 03:44:32 +03:00
static inline void zap_deposited_table ( struct mm_struct * mm , pmd_t * pmd )
{
pgtable_t pgtable ;
pgtable = pgtable_trans_huge_withdraw ( mm , pmd ) ;
pte_free ( mm , pgtable ) ;
2017-11-16 04:35:37 +03:00
mm_dec_nr_ptes ( mm ) ;
2016-12-13 03:44:32 +03:00
}
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
int zap_huge_pmd ( struct mmu_gather * tlb , struct vm_area_struct * vma ,
2012-01-13 05:19:16 +04:00
pmd_t * pmd , unsigned long addr )
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
{
2015-09-09 00:59:31 +03:00
pmd_t orig_pmd ;
2013-11-15 02:30:54 +04:00
spinlock_t * ptl ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
2018-08-31 15:46:08 +03:00
tlb_change_page_size ( tlb , HPAGE_PMD_SIZE ) ;
2016-12-13 03:42:40 +03:00
2016-01-22 03:40:25 +03:00
ptl = __pmd_trans_huge_lock ( pmd , vma ) ;
if ( ! ptl )
2015-09-09 00:59:31 +03:00
return 0 ;
/*
* For architectures like ppc64 we look at deposited pgtable
* when calling pmdp_huge_get_and_clear . So do the
* pgtable_trans_huge_withdraw after finishing pmdp related
* operations .
*/
2020-05-05 10:17:28 +03:00
orig_pmd = pmdp_huge_get_and_clear_full ( vma , addr , pmd ,
tlb - > fullmm ) ;
2015-09-09 00:59:31 +03:00
tlb_remove_pmd_tlb_entry ( tlb , pmd , addr ) ;
2020-03-24 20:47:17 +03:00
if ( vma_is_special_huge ( vma ) ) {
2017-05-09 01:59:43 +03:00
if ( arch_needs_pgtable_deposit ( ) )
zap_deposited_table ( tlb - > mm , pmd ) ;
2015-09-09 00:59:31 +03:00
spin_unlock ( ptl ) ;
} else if ( is_huge_zero_pmd ( orig_pmd ) ) {
2017-05-09 01:59:40 +03:00
zap_deposited_table ( tlb - > mm , pmd ) ;
2015-09-09 00:59:31 +03:00
spin_unlock ( ptl ) ;
} else {
2017-09-09 02:10:57 +03:00
struct page * page = NULL ;
int flush_needed = 1 ;
if ( pmd_present ( orig_pmd ) ) {
page = pmd_page ( orig_pmd ) ;
mm/munlock: rmap call mlock_vma_page() munlock_vma_page()
Add vma argument to mlock_vma_page() and munlock_vma_page(), make them
inline functions which check (vma->vm_flags & VM_LOCKED) before calling
mlock_page() and munlock_page() in mm/mlock.c.
Add bool compound to mlock_vma_page() and munlock_vma_page(): this is
because we have understandable difficulty in accounting pte maps of THPs,
and if passed a PageHead page, mlock_page() and munlock_page() cannot
tell whether it's a pmd map to be counted or a pte map to be ignored.
Add vma arg to page_add_file_rmap() and page_remove_rmap(), like the
others, and use that to call mlock_vma_page() at the end of the page
adds, and munlock_vma_page() at the end of page_remove_rmap() (end or
beginning? unimportant, but end was easier for assertions in testing).
No page lock is required (although almost all adds happen to hold it):
delete the "Serialize with page migration" BUG_ON(!PageLocked(page))s.
Certainly page lock did serialize with page migration, but I'm having
difficulty explaining why that was ever important.
Mlock accounting on THPs has been hard to define, differed between anon
and file, involved PageDoubleMap in some places and not others, required
clear_page_mlock() at some points. Keep it simple now: just count the
pmds and ignore the ptes, there is no reason for ptes to undo pmd mlocks.
page_add_new_anon_rmap() callers unchanged: they have long been calling
lru_cache_add_inactive_or_unevictable(), which does its own VM_LOCKED
handling (it also checks for not VM_SPECIAL: I think that's overcautious,
and inconsistent with other checks, that mmap_region() already prevents
VM_LOCKED on VM_SPECIAL; but haven't quite convinced myself to change it).
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
2022-02-15 05:26:39 +03:00
page_remove_rmap ( page , vma , true ) ;
2017-09-09 02:10:57 +03:00
VM_BUG_ON_PAGE ( page_mapcount ( page ) < 0 , page ) ;
VM_BUG_ON_PAGE ( ! PageHead ( page ) , page ) ;
} else if ( thp_migration_supported ( ) ) {
swp_entry_t entry ;
VM_BUG_ON ( ! is_pmd_migration_entry ( orig_pmd ) ) ;
entry = pmd_to_swp_entry ( orig_pmd ) ;
2021-07-01 04:54:06 +03:00
page = pfn_swap_entry_to_page ( entry ) ;
2017-09-09 02:10:57 +03:00
flush_needed = 0 ;
} else
WARN_ONCE ( 1 , " Non present huge pmd without pmd migration enabled! " ) ;
2016-07-27 01:25:34 +03:00
if ( PageAnon ( page ) ) {
2017-05-09 01:59:40 +03:00
zap_deposited_table ( tlb - > mm , pmd ) ;
2016-07-27 01:25:34 +03:00
add_mm_counter ( tlb - > mm , MM_ANONPAGES , - HPAGE_PMD_NR ) ;
} else {
2016-12-13 03:44:32 +03:00
if ( arch_needs_pgtable_deposit ( ) )
zap_deposited_table ( tlb - > mm , pmd ) ;
2018-08-18 01:44:55 +03:00
add_mm_counter ( tlb - > mm , mm_counter_file ( page ) , - HPAGE_PMD_NR ) ;
2016-07-27 01:25:34 +03:00
}
2017-09-09 02:10:57 +03:00
2015-09-09 00:59:31 +03:00
spin_unlock ( ptl ) ;
2017-09-09 02:10:57 +03:00
if ( flush_needed )
tlb_remove_page_size ( tlb , page , HPAGE_PMD_SIZE ) ;
2012-03-22 03:33:57 +04:00
}
2015-09-09 00:59:31 +03:00
return 1 ;
thp: transparent hugepage core
Lately I've been working to make KVM use hugepages transparently without
the usual restrictions of hugetlbfs. Some of the restrictions I'd like to
see removed:
1) hugepages have to be swappable or the guest physical memory remains
locked in RAM and can't be paged out to swap
2) if a hugepage allocation fails, regular pages should be allocated
instead and mixed in the same vma without any failure and without
userland noticing
3) if some task quits and more hugepages become available in the
buddy, guest physical memory backed by regular pages should be
relocated on hugepages automatically in regions under
madvise(MADV_HUGEPAGE) (ideally event driven by waking up the
kernel deamon if the order=HPAGE_PMD_SHIFT-PAGE_SHIFT list becomes
not null)
4) avoidance of reservation and maximization of use of hugepages whenever
possible. Reservation (needed to avoid runtime fatal faliures) may be ok for
1 machine with 1 database with 1 database cache with 1 database cache size
known at boot time. It's definitely not feasible with a virtualization
hypervisor usage like RHEV-H that runs an unknown number of virtual machines
with an unknown size of each virtual machine with an unknown amount of
pagecache that could be potentially useful in the host for guest not using
O_DIRECT (aka cache=off).
hugepages in the virtualization hypervisor (and also in the guest!) are
much more important than in a regular host not using virtualization,
becasue with NPT/EPT they decrease the tlb-miss cacheline accesses from 24
to 19 in case only the hypervisor uses transparent hugepages, and they
decrease the tlb-miss cacheline accesses from 19 to 15 in case both the
linux hypervisor and the linux guest both uses this patch (though the
guest will limit the addition speedup to anonymous regions only for
now...). Even more important is that the tlb miss handler is much slower
on a NPT/EPT guest than for a regular shadow paging or no-virtualization
scenario. So maximizing the amount of virtual memory cached by the TLB
pays off significantly more with NPT/EPT than without (even if there would
be no significant speedup in the tlb-miss runtime).
The first (and more tedious) part of this work requires allowing the VM to
handle anonymous hugepages mixed with regular pages transparently on
regular anonymous vmas. This is what this patch tries to achieve in the
least intrusive possible way. We want hugepages and hugetlb to be used in
a way so that all applications can benefit without changes (as usual we
leverage the KVM virtualization design: by improving the Linux VM at
large, KVM gets the performance boost too).
The most important design choice is: always fallback to 4k allocation if
the hugepage allocation fails! This is the _very_ opposite of some large
pagecache patches that failed with -EIO back then if a 64k (or similar)
allocation failed...
Second important decision (to reduce the impact of the feature on the
existing pagetable handling code) is that at any time we can split an
hugepage into 512 regular pages and it has to be done with an operation
that can't fail. This way the reliability of the swapping isn't decreased
(no need to allocate memory when we are short on memory to swap) and it's
trivial to plug a split_huge_page* one-liner where needed without
polluting the VM. Over time we can teach mprotect, mremap and friends to
handle pmd_trans_huge natively without calling split_huge_page*. The fact
it can't fail isn't just for swap: if split_huge_page would return -ENOMEM
(instead of the current void) we'd need to rollback the mprotect from the
middle of it (ideally including undoing the split_vma) which would be a
big change and in the very wrong direction (it'd likely be simpler not to
call split_huge_page at all and to teach mprotect and friends to handle
hugepages instead of rolling them back from the middle). In short the
very value of split_huge_page is that it can't fail.
The collapsing and madvise(MADV_HUGEPAGE) part will remain separated and
incremental and it'll just be an "harmless" addition later if this initial
part is agreed upon. It also should be noted that locking-wise replacing
regular pages with hugepages is going to be very easy if compared to what
I'm doing below in split_huge_page, as it will only happen when
page_count(page) matches page_mapcount(page) if we can take the PG_lock
and mmap_sem in write mode. collapse_huge_page will be a "best effort"
that (unlike split_huge_page) can fail at the minimal sign of trouble and
we can try again later. collapse_huge_page will be similar to how KSM
works and the madvise(MADV_HUGEPAGE) will work similar to
madvise(MADV_MERGEABLE).
The default I like is that transparent hugepages are used at page fault
time. This can be changed with
/sys/kernel/mm/transparent_hugepage/enabled. The control knob can be set
to three values "always", "madvise", "never" which mean respectively that
hugepages are always used, or only inside madvise(MADV_HUGEPAGE) regions,
or never used. /sys/kernel/mm/transparent_hugepage/defrag instead
controls if the hugepage allocation should defrag memory aggressively
"always", only inside "madvise" regions, or "never".
The pmd_trans_splitting/pmd_trans_huge locking is very solid. The
put_page (from get_user_page users that can't use mmu notifier like
O_DIRECT) that runs against a __split_huge_page_refcount instead was a
pain to serialize in a way that would result always in a coherent page
count for both tail and head. I think my locking solution with a
compound_lock taken only after the page_first is valid and is still a
PageHead should be safe but it surely needs review from SMP race point of
view. In short there is no current existing way to serialize the O_DIRECT
final put_page against split_huge_page_refcount so I had to invent a new
one (O_DIRECT loses knowledge on the mapping status by the time gup_fast
returns so...). And I didn't want to impact all gup/gup_fast users for
now, maybe if we change the gup interface substantially we can avoid this
locking, I admit I didn't think too much about it because changing the gup
unpinning interface would be invasive.
If we ignored O_DIRECT we could stick to the existing compound refcounting
code, by simply adding a get_user_pages_fast_flags(foll_flags) where KVM
(and any other mmu notifier user) would call it without FOLL_GET (and if
FOLL_GET isn't set we'd just BUG_ON if nobody registered itself in the
current task mmu notifier list yet). But O_DIRECT is fundamental for
decent performance of virtualized I/O on fast storage so we can't avoid it
to solve the race of put_page against split_huge_page_refcount to achieve
a complete hugepage feature for KVM.
Swap and oom works fine (well just like with regular pages ;). MMU
notifier is handled transparently too, with the exception of the young bit
on the pmd, that didn't have a range check but I think KVM will be fine
because the whole point of hugepages is that EPT/NPT will also use a huge
pmd when they notice gup returns pages with PageCompound set, so they
won't care of a range and there's just the pmd young bit to check in that
case.
NOTE: in some cases if the L2 cache is small, this may slowdown and waste
memory during COWs because 4M of memory are accessed in a single fault
instead of 8k (the payoff is that after COW the program can run faster).
So we might want to switch the copy_huge_page (and clear_huge_page too) to
not temporal stores. I also extensively researched ways to avoid this
cache trashing with a full prefault logic that would cow in 8k/16k/32k/64k
up to 1M (I can send those patches that fully implemented prefault) but I
concluded they're not worth it and they add an huge additional complexity
and they remove all tlb benefits until the full hugepage has been faulted
in, to save a little bit of memory and some cache during app startup, but
they still don't improve substantially the cache-trashing during startup
if the prefault happens in >4k chunks. One reason is that those 4k pte
entries copied are still mapped on a perfectly cache-colored hugepage, so
the trashing is the worst one can generate in those copies (cow of 4k page
copies aren't so well colored so they trashes less, but again this results
in software running faster after the page fault). Those prefault patches
allowed things like a pte where post-cow pages were local 4k regular anon
pages and the not-yet-cowed pte entries were pointing in the middle of
some hugepage mapped read-only. If it doesn't payoff substantially with
todays hardware it will payoff even less in the future with larger l2
caches, and the prefault logic would blot the VM a lot. If one is
emebdded transparent_hugepage can be disabled during boot with sysfs or
with the boot commandline parameter transparent_hugepage=0 (or
transparent_hugepage=2 to restrict hugepages inside madvise regions) that
will ensure not a single hugepage is allocated at boot time. It is simple
enough to just disable transparent hugepage globally and let transparent
hugepages be allocated selectively by applications in the MADV_HUGEPAGE
region (both at page fault time, and if enabled with the
collapse_huge_page too through the kernel daemon).
This patch supports only hugepages mapped in the pmd, archs that have
smaller hugepages will not fit in this patch alone. Also some archs like
power have certain tlb limits that prevents mixing different page size in
the same regions so they will not fit in this framework that requires
"graceful fallback" to basic PAGE_SIZE in case of physical memory
fragmentation. hugetlbfs remains a perfect fit for those because its
software limits happen to match the hardware limits. hugetlbfs also
remains a perfect fit for hugepage sizes like 1GByte that cannot be hoped
to be found not fragmented after a certain system uptime and that would be
very expensive to defragment with relocation, so requiring reservation.
hugetlbfs is the "reservation way", the point of transparent hugepages is
not to have any reservation at all and maximizing the use of cache and
hugepages at all times automatically.
Some performance result:
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largep
ages3
memset page fault 1566023
memset tlb miss 453854
memset second tlb miss 453321
random access tlb miss 41635
random access second tlb miss 41658
vmx andrea # LD_PRELOAD=/usr/lib64/libhugetlbfs.so HUGETLB_MORECORE=yes HUGETLB_PATH=/mnt/huge/ ./largepages3
memset page fault 1566471
memset tlb miss 453375
memset second tlb miss 453320
random access tlb miss 41636
random access second tlb miss 41637
vmx andrea # ./largepages3
memset page fault 1566642
memset tlb miss 453417
memset second tlb miss 453313
random access tlb miss 41630
random access second tlb miss 41647
vmx andrea # ./largepages3
memset page fault 1566872
memset tlb miss 453418
memset second tlb miss 453315
random access tlb miss 41618
random access second tlb miss 41659
vmx andrea # echo 0 > /proc/sys/vm/transparent_hugepage
vmx andrea # ./largepages3
memset page fault 2182476
memset tlb miss 460305
memset second tlb miss 460179
random access tlb miss 44483
random access second tlb miss 44186
vmx andrea # ./largepages3
memset page fault 2182791
memset tlb miss 460742
memset second tlb miss 459962
random access tlb miss 43981
random access second tlb miss 43988
============
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (3UL*1024*1024*1024)
int main()
{
char *p = malloc(SIZE), *p2;
struct timeval before, after;
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset page fault %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
memset(p, 0, SIZE);
gettimeofday(&after, NULL);
printf("memset second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
gettimeofday(&before, NULL);
for (p2 = p; p2 < p+SIZE; p2 += 4096)
*p2 = 0;
gettimeofday(&after, NULL);
printf("random access second tlb miss %Lu\n",
(after.tv_sec-before.tv_sec)*1000000UL +
after.tv_usec-before.tv_usec);
return 0;
}
============
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-14 02:46:52 +03:00
}
2016-12-13 03:44:29 +03:00
# ifndef pmd_move_must_withdraw
static inline int pmd_move_must_withdraw ( spinlock_t * new_pmd_ptl ,
spinlock_t * old_pmd_ptl ,
struct vm_area_struct * vma )
{
/*
* With split pmd lock we also need to move preallocated
* PTE page table if new_pmd is on different PMD page table .
*
* We also don ' t deposit and withdraw tables for file pages .
*/
return ( new_pmd_ptl ! = old_pmd_ptl ) & & vma_is_anonymous ( vma ) ;
}
# endif
2017-09-09 02:11:04 +03:00
static pmd_t move_soft_dirty_pmd ( pmd_t pmd )
{
# ifdef CONFIG_MEM_SOFT_DIRTY
if ( unlikely ( is_pmd_migration_entry ( pmd ) ) )
pmd = pmd_swp_mksoft_dirty ( pmd ) ;
else if ( pmd_present ( pmd ) )
pmd = pmd_mksoft_dirty ( pmd ) ;
# endif
return pmd ;
}
2016-05-20 03:12:54 +03:00
bool move_huge_pmd ( struct vm_area_struct * vma , unsigned long old_addr ,
2020-08-07 09:23:40 +03:00
unsigned long new_addr , pmd_t * old_pmd , pmd_t * new_pmd )
thp: mremap support and TLB optimization
This adds THP support to mremap (decreases the number of split_huge_page()
calls).
Here are also some benchmarks with a proggy like this:
===
#define _GNU_SOURCE
#include <sys/mman.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (5UL*1024*1024*1024)
int main()
{
static struct timeval oldstamp, newstamp;
long diffsec;
char *p, *p2, *p3, *p4;
if (posix_memalign((void **)&p, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p2, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p3, 2*1024*1024, 4096))
perror("memalign"), exit(1);
memset(p, 0xff, SIZE);
memset(p2, 0xff, SIZE);
memset(p3, 0x77, 4096);
gettimeofday(&oldstamp, NULL);
p4 = mremap(p, SIZE, SIZE, MREMAP_FIXED|MREMAP_MAYMOVE, p3);
gettimeofday(&newstamp, NULL);
diffsec = newstamp.tv_sec - oldstamp.tv_sec;
diffsec = newstamp.tv_usec - oldstamp.tv_usec + 1000000 * diffsec;
printf("usec %ld\n", diffsec);
if (p == MAP_FAILED || p4 != p3)
//if (p == MAP_FAILED)
perror("mremap"), exit(1);
if (memcmp(p4, p2, SIZE))
printf("mremap bug\n"), exit(1);
printf("ok\n");
return 0;
}
===
THP on
Performance counter stats for './largepage13' (3 runs):
69195836 dTLB-loads ( +- 3.546% ) (scaled from 50.30%)
60708 dTLB-load-misses ( +- 11.776% ) (scaled from 52.62%)
676266476 dTLB-stores ( +- 5.654% ) (scaled from 69.54%)
29856 dTLB-store-misses ( +- 4.081% ) (scaled from 89.22%)
1055848782 iTLB-loads ( +- 4.526% ) (scaled from 80.18%)
8689 iTLB-load-misses ( +- 2.987% ) (scaled from 58.20%)
7.314454164 seconds time elapsed ( +- 0.023% )
THP off
Performance counter stats for './largepage13' (3 runs):
1967379311 dTLB-loads ( +- 0.506% ) (scaled from 60.59%)
9238687 dTLB-load-misses ( +- 22.547% ) (scaled from 61.87%)
2014239444 dTLB-stores ( +- 0.692% ) (scaled from 60.40%)
3312335 dTLB-store-misses ( +- 7.304% ) (scaled from 67.60%)
6764372065 iTLB-loads ( +- 0.925% ) (scaled from 79.00%)
8202 iTLB-load-misses ( +- 0.475% ) (scaled from 70.55%)
9.693655243 seconds time elapsed ( +- 0.069% )
grep thp /proc/vmstat
thp_fault_alloc 35849
thp_fault_fallback 0
thp_collapse_alloc 3
thp_collapse_alloc_failed 0
thp_split 0
thp_split 0 confirms no thp split despite plenty of hugepages allocated.
The measurement of only the mremap time (so excluding the 3 long
memset and final long 10GB memory accessing memcmp):
THP on
usec 14824
usec 14862
usec 14859
THP off
usec 256416
usec 255981
usec 255847
With an older kernel without the mremap optimizations (the below patch
optimizes the non THP version too).
THP on
usec 392107
usec 390237
usec 404124
THP off
usec 444294
usec 445237
usec 445820
I guess with a threaded program that sends more IPI on large SMP it'd
create an even larger difference.
All debug options are off except DEBUG_VM to avoid skewing the
results.
The only problem for native 2M mremap like it happens above both the
source and destination address must be 2M aligned or the hugepmd can't be
moved without a split but that is an hardware limitation.
[akpm@linux-foundation.org: coding-style nitpicking]
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Johannes Weiner <jweiner@redhat.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-11-01 04:08:30 +04:00
{
2013-11-15 02:30:54 +04:00
spinlock_t * old_ptl , * new_ptl ;
thp: mremap support and TLB optimization
This adds THP support to mremap (decreases the number of split_huge_page()
calls).
Here are also some benchmarks with a proggy like this:
===
#define _GNU_SOURCE
#include <sys/mman.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (5UL*1024*1024*1024)
int main()
{
static struct timeval oldstamp, newstamp;
long diffsec;
char *p, *p2, *p3, *p4;
if (posix_memalign((void **)&p, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p2, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p3, 2*1024*1024, 4096))
perror("memalign"), exit(1);
memset(p, 0xff, SIZE);
memset(p2, 0xff, SIZE);
memset(p3, 0x77, 4096);
gettimeofday(&oldstamp, NULL);
p4 = mremap(p, SIZE, SIZE, MREMAP_FIXED|MREMAP_MAYMOVE, p3);
gettimeofday(&newstamp, NULL);
diffsec = newstamp.tv_sec - oldstamp.tv_sec;
diffsec = newstamp.tv_usec - oldstamp.tv_usec + 1000000 * diffsec;
printf("usec %ld\n", diffsec);
if (p == MAP_FAILED || p4 != p3)
//if (p == MAP_FAILED)
perror("mremap"), exit(1);
if (memcmp(p4, p2, SIZE))
printf("mremap bug\n"), exit(1);
printf("ok\n");
return 0;
}
===
THP on
Performance counter stats for './largepage13' (3 runs):
69195836 dTLB-loads ( +- 3.546% ) (scaled from 50.30%)
60708 dTLB-load-misses ( +- 11.776% ) (scaled from 52.62%)
676266476 dTLB-stores ( +- 5.654% ) (scaled from 69.54%)
29856 dTLB-store-misses ( +- 4.081% ) (scaled from 89.22%)
1055848782 iTLB-loads ( +- 4.526% ) (scaled from 80.18%)
8689 iTLB-load-misses ( +- 2.987% ) (scaled from 58.20%)
7.314454164 seconds time elapsed ( +- 0.023% )
THP off
Performance counter stats for './largepage13' (3 runs):
1967379311 dTLB-loads ( +- 0.506% ) (scaled from 60.59%)
9238687 dTLB-load-misses ( +- 22.547% ) (scaled from 61.87%)
2014239444 dTLB-stores ( +- 0.692% ) (scaled from 60.40%)
3312335 dTLB-store-misses ( +- 7.304% ) (scaled from 67.60%)
6764372065 iTLB-loads ( +- 0.925% ) (scaled from 79.00%)
8202 iTLB-load-misses ( +- 0.475% ) (scaled from 70.55%)
9.693655243 seconds time elapsed ( +- 0.069% )
grep thp /proc/vmstat
thp_fault_alloc 35849
thp_fault_fallback 0
thp_collapse_alloc 3
thp_collapse_alloc_failed 0
thp_split 0
thp_split 0 confirms no thp split despite plenty of hugepages allocated.
The measurement of only the mremap time (so excluding the 3 long
memset and final long 10GB memory accessing memcmp):
THP on
usec 14824
usec 14862
usec 14859
THP off
usec 256416
usec 255981
usec 255847
With an older kernel without the mremap optimizations (the below patch
optimizes the non THP version too).
THP on
usec 392107
usec 390237
usec 404124
THP off
usec 444294
usec 445237
usec 445820
I guess with a threaded program that sends more IPI on large SMP it'd
create an even larger difference.
All debug options are off except DEBUG_VM to avoid skewing the
results.
The only problem for native 2M mremap like it happens above both the
source and destination address must be 2M aligned or the hugepmd can't be
moved without a split but that is an hardware limitation.
[akpm@linux-foundation.org: coding-style nitpicking]
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Johannes Weiner <jweiner@redhat.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-11-01 04:08:30 +04:00
pmd_t pmd ;
struct mm_struct * mm = vma - > vm_mm ;
mremap: fix race between mremap() and page cleanning
Prior to 3.15, there was a race between zap_pte_range() and
page_mkclean() where writes to a page could be lost. Dave Hansen
discovered by inspection that there is a similar race between
move_ptes() and page_mkclean().
We've been able to reproduce the issue by enlarging the race window with
a msleep(), but have not been able to hit it without modifying the code.
So, we think it's a real issue, but is difficult or impossible to hit in
practice.
The zap_pte_range() issue is fixed by commit 1cf35d47712d("mm: split
'tlb_flush_mmu()' into tlb flushing and memory freeing parts"). And
this patch is to fix the race between page_mkclean() and mremap().
Here is one possible way to hit the race: suppose a process mmapped a
file with READ | WRITE and SHARED, it has two threads and they are bound
to 2 different CPUs, e.g. CPU1 and CPU2. mmap returned X, then thread
1 did a write to addr X so that CPU1 now has a writable TLB for addr X
on it. Thread 2 starts mremaping from addr X to Y while thread 1
cleaned the page and then did another write to the old addr X again.
The 2nd write from thread 1 could succeed but the value will get lost.
thread 1 thread 2
(bound to CPU1) (bound to CPU2)
1: write 1 to addr X to get a
writeable TLB on this CPU
2: mremap starts
3: move_ptes emptied PTE for addr X
and setup new PTE for addr Y and
then dropped PTL for X and Y
4: page laundering for N by doing
fadvise FADV_DONTNEED. When done,
pageframe N is deemed clean.
5: *write 2 to addr X
6: tlb flush for addr X
7: munmap (Y, pagesize) to make the
page unmapped
8: fadvise with FADV_DONTNEED again
to kick the page off the pagecache
9: pread the page from file to verify
the value. If 1 is there, it means
we have lost the written 2.
*the write may or may not cause segmentation fault, it depends on
if the TLB is still on the CPU.
Please note that this is only one specific way of how the race could
occur, it didn't mean that the race could only occur in exact the above
config, e.g. more than 2 threads could be involved and fadvise() could
be done in another thread, etc.
For anonymous pages, they could race between mremap() and page reclaim:
THP: a huge PMD is moved by mremap to a new huge PMD, then the new huge
PMD gets unmapped/splitted/pagedout before the flush tlb happened for
the old huge PMD in move_page_tables() and we could still write data to
it. The normal anonymous page has similar situation.
To fix this, check for any dirty PTE in move_ptes()/move_huge_pmd() and
if any, did the flush before dropping the PTL. If we did the flush for
every move_ptes()/move_huge_pmd() call then we do not need to do the
flush in move_pages_tables() for the whole range. But if we didn't, we
still need to do the whole range flush.
Alternatively, we can track which part of the range is flushed in
move_ptes()/move_huge_pmd() and which didn't to avoid flushing the whole
range in move_page_tables(). But that would require multiple tlb
flushes for the different sub-ranges and should be less efficient than
the single whole range flush.
KBuild test on my Sandybridge desktop doesn't show any noticeable change.
v4.9-rc4:
real 5m14.048s
user 32m19.800s
sys 4m50.320s
With this commit:
real 5m13.888s
user 32m19.330s
sys 4m51.200s
Reported-by: Dave Hansen <dave.hansen@intel.com>
Signed-off-by: Aaron Lu <aaron.lu@intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-11-10 12:16:33 +03:00
bool force_flush = false ;
thp: mremap support and TLB optimization
This adds THP support to mremap (decreases the number of split_huge_page()
calls).
Here are also some benchmarks with a proggy like this:
===
#define _GNU_SOURCE
#include <sys/mman.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (5UL*1024*1024*1024)
int main()
{
static struct timeval oldstamp, newstamp;
long diffsec;
char *p, *p2, *p3, *p4;
if (posix_memalign((void **)&p, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p2, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p3, 2*1024*1024, 4096))
perror("memalign"), exit(1);
memset(p, 0xff, SIZE);
memset(p2, 0xff, SIZE);
memset(p3, 0x77, 4096);
gettimeofday(&oldstamp, NULL);
p4 = mremap(p, SIZE, SIZE, MREMAP_FIXED|MREMAP_MAYMOVE, p3);
gettimeofday(&newstamp, NULL);
diffsec = newstamp.tv_sec - oldstamp.tv_sec;
diffsec = newstamp.tv_usec - oldstamp.tv_usec + 1000000 * diffsec;
printf("usec %ld\n", diffsec);
if (p == MAP_FAILED || p4 != p3)
//if (p == MAP_FAILED)
perror("mremap"), exit(1);
if (memcmp(p4, p2, SIZE))
printf("mremap bug\n"), exit(1);
printf("ok\n");
return 0;
}
===
THP on
Performance counter stats for './largepage13' (3 runs):
69195836 dTLB-loads ( +- 3.546% ) (scaled from 50.30%)
60708 dTLB-load-misses ( +- 11.776% ) (scaled from 52.62%)
676266476 dTLB-stores ( +- 5.654% ) (scaled from 69.54%)
29856 dTLB-store-misses ( +- 4.081% ) (scaled from 89.22%)
1055848782 iTLB-loads ( +- 4.526% ) (scaled from 80.18%)
8689 iTLB-load-misses ( +- 2.987% ) (scaled from 58.20%)
7.314454164 seconds time elapsed ( +- 0.023% )
THP off
Performance counter stats for './largepage13' (3 runs):
1967379311 dTLB-loads ( +- 0.506% ) (scaled from 60.59%)
9238687 dTLB-load-misses ( +- 22.547% ) (scaled from 61.87%)
2014239444 dTLB-stores ( +- 0.692% ) (scaled from 60.40%)
3312335 dTLB-store-misses ( +- 7.304% ) (scaled from 67.60%)
6764372065 iTLB-loads ( +- 0.925% ) (scaled from 79.00%)
8202 iTLB-load-misses ( +- 0.475% ) (scaled from 70.55%)
9.693655243 seconds time elapsed ( +- 0.069% )
grep thp /proc/vmstat
thp_fault_alloc 35849
thp_fault_fallback 0
thp_collapse_alloc 3
thp_collapse_alloc_failed 0
thp_split 0
thp_split 0 confirms no thp split despite plenty of hugepages allocated.
The measurement of only the mremap time (so excluding the 3 long
memset and final long 10GB memory accessing memcmp):
THP on
usec 14824
usec 14862
usec 14859
THP off
usec 256416
usec 255981
usec 255847
With an older kernel without the mremap optimizations (the below patch
optimizes the non THP version too).
THP on
usec 392107
usec 390237
usec 404124
THP off
usec 444294
usec 445237
usec 445820
I guess with a threaded program that sends more IPI on large SMP it'd
create an even larger difference.
All debug options are off except DEBUG_VM to avoid skewing the
results.
The only problem for native 2M mremap like it happens above both the
source and destination address must be 2M aligned or the hugepmd can't be
moved without a split but that is an hardware limitation.
[akpm@linux-foundation.org: coding-style nitpicking]
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Johannes Weiner <jweiner@redhat.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-11-01 04:08:30 +04:00
/*
* The destination pmd shouldn ' t be established , free_pgtables ( )
* should have release it .
*/
if ( WARN_ON ( ! pmd_none ( * new_pmd ) ) ) {
VM_BUG_ON ( pmd_trans_huge ( * new_pmd ) ) ;
2016-01-16 03:53:39 +03:00
return false ;
thp: mremap support and TLB optimization
This adds THP support to mremap (decreases the number of split_huge_page()
calls).
Here are also some benchmarks with a proggy like this:
===
#define _GNU_SOURCE
#include <sys/mman.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (5UL*1024*1024*1024)
int main()
{
static struct timeval oldstamp, newstamp;
long diffsec;
char *p, *p2, *p3, *p4;
if (posix_memalign((void **)&p, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p2, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p3, 2*1024*1024, 4096))
perror("memalign"), exit(1);
memset(p, 0xff, SIZE);
memset(p2, 0xff, SIZE);
memset(p3, 0x77, 4096);
gettimeofday(&oldstamp, NULL);
p4 = mremap(p, SIZE, SIZE, MREMAP_FIXED|MREMAP_MAYMOVE, p3);
gettimeofday(&newstamp, NULL);
diffsec = newstamp.tv_sec - oldstamp.tv_sec;
diffsec = newstamp.tv_usec - oldstamp.tv_usec + 1000000 * diffsec;
printf("usec %ld\n", diffsec);
if (p == MAP_FAILED || p4 != p3)
//if (p == MAP_FAILED)
perror("mremap"), exit(1);
if (memcmp(p4, p2, SIZE))
printf("mremap bug\n"), exit(1);
printf("ok\n");
return 0;
}
===
THP on
Performance counter stats for './largepage13' (3 runs):
69195836 dTLB-loads ( +- 3.546% ) (scaled from 50.30%)
60708 dTLB-load-misses ( +- 11.776% ) (scaled from 52.62%)
676266476 dTLB-stores ( +- 5.654% ) (scaled from 69.54%)
29856 dTLB-store-misses ( +- 4.081% ) (scaled from 89.22%)
1055848782 iTLB-loads ( +- 4.526% ) (scaled from 80.18%)
8689 iTLB-load-misses ( +- 2.987% ) (scaled from 58.20%)
7.314454164 seconds time elapsed ( +- 0.023% )
THP off
Performance counter stats for './largepage13' (3 runs):
1967379311 dTLB-loads ( +- 0.506% ) (scaled from 60.59%)
9238687 dTLB-load-misses ( +- 22.547% ) (scaled from 61.87%)
2014239444 dTLB-stores ( +- 0.692% ) (scaled from 60.40%)
3312335 dTLB-store-misses ( +- 7.304% ) (scaled from 67.60%)
6764372065 iTLB-loads ( +- 0.925% ) (scaled from 79.00%)
8202 iTLB-load-misses ( +- 0.475% ) (scaled from 70.55%)
9.693655243 seconds time elapsed ( +- 0.069% )
grep thp /proc/vmstat
thp_fault_alloc 35849
thp_fault_fallback 0
thp_collapse_alloc 3
thp_collapse_alloc_failed 0
thp_split 0
thp_split 0 confirms no thp split despite plenty of hugepages allocated.
The measurement of only the mremap time (so excluding the 3 long
memset and final long 10GB memory accessing memcmp):
THP on
usec 14824
usec 14862
usec 14859
THP off
usec 256416
usec 255981
usec 255847
With an older kernel without the mremap optimizations (the below patch
optimizes the non THP version too).
THP on
usec 392107
usec 390237
usec 404124
THP off
usec 444294
usec 445237
usec 445820
I guess with a threaded program that sends more IPI on large SMP it'd
create an even larger difference.
All debug options are off except DEBUG_VM to avoid skewing the
results.
The only problem for native 2M mremap like it happens above both the
source and destination address must be 2M aligned or the hugepmd can't be
moved without a split but that is an hardware limitation.
[akpm@linux-foundation.org: coding-style nitpicking]
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Johannes Weiner <jweiner@redhat.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-11-01 04:08:30 +04:00
}
2013-11-15 02:30:54 +04:00
/*
* We don ' t have to worry about the ordering of src and dst
2020-06-09 07:33:54 +03:00
* ptlocks because exclusive mmap_lock prevents deadlock .
2013-11-15 02:30:54 +04:00
*/
2016-01-22 03:40:25 +03:00
old_ptl = __pmd_trans_huge_lock ( old_pmd , vma ) ;
if ( old_ptl ) {
2013-11-15 02:30:54 +04:00
new_ptl = pmd_lockptr ( mm , new_pmd ) ;
if ( new_ptl ! = old_ptl )
spin_lock_nested ( new_ptl , SINGLE_DEPTH_NESTING ) ;
2015-06-25 02:57:44 +03:00
pmd = pmdp_huge_get_and_clear ( mm , old_addr , old_pmd ) ;
2018-10-13 01:22:59 +03:00
if ( pmd_present ( pmd ) )
2016-11-29 08:27:31 +03:00
force_flush = true ;
2012-03-22 03:33:57 +04:00
VM_BUG_ON ( ! pmd_none ( * new_pmd ) ) ;
thp: move preallocated PTE page table on move_huge_pmd()
Andrey Wagin reported crash on VM_BUG_ON() in pgtable_pmd_page_dtor() with
fallowing backtrace:
free_pgd_range+0x2bf/0x410
free_pgtables+0xce/0x120
unmap_region+0xe0/0x120
do_munmap+0x249/0x360
move_vma+0x144/0x270
SyS_mremap+0x3b9/0x510
system_call_fastpath+0x16/0x1b
The crash can be reproduce with this test case:
#define _GNU_SOURCE
#include <sys/mman.h>
#include <stdio.h>
#include <unistd.h>
#define MB (1024 * 1024UL)
#define GB (1024 * MB)
int main(int argc, char **argv)
{
char *p;
int i;
p = mmap((void *) GB, 10 * MB, PROT_READ | PROT_WRITE,
MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED, -1, 0);
for (i = 0; i < 10 * MB; i += 4096)
p[i] = 1;
mremap(p, 10 * MB, 10 * MB, MREMAP_FIXED | MREMAP_MAYMOVE, 2 * GB);
return 0;
}
Due to split PMD lock, we now store preallocated PTE tables for THP
pages per-PMD table. It means we need to move them to other PMD table
if huge PMD moved there.
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reported-by: Andrey Vagin <avagin@openvz.org>
Tested-by: Andrey Vagin <avagin@openvz.org>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-12-13 05:12:33 +04:00
2016-12-13 03:44:29 +03:00
if ( pmd_move_must_withdraw ( new_ptl , old_ptl , vma ) ) {
2014-01-13 10:04:24 +04:00
pgtable_t pgtable ;
thp: move preallocated PTE page table on move_huge_pmd()
Andrey Wagin reported crash on VM_BUG_ON() in pgtable_pmd_page_dtor() with
fallowing backtrace:
free_pgd_range+0x2bf/0x410
free_pgtables+0xce/0x120
unmap_region+0xe0/0x120
do_munmap+0x249/0x360
move_vma+0x144/0x270
SyS_mremap+0x3b9/0x510
system_call_fastpath+0x16/0x1b
The crash can be reproduce with this test case:
#define _GNU_SOURCE
#include <sys/mman.h>
#include <stdio.h>
#include <unistd.h>
#define MB (1024 * 1024UL)
#define GB (1024 * MB)
int main(int argc, char **argv)
{
char *p;
int i;
p = mmap((void *) GB, 10 * MB, PROT_READ | PROT_WRITE,
MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED, -1, 0);
for (i = 0; i < 10 * MB; i += 4096)
p[i] = 1;
mremap(p, 10 * MB, 10 * MB, MREMAP_FIXED | MREMAP_MAYMOVE, 2 * GB);
return 0;
}
Due to split PMD lock, we now store preallocated PTE tables for THP
pages per-PMD table. It means we need to move them to other PMD table
if huge PMD moved there.
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reported-by: Andrey Vagin <avagin@openvz.org>
Tested-by: Andrey Vagin <avagin@openvz.org>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-12-13 05:12:33 +04:00
pgtable = pgtable_trans_huge_withdraw ( mm , old_pmd ) ;
pgtable_trans_huge_deposit ( mm , new_pmd , pgtable ) ;
}
2017-09-09 02:11:04 +03:00
pmd = move_soft_dirty_pmd ( pmd ) ;
set_pmd_at ( mm , new_addr , new_pmd , pmd ) ;
mremap: fix race between mremap() and page cleanning
Prior to 3.15, there was a race between zap_pte_range() and
page_mkclean() where writes to a page could be lost. Dave Hansen
discovered by inspection that there is a similar race between
move_ptes() and page_mkclean().
We've been able to reproduce the issue by enlarging the race window with
a msleep(), but have not been able to hit it without modifying the code.
So, we think it's a real issue, but is difficult or impossible to hit in
practice.
The zap_pte_range() issue is fixed by commit 1cf35d47712d("mm: split
'tlb_flush_mmu()' into tlb flushing and memory freeing parts"). And
this patch is to fix the race between page_mkclean() and mremap().
Here is one possible way to hit the race: suppose a process mmapped a
file with READ | WRITE and SHARED, it has two threads and they are bound
to 2 different CPUs, e.g. CPU1 and CPU2. mmap returned X, then thread
1 did a write to addr X so that CPU1 now has a writable TLB for addr X
on it. Thread 2 starts mremaping from addr X to Y while thread 1
cleaned the page and then did another write to the old addr X again.
The 2nd write from thread 1 could succeed but the value will get lost.
thread 1 thread 2
(bound to CPU1) (bound to CPU2)
1: write 1 to addr X to get a
writeable TLB on this CPU
2: mremap starts
3: move_ptes emptied PTE for addr X
and setup new PTE for addr Y and
then dropped PTL for X and Y
4: page laundering for N by doing
fadvise FADV_DONTNEED. When done,
pageframe N is deemed clean.
5: *write 2 to addr X
6: tlb flush for addr X
7: munmap (Y, pagesize) to make the
page unmapped
8: fadvise with FADV_DONTNEED again
to kick the page off the pagecache
9: pread the page from file to verify
the value. If 1 is there, it means
we have lost the written 2.
*the write may or may not cause segmentation fault, it depends on
if the TLB is still on the CPU.
Please note that this is only one specific way of how the race could
occur, it didn't mean that the race could only occur in exact the above
config, e.g. more than 2 threads could be involved and fadvise() could
be done in another thread, etc.
For anonymous pages, they could race between mremap() and page reclaim:
THP: a huge PMD is moved by mremap to a new huge PMD, then the new huge
PMD gets unmapped/splitted/pagedout before the flush tlb happened for
the old huge PMD in move_page_tables() and we could still write data to
it. The normal anonymous page has similar situation.
To fix this, check for any dirty PTE in move_ptes()/move_huge_pmd() and
if any, did the flush before dropping the PTL. If we did the flush for
every move_ptes()/move_huge_pmd() call then we do not need to do the
flush in move_pages_tables() for the whole range. But if we didn't, we
still need to do the whole range flush.
Alternatively, we can track which part of the range is flushed in
move_ptes()/move_huge_pmd() and which didn't to avoid flushing the whole
range in move_page_tables(). But that would require multiple tlb
flushes for the different sub-ranges and should be less efficient than
the single whole range flush.
KBuild test on my Sandybridge desktop doesn't show any noticeable change.
v4.9-rc4:
real 5m14.048s
user 32m19.800s
sys 4m50.320s
With this commit:
real 5m13.888s
user 32m19.330s
sys 4m51.200s
Reported-by: Dave Hansen <dave.hansen@intel.com>
Signed-off-by: Aaron Lu <aaron.lu@intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-11-10 12:16:33 +03:00
if ( force_flush )
2022-07-04 16:21:46 +03:00
flush_pmd_tlb_range ( vma , old_addr , old_addr + PMD_SIZE ) ;
2018-10-13 01:22:59 +03:00
if ( new_ptl ! = old_ptl )
spin_unlock ( new_ptl ) ;
2013-11-15 02:30:54 +04:00
spin_unlock ( old_ptl ) ;
2016-01-16 03:53:39 +03:00
return true ;
thp: mremap support and TLB optimization
This adds THP support to mremap (decreases the number of split_huge_page()
calls).
Here are also some benchmarks with a proggy like this:
===
#define _GNU_SOURCE
#include <sys/mman.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (5UL*1024*1024*1024)
int main()
{
static struct timeval oldstamp, newstamp;
long diffsec;
char *p, *p2, *p3, *p4;
if (posix_memalign((void **)&p, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p2, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p3, 2*1024*1024, 4096))
perror("memalign"), exit(1);
memset(p, 0xff, SIZE);
memset(p2, 0xff, SIZE);
memset(p3, 0x77, 4096);
gettimeofday(&oldstamp, NULL);
p4 = mremap(p, SIZE, SIZE, MREMAP_FIXED|MREMAP_MAYMOVE, p3);
gettimeofday(&newstamp, NULL);
diffsec = newstamp.tv_sec - oldstamp.tv_sec;
diffsec = newstamp.tv_usec - oldstamp.tv_usec + 1000000 * diffsec;
printf("usec %ld\n", diffsec);
if (p == MAP_FAILED || p4 != p3)
//if (p == MAP_FAILED)
perror("mremap"), exit(1);
if (memcmp(p4, p2, SIZE))
printf("mremap bug\n"), exit(1);
printf("ok\n");
return 0;
}
===
THP on
Performance counter stats for './largepage13' (3 runs):
69195836 dTLB-loads ( +- 3.546% ) (scaled from 50.30%)
60708 dTLB-load-misses ( +- 11.776% ) (scaled from 52.62%)
676266476 dTLB-stores ( +- 5.654% ) (scaled from 69.54%)
29856 dTLB-store-misses ( +- 4.081% ) (scaled from 89.22%)
1055848782 iTLB-loads ( +- 4.526% ) (scaled from 80.18%)
8689 iTLB-load-misses ( +- 2.987% ) (scaled from 58.20%)
7.314454164 seconds time elapsed ( +- 0.023% )
THP off
Performance counter stats for './largepage13' (3 runs):
1967379311 dTLB-loads ( +- 0.506% ) (scaled from 60.59%)
9238687 dTLB-load-misses ( +- 22.547% ) (scaled from 61.87%)
2014239444 dTLB-stores ( +- 0.692% ) (scaled from 60.40%)
3312335 dTLB-store-misses ( +- 7.304% ) (scaled from 67.60%)
6764372065 iTLB-loads ( +- 0.925% ) (scaled from 79.00%)
8202 iTLB-load-misses ( +- 0.475% ) (scaled from 70.55%)
9.693655243 seconds time elapsed ( +- 0.069% )
grep thp /proc/vmstat
thp_fault_alloc 35849
thp_fault_fallback 0
thp_collapse_alloc 3
thp_collapse_alloc_failed 0
thp_split 0
thp_split 0 confirms no thp split despite plenty of hugepages allocated.
The measurement of only the mremap time (so excluding the 3 long
memset and final long 10GB memory accessing memcmp):
THP on
usec 14824
usec 14862
usec 14859
THP off
usec 256416
usec 255981
usec 255847
With an older kernel without the mremap optimizations (the below patch
optimizes the non THP version too).
THP on
usec 392107
usec 390237
usec 404124
THP off
usec 444294
usec 445237
usec 445820
I guess with a threaded program that sends more IPI on large SMP it'd
create an even larger difference.
All debug options are off except DEBUG_VM to avoid skewing the
results.
The only problem for native 2M mremap like it happens above both the
source and destination address must be 2M aligned or the hugepmd can't be
moved without a split but that is an hardware limitation.
[akpm@linux-foundation.org: coding-style nitpicking]
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Johannes Weiner <jweiner@redhat.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-11-01 04:08:30 +04:00
}
2016-01-16 03:53:39 +03:00
return false ;
thp: mremap support and TLB optimization
This adds THP support to mremap (decreases the number of split_huge_page()
calls).
Here are also some benchmarks with a proggy like this:
===
#define _GNU_SOURCE
#include <sys/mman.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <sys/time.h>
#define SIZE (5UL*1024*1024*1024)
int main()
{
static struct timeval oldstamp, newstamp;
long diffsec;
char *p, *p2, *p3, *p4;
if (posix_memalign((void **)&p, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p2, 2*1024*1024, SIZE))
perror("memalign"), exit(1);
if (posix_memalign((void **)&p3, 2*1024*1024, 4096))
perror("memalign"), exit(1);
memset(p, 0xff, SIZE);
memset(p2, 0xff, SIZE);
memset(p3, 0x77, 4096);
gettimeofday(&oldstamp, NULL);
p4 = mremap(p, SIZE, SIZE, MREMAP_FIXED|MREMAP_MAYMOVE, p3);
gettimeofday(&newstamp, NULL);
diffsec = newstamp.tv_sec - oldstamp.tv_sec;
diffsec = newstamp.tv_usec - oldstamp.tv_usec + 1000000 * diffsec;
printf("usec %ld\n", diffsec);
if (p == MAP_FAILED || p4 != p3)
//if (p == MAP_FAILED)
perror("mremap"), exit(1);
if (memcmp(p4, p2, SIZE))
printf("mremap bug\n"), exit(1);
printf("ok\n");
return 0;
}
===
THP on
Performance counter stats for './largepage13' (3 runs):
69195836 dTLB-loads ( +- 3.546% ) (scaled from 50.30%)
60708 dTLB-load-misses ( +- 11.776% ) (scaled from 52.62%)
676266476 dTLB-stores ( +- 5.654% ) (scaled from 69.54%)
29856 dTLB-store-misses ( +- 4.081% ) (scaled from 89.22%)
1055848782 iTLB-loads ( +- 4.526% ) (scaled from 80.18%)
8689 iTLB-load-misses ( +- 2.987% ) (scaled from 58.20%)
7.314454164 seconds time elapsed ( +- 0.023% )
THP off
Performance counter stats for './largepage13' (3 runs):
1967379311 dTLB-loads ( +- 0.506% ) (scaled from 60.59%)
9238687 dTLB-load-misses ( +- 22.547% ) (scaled from 61.87%)
2014239444 dTLB-stores ( +- 0.692% ) (scaled from 60.40%)
3312335 dTLB-store-misses ( +- 7.304% ) (scaled from 67.60%)
6764372065 iTLB-loads ( +- 0.925% ) (scaled from 79.00%)
8202 iTLB-load-misses ( +- 0.475% ) (scaled from 70.55%)
9.693655243 seconds time elapsed ( +- 0.069% )
grep thp /proc/vmstat
thp_fault_alloc 35849
thp_fault_fallback 0
thp_collapse_alloc 3
thp_collapse_alloc_failed 0
thp_split 0
thp_split 0 confirms no thp split despite plenty of hugepages allocated.
The measurement of only the mremap time (so excluding the 3 long
memset and final long 10GB memory accessing memcmp):
THP on
usec 14824
usec 14862
usec 14859
THP off
usec 256416
usec 255981
usec 255847
With an older kernel without the mremap optimizations (the below patch
optimizes the non THP version too).
THP on
usec 392107
usec 390237
usec 404124
THP off
usec 444294
usec 445237
usec 445820
I guess with a threaded program that sends more IPI on large SMP it'd
create an even larger difference.
All debug options are off except DEBUG_VM to avoid skewing the
results.
The only problem for native 2M mremap like it happens above both the
source and destination address must be 2M aligned or the hugepmd can't be
moved without a split but that is an hardware limitation.
[akpm@linux-foundation.org: coding-style nitpicking]
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Johannes Weiner <jweiner@redhat.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-11-01 04:08:30 +04:00
}
2013-10-07 14:28:49 +04:00
/*
* Returns
* - 0 if PMD could not be locked
2021-05-07 04:06:47 +03:00
* - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
2021-07-01 04:51:55 +03:00
* or if prot_numa but THP migration is not supported
2021-05-07 04:06:47 +03:00
* - HPAGE_PMD_NR if protections changed and TLB flush necessary
2013-10-07 14:28:49 +04:00
*/
mm/mprotect: use mmu_gather
Patch series "mm/mprotect: avoid unnecessary TLB flushes", v6.
This patchset is intended to remove unnecessary TLB flushes during
mprotect() syscalls. Once this patch-set make it through, similar and
further optimizations for MADV_COLD and userfaultfd would be possible.
Basically, there are 3 optimizations in this patch-set:
1. Use TLB batching infrastructure to batch flushes across VMAs and do
better/fewer flushes. This would also be handy for later userfaultfd
enhancements.
2. Avoid unnecessary TLB flushes. This optimization is the one that
provides most of the performance benefits. Unlike previous versions,
we now only avoid flushes that would not result in spurious
page-faults.
3. Avoiding TLB flushes on change_huge_pmd() that are only needed to
prevent the A/D bits from changing.
Andrew asked for some benchmark numbers. I do not have an easy
determinate macrobenchmark in which it is easy to show benefit. I
therefore ran a microbenchmark: a loop that does the following on
anonymous memory, just as a sanity check to see that time is saved by
avoiding TLB flushes. The loop goes:
mprotect(p, PAGE_SIZE, PROT_READ)
mprotect(p, PAGE_SIZE, PROT_READ|PROT_WRITE)
*p = 0; // make the page writable
The test was run in KVM guest with 1 or 2 threads (the second thread was
busy-looping). I measured the time (cycles) of each operation:
1 thread 2 threads
mmots +patch mmots +patch
PROT_READ 3494 2725 (-22%) 8630 7788 (-10%)
PROT_READ|WRITE 3952 2724 (-31%) 9075 2865 (-68%)
[ mmots = v5.17-rc6-mmots-2022-03-06-20-38 ]
The exact numbers are really meaningless, but the benefit is clear. There
are 2 interesting results though.
(1) PROT_READ is cheaper, while one can expect it not to be affected.
This is presumably due to TLB miss that is saved
(2) Without memory access (*p = 0), the speedup of the patch is even
greater. In that scenario mprotect(PROT_READ) also avoids the TLB flush.
As a result both operations on the patched kernel take roughly ~1500
cycles (with either 1 or 2 threads), whereas on mmotm their cost is as
high as presented in the table.
This patch (of 3):
change_pXX_range() currently does not use mmu_gather, but instead
implements its own deferred TLB flushes scheme. This both complicates the
code, as developers need to be aware of different invalidation schemes,
and prevents opportunities to avoid TLB flushes or perform them in finer
granularity.
The use of mmu_gather for modified PTEs has benefits in various scenarios
even if pages are not released. For instance, if only a single page needs
to be flushed out of a range of many pages, only that page would be
flushed. If a THP page is flushed, on x86 a single TLB invlpg instruction
can be used instead of 512 instructions (or a full TLB flush, which would
Linux would actually use by default). mprotect() over multiple VMAs
requires a single flush.
Use mmu_gather in change_pXX_range(). As the pages are not released, only
record the flushed range using tlb_flush_pXX_range().
Handle THP similarly and get rid of flush_cache_range() which becomes
redundant since tlb_start_vma() calls it when needed.
Link: https://lkml.kernel.org/r/20220401180821.1986781-1-namit@vmware.com
Link: https://lkml.kernel.org/r/20220401180821.1986781-2-namit@vmware.com
Signed-off-by: Nadav Amit <namit@vmware.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrew Cooper <andrew.cooper3@citrix.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Will Deacon <will@kernel.org>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Nick Piggin <npiggin@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:50 +03:00
int change_huge_pmd ( struct mmu_gather * tlb , struct vm_area_struct * vma ,
pmd_t * pmd , unsigned long addr , pgprot_t newprot ,
unsigned long cp_flags )
2011-01-14 02:47:04 +03:00
{
struct mm_struct * mm = vma - > vm_mm ;
2013-11-15 02:30:54 +04:00
spinlock_t * ptl ;
2022-05-10 04:20:50 +03:00
pmd_t oldpmd , entry ;
2020-04-07 06:05:45 +03:00
bool prot_numa = cp_flags & MM_CP_PROT_NUMA ;
userfaultfd: wp: apply _PAGE_UFFD_WP bit
Firstly, introduce two new flags MM_CP_UFFD_WP[_RESOLVE] for
change_protection() when used with uffd-wp and make sure the two new flags
are exclusively used. Then,
- For MM_CP_UFFD_WP: apply the _PAGE_UFFD_WP bit and remove _PAGE_RW
when a range of memory is write protected by uffd
- For MM_CP_UFFD_WP_RESOLVE: remove the _PAGE_UFFD_WP bit and recover
_PAGE_RW when write protection is resolved from userspace
And use this new interface in mwriteprotect_range() to replace the old
MM_CP_DIRTY_ACCT.
Do this change for both PTEs and huge PMDs. Then we can start to identify
which PTE/PMD is write protected by general (e.g., COW or soft dirty
tracking), and which is for userfaultfd-wp.
Since we should keep the _PAGE_UFFD_WP when doing pte_modify(), add it
into _PAGE_CHG_MASK as well. Meanwhile, since we have this new bit, we
can be even more strict when detecting uffd-wp page faults in either
do_wp_page() or wp_huge_pmd().
After we're with _PAGE_UFFD_WP, a special case is when a page is both
protected by the general COW logic and also userfault-wp. Here the
userfault-wp will have higher priority and will be handled first. Only
after the uffd-wp bit is cleared on the PTE/PMD will we continue to handle
the general COW. These are the steps on what will happen with such a
page:
1. CPU accesses write protected shared page (so both protected by
general COW and uffd-wp), blocked by uffd-wp first because in
do_wp_page we'll handle uffd-wp first, so it has higher priority
than general COW.
2. Uffd service thread receives the request, do UFFDIO_WRITEPROTECT
to remove the uffd-wp bit upon the PTE/PMD. However here we
still keep the write bit cleared. Notify the blocked CPU.
3. The blocked CPU resumes the page fault process with a fault
retry, during retry it'll notice it was not with the uffd-wp bit
this time but it is still write protected by general COW, then
it'll go though the COW path in the fault handler, copy the page,
apply write bit where necessary, and retry again.
4. The CPU will be able to access this page with write bit set.
Suggested-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Brian Geffon <bgeffon@google.com>
Cc: Pavel Emelyanov <xemul@openvz.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Martin Cracauer <cracauer@cons.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Bobby Powers <bobbypowers@gmail.com>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: "Kirill A . Shutemov" <kirill@shutemov.name>
Cc: Maya Gokhale <gokhale2@llnl.gov>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Marty McFadden <mcfadden8@llnl.gov>
Cc: Denis Plotnikov <dplotnikov@virtuozzo.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Shaohua Li <shli@fb.com>
Link: http://lkml.kernel.org/r/20200220163112.11409-8-peterx@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-07 06:05:49 +03:00
bool uffd_wp = cp_flags & MM_CP_UFFD_WP ;
bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE ;
2022-11-08 20:46:50 +03:00
int ret = 1 ;
2011-01-14 02:47:04 +03:00
mm/mprotect: use mmu_gather
Patch series "mm/mprotect: avoid unnecessary TLB flushes", v6.
This patchset is intended to remove unnecessary TLB flushes during
mprotect() syscalls. Once this patch-set make it through, similar and
further optimizations for MADV_COLD and userfaultfd would be possible.
Basically, there are 3 optimizations in this patch-set:
1. Use TLB batching infrastructure to batch flushes across VMAs and do
better/fewer flushes. This would also be handy for later userfaultfd
enhancements.
2. Avoid unnecessary TLB flushes. This optimization is the one that
provides most of the performance benefits. Unlike previous versions,
we now only avoid flushes that would not result in spurious
page-faults.
3. Avoiding TLB flushes on change_huge_pmd() that are only needed to
prevent the A/D bits from changing.
Andrew asked for some benchmark numbers. I do not have an easy
determinate macrobenchmark in which it is easy to show benefit. I
therefore ran a microbenchmark: a loop that does the following on
anonymous memory, just as a sanity check to see that time is saved by
avoiding TLB flushes. The loop goes:
mprotect(p, PAGE_SIZE, PROT_READ)
mprotect(p, PAGE_SIZE, PROT_READ|PROT_WRITE)
*p = 0; // make the page writable
The test was run in KVM guest with 1 or 2 threads (the second thread was
busy-looping). I measured the time (cycles) of each operation:
1 thread 2 threads
mmots +patch mmots +patch
PROT_READ 3494 2725 (-22%) 8630 7788 (-10%)
PROT_READ|WRITE 3952 2724 (-31%) 9075 2865 (-68%)
[ mmots = v5.17-rc6-mmots-2022-03-06-20-38 ]
The exact numbers are really meaningless, but the benefit is clear. There
are 2 interesting results though.
(1) PROT_READ is cheaper, while one can expect it not to be affected.
This is presumably due to TLB miss that is saved
(2) Without memory access (*p = 0), the speedup of the patch is even
greater. In that scenario mprotect(PROT_READ) also avoids the TLB flush.
As a result both operations on the patched kernel take roughly ~1500
cycles (with either 1 or 2 threads), whereas on mmotm their cost is as
high as presented in the table.
This patch (of 3):
change_pXX_range() currently does not use mmu_gather, but instead
implements its own deferred TLB flushes scheme. This both complicates the
code, as developers need to be aware of different invalidation schemes,
and prevents opportunities to avoid TLB flushes or perform them in finer
granularity.
The use of mmu_gather for modified PTEs has benefits in various scenarios
even if pages are not released. For instance, if only a single page needs
to be flushed out of a range of many pages, only that page would be
flushed. If a THP page is flushed, on x86 a single TLB invlpg instruction
can be used instead of 512 instructions (or a full TLB flush, which would
Linux would actually use by default). mprotect() over multiple VMAs
requires a single flush.
Use mmu_gather in change_pXX_range(). As the pages are not released, only
record the flushed range using tlb_flush_pXX_range().
Handle THP similarly and get rid of flush_cache_range() which becomes
redundant since tlb_start_vma() calls it when needed.
Link: https://lkml.kernel.org/r/20220401180821.1986781-1-namit@vmware.com
Link: https://lkml.kernel.org/r/20220401180821.1986781-2-namit@vmware.com
Signed-off-by: Nadav Amit <namit@vmware.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrew Cooper <andrew.cooper3@citrix.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Will Deacon <will@kernel.org>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Nick Piggin <npiggin@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:50 +03:00
tlb_change_page_size ( tlb , HPAGE_PMD_SIZE ) ;
2021-07-01 04:51:55 +03:00
if ( prot_numa & & ! thp_migration_supported ( ) )
return 1 ;
2016-01-22 03:40:25 +03:00
ptl = __pmd_trans_huge_lock ( pmd , vma ) ;
2017-04-14 00:56:17 +03:00
if ( ! ptl )
return 0 ;
2015-02-13 01:58:35 +03:00
mm: thp: check pmd migration entry in common path
When THP migration is being used, memory management code needs to handle
pmd migration entries properly. This patch uses !pmd_present() or
is_swap_pmd() (depending on whether pmd_none() needs separate code or
not) to check pmd migration entries at the places where a pmd entry is
present.
Since pmd-related code uses split_huge_page(), split_huge_pmd(),
pmd_trans_huge(), pmd_trans_unstable(), or
pmd_none_or_trans_huge_or_clear_bad(), this patch:
1. adds pmd migration entry split code in split_huge_pmd(),
2. takes care of pmd migration entries whenever pmd_trans_huge() is present,
3. makes pmd_none_or_trans_huge_or_clear_bad() pmd migration entry aware.
Since split_huge_page() uses split_huge_pmd() and pmd_trans_unstable()
is equivalent to pmd_none_or_trans_huge_or_clear_bad(), we do not change
them.
Until this commit, a pmd entry should be:
1. pointing to a pte page,
2. is_swap_pmd(),
3. pmd_trans_huge(),
4. pmd_devmap(), or
5. pmd_none().
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-09 02:11:01 +03:00
# ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
if ( is_swap_pmd ( * pmd ) ) {
swp_entry_t entry = pmd_to_swp_entry ( * pmd ) ;
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
struct page * page = pfn_swap_entry_to_page ( entry ) ;
mm: thp: check pmd migration entry in common path
When THP migration is being used, memory management code needs to handle
pmd migration entries properly. This patch uses !pmd_present() or
is_swap_pmd() (depending on whether pmd_none() needs separate code or
not) to check pmd migration entries at the places where a pmd entry is
present.
Since pmd-related code uses split_huge_page(), split_huge_pmd(),
pmd_trans_huge(), pmd_trans_unstable(), or
pmd_none_or_trans_huge_or_clear_bad(), this patch:
1. adds pmd migration entry split code in split_huge_pmd(),
2. takes care of pmd migration entries whenever pmd_trans_huge() is present,
3. makes pmd_none_or_trans_huge_or_clear_bad() pmd migration entry aware.
Since split_huge_page() uses split_huge_pmd() and pmd_trans_unstable()
is equivalent to pmd_none_or_trans_huge_or_clear_bad(), we do not change
them.
Until this commit, a pmd entry should be:
1. pointing to a pte page,
2. is_swap_pmd(),
3. pmd_trans_huge(),
4. pmd_devmap(), or
5. pmd_none().
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-09 02:11:01 +03:00
VM_BUG_ON ( ! is_pmd_migration_entry ( * pmd ) ) ;
2021-07-01 04:54:09 +03:00
if ( is_writable_migration_entry ( entry ) ) {
mm: thp: check pmd migration entry in common path
When THP migration is being used, memory management code needs to handle
pmd migration entries properly. This patch uses !pmd_present() or
is_swap_pmd() (depending on whether pmd_none() needs separate code or
not) to check pmd migration entries at the places where a pmd entry is
present.
Since pmd-related code uses split_huge_page(), split_huge_pmd(),
pmd_trans_huge(), pmd_trans_unstable(), or
pmd_none_or_trans_huge_or_clear_bad(), this patch:
1. adds pmd migration entry split code in split_huge_pmd(),
2. takes care of pmd migration entries whenever pmd_trans_huge() is present,
3. makes pmd_none_or_trans_huge_or_clear_bad() pmd migration entry aware.
Since split_huge_page() uses split_huge_pmd() and pmd_trans_unstable()
is equivalent to pmd_none_or_trans_huge_or_clear_bad(), we do not change
them.
Until this commit, a pmd entry should be:
1. pointing to a pte page,
2. is_swap_pmd(),
3. pmd_trans_huge(),
4. pmd_devmap(), or
5. pmd_none().
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-09 02:11:01 +03:00
pmd_t newpmd ;
/*
* A protection check is difficult so
* just be safe and disable write
*/
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
if ( PageAnon ( page ) )
entry = make_readable_exclusive_migration_entry ( swp_offset ( entry ) ) ;
else
entry = make_readable_migration_entry ( swp_offset ( entry ) ) ;
mm: thp: check pmd migration entry in common path
When THP migration is being used, memory management code needs to handle
pmd migration entries properly. This patch uses !pmd_present() or
is_swap_pmd() (depending on whether pmd_none() needs separate code or
not) to check pmd migration entries at the places where a pmd entry is
present.
Since pmd-related code uses split_huge_page(), split_huge_pmd(),
pmd_trans_huge(), pmd_trans_unstable(), or
pmd_none_or_trans_huge_or_clear_bad(), this patch:
1. adds pmd migration entry split code in split_huge_pmd(),
2. takes care of pmd migration entries whenever pmd_trans_huge() is present,
3. makes pmd_none_or_trans_huge_or_clear_bad() pmd migration entry aware.
Since split_huge_page() uses split_huge_pmd() and pmd_trans_unstable()
is equivalent to pmd_none_or_trans_huge_or_clear_bad(), we do not change
them.
Until this commit, a pmd entry should be:
1. pointing to a pte page,
2. is_swap_pmd(),
3. pmd_trans_huge(),
4. pmd_devmap(), or
5. pmd_none().
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-09 02:11:01 +03:00
newpmd = swp_entry_to_pmd ( entry ) ;
2017-09-09 02:11:04 +03:00
if ( pmd_swp_soft_dirty ( * pmd ) )
newpmd = pmd_swp_mksoft_dirty ( newpmd ) ;
mm/userfaultfd: fix uffd-wp special cases for fork()
We tried to do something similar in b569a1760782 ("userfaultfd: wp: drop
_PAGE_UFFD_WP properly when fork") previously, but it's not doing it all
right.. A few fixes around the code path:
1. We were referencing VM_UFFD_WP vm_flags on the _old_ vma rather
than the new vma. That's overlooked in b569a1760782, so it won't work
as expected. Thanks to the recent rework on fork code
(7a4830c380f3a8b3), we can easily get the new vma now, so switch the
checks to that.
2. Dropping the uffd-wp bit in copy_huge_pmd() could be wrong if the
huge pmd is a migration huge pmd. When it happens, instead of using
pmd_uffd_wp(), we should use pmd_swp_uffd_wp(). The fix is simply to
handle them separately.
3. Forget to carry over uffd-wp bit for a write migration huge pmd
entry. This also happens in copy_huge_pmd(), where we converted a
write huge migration entry into a read one.
4. In copy_nonpresent_pte(), drop uffd-wp if necessary for swap ptes.
5. In copy_present_page() when COW is enforced when fork(), we also
need to pass over the uffd-wp bit if VM_UFFD_WP is armed on the new
vma, and when the pte to be copied has uffd-wp bit set.
Remove the comment in copy_present_pte() about this. It won't help a huge
lot to only comment there, but comment everywhere would be an overkill.
Let's assume the commit messages would help.
[peterx@redhat.com: fix a few thp pmd missing uffd-wp bit]
Link: https://lkml.kernel.org/r/20210428225030.9708-4-peterx@redhat.com
Link: https://lkml.kernel.org/r/20210428225030.9708-3-peterx@redhat.com
Fixes: b569a1760782f ("userfaultfd: wp: drop _PAGE_UFFD_WP properly when fork")
Signed-off-by: Peter Xu <peterx@redhat.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Brian Geffon <bgeffon@google.com>
Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Joe Perches <joe@perches.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Lokesh Gidra <lokeshgidra@google.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Oliver Upton <oupton@google.com>
Cc: Shaohua Li <shli@fb.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Wang Qing <wangqing@vivo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-01 04:49:02 +03:00
if ( pmd_swp_uffd_wp ( * pmd ) )
newpmd = pmd_swp_mkuffd_wp ( newpmd ) ;
mm: thp: check pmd migration entry in common path
When THP migration is being used, memory management code needs to handle
pmd migration entries properly. This patch uses !pmd_present() or
is_swap_pmd() (depending on whether pmd_none() needs separate code or
not) to check pmd migration entries at the places where a pmd entry is
present.
Since pmd-related code uses split_huge_page(), split_huge_pmd(),
pmd_trans_huge(), pmd_trans_unstable(), or
pmd_none_or_trans_huge_or_clear_bad(), this patch:
1. adds pmd migration entry split code in split_huge_pmd(),
2. takes care of pmd migration entries whenever pmd_trans_huge() is present,
3. makes pmd_none_or_trans_huge_or_clear_bad() pmd migration entry aware.
Since split_huge_page() uses split_huge_pmd() and pmd_trans_unstable()
is equivalent to pmd_none_or_trans_huge_or_clear_bad(), we do not change
them.
Until this commit, a pmd entry should be:
1. pointing to a pte page,
2. is_swap_pmd(),
3. pmd_trans_huge(),
4. pmd_devmap(), or
5. pmd_none().
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-09 02:11:01 +03:00
set_pmd_at ( mm , addr , pmd , newpmd ) ;
}
goto unlock ;
}
# endif
2022-03-23 00:46:27 +03:00
if ( prot_numa ) {
struct page * page ;
memory tiering: hot page selection with hint page fault latency
Patch series "memory tiering: hot page selection", v4.
To optimize page placement in a memory tiering system with NUMA balancing,
the hot pages in the slow memory nodes need to be identified.
Essentially, the original NUMA balancing implementation selects the mostly
recently accessed (MRU) pages to promote. But this isn't a perfect
algorithm to identify the hot pages. Because the pages with quite low
access frequency may be accessed eventually given the NUMA balancing page
table scanning period could be quite long (e.g. 60 seconds). So in this
patchset, we implement a new hot page identification algorithm based on
the latency between NUMA balancing page table scanning and hint page
fault. Which is a kind of mostly frequently accessed (MFU) algorithm.
In NUMA balancing memory tiering mode, if there are hot pages in slow
memory node and cold pages in fast memory node, we need to promote/demote
hot/cold pages between the fast and cold memory nodes.
A choice is to promote/demote as fast as possible. But the CPU cycles and
memory bandwidth consumed by the high promoting/demoting throughput will
hurt the latency of some workload because of accessing inflating and slow
memory bandwidth contention.
A way to resolve this issue is to restrict the max promoting/demoting
throughput. It will take longer to finish the promoting/demoting. But
the workload latency will be better. This is implemented in this patchset
as the page promotion rate limit mechanism.
The promotion hot threshold is workload and system configuration
dependent. So in this patchset, a method to adjust the hot threshold
automatically is implemented. The basic idea is to control the number of
the candidate promotion pages to match the promotion rate limit.
We used the pmbench memory accessing benchmark tested the patchset on a
2-socket server system with DRAM and PMEM installed. The test results are
as follows,
pmbench score promote rate
(accesses/s) MB/s
------------- ------------
base 146887704.1 725.6
hot selection 165695601.2 544.0
rate limit 162814569.8 165.2
auto adjustment 170495294.0 136.9
From the results above,
With hot page selection patch [1/3], the pmbench score increases about
12.8%, and promote rate (overhead) decreases about 25.0%, compared with
base kernel.
With rate limit patch [2/3], pmbench score decreases about 1.7%, and
promote rate decreases about 69.6%, compared with hot page selection
patch.
With threshold auto adjustment patch [3/3], pmbench score increases about
4.7%, and promote rate decrease about 17.1%, compared with rate limit
patch.
Baolin helped to test the patchset with MySQL on a machine which contains
1 DRAM node (30G) and 1 PMEM node (126G).
sysbench /usr/share/sysbench/oltp_read_write.lua \
......
--tables=200 \
--table-size=1000000 \
--report-interval=10 \
--threads=16 \
--time=120
The tps can be improved about 5%.
This patch (of 3):
To optimize page placement in a memory tiering system with NUMA balancing,
the hot pages in the slow memory node need to be identified. Essentially,
the original NUMA balancing implementation selects the mostly recently
accessed (MRU) pages to promote. But this isn't a perfect algorithm to
identify the hot pages. Because the pages with quite low access frequency
may be accessed eventually given the NUMA balancing page table scanning
period could be quite long (e.g. 60 seconds). The most frequently
accessed (MFU) algorithm is better.
So, in this patch we implemented a better hot page selection algorithm.
Which is based on NUMA balancing page table scanning and hint page fault
as follows,
- When the page tables of the processes are scanned to change PTE/PMD
to be PROT_NONE, the current time is recorded in struct page as scan
time.
- When the page is accessed, hint page fault will occur. The scan
time is gotten from the struct page. And The hint page fault
latency is defined as
hint page fault time - scan time
The shorter the hint page fault latency of a page is, the higher the
probability of their access frequency to be higher. So the hint page
fault latency is a better estimation of the page hot/cold.
It's hard to find some extra space in struct page to hold the scan time.
Fortunately, we can reuse some bits used by the original NUMA balancing.
NUMA balancing uses some bits in struct page to store the page accessing
CPU and PID (referring to page_cpupid_xchg_last()). Which is used by the
multi-stage node selection algorithm to avoid to migrate pages shared
accessed by the NUMA nodes back and forth. But for pages in the slow
memory node, even if they are shared accessed by multiple NUMA nodes, as
long as the pages are hot, they need to be promoted to the fast memory
node. So the accessing CPU and PID information are unnecessary for the
slow memory pages. We can reuse these bits in struct page to record the
scan time. For the fast memory pages, these bits are used as before.
For the hot threshold, the default value is 1 second, which works well in
our performance test. All pages with hint page fault latency < hot
threshold will be considered hot.
It's hard for users to determine the hot threshold. So we don't provide a
kernel ABI to set it, just provide a debugfs interface for advanced users
to experiment. We will continue to work on a hot threshold automatic
adjustment mechanism.
The downside of the above method is that the response time to the workload
hot spot changing may be much longer. For example,
- A previous cold memory area becomes hot
- The hint page fault will be triggered. But the hint page fault
latency isn't shorter than the hot threshold. So the pages will
not be promoted.
- When the memory area is scanned again, maybe after a scan period,
the hint page fault latency measured will be shorter than the hot
threshold and the pages will be promoted.
To mitigate this, if there are enough free space in the fast memory node,
the hot threshold will not be used, all pages will be promoted upon the
hint page fault for fast response.
Thanks Zhong Jiang reported and tested the fix for a bug when disabling
memory tiering mode dynamically.
Link: https://lkml.kernel.org/r/20220713083954.34196-1-ying.huang@intel.com
Link: https://lkml.kernel.org/r/20220713083954.34196-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Wei Xu <weixugc@google.com>
Cc: osalvador <osalvador@suse.de>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Zhong Jiang <zhongjiang-ali@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-07-13 11:39:51 +03:00
bool toptier ;
2022-03-23 00:46:27 +03:00
/*
* Avoid trapping faults against the zero page . The read - only
* data is likely to be read - cached on the local CPU and
* local / remote hits to the zero page are not interesting .
*/
if ( is_huge_zero_pmd ( * pmd ) )
goto unlock ;
2012-03-22 03:33:57 +04:00
2022-03-23 00:46:27 +03:00
if ( pmd_protnone ( * pmd ) )
goto unlock ;
2017-04-14 00:56:17 +03:00
2022-03-23 00:46:27 +03:00
page = pmd_page ( * pmd ) ;
memory tiering: hot page selection with hint page fault latency
Patch series "memory tiering: hot page selection", v4.
To optimize page placement in a memory tiering system with NUMA balancing,
the hot pages in the slow memory nodes need to be identified.
Essentially, the original NUMA balancing implementation selects the mostly
recently accessed (MRU) pages to promote. But this isn't a perfect
algorithm to identify the hot pages. Because the pages with quite low
access frequency may be accessed eventually given the NUMA balancing page
table scanning period could be quite long (e.g. 60 seconds). So in this
patchset, we implement a new hot page identification algorithm based on
the latency between NUMA balancing page table scanning and hint page
fault. Which is a kind of mostly frequently accessed (MFU) algorithm.
In NUMA balancing memory tiering mode, if there are hot pages in slow
memory node and cold pages in fast memory node, we need to promote/demote
hot/cold pages between the fast and cold memory nodes.
A choice is to promote/demote as fast as possible. But the CPU cycles and
memory bandwidth consumed by the high promoting/demoting throughput will
hurt the latency of some workload because of accessing inflating and slow
memory bandwidth contention.
A way to resolve this issue is to restrict the max promoting/demoting
throughput. It will take longer to finish the promoting/demoting. But
the workload latency will be better. This is implemented in this patchset
as the page promotion rate limit mechanism.
The promotion hot threshold is workload and system configuration
dependent. So in this patchset, a method to adjust the hot threshold
automatically is implemented. The basic idea is to control the number of
the candidate promotion pages to match the promotion rate limit.
We used the pmbench memory accessing benchmark tested the patchset on a
2-socket server system with DRAM and PMEM installed. The test results are
as follows,
pmbench score promote rate
(accesses/s) MB/s
------------- ------------
base 146887704.1 725.6
hot selection 165695601.2 544.0
rate limit 162814569.8 165.2
auto adjustment 170495294.0 136.9
From the results above,
With hot page selection patch [1/3], the pmbench score increases about
12.8%, and promote rate (overhead) decreases about 25.0%, compared with
base kernel.
With rate limit patch [2/3], pmbench score decreases about 1.7%, and
promote rate decreases about 69.6%, compared with hot page selection
patch.
With threshold auto adjustment patch [3/3], pmbench score increases about
4.7%, and promote rate decrease about 17.1%, compared with rate limit
patch.
Baolin helped to test the patchset with MySQL on a machine which contains
1 DRAM node (30G) and 1 PMEM node (126G).
sysbench /usr/share/sysbench/oltp_read_write.lua \
......
--tables=200 \
--table-size=1000000 \
--report-interval=10 \
--threads=16 \
--time=120
The tps can be improved about 5%.
This patch (of 3):
To optimize page placement in a memory tiering system with NUMA balancing,
the hot pages in the slow memory node need to be identified. Essentially,
the original NUMA balancing implementation selects the mostly recently
accessed (MRU) pages to promote. But this isn't a perfect algorithm to
identify the hot pages. Because the pages with quite low access frequency
may be accessed eventually given the NUMA balancing page table scanning
period could be quite long (e.g. 60 seconds). The most frequently
accessed (MFU) algorithm is better.
So, in this patch we implemented a better hot page selection algorithm.
Which is based on NUMA balancing page table scanning and hint page fault
as follows,
- When the page tables of the processes are scanned to change PTE/PMD
to be PROT_NONE, the current time is recorded in struct page as scan
time.
- When the page is accessed, hint page fault will occur. The scan
time is gotten from the struct page. And The hint page fault
latency is defined as
hint page fault time - scan time
The shorter the hint page fault latency of a page is, the higher the
probability of their access frequency to be higher. So the hint page
fault latency is a better estimation of the page hot/cold.
It's hard to find some extra space in struct page to hold the scan time.
Fortunately, we can reuse some bits used by the original NUMA balancing.
NUMA balancing uses some bits in struct page to store the page accessing
CPU and PID (referring to page_cpupid_xchg_last()). Which is used by the
multi-stage node selection algorithm to avoid to migrate pages shared
accessed by the NUMA nodes back and forth. But for pages in the slow
memory node, even if they are shared accessed by multiple NUMA nodes, as
long as the pages are hot, they need to be promoted to the fast memory
node. So the accessing CPU and PID information are unnecessary for the
slow memory pages. We can reuse these bits in struct page to record the
scan time. For the fast memory pages, these bits are used as before.
For the hot threshold, the default value is 1 second, which works well in
our performance test. All pages with hint page fault latency < hot
threshold will be considered hot.
It's hard for users to determine the hot threshold. So we don't provide a
kernel ABI to set it, just provide a debugfs interface for advanced users
to experiment. We will continue to work on a hot threshold automatic
adjustment mechanism.
The downside of the above method is that the response time to the workload
hot spot changing may be much longer. For example,
- A previous cold memory area becomes hot
- The hint page fault will be triggered. But the hint page fault
latency isn't shorter than the hot threshold. So the pages will
not be promoted.
- When the memory area is scanned again, maybe after a scan period,
the hint page fault latency measured will be shorter than the hot
threshold and the pages will be promoted.
To mitigate this, if there are enough free space in the fast memory node,
the hot threshold will not be used, all pages will be promoted upon the
hint page fault for fast response.
Thanks Zhong Jiang reported and tested the fix for a bug when disabling
memory tiering mode dynamically.
Link: https://lkml.kernel.org/r/20220713083954.34196-1-ying.huang@intel.com
Link: https://lkml.kernel.org/r/20220713083954.34196-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Wei Xu <weixugc@google.com>
Cc: osalvador <osalvador@suse.de>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Zhong Jiang <zhongjiang-ali@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-07-13 11:39:51 +03:00
toptier = node_is_toptier ( page_to_nid ( page ) ) ;
2022-03-23 00:46:27 +03:00
/*
* Skip scanning top tier node if normal numa
* balancing is disabled
*/
if ( ! ( sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL ) & &
memory tiering: hot page selection with hint page fault latency
Patch series "memory tiering: hot page selection", v4.
To optimize page placement in a memory tiering system with NUMA balancing,
the hot pages in the slow memory nodes need to be identified.
Essentially, the original NUMA balancing implementation selects the mostly
recently accessed (MRU) pages to promote. But this isn't a perfect
algorithm to identify the hot pages. Because the pages with quite low
access frequency may be accessed eventually given the NUMA balancing page
table scanning period could be quite long (e.g. 60 seconds). So in this
patchset, we implement a new hot page identification algorithm based on
the latency between NUMA balancing page table scanning and hint page
fault. Which is a kind of mostly frequently accessed (MFU) algorithm.
In NUMA balancing memory tiering mode, if there are hot pages in slow
memory node and cold pages in fast memory node, we need to promote/demote
hot/cold pages between the fast and cold memory nodes.
A choice is to promote/demote as fast as possible. But the CPU cycles and
memory bandwidth consumed by the high promoting/demoting throughput will
hurt the latency of some workload because of accessing inflating and slow
memory bandwidth contention.
A way to resolve this issue is to restrict the max promoting/demoting
throughput. It will take longer to finish the promoting/demoting. But
the workload latency will be better. This is implemented in this patchset
as the page promotion rate limit mechanism.
The promotion hot threshold is workload and system configuration
dependent. So in this patchset, a method to adjust the hot threshold
automatically is implemented. The basic idea is to control the number of
the candidate promotion pages to match the promotion rate limit.
We used the pmbench memory accessing benchmark tested the patchset on a
2-socket server system with DRAM and PMEM installed. The test results are
as follows,
pmbench score promote rate
(accesses/s) MB/s
------------- ------------
base 146887704.1 725.6
hot selection 165695601.2 544.0
rate limit 162814569.8 165.2
auto adjustment 170495294.0 136.9
From the results above,
With hot page selection patch [1/3], the pmbench score increases about
12.8%, and promote rate (overhead) decreases about 25.0%, compared with
base kernel.
With rate limit patch [2/3], pmbench score decreases about 1.7%, and
promote rate decreases about 69.6%, compared with hot page selection
patch.
With threshold auto adjustment patch [3/3], pmbench score increases about
4.7%, and promote rate decrease about 17.1%, compared with rate limit
patch.
Baolin helped to test the patchset with MySQL on a machine which contains
1 DRAM node (30G) and 1 PMEM node (126G).
sysbench /usr/share/sysbench/oltp_read_write.lua \
......
--tables=200 \
--table-size=1000000 \
--report-interval=10 \
--threads=16 \
--time=120
The tps can be improved about 5%.
This patch (of 3):
To optimize page placement in a memory tiering system with NUMA balancing,
the hot pages in the slow memory node need to be identified. Essentially,
the original NUMA balancing implementation selects the mostly recently
accessed (MRU) pages to promote. But this isn't a perfect algorithm to
identify the hot pages. Because the pages with quite low access frequency
may be accessed eventually given the NUMA balancing page table scanning
period could be quite long (e.g. 60 seconds). The most frequently
accessed (MFU) algorithm is better.
So, in this patch we implemented a better hot page selection algorithm.
Which is based on NUMA balancing page table scanning and hint page fault
as follows,
- When the page tables of the processes are scanned to change PTE/PMD
to be PROT_NONE, the current time is recorded in struct page as scan
time.
- When the page is accessed, hint page fault will occur. The scan
time is gotten from the struct page. And The hint page fault
latency is defined as
hint page fault time - scan time
The shorter the hint page fault latency of a page is, the higher the
probability of their access frequency to be higher. So the hint page
fault latency is a better estimation of the page hot/cold.
It's hard to find some extra space in struct page to hold the scan time.
Fortunately, we can reuse some bits used by the original NUMA balancing.
NUMA balancing uses some bits in struct page to store the page accessing
CPU and PID (referring to page_cpupid_xchg_last()). Which is used by the
multi-stage node selection algorithm to avoid to migrate pages shared
accessed by the NUMA nodes back and forth. But for pages in the slow
memory node, even if they are shared accessed by multiple NUMA nodes, as
long as the pages are hot, they need to be promoted to the fast memory
node. So the accessing CPU and PID information are unnecessary for the
slow memory pages. We can reuse these bits in struct page to record the
scan time. For the fast memory pages, these bits are used as before.
For the hot threshold, the default value is 1 second, which works well in
our performance test. All pages with hint page fault latency < hot
threshold will be considered hot.
It's hard for users to determine the hot threshold. So we don't provide a
kernel ABI to set it, just provide a debugfs interface for advanced users
to experiment. We will continue to work on a hot threshold automatic
adjustment mechanism.
The downside of the above method is that the response time to the workload
hot spot changing may be much longer. For example,
- A previous cold memory area becomes hot
- The hint page fault will be triggered. But the hint page fault
latency isn't shorter than the hot threshold. So the pages will
not be promoted.
- When the memory area is scanned again, maybe after a scan period,
the hint page fault latency measured will be shorter than the hot
threshold and the pages will be promoted.
To mitigate this, if there are enough free space in the fast memory node,
the hot threshold will not be used, all pages will be promoted upon the
hint page fault for fast response.
Thanks Zhong Jiang reported and tested the fix for a bug when disabling
memory tiering mode dynamically.
Link: https://lkml.kernel.org/r/20220713083954.34196-1-ying.huang@intel.com
Link: https://lkml.kernel.org/r/20220713083954.34196-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Wei Xu <weixugc@google.com>
Cc: osalvador <osalvador@suse.de>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Zhong Jiang <zhongjiang-ali@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-07-13 11:39:51 +03:00
toptier )
2022-03-23 00:46:27 +03:00
goto unlock ;
memory tiering: hot page selection with hint page fault latency
Patch series "memory tiering: hot page selection", v4.
To optimize page placement in a memory tiering system with NUMA balancing,
the hot pages in the slow memory nodes need to be identified.
Essentially, the original NUMA balancing implementation selects the mostly
recently accessed (MRU) pages to promote. But this isn't a perfect
algorithm to identify the hot pages. Because the pages with quite low
access frequency may be accessed eventually given the NUMA balancing page
table scanning period could be quite long (e.g. 60 seconds). So in this
patchset, we implement a new hot page identification algorithm based on
the latency between NUMA balancing page table scanning and hint page
fault. Which is a kind of mostly frequently accessed (MFU) algorithm.
In NUMA balancing memory tiering mode, if there are hot pages in slow
memory node and cold pages in fast memory node, we need to promote/demote
hot/cold pages between the fast and cold memory nodes.
A choice is to promote/demote as fast as possible. But the CPU cycles and
memory bandwidth consumed by the high promoting/demoting throughput will
hurt the latency of some workload because of accessing inflating and slow
memory bandwidth contention.
A way to resolve this issue is to restrict the max promoting/demoting
throughput. It will take longer to finish the promoting/demoting. But
the workload latency will be better. This is implemented in this patchset
as the page promotion rate limit mechanism.
The promotion hot threshold is workload and system configuration
dependent. So in this patchset, a method to adjust the hot threshold
automatically is implemented. The basic idea is to control the number of
the candidate promotion pages to match the promotion rate limit.
We used the pmbench memory accessing benchmark tested the patchset on a
2-socket server system with DRAM and PMEM installed. The test results are
as follows,
pmbench score promote rate
(accesses/s) MB/s
------------- ------------
base 146887704.1 725.6
hot selection 165695601.2 544.0
rate limit 162814569.8 165.2
auto adjustment 170495294.0 136.9
From the results above,
With hot page selection patch [1/3], the pmbench score increases about
12.8%, and promote rate (overhead) decreases about 25.0%, compared with
base kernel.
With rate limit patch [2/3], pmbench score decreases about 1.7%, and
promote rate decreases about 69.6%, compared with hot page selection
patch.
With threshold auto adjustment patch [3/3], pmbench score increases about
4.7%, and promote rate decrease about 17.1%, compared with rate limit
patch.
Baolin helped to test the patchset with MySQL on a machine which contains
1 DRAM node (30G) and 1 PMEM node (126G).
sysbench /usr/share/sysbench/oltp_read_write.lua \
......
--tables=200 \
--table-size=1000000 \
--report-interval=10 \
--threads=16 \
--time=120
The tps can be improved about 5%.
This patch (of 3):
To optimize page placement in a memory tiering system with NUMA balancing,
the hot pages in the slow memory node need to be identified. Essentially,
the original NUMA balancing implementation selects the mostly recently
accessed (MRU) pages to promote. But this isn't a perfect algorithm to
identify the hot pages. Because the pages with quite low access frequency
may be accessed eventually given the NUMA balancing page table scanning
period could be quite long (e.g. 60 seconds). The most frequently
accessed (MFU) algorithm is better.
So, in this patch we implemented a better hot page selection algorithm.
Which is based on NUMA balancing page table scanning and hint page fault
as follows,
- When the page tables of the processes are scanned to change PTE/PMD
to be PROT_NONE, the current time is recorded in struct page as scan
time.
- When the page is accessed, hint page fault will occur. The scan
time is gotten from the struct page. And The hint page fault
latency is defined as
hint page fault time - scan time
The shorter the hint page fault latency of a page is, the higher the
probability of their access frequency to be higher. So the hint page
fault latency is a better estimation of the page hot/cold.
It's hard to find some extra space in struct page to hold the scan time.
Fortunately, we can reuse some bits used by the original NUMA balancing.
NUMA balancing uses some bits in struct page to store the page accessing
CPU and PID (referring to page_cpupid_xchg_last()). Which is used by the
multi-stage node selection algorithm to avoid to migrate pages shared
accessed by the NUMA nodes back and forth. But for pages in the slow
memory node, even if they are shared accessed by multiple NUMA nodes, as
long as the pages are hot, they need to be promoted to the fast memory
node. So the accessing CPU and PID information are unnecessary for the
slow memory pages. We can reuse these bits in struct page to record the
scan time. For the fast memory pages, these bits are used as before.
For the hot threshold, the default value is 1 second, which works well in
our performance test. All pages with hint page fault latency < hot
threshold will be considered hot.
It's hard for users to determine the hot threshold. So we don't provide a
kernel ABI to set it, just provide a debugfs interface for advanced users
to experiment. We will continue to work on a hot threshold automatic
adjustment mechanism.
The downside of the above method is that the response time to the workload
hot spot changing may be much longer. For example,
- A previous cold memory area becomes hot
- The hint page fault will be triggered. But the hint page fault
latency isn't shorter than the hot threshold. So the pages will
not be promoted.
- When the memory area is scanned again, maybe after a scan period,
the hint page fault latency measured will be shorter than the hot
threshold and the pages will be promoted.
To mitigate this, if there are enough free space in the fast memory node,
the hot threshold will not be used, all pages will be promoted upon the
hint page fault for fast response.
Thanks Zhong Jiang reported and tested the fix for a bug when disabling
memory tiering mode dynamically.
Link: https://lkml.kernel.org/r/20220713083954.34196-1-ying.huang@intel.com
Link: https://lkml.kernel.org/r/20220713083954.34196-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Tested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Wei Xu <weixugc@google.com>
Cc: osalvador <osalvador@suse.de>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Zhong Jiang <zhongjiang-ali@linux.alibaba.com>
Cc: Oscar Salvador <osalvador@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-07-13 11:39:51 +03:00
if ( sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING & &
! toptier )
xchg_page_access_time ( page , jiffies_to_msecs ( jiffies ) ) ;
2022-03-23 00:46:27 +03:00
}
2017-04-14 00:56:20 +03:00
/*
2020-06-09 07:33:51 +03:00
* In case prot_numa , we are under mmap_read_lock ( mm ) . It ' s critical
2017-04-14 00:56:20 +03:00
* to not clear pmd intermittently to avoid race with MADV_DONTNEED
2020-06-09 07:33:51 +03:00
* which is also under mmap_read_lock ( mm ) :
2017-04-14 00:56:20 +03:00
*
* CPU0 : CPU1 :
* change_huge_pmd ( prot_numa = 1 )
* pmdp_huge_get_and_clear_notify ( )
* madvise_dontneed ( )
* zap_pmd_range ( )
* pmd_trans_huge ( * pmd ) = = 0 ( without ptl )
* // skip the pmd
* set_pmd_at ( ) ;
* // pmd is re-established
*
* The race makes MADV_DONTNEED miss the huge pmd and don ' t clear it
* which may break userspace .
*
2022-05-10 04:20:50 +03:00
* pmdp_invalidate_ad ( ) is required to make sure we don ' t miss
2017-04-14 00:56:20 +03:00
* dirty / young flags set by hardware .
*/
2022-05-10 04:20:50 +03:00
oldpmd = pmdp_invalidate_ad ( vma , addr , pmd ) ;
2017-04-14 00:56:20 +03:00
2022-05-10 04:20:50 +03:00
entry = pmd_modify ( oldpmd , newprot ) ;
userfaultfd: wp: apply _PAGE_UFFD_WP bit
Firstly, introduce two new flags MM_CP_UFFD_WP[_RESOLVE] for
change_protection() when used with uffd-wp and make sure the two new flags
are exclusively used. Then,
- For MM_CP_UFFD_WP: apply the _PAGE_UFFD_WP bit and remove _PAGE_RW
when a range of memory is write protected by uffd
- For MM_CP_UFFD_WP_RESOLVE: remove the _PAGE_UFFD_WP bit and recover
_PAGE_RW when write protection is resolved from userspace
And use this new interface in mwriteprotect_range() to replace the old
MM_CP_DIRTY_ACCT.
Do this change for both PTEs and huge PMDs. Then we can start to identify
which PTE/PMD is write protected by general (e.g., COW or soft dirty
tracking), and which is for userfaultfd-wp.
Since we should keep the _PAGE_UFFD_WP when doing pte_modify(), add it
into _PAGE_CHG_MASK as well. Meanwhile, since we have this new bit, we
can be even more strict when detecting uffd-wp page faults in either
do_wp_page() or wp_huge_pmd().
After we're with _PAGE_UFFD_WP, a special case is when a page is both
protected by the general COW logic and also userfault-wp. Here the
userfault-wp will have higher priority and will be handled first. Only
after the uffd-wp bit is cleared on the PTE/PMD will we continue to handle
the general COW. These are the steps on what will happen with such a
page:
1. CPU accesses write protected shared page (so both protected by
general COW and uffd-wp), blocked by uffd-wp first because in
do_wp_page we'll handle uffd-wp first, so it has higher priority
than general COW.
2. Uffd service thread receives the request, do UFFDIO_WRITEPROTECT
to remove the uffd-wp bit upon the PTE/PMD. However here we
still keep the write bit cleared. Notify the blocked CPU.
3. The blocked CPU resumes the page fault process with a fault
retry, during retry it'll notice it was not with the uffd-wp bit
this time but it is still write protected by general COW, then
it'll go though the COW path in the fault handler, copy the page,
apply write bit where necessary, and retry again.
4. The CPU will be able to access this page with write bit set.
Suggested-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Brian Geffon <bgeffon@google.com>
Cc: Pavel Emelyanov <xemul@openvz.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Martin Cracauer <cracauer@cons.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Bobby Powers <bobbypowers@gmail.com>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: "Kirill A . Shutemov" <kirill@shutemov.name>
Cc: Maya Gokhale <gokhale2@llnl.gov>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Marty McFadden <mcfadden8@llnl.gov>
Cc: Denis Plotnikov <dplotnikov@virtuozzo.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Shaohua Li <shli@fb.com>
Link: http://lkml.kernel.org/r/20200220163112.11409-8-peterx@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-07 06:05:49 +03:00
if ( uffd_wp ) {
entry = pmd_wrprotect ( entry ) ;
entry = pmd_mkuffd_wp ( entry ) ;
} else if ( uffd_wp_resolve ) {
/*
* Leave the write bit to be handled by PF interrupt
* handler , then things like COW could be properly
* handled .
*/
entry = pmd_clear_uffd_wp ( entry ) ;
}
2022-11-08 20:46:48 +03:00
/* See change_pte_range(). */
if ( ( cp_flags & MM_CP_TRY_CHANGE_WRITABLE ) & & ! pmd_write ( entry ) & &
can_change_pmd_writable ( vma , addr , entry ) )
entry = pmd_mkwrite ( entry ) ;
2017-04-14 00:56:17 +03:00
ret = HPAGE_PMD_NR ;
set_pmd_at ( mm , addr , pmd , entry ) ;
mm/mprotect: use mmu_gather
Patch series "mm/mprotect: avoid unnecessary TLB flushes", v6.
This patchset is intended to remove unnecessary TLB flushes during
mprotect() syscalls. Once this patch-set make it through, similar and
further optimizations for MADV_COLD and userfaultfd would be possible.
Basically, there are 3 optimizations in this patch-set:
1. Use TLB batching infrastructure to batch flushes across VMAs and do
better/fewer flushes. This would also be handy for later userfaultfd
enhancements.
2. Avoid unnecessary TLB flushes. This optimization is the one that
provides most of the performance benefits. Unlike previous versions,
we now only avoid flushes that would not result in spurious
page-faults.
3. Avoiding TLB flushes on change_huge_pmd() that are only needed to
prevent the A/D bits from changing.
Andrew asked for some benchmark numbers. I do not have an easy
determinate macrobenchmark in which it is easy to show benefit. I
therefore ran a microbenchmark: a loop that does the following on
anonymous memory, just as a sanity check to see that time is saved by
avoiding TLB flushes. The loop goes:
mprotect(p, PAGE_SIZE, PROT_READ)
mprotect(p, PAGE_SIZE, PROT_READ|PROT_WRITE)
*p = 0; // make the page writable
The test was run in KVM guest with 1 or 2 threads (the second thread was
busy-looping). I measured the time (cycles) of each operation:
1 thread 2 threads
mmots +patch mmots +patch
PROT_READ 3494 2725 (-22%) 8630 7788 (-10%)
PROT_READ|WRITE 3952 2724 (-31%) 9075 2865 (-68%)
[ mmots = v5.17-rc6-mmots-2022-03-06-20-38 ]
The exact numbers are really meaningless, but the benefit is clear. There
are 2 interesting results though.
(1) PROT_READ is cheaper, while one can expect it not to be affected.
This is presumably due to TLB miss that is saved
(2) Without memory access (*p = 0), the speedup of the patch is even
greater. In that scenario mprotect(PROT_READ) also avoids the TLB flush.
As a result both operations on the patched kernel take roughly ~1500
cycles (with either 1 or 2 threads), whereas on mmotm their cost is as
high as presented in the table.
This patch (of 3):
change_pXX_range() currently does not use mmu_gather, but instead
implements its own deferred TLB flushes scheme. This both complicates the
code, as developers need to be aware of different invalidation schemes,
and prevents opportunities to avoid TLB flushes or perform them in finer
granularity.
The use of mmu_gather for modified PTEs has benefits in various scenarios
even if pages are not released. For instance, if only a single page needs
to be flushed out of a range of many pages, only that page would be
flushed. If a THP page is flushed, on x86 a single TLB invlpg instruction
can be used instead of 512 instructions (or a full TLB flush, which would
Linux would actually use by default). mprotect() over multiple VMAs
requires a single flush.
Use mmu_gather in change_pXX_range(). As the pages are not released, only
record the flushed range using tlb_flush_pXX_range().
Handle THP similarly and get rid of flush_cache_range() which becomes
redundant since tlb_start_vma() calls it when needed.
Link: https://lkml.kernel.org/r/20220401180821.1986781-1-namit@vmware.com
Link: https://lkml.kernel.org/r/20220401180821.1986781-2-namit@vmware.com
Signed-off-by: Nadav Amit <namit@vmware.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrew Cooper <andrew.cooper3@citrix.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Will Deacon <will@kernel.org>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Nick Piggin <npiggin@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:50 +03:00
2022-05-10 04:20:50 +03:00
if ( huge_pmd_needs_flush ( oldpmd , entry ) )
tlb_flush_pmd_range ( tlb , addr , HPAGE_PMD_SIZE ) ;
2017-04-14 00:56:17 +03:00
unlock :
spin_unlock ( ptl ) ;
2012-03-22 03:33:57 +04:00
return ret ;
}
/*
2016-07-27 01:27:04 +03:00
* Returns page table lock pointer if a given pmd maps a thp , NULL otherwise .
2012-03-22 03:33:57 +04:00
*
2016-07-27 01:27:04 +03:00
* Note that if it returns page table lock pointer , this routine returns without
* unlocking page table lock . So callers must unlock it .
2012-03-22 03:33:57 +04:00
*/
2016-01-22 03:40:25 +03:00
spinlock_t * __pmd_trans_huge_lock ( pmd_t * pmd , struct vm_area_struct * vma )
2012-03-22 03:33:57 +04:00
{
2016-01-22 03:40:25 +03:00
spinlock_t * ptl ;
ptl = pmd_lock ( vma - > vm_mm , pmd ) ;
mm: thp: check pmd migration entry in common path
When THP migration is being used, memory management code needs to handle
pmd migration entries properly. This patch uses !pmd_present() or
is_swap_pmd() (depending on whether pmd_none() needs separate code or
not) to check pmd migration entries at the places where a pmd entry is
present.
Since pmd-related code uses split_huge_page(), split_huge_pmd(),
pmd_trans_huge(), pmd_trans_unstable(), or
pmd_none_or_trans_huge_or_clear_bad(), this patch:
1. adds pmd migration entry split code in split_huge_pmd(),
2. takes care of pmd migration entries whenever pmd_trans_huge() is present,
3. makes pmd_none_or_trans_huge_or_clear_bad() pmd migration entry aware.
Since split_huge_page() uses split_huge_pmd() and pmd_trans_unstable()
is equivalent to pmd_none_or_trans_huge_or_clear_bad(), we do not change
them.
Until this commit, a pmd entry should be:
1. pointing to a pte page,
2. is_swap_pmd(),
3. pmd_trans_huge(),
4. pmd_devmap(), or
5. pmd_none().
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-09 02:11:01 +03:00
if ( likely ( is_swap_pmd ( * pmd ) | | pmd_trans_huge ( * pmd ) | |
pmd_devmap ( * pmd ) ) )
2016-01-22 03:40:25 +03:00
return ptl ;
spin_unlock ( ptl ) ;
return NULL ;
2011-01-14 02:47:04 +03:00
}
2017-02-25 01:57:02 +03:00
/*
2022-07-04 16:21:48 +03:00
* Returns page table lock pointer if a given pud maps a thp , NULL otherwise .
2017-02-25 01:57:02 +03:00
*
2022-07-04 16:21:48 +03:00
* Note that if it returns page table lock pointer , this routine returns without
* unlocking page table lock . So callers must unlock it .
2017-02-25 01:57:02 +03:00
*/
spinlock_t * __pud_trans_huge_lock ( pud_t * pud , struct vm_area_struct * vma )
{
spinlock_t * ptl ;
ptl = pud_lock ( vma - > vm_mm , pud ) ;
if ( likely ( pud_trans_huge ( * pud ) | | pud_devmap ( * pud ) ) )
return ptl ;
spin_unlock ( ptl ) ;
return NULL ;
}
# ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
int zap_huge_pud ( struct mmu_gather * tlb , struct vm_area_struct * vma ,
pud_t * pud , unsigned long addr )
{
spinlock_t * ptl ;
ptl = __pud_trans_huge_lock ( pud , vma ) ;
if ( ! ptl )
return 0 ;
2022-07-04 16:21:54 +03:00
2019-03-06 02:50:00 +03:00
pudp_huge_get_and_clear_full ( tlb - > mm , addr , pud , tlb - > fullmm ) ;
2017-02-25 01:57:02 +03:00
tlb_remove_pud_tlb_entry ( tlb , pud , addr ) ;
2020-03-24 20:47:17 +03:00
if ( vma_is_special_huge ( vma ) ) {
2017-02-25 01:57:02 +03:00
spin_unlock ( ptl ) ;
/* No zero page support yet */
} else {
/* No support for anonymous PUD pages yet */
BUG ( ) ;
}
return 1 ;
}
static void __split_huge_pud_locked ( struct vm_area_struct * vma , pud_t * pud ,
unsigned long haddr )
{
VM_BUG_ON ( haddr & ~ HPAGE_PUD_MASK ) ;
VM_BUG_ON_VMA ( vma - > vm_start > haddr , vma ) ;
VM_BUG_ON_VMA ( vma - > vm_end < haddr + HPAGE_PUD_SIZE , vma ) ;
VM_BUG_ON ( ! pud_trans_huge ( * pud ) & & ! pud_devmap ( * pud ) ) ;
2017-03-10 03:17:00 +03:00
count_vm_event ( THP_SPLIT_PUD ) ;
2017-02-25 01:57:02 +03:00
pudp_huge_clear_flush_notify ( vma , haddr , pud ) ;
}
void __split_huge_pud ( struct vm_area_struct * vma , pud_t * pud ,
unsigned long address )
{
spinlock_t * ptl ;
2018-12-28 11:38:09 +03:00
struct mmu_notifier_range range ;
2017-02-25 01:57:02 +03:00
2019-05-14 03:20:53 +03:00
mmu_notifier_range_init ( & range , MMU_NOTIFY_CLEAR , 0 , vma , vma - > vm_mm ,
mm/mmu_notifier: contextual information for event triggering invalidation
CPU page table update can happens for many reasons, not only as a result
of a syscall (munmap(), mprotect(), mremap(), madvise(), ...) but also as
a result of kernel activities (memory compression, reclaim, migration,
...).
Users of mmu notifier API track changes to the CPU page table and take
specific action for them. While current API only provide range of virtual
address affected by the change, not why the changes is happening.
This patchset do the initial mechanical convertion of all the places that
calls mmu_notifier_range_init to also provide the default MMU_NOTIFY_UNMAP
event as well as the vma if it is know (most invalidation happens against
a given vma). Passing down the vma allows the users of mmu notifier to
inspect the new vma page protection.
The MMU_NOTIFY_UNMAP is always the safe default as users of mmu notifier
should assume that every for the range is going away when that event
happens. A latter patch do convert mm call path to use a more appropriate
events for each call.
This is done as 2 patches so that no call site is forgotten especialy
as it uses this following coccinelle patch:
%<----------------------------------------------------------------------
@@
identifier I1, I2, I3, I4;
@@
static inline void mmu_notifier_range_init(struct mmu_notifier_range *I1,
+enum mmu_notifier_event event,
+unsigned flags,
+struct vm_area_struct *vma,
struct mm_struct *I2, unsigned long I3, unsigned long I4) { ... }
@@
@@
-#define mmu_notifier_range_init(range, mm, start, end)
+#define mmu_notifier_range_init(range, event, flags, vma, mm, start, end)
@@
expression E1, E3, E4;
identifier I1;
@@
<...
mmu_notifier_range_init(E1,
+MMU_NOTIFY_UNMAP, 0, I1,
I1->vm_mm, E3, E4)
...>
@@
expression E1, E2, E3, E4;
identifier FN, VMA;
@@
FN(..., struct vm_area_struct *VMA, ...) {
<...
mmu_notifier_range_init(E1,
+MMU_NOTIFY_UNMAP, 0, VMA,
E2, E3, E4)
...> }
@@
expression E1, E2, E3, E4;
identifier FN, VMA;
@@
FN(...) {
struct vm_area_struct *VMA;
<...
mmu_notifier_range_init(E1,
+MMU_NOTIFY_UNMAP, 0, VMA,
E2, E3, E4)
...> }
@@
expression E1, E2, E3, E4;
identifier FN;
@@
FN(...) {
<...
mmu_notifier_range_init(E1,
+MMU_NOTIFY_UNMAP, 0, NULL,
E2, E3, E4)
...> }
---------------------------------------------------------------------->%
Applied with:
spatch --all-includes --sp-file mmu-notifier.spatch fs/proc/task_mmu.c --in-place
spatch --sp-file mmu-notifier.spatch --dir kernel/events/ --in-place
spatch --sp-file mmu-notifier.spatch --dir mm --in-place
Link: http://lkml.kernel.org/r/20190326164747.24405-6-jglisse@redhat.com
Signed-off-by: Jérôme Glisse <jglisse@redhat.com>
Reviewed-by: Ralph Campbell <rcampbell@nvidia.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Cc: Christian König <christian.koenig@amd.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Jani Nikula <jani.nikula@linux.intel.com>
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Felix Kuehling <Felix.Kuehling@amd.com>
Cc: Jason Gunthorpe <jgg@mellanox.com>
Cc: Ross Zwisler <zwisler@kernel.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krcmar <rkrcmar@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Christian Koenig <christian.koenig@amd.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 03:20:49 +03:00
address & HPAGE_PUD_MASK ,
2018-12-28 11:38:09 +03:00
( address & HPAGE_PUD_MASK ) + HPAGE_PUD_SIZE ) ;
mmu_notifier_invalidate_range_start ( & range ) ;
ptl = pud_lock ( vma - > vm_mm , pud ) ;
2017-02-25 01:57:02 +03:00
if ( unlikely ( ! pud_trans_huge ( * pud ) & & ! pud_devmap ( * pud ) ) )
goto out ;
2018-12-28 11:38:09 +03:00
__split_huge_pud_locked ( vma , pud , range . start ) ;
2017-02-25 01:57:02 +03:00
out :
spin_unlock ( ptl ) ;
2017-11-16 04:34:11 +03:00
/*
* No need to double call mmu_notifier - > invalidate_range ( ) callback as
* the above pudp_huge_clear_flush_notify ( ) did already call it .
*/
2018-12-28 11:38:09 +03:00
mmu_notifier_invalidate_range_only_end ( & range ) ;
2017-02-25 01:57:02 +03:00
}
# endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2016-01-16 03:53:53 +03:00
static void __split_huge_zero_page_pmd ( struct vm_area_struct * vma ,
unsigned long haddr , pmd_t * pmd )
{
struct mm_struct * mm = vma - > vm_mm ;
pgtable_t pgtable ;
pmd_t _pmd ;
int i ;
2017-11-16 04:34:07 +03:00
/*
* Leave pmd empty until pte is filled note that it is fine to delay
* notification until mmu_notifier_invalidate_range_end ( ) as we are
* replacing a zero pmd write protected page with a zero pte write
* protected page .
*
2022-06-27 09:00:26 +03:00
* See Documentation / mm / mmu_notifier . rst
2017-11-16 04:34:07 +03:00
*/
pmdp_huge_clear_flush ( vma , haddr , pmd ) ;
2016-01-16 03:53:53 +03:00
pgtable = pgtable_trans_huge_withdraw ( mm , pmd ) ;
pmd_populate ( mm , & _pmd , pgtable ) ;
for ( i = 0 ; i < HPAGE_PMD_NR ; i + + , haddr + = PAGE_SIZE ) {
pte_t * pte , entry ;
entry = pfn_pte ( my_zero_pfn ( haddr ) , vma - > vm_page_prot ) ;
entry = pte_mkspecial ( entry ) ;
pte = pte_offset_map ( & _pmd , haddr ) ;
VM_BUG_ON ( ! pte_none ( * pte ) ) ;
set_pte_at ( mm , haddr , pte , entry ) ;
pte_unmap ( pte ) ;
}
smp_wmb ( ) ; /* make pte visible before pmd */
pmd_populate ( mm , pmd , pgtable ) ;
}
static void __split_huge_pmd_locked ( struct vm_area_struct * vma , pmd_t * pmd ,
2016-01-16 03:53:56 +03:00
unsigned long haddr , bool freeze )
2016-01-16 03:53:53 +03:00
{
struct mm_struct * mm = vma - > vm_mm ;
struct page * page ;
pgtable_t pgtable ;
2018-02-01 03:18:24 +03:00
pmd_t old_pmd , _pmd ;
userfaultfd: wp: apply _PAGE_UFFD_WP bit
Firstly, introduce two new flags MM_CP_UFFD_WP[_RESOLVE] for
change_protection() when used with uffd-wp and make sure the two new flags
are exclusively used. Then,
- For MM_CP_UFFD_WP: apply the _PAGE_UFFD_WP bit and remove _PAGE_RW
when a range of memory is write protected by uffd
- For MM_CP_UFFD_WP_RESOLVE: remove the _PAGE_UFFD_WP bit and recover
_PAGE_RW when write protection is resolved from userspace
And use this new interface in mwriteprotect_range() to replace the old
MM_CP_DIRTY_ACCT.
Do this change for both PTEs and huge PMDs. Then we can start to identify
which PTE/PMD is write protected by general (e.g., COW or soft dirty
tracking), and which is for userfaultfd-wp.
Since we should keep the _PAGE_UFFD_WP when doing pte_modify(), add it
into _PAGE_CHG_MASK as well. Meanwhile, since we have this new bit, we
can be even more strict when detecting uffd-wp page faults in either
do_wp_page() or wp_huge_pmd().
After we're with _PAGE_UFFD_WP, a special case is when a page is both
protected by the general COW logic and also userfault-wp. Here the
userfault-wp will have higher priority and will be handled first. Only
after the uffd-wp bit is cleared on the PTE/PMD will we continue to handle
the general COW. These are the steps on what will happen with such a
page:
1. CPU accesses write protected shared page (so both protected by
general COW and uffd-wp), blocked by uffd-wp first because in
do_wp_page we'll handle uffd-wp first, so it has higher priority
than general COW.
2. Uffd service thread receives the request, do UFFDIO_WRITEPROTECT
to remove the uffd-wp bit upon the PTE/PMD. However here we
still keep the write bit cleared. Notify the blocked CPU.
3. The blocked CPU resumes the page fault process with a fault
retry, during retry it'll notice it was not with the uffd-wp bit
this time but it is still write protected by general COW, then
it'll go though the COW path in the fault handler, copy the page,
apply write bit where necessary, and retry again.
4. The CPU will be able to access this page with write bit set.
Suggested-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Brian Geffon <bgeffon@google.com>
Cc: Pavel Emelyanov <xemul@openvz.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Martin Cracauer <cracauer@cons.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Bobby Powers <bobbypowers@gmail.com>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: "Kirill A . Shutemov" <kirill@shutemov.name>
Cc: Maya Gokhale <gokhale2@llnl.gov>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Marty McFadden <mcfadden8@llnl.gov>
Cc: Denis Plotnikov <dplotnikov@virtuozzo.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Shaohua Li <shli@fb.com>
Link: http://lkml.kernel.org/r/20200220163112.11409-8-peterx@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-07 06:05:49 +03:00
bool young , write , soft_dirty , pmd_migration = false , uffd_wp = false ;
2022-08-11 19:13:28 +03:00
bool anon_exclusive = false , dirty = false ;
2016-02-24 18:58:03 +03:00
unsigned long addr ;
2016-01-16 03:53:53 +03:00
int i ;
VM_BUG_ON ( haddr & ~ HPAGE_PMD_MASK ) ;
VM_BUG_ON_VMA ( vma - > vm_start > haddr , vma ) ;
VM_BUG_ON_VMA ( vma - > vm_end < haddr + HPAGE_PMD_SIZE , vma ) ;
mm: thp: check pmd migration entry in common path
When THP migration is being used, memory management code needs to handle
pmd migration entries properly. This patch uses !pmd_present() or
is_swap_pmd() (depending on whether pmd_none() needs separate code or
not) to check pmd migration entries at the places where a pmd entry is
present.
Since pmd-related code uses split_huge_page(), split_huge_pmd(),
pmd_trans_huge(), pmd_trans_unstable(), or
pmd_none_or_trans_huge_or_clear_bad(), this patch:
1. adds pmd migration entry split code in split_huge_pmd(),
2. takes care of pmd migration entries whenever pmd_trans_huge() is present,
3. makes pmd_none_or_trans_huge_or_clear_bad() pmd migration entry aware.
Since split_huge_page() uses split_huge_pmd() and pmd_trans_unstable()
is equivalent to pmd_none_or_trans_huge_or_clear_bad(), we do not change
them.
Until this commit, a pmd entry should be:
1. pointing to a pte page,
2. is_swap_pmd(),
3. pmd_trans_huge(),
4. pmd_devmap(), or
5. pmd_none().
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-09 02:11:01 +03:00
VM_BUG_ON ( ! is_pmd_migration_entry ( * pmd ) & & ! pmd_trans_huge ( * pmd )
& & ! pmd_devmap ( * pmd ) ) ;
2016-01-16 03:53:53 +03:00
count_vm_event ( THP_SPLIT_PMD ) ;
2016-07-27 01:25:37 +03:00
if ( ! vma_is_anonymous ( vma ) ) {
mm/thp: fix __split_huge_pmd_locked() on shmem migration entry
Patch series "mm/thp: fix THP splitting unmap BUGs and related", v10.
Here is v2 batch of long-standing THP bug fixes that I had not got
around to sending before, but prompted now by Wang Yugui's report
https://lore.kernel.org/linux-mm/20210412180659.B9E3.409509F4@e16-tech.com/
Wang Yugui has tested a rollup of these fixes applied to 5.10.39, and
they have done no harm, but have *not* fixed that issue: something more
is needed and I have no idea of what.
This patch (of 7):
Stressing huge tmpfs page migration racing hole punch often crashed on
the VM_BUG_ON(!pmd_present) in pmdp_huge_clear_flush(), with DEBUG_VM=y
kernel; or shortly afterwards, on a bad dereference in
__split_huge_pmd_locked() when DEBUG_VM=n. They forgot to allow for pmd
migration entries in the non-anonymous case.
Full disclosure: those particular experiments were on a kernel with more
relaxed mmap_lock and i_mmap_rwsem locking, and were not repeated on the
vanilla kernel: it is conceivable that stricter locking happens to avoid
those cases, or makes them less likely; but __split_huge_pmd_locked()
already allowed for pmd migration entries when handling anonymous THPs,
so this commit brings the shmem and file THP handling into line.
And while there: use old_pmd rather than _pmd, as in the following
blocks; and make it clearer to the eye that the !vma_is_anonymous()
block is self-contained, making an early return after accounting for
unmapping.
Link: https://lkml.kernel.org/r/af88612-1473-2eaa-903-8d1a448b26@google.com
Link: https://lkml.kernel.org/r/dd221a99-efb3-cd1d-6256-7e646af29314@google.com
Fixes: e71769ae5260 ("mm: enable thp migration for shmem thp")
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Wang Yugui <wangyugui@e16-tech.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Ralph Campbell <rcampbell@nvidia.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Jue Wang <juew@google.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-16 04:23:45 +03:00
old_pmd = pmdp_huge_clear_flush_notify ( vma , haddr , pmd ) ;
2016-12-13 03:44:32 +03:00
/*
* We are going to unmap this huge page . So
* just go ahead and zap it
*/
if ( arch_needs_pgtable_deposit ( ) )
zap_deposited_table ( mm , pmd ) ;
2020-03-24 20:47:17 +03:00
if ( vma_is_special_huge ( vma ) )
2016-07-27 01:25:37 +03:00
return ;
mm/thp: fix __split_huge_pmd_locked() on shmem migration entry
Patch series "mm/thp: fix THP splitting unmap BUGs and related", v10.
Here is v2 batch of long-standing THP bug fixes that I had not got
around to sending before, but prompted now by Wang Yugui's report
https://lore.kernel.org/linux-mm/20210412180659.B9E3.409509F4@e16-tech.com/
Wang Yugui has tested a rollup of these fixes applied to 5.10.39, and
they have done no harm, but have *not* fixed that issue: something more
is needed and I have no idea of what.
This patch (of 7):
Stressing huge tmpfs page migration racing hole punch often crashed on
the VM_BUG_ON(!pmd_present) in pmdp_huge_clear_flush(), with DEBUG_VM=y
kernel; or shortly afterwards, on a bad dereference in
__split_huge_pmd_locked() when DEBUG_VM=n. They forgot to allow for pmd
migration entries in the non-anonymous case.
Full disclosure: those particular experiments were on a kernel with more
relaxed mmap_lock and i_mmap_rwsem locking, and were not repeated on the
vanilla kernel: it is conceivable that stricter locking happens to avoid
those cases, or makes them less likely; but __split_huge_pmd_locked()
already allowed for pmd migration entries when handling anonymous THPs,
so this commit brings the shmem and file THP handling into line.
And while there: use old_pmd rather than _pmd, as in the following
blocks; and make it clearer to the eye that the !vma_is_anonymous()
block is self-contained, making an early return after accounting for
unmapping.
Link: https://lkml.kernel.org/r/af88612-1473-2eaa-903-8d1a448b26@google.com
Link: https://lkml.kernel.org/r/dd221a99-efb3-cd1d-6256-7e646af29314@google.com
Fixes: e71769ae5260 ("mm: enable thp migration for shmem thp")
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Wang Yugui <wangyugui@e16-tech.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Ralph Campbell <rcampbell@nvidia.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Jue Wang <juew@google.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-16 04:23:45 +03:00
if ( unlikely ( is_pmd_migration_entry ( old_pmd ) ) ) {
swp_entry_t entry ;
entry = pmd_to_swp_entry ( old_pmd ) ;
2021-07-01 04:54:06 +03:00
page = pfn_swap_entry_to_page ( entry ) ;
mm/thp: fix __split_huge_pmd_locked() on shmem migration entry
Patch series "mm/thp: fix THP splitting unmap BUGs and related", v10.
Here is v2 batch of long-standing THP bug fixes that I had not got
around to sending before, but prompted now by Wang Yugui's report
https://lore.kernel.org/linux-mm/20210412180659.B9E3.409509F4@e16-tech.com/
Wang Yugui has tested a rollup of these fixes applied to 5.10.39, and
they have done no harm, but have *not* fixed that issue: something more
is needed and I have no idea of what.
This patch (of 7):
Stressing huge tmpfs page migration racing hole punch often crashed on
the VM_BUG_ON(!pmd_present) in pmdp_huge_clear_flush(), with DEBUG_VM=y
kernel; or shortly afterwards, on a bad dereference in
__split_huge_pmd_locked() when DEBUG_VM=n. They forgot to allow for pmd
migration entries in the non-anonymous case.
Full disclosure: those particular experiments were on a kernel with more
relaxed mmap_lock and i_mmap_rwsem locking, and were not repeated on the
vanilla kernel: it is conceivable that stricter locking happens to avoid
those cases, or makes them less likely; but __split_huge_pmd_locked()
already allowed for pmd migration entries when handling anonymous THPs,
so this commit brings the shmem and file THP handling into line.
And while there: use old_pmd rather than _pmd, as in the following
blocks; and make it clearer to the eye that the !vma_is_anonymous()
block is self-contained, making an early return after accounting for
unmapping.
Link: https://lkml.kernel.org/r/af88612-1473-2eaa-903-8d1a448b26@google.com
Link: https://lkml.kernel.org/r/dd221a99-efb3-cd1d-6256-7e646af29314@google.com
Fixes: e71769ae5260 ("mm: enable thp migration for shmem thp")
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Wang Yugui <wangyugui@e16-tech.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Ralph Campbell <rcampbell@nvidia.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Jue Wang <juew@google.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-16 04:23:45 +03:00
} else {
page = pmd_page ( old_pmd ) ;
if ( ! PageDirty ( page ) & & pmd_dirty ( old_pmd ) )
set_page_dirty ( page ) ;
if ( ! PageReferenced ( page ) & & pmd_young ( old_pmd ) )
SetPageReferenced ( page ) ;
mm/munlock: rmap call mlock_vma_page() munlock_vma_page()
Add vma argument to mlock_vma_page() and munlock_vma_page(), make them
inline functions which check (vma->vm_flags & VM_LOCKED) before calling
mlock_page() and munlock_page() in mm/mlock.c.
Add bool compound to mlock_vma_page() and munlock_vma_page(): this is
because we have understandable difficulty in accounting pte maps of THPs,
and if passed a PageHead page, mlock_page() and munlock_page() cannot
tell whether it's a pmd map to be counted or a pte map to be ignored.
Add vma arg to page_add_file_rmap() and page_remove_rmap(), like the
others, and use that to call mlock_vma_page() at the end of the page
adds, and munlock_vma_page() at the end of page_remove_rmap() (end or
beginning? unimportant, but end was easier for assertions in testing).
No page lock is required (although almost all adds happen to hold it):
delete the "Serialize with page migration" BUG_ON(!PageLocked(page))s.
Certainly page lock did serialize with page migration, but I'm having
difficulty explaining why that was ever important.
Mlock accounting on THPs has been hard to define, differed between anon
and file, involved PageDoubleMap in some places and not others, required
clear_page_mlock() at some points. Keep it simple now: just count the
pmds and ignore the ptes, there is no reason for ptes to undo pmd mlocks.
page_add_new_anon_rmap() callers unchanged: they have long been calling
lru_cache_add_inactive_or_unevictable(), which does its own VM_LOCKED
handling (it also checks for not VM_SPECIAL: I think that's overcautious,
and inconsistent with other checks, that mmap_region() already prevents
VM_LOCKED on VM_SPECIAL; but haven't quite convinced myself to change it).
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
2022-02-15 05:26:39 +03:00
page_remove_rmap ( page , vma , true ) ;
mm/thp: fix __split_huge_pmd_locked() on shmem migration entry
Patch series "mm/thp: fix THP splitting unmap BUGs and related", v10.
Here is v2 batch of long-standing THP bug fixes that I had not got
around to sending before, but prompted now by Wang Yugui's report
https://lore.kernel.org/linux-mm/20210412180659.B9E3.409509F4@e16-tech.com/
Wang Yugui has tested a rollup of these fixes applied to 5.10.39, and
they have done no harm, but have *not* fixed that issue: something more
is needed and I have no idea of what.
This patch (of 7):
Stressing huge tmpfs page migration racing hole punch often crashed on
the VM_BUG_ON(!pmd_present) in pmdp_huge_clear_flush(), with DEBUG_VM=y
kernel; or shortly afterwards, on a bad dereference in
__split_huge_pmd_locked() when DEBUG_VM=n. They forgot to allow for pmd
migration entries in the non-anonymous case.
Full disclosure: those particular experiments were on a kernel with more
relaxed mmap_lock and i_mmap_rwsem locking, and were not repeated on the
vanilla kernel: it is conceivable that stricter locking happens to avoid
those cases, or makes them less likely; but __split_huge_pmd_locked()
already allowed for pmd migration entries when handling anonymous THPs,
so this commit brings the shmem and file THP handling into line.
And while there: use old_pmd rather than _pmd, as in the following
blocks; and make it clearer to the eye that the !vma_is_anonymous()
block is self-contained, making an early return after accounting for
unmapping.
Link: https://lkml.kernel.org/r/af88612-1473-2eaa-903-8d1a448b26@google.com
Link: https://lkml.kernel.org/r/dd221a99-efb3-cd1d-6256-7e646af29314@google.com
Fixes: e71769ae5260 ("mm: enable thp migration for shmem thp")
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Wang Yugui <wangyugui@e16-tech.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Ralph Campbell <rcampbell@nvidia.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Jue Wang <juew@google.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-16 04:23:45 +03:00
put_page ( page ) ;
}
2018-08-18 01:44:55 +03:00
add_mm_counter ( mm , mm_counter_file ( page ) , - HPAGE_PMD_NR ) ;
2016-01-16 03:53:53 +03:00
return ;
mm/thp: fix __split_huge_pmd_locked() on shmem migration entry
Patch series "mm/thp: fix THP splitting unmap BUGs and related", v10.
Here is v2 batch of long-standing THP bug fixes that I had not got
around to sending before, but prompted now by Wang Yugui's report
https://lore.kernel.org/linux-mm/20210412180659.B9E3.409509F4@e16-tech.com/
Wang Yugui has tested a rollup of these fixes applied to 5.10.39, and
they have done no harm, but have *not* fixed that issue: something more
is needed and I have no idea of what.
This patch (of 7):
Stressing huge tmpfs page migration racing hole punch often crashed on
the VM_BUG_ON(!pmd_present) in pmdp_huge_clear_flush(), with DEBUG_VM=y
kernel; or shortly afterwards, on a bad dereference in
__split_huge_pmd_locked() when DEBUG_VM=n. They forgot to allow for pmd
migration entries in the non-anonymous case.
Full disclosure: those particular experiments were on a kernel with more
relaxed mmap_lock and i_mmap_rwsem locking, and were not repeated on the
vanilla kernel: it is conceivable that stricter locking happens to avoid
those cases, or makes them less likely; but __split_huge_pmd_locked()
already allowed for pmd migration entries when handling anonymous THPs,
so this commit brings the shmem and file THP handling into line.
And while there: use old_pmd rather than _pmd, as in the following
blocks; and make it clearer to the eye that the !vma_is_anonymous()
block is self-contained, making an early return after accounting for
unmapping.
Link: https://lkml.kernel.org/r/af88612-1473-2eaa-903-8d1a448b26@google.com
Link: https://lkml.kernel.org/r/dd221a99-efb3-cd1d-6256-7e646af29314@google.com
Fixes: e71769ae5260 ("mm: enable thp migration for shmem thp")
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Wang Yugui <wangyugui@e16-tech.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Ralph Campbell <rcampbell@nvidia.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Jue Wang <juew@google.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-16 04:23:45 +03:00
}
2021-06-16 04:23:49 +03:00
if ( is_huge_zero_pmd ( * pmd ) ) {
2017-11-16 04:34:11 +03:00
/*
* FIXME : Do we want to invalidate secondary mmu by calling
* mmu_notifier_invalidate_range ( ) see comments below inside
* __split_huge_pmd ( ) ?
*
* We are going from a zero huge page write protected to zero
* small page also write protected so it does not seems useful
* to invalidate secondary mmu at this time .
*/
2016-01-16 03:53:53 +03:00
return __split_huge_zero_page_pmd ( vma , haddr , pmd ) ;
}
2018-02-01 03:18:24 +03:00
/*
* Up to this point the pmd is present and huge and userland has the
* whole access to the hugepage during the split ( which happens in
* place ) . If we overwrite the pmd with the not - huge version pointing
* to the pte here ( which of course we could if all CPUs were bug
* free ) , userland could trigger a small page size TLB miss on the
* small sized TLB while the hugepage TLB entry is still established in
* the huge TLB . Some CPU doesn ' t like that .
2020-08-07 09:26:08 +03:00
* See http : //support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
* 383 on page 105. Intel should be safe but is also warns that it ' s
2018-02-01 03:18:24 +03:00
* only safe if the permission and cache attributes of the two entries
* loaded in the two TLB is identical ( which should be the case here ) .
* But it is generally safer to never allow small and huge TLB entries
* for the same virtual address to be loaded simultaneously . So instead
* of doing " pmd_populate(); flush_pmd_tlb_range(); " we first mark the
* current pmd notpresent ( atomically because here the pmd_trans_huge
* must remain set at all times on the pmd until the split is complete
* for this pmd ) , then we flush the SMP TLB and finally we write the
* non - huge version of the pmd entry with pmd_populate .
*/
old_pmd = pmdp_invalidate ( vma , haddr , pmd ) ;
pmd_migration = is_pmd_migration_entry ( old_pmd ) ;
2018-12-22 01:30:50 +03:00
if ( unlikely ( pmd_migration ) ) {
mm: thp: check pmd migration entry in common path
When THP migration is being used, memory management code needs to handle
pmd migration entries properly. This patch uses !pmd_present() or
is_swap_pmd() (depending on whether pmd_none() needs separate code or
not) to check pmd migration entries at the places where a pmd entry is
present.
Since pmd-related code uses split_huge_page(), split_huge_pmd(),
pmd_trans_huge(), pmd_trans_unstable(), or
pmd_none_or_trans_huge_or_clear_bad(), this patch:
1. adds pmd migration entry split code in split_huge_pmd(),
2. takes care of pmd migration entries whenever pmd_trans_huge() is present,
3. makes pmd_none_or_trans_huge_or_clear_bad() pmd migration entry aware.
Since split_huge_page() uses split_huge_pmd() and pmd_trans_unstable()
is equivalent to pmd_none_or_trans_huge_or_clear_bad(), we do not change
them.
Until this commit, a pmd entry should be:
1. pointing to a pte page,
2. is_swap_pmd(),
3. pmd_trans_huge(),
4. pmd_devmap(), or
5. pmd_none().
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-09 02:11:01 +03:00
swp_entry_t entry ;
2018-02-01 03:18:24 +03:00
entry = pmd_to_swp_entry ( old_pmd ) ;
2021-07-01 04:54:06 +03:00
page = pfn_swap_entry_to_page ( entry ) ;
2021-07-01 04:54:09 +03:00
write = is_writable_migration_entry ( entry ) ;
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
if ( PageAnon ( page ) )
anon_exclusive = is_readable_exclusive_migration_entry ( entry ) ;
2022-08-11 19:13:29 +03:00
young = is_migration_entry_young ( entry ) ;
dirty = is_migration_entry_dirty ( entry ) ;
2018-12-22 01:30:50 +03:00
soft_dirty = pmd_swp_soft_dirty ( old_pmd ) ;
2020-04-07 06:06:01 +03:00
uffd_wp = pmd_swp_uffd_wp ( old_pmd ) ;
2018-12-22 01:30:50 +03:00
} else {
2018-02-01 03:18:24 +03:00
page = pmd_page ( old_pmd ) ;
2022-08-11 19:13:28 +03:00
if ( pmd_dirty ( old_pmd ) ) {
dirty = true ;
2018-12-22 01:30:50 +03:00
SetPageDirty ( page ) ;
2022-08-11 19:13:28 +03:00
}
2018-12-22 01:30:50 +03:00
write = pmd_write ( old_pmd ) ;
young = pmd_young ( old_pmd ) ;
soft_dirty = pmd_soft_dirty ( old_pmd ) ;
userfaultfd: wp: apply _PAGE_UFFD_WP bit
Firstly, introduce two new flags MM_CP_UFFD_WP[_RESOLVE] for
change_protection() when used with uffd-wp and make sure the two new flags
are exclusively used. Then,
- For MM_CP_UFFD_WP: apply the _PAGE_UFFD_WP bit and remove _PAGE_RW
when a range of memory is write protected by uffd
- For MM_CP_UFFD_WP_RESOLVE: remove the _PAGE_UFFD_WP bit and recover
_PAGE_RW when write protection is resolved from userspace
And use this new interface in mwriteprotect_range() to replace the old
MM_CP_DIRTY_ACCT.
Do this change for both PTEs and huge PMDs. Then we can start to identify
which PTE/PMD is write protected by general (e.g., COW or soft dirty
tracking), and which is for userfaultfd-wp.
Since we should keep the _PAGE_UFFD_WP when doing pte_modify(), add it
into _PAGE_CHG_MASK as well. Meanwhile, since we have this new bit, we
can be even more strict when detecting uffd-wp page faults in either
do_wp_page() or wp_huge_pmd().
After we're with _PAGE_UFFD_WP, a special case is when a page is both
protected by the general COW logic and also userfault-wp. Here the
userfault-wp will have higher priority and will be handled first. Only
after the uffd-wp bit is cleared on the PTE/PMD will we continue to handle
the general COW. These are the steps on what will happen with such a
page:
1. CPU accesses write protected shared page (so both protected by
general COW and uffd-wp), blocked by uffd-wp first because in
do_wp_page we'll handle uffd-wp first, so it has higher priority
than general COW.
2. Uffd service thread receives the request, do UFFDIO_WRITEPROTECT
to remove the uffd-wp bit upon the PTE/PMD. However here we
still keep the write bit cleared. Notify the blocked CPU.
3. The blocked CPU resumes the page fault process with a fault
retry, during retry it'll notice it was not with the uffd-wp bit
this time but it is still write protected by general COW, then
it'll go though the COW path in the fault handler, copy the page,
apply write bit where necessary, and retry again.
4. The CPU will be able to access this page with write bit set.
Suggested-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Brian Geffon <bgeffon@google.com>
Cc: Pavel Emelyanov <xemul@openvz.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Martin Cracauer <cracauer@cons.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Bobby Powers <bobbypowers@gmail.com>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: "Kirill A . Shutemov" <kirill@shutemov.name>
Cc: Maya Gokhale <gokhale2@llnl.gov>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Marty McFadden <mcfadden8@llnl.gov>
Cc: Denis Plotnikov <dplotnikov@virtuozzo.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Shaohua Li <shli@fb.com>
Link: http://lkml.kernel.org/r/20200220163112.11409-8-peterx@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-07 06:05:49 +03:00
uffd_wp = pmd_uffd_wp ( old_pmd ) ;
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
2022-03-23 00:46:11 +03:00
VM_BUG_ON_PAGE ( ! page_count ( page ) , page ) ;
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
/*
* Without " freeze " , we ' ll simply split the PMD , propagating the
* PageAnonExclusive ( ) flag for each PTE by setting it for
* each subpage - - no need to ( temporarily ) clear .
*
* With " freeze " we want to replace mapped pages by
* migration entries right away . This is only possible if we
* managed to clear PageAnonExclusive ( ) - - see
* set_pmd_migration_entry ( ) .
*
* In case we cannot clear PageAnonExclusive ( ) , split the PMD
* only and let try_to_migrate_one ( ) fail later .
mm: fix PageAnonExclusive clearing racing with concurrent RCU GUP-fast
commit 6c287605fd56 ("mm: remember exclusively mapped anonymous pages with
PG_anon_exclusive") made sure that when PageAnonExclusive() has to be
cleared during temporary unmapping of a page, that the PTE is
cleared/invalidated and that the TLB is flushed.
What we want to achieve in all cases is that we cannot end up with a pin on
an anonymous page that may be shared, because such pins would be
unreliable and could result in memory corruptions when the mapped page
and the pin go out of sync due to a write fault.
That TLB flush handling was inspired by an outdated comment in
mm/ksm.c:write_protect_page(), which similarly required the TLB flush in
the past to synchronize with GUP-fast. However, ever since general RCU GUP
fast was introduced in commit 2667f50e8b81 ("mm: introduce a general RCU
get_user_pages_fast()"), a TLB flush is no longer sufficient to handle
concurrent GUP-fast in all cases -- it only handles traditional IPI-based
GUP-fast correctly.
Peter Xu (thankfully) questioned whether that TLB flush is really
required. On architectures that send an IPI broadcast on TLB flush,
it works as expected. To synchronize with RCU GUP-fast properly, we're
conceptually fine, however, we have to enforce a certain memory order and
are missing memory barriers.
Let's document that, avoid the TLB flush where possible and use proper
explicit memory barriers where required. We shouldn't really care about the
additional memory barriers here, as we're not on extremely hot paths --
and we're getting rid of some TLB flushes.
We use a smp_mb() pair for handling concurrent pinning and a
smp_rmb()/smp_wmb() pair for handling the corner case of only temporary
PTE changes but permanent PageAnonExclusive changes.
One extreme example, whereby GUP-fast takes a R/O pin and KSM wants to
convert an exclusive anonymous page to a KSM page, and that page is already
mapped write-protected (-> no PTE change) would be:
Thread 0 (KSM) Thread 1 (GUP-fast)
(B1) Read the PTE
# (B2) skipped without FOLL_WRITE
(A1) Clear PTE
smp_mb()
(A2) Check pinned
(B3) Pin the mapped page
smp_mb()
(A3) Clear PageAnonExclusive
smp_wmb()
(A4) Restore PTE
(B4) Check if the PTE changed
smp_rmb()
(B5) Check PageAnonExclusive
Thread 1 will properly detect that PageAnonExclusive was cleared and
back off.
Note that we don't need a memory barrier between checking if the page is
pinned and clearing PageAnonExclusive, because stores are not
speculated.
The possible issues due to reordering are of theoretical nature so far
and attempts to reproduce the race failed.
Especially the "no PTE change" case isn't the common case, because we'd
need an exclusive anonymous page that's mapped R/O and the PTE is clean
in KSM code -- and using KSM with page pinning isn't extremely common.
Further, the clear+TLB flush we used for now implies a memory barrier.
So the problematic missing part should be the missing memory barrier
after pinning but before checking if the PTE changed.
Link: https://lkml.kernel.org/r/20220901083559.67446-1-david@redhat.com
Fixes: 6c287605fd56 ("mm: remember exclusively mapped anonymous pages with PG_anon_exclusive")
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Andrea Parri <parri.andrea@gmail.com>
Cc: Will Deacon <will@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: "Paul E. McKenney" <paulmck@kernel.org>
Cc: Christoph von Recklinghausen <crecklin@redhat.com>
Cc: Don Dutile <ddutile@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-01 11:35:59 +03:00
*
* See page_try_share_anon_rmap ( ) : invalidate PMD first .
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
*/
anon_exclusive = PageAnon ( page ) & & PageAnonExclusive ( page ) ;
if ( freeze & & anon_exclusive & & page_try_share_anon_rmap ( page ) )
freeze = false ;
2022-11-22 12:51:50 +03:00
if ( ! freeze )
page_ref_add ( page , HPAGE_PMD_NR - 1 ) ;
2018-12-22 01:30:50 +03:00
}
2016-01-16 03:53:53 +03:00
2018-02-01 03:18:24 +03:00
/*
* Withdraw the table only after we mark the pmd entry invalid .
* This ' s critical for some architectures ( Power ) .
*/
2016-01-16 03:53:53 +03:00
pgtable = pgtable_trans_huge_withdraw ( mm , pmd ) ;
pmd_populate ( mm , & _pmd , pgtable ) ;
2016-02-24 18:58:03 +03:00
for ( i = 0 , addr = haddr ; i < HPAGE_PMD_NR ; i + + , addr + = PAGE_SIZE ) {
2016-01-16 03:53:53 +03:00
pte_t entry , * pte ;
/*
* Note that NUMA hinting access restrictions are not
* transferred to avoid any possibility of altering
* permissions across VMAs .
*/
mm: thp: check pmd migration entry in common path
When THP migration is being used, memory management code needs to handle
pmd migration entries properly. This patch uses !pmd_present() or
is_swap_pmd() (depending on whether pmd_none() needs separate code or
not) to check pmd migration entries at the places where a pmd entry is
present.
Since pmd-related code uses split_huge_page(), split_huge_pmd(),
pmd_trans_huge(), pmd_trans_unstable(), or
pmd_none_or_trans_huge_or_clear_bad(), this patch:
1. adds pmd migration entry split code in split_huge_pmd(),
2. takes care of pmd migration entries whenever pmd_trans_huge() is present,
3. makes pmd_none_or_trans_huge_or_clear_bad() pmd migration entry aware.
Since split_huge_page() uses split_huge_pmd() and pmd_trans_unstable()
is equivalent to pmd_none_or_trans_huge_or_clear_bad(), we do not change
them.
Until this commit, a pmd entry should be:
1. pointing to a pte page,
2. is_swap_pmd(),
3. pmd_trans_huge(),
4. pmd_devmap(), or
5. pmd_none().
Signed-off-by: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-09 02:11:01 +03:00
if ( freeze | | pmd_migration ) {
2016-01-16 03:53:56 +03:00
swp_entry_t swp_entry ;
2021-07-01 04:54:09 +03:00
if ( write )
swp_entry = make_writable_migration_entry (
page_to_pfn ( page + i ) ) ;
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
else if ( anon_exclusive )
swp_entry = make_readable_exclusive_migration_entry (
page_to_pfn ( page + i ) ) ;
2021-07-01 04:54:09 +03:00
else
swp_entry = make_readable_migration_entry (
page_to_pfn ( page + i ) ) ;
2022-08-11 19:13:29 +03:00
if ( young )
swp_entry = make_migration_entry_young ( swp_entry ) ;
if ( dirty )
swp_entry = make_migration_entry_dirty ( swp_entry ) ;
2016-01-16 03:53:56 +03:00
entry = swp_entry_to_pte ( swp_entry ) ;
2016-08-26 01:16:57 +03:00
if ( soft_dirty )
entry = pte_swp_mksoft_dirty ( entry ) ;
2020-04-07 06:06:01 +03:00
if ( uffd_wp )
entry = pte_swp_mkuffd_wp ( entry ) ;
2016-01-16 03:53:56 +03:00
} else {
2016-10-08 03:01:22 +03:00
entry = mk_pte ( page + i , READ_ONCE ( vma - > vm_page_prot ) ) ;
mm/huge_memory.c: don't split THP page when MADV_FREE syscall is called
We don't need to split THP page when MADV_FREE syscall is called if
[start, len] is aligned with THP size. The split could be done when VM
decide to free it in reclaim path if memory pressure is heavy. With
that, we could avoid unnecessary THP split.
For the feature, this patch changes pte dirtness marking logic of THP.
Now, it marks every ptes of pages dirty unconditionally in splitting,
which makes MADV_FREE void. So, instead, this patch propagates pmd
dirtiness to all pages via PG_dirty and restores pte dirtiness from
PG_dirty. With this, if pmd is clean(ie, MADV_FREEed) when split
happens(e,g, shrink_page_list), all of pages are clean too so we could
discard them.
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Shaohua Li <shli@kernel.org>
Cc: <yalin.wang2010@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Daniel Micay <danielmicay@gmail.com>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Helge Deller <deller@gmx.de>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: Jason Evans <je@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttil <mika.penttila@nextfour.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: Roland Dreier <roland@kernel.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Shaohua Li <shli@kernel.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-16 03:55:42 +03:00
entry = maybe_mkwrite ( entry , vma ) ;
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
if ( anon_exclusive )
SetPageAnonExclusive ( page + i ) ;
2016-01-16 03:53:56 +03:00
if ( ! young )
entry = pte_mkold ( entry ) ;
2022-11-25 21:58:57 +03:00
/* NOTE: this may set soft-dirty too on some archs */
if ( dirty )
entry = pte_mkdirty ( entry ) ;
2022-11-09 02:57:23 +03:00
/*
2022-11-25 21:58:57 +03:00
* NOTE : this needs to happen after pte_mkdirty ,
* because some archs ( sparc64 , loongarch ) could
* set hw write bit when mkdirty .
2022-11-09 02:57:23 +03:00
*/
2022-11-25 21:58:57 +03:00
if ( ! write )
entry = pte_wrprotect ( entry ) ;
2016-08-26 01:16:57 +03:00
if ( soft_dirty )
entry = pte_mksoft_dirty ( entry ) ;
userfaultfd: wp: apply _PAGE_UFFD_WP bit
Firstly, introduce two new flags MM_CP_UFFD_WP[_RESOLVE] for
change_protection() when used with uffd-wp and make sure the two new flags
are exclusively used. Then,
- For MM_CP_UFFD_WP: apply the _PAGE_UFFD_WP bit and remove _PAGE_RW
when a range of memory is write protected by uffd
- For MM_CP_UFFD_WP_RESOLVE: remove the _PAGE_UFFD_WP bit and recover
_PAGE_RW when write protection is resolved from userspace
And use this new interface in mwriteprotect_range() to replace the old
MM_CP_DIRTY_ACCT.
Do this change for both PTEs and huge PMDs. Then we can start to identify
which PTE/PMD is write protected by general (e.g., COW or soft dirty
tracking), and which is for userfaultfd-wp.
Since we should keep the _PAGE_UFFD_WP when doing pte_modify(), add it
into _PAGE_CHG_MASK as well. Meanwhile, since we have this new bit, we
can be even more strict when detecting uffd-wp page faults in either
do_wp_page() or wp_huge_pmd().
After we're with _PAGE_UFFD_WP, a special case is when a page is both
protected by the general COW logic and also userfault-wp. Here the
userfault-wp will have higher priority and will be handled first. Only
after the uffd-wp bit is cleared on the PTE/PMD will we continue to handle
the general COW. These are the steps on what will happen with such a
page:
1. CPU accesses write protected shared page (so both protected by
general COW and uffd-wp), blocked by uffd-wp first because in
do_wp_page we'll handle uffd-wp first, so it has higher priority
than general COW.
2. Uffd service thread receives the request, do UFFDIO_WRITEPROTECT
to remove the uffd-wp bit upon the PTE/PMD. However here we
still keep the write bit cleared. Notify the blocked CPU.
3. The blocked CPU resumes the page fault process with a fault
retry, during retry it'll notice it was not with the uffd-wp bit
this time but it is still write protected by general COW, then
it'll go though the COW path in the fault handler, copy the page,
apply write bit where necessary, and retry again.
4. The CPU will be able to access this page with write bit set.
Suggested-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Brian Geffon <bgeffon@google.com>
Cc: Pavel Emelyanov <xemul@openvz.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Martin Cracauer <cracauer@cons.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Bobby Powers <bobbypowers@gmail.com>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: "Kirill A . Shutemov" <kirill@shutemov.name>
Cc: Maya Gokhale <gokhale2@llnl.gov>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Marty McFadden <mcfadden8@llnl.gov>
Cc: Denis Plotnikov <dplotnikov@virtuozzo.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Shaohua Li <shli@fb.com>
Link: http://lkml.kernel.org/r/20200220163112.11409-8-peterx@redhat.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-07 06:05:49 +03:00
if ( uffd_wp )
entry = pte_mkuffd_wp ( entry ) ;
2022-11-22 12:51:50 +03:00
page_add_anon_rmap ( page + i , vma , addr , false ) ;
2016-01-16 03:53:56 +03:00
}
2016-02-24 18:58:03 +03:00
pte = pte_offset_map ( & _pmd , addr ) ;
2016-01-16 03:53:53 +03:00
BUG_ON ( ! pte_none ( * pte ) ) ;
2016-02-24 18:58:03 +03:00
set_pte_at ( mm , addr , pte , entry ) ;
2020-09-19 07:20:24 +03:00
pte_unmap ( pte ) ;
2016-01-16 03:53:53 +03:00
}
mm,thp,rmap: simplify compound page mapcount handling
Compound page (folio) mapcount calculations have been different for anon
and file (or shmem) THPs, and involved the obscure PageDoubleMap flag.
And each huge mapping and unmapping of a file (or shmem) THP involved
atomically incrementing and decrementing the mapcount of every subpage of
that huge page, dirtying many struct page cachelines.
Add subpages_mapcount field to the struct folio and first tail page, so
that the total of subpage mapcounts is available in one place near the
head: then page_mapcount() and total_mapcount() and page_mapped(), and
their folio equivalents, are so quick that anon and file and hugetlb don't
need to be optimized differently. Delete the unloved PageDoubleMap.
page_add and page_remove rmap functions must now maintain the
subpages_mapcount as well as the subpage _mapcount, when dealing with pte
mappings of huge pages; and correct maintenance of NR_ANON_MAPPED and
NR_FILE_MAPPED statistics still needs reading through the subpages, using
nr_subpages_unmapped() - but only when first or last pmd mapping finds
subpages_mapcount raised (double-map case, not the common case).
But are those counts (used to decide when to split an anon THP, and in
vmscan's pagecache_reclaimable heuristic) correctly maintained? Not
quite: since page_remove_rmap() (and also split_huge_pmd()) is often
called without page lock, there can be races when a subpage pte mapcount
0<->1 while compound pmd mapcount 0<->1 is scanning - races which the
previous implementation had prevented. The statistics might become
inaccurate, and even drift down until they underflow through 0. That is
not good enough, but is better dealt with in a followup patch.
Update a few comments on first and second tail page overlaid fields.
hugepage_add_new_anon_rmap() has to "increment" compound_mapcount, but
subpages_mapcount and compound_pincount are already correctly at 0, so
delete its reinitialization of compound_pincount.
A simple 100 X munmap(mmap(2GB, MAP_SHARED|MAP_POPULATE, tmpfs), 2GB) took
18 seconds on small pages, and used to take 1 second on huge pages, but
now takes 119 milliseconds on huge pages. Mapping by pmds a second time
used to take 860ms and now takes 92ms; mapping by pmds after mapping by
ptes (when the scan is needed) used to take 870ms and now takes 495ms.
But there might be some benchmarks which would show a slowdown, because
tail struct pages now fall out of cache until final freeing checks them.
Link: https://lkml.kernel.org/r/47ad693-717-79c8-e1ba-46c3a6602e48@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: James Houghton <jthoughton@google.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Naoya Horiguchi <naoya.horiguchi@linux.dev>
Cc: Peter Xu <peterx@redhat.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zach O'Keefe <zokeefe@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-03 04:51:38 +03:00
if ( ! pmd_migration )
page_remove_rmap ( page , vma , true ) ;
2022-11-22 12:51:50 +03:00
if ( freeze )
put_page ( page ) ;
2016-01-16 03:53:53 +03:00
smp_wmb ( ) ; /* make pte visible before pmd */
pmd_populate ( mm , pmd , pgtable ) ;
}
void __split_huge_pmd ( struct vm_area_struct * vma , pmd_t * pmd ,
2022-01-21 18:44:52 +03:00
unsigned long address , bool freeze , struct folio * folio )
2016-01-16 03:53:53 +03:00
{
spinlock_t * ptl ;
2018-12-28 11:38:09 +03:00
struct mmu_notifier_range range ;
2016-01-16 03:53:53 +03:00
2019-05-14 03:20:53 +03:00
mmu_notifier_range_init ( & range , MMU_NOTIFY_CLEAR , 0 , vma , vma - > vm_mm ,
mm/mmu_notifier: contextual information for event triggering invalidation
CPU page table update can happens for many reasons, not only as a result
of a syscall (munmap(), mprotect(), mremap(), madvise(), ...) but also as
a result of kernel activities (memory compression, reclaim, migration,
...).
Users of mmu notifier API track changes to the CPU page table and take
specific action for them. While current API only provide range of virtual
address affected by the change, not why the changes is happening.
This patchset do the initial mechanical convertion of all the places that
calls mmu_notifier_range_init to also provide the default MMU_NOTIFY_UNMAP
event as well as the vma if it is know (most invalidation happens against
a given vma). Passing down the vma allows the users of mmu notifier to
inspect the new vma page protection.
The MMU_NOTIFY_UNMAP is always the safe default as users of mmu notifier
should assume that every for the range is going away when that event
happens. A latter patch do convert mm call path to use a more appropriate
events for each call.
This is done as 2 patches so that no call site is forgotten especialy
as it uses this following coccinelle patch:
%<----------------------------------------------------------------------
@@
identifier I1, I2, I3, I4;
@@
static inline void mmu_notifier_range_init(struct mmu_notifier_range *I1,
+enum mmu_notifier_event event,
+unsigned flags,
+struct vm_area_struct *vma,
struct mm_struct *I2, unsigned long I3, unsigned long I4) { ... }
@@
@@
-#define mmu_notifier_range_init(range, mm, start, end)
+#define mmu_notifier_range_init(range, event, flags, vma, mm, start, end)
@@
expression E1, E3, E4;
identifier I1;
@@
<...
mmu_notifier_range_init(E1,
+MMU_NOTIFY_UNMAP, 0, I1,
I1->vm_mm, E3, E4)
...>
@@
expression E1, E2, E3, E4;
identifier FN, VMA;
@@
FN(..., struct vm_area_struct *VMA, ...) {
<...
mmu_notifier_range_init(E1,
+MMU_NOTIFY_UNMAP, 0, VMA,
E2, E3, E4)
...> }
@@
expression E1, E2, E3, E4;
identifier FN, VMA;
@@
FN(...) {
struct vm_area_struct *VMA;
<...
mmu_notifier_range_init(E1,
+MMU_NOTIFY_UNMAP, 0, VMA,
E2, E3, E4)
...> }
@@
expression E1, E2, E3, E4;
identifier FN;
@@
FN(...) {
<...
mmu_notifier_range_init(E1,
+MMU_NOTIFY_UNMAP, 0, NULL,
E2, E3, E4)
...> }
---------------------------------------------------------------------->%
Applied with:
spatch --all-includes --sp-file mmu-notifier.spatch fs/proc/task_mmu.c --in-place
spatch --sp-file mmu-notifier.spatch --dir kernel/events/ --in-place
spatch --sp-file mmu-notifier.spatch --dir mm --in-place
Link: http://lkml.kernel.org/r/20190326164747.24405-6-jglisse@redhat.com
Signed-off-by: Jérôme Glisse <jglisse@redhat.com>
Reviewed-by: Ralph Campbell <rcampbell@nvidia.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Cc: Christian König <christian.koenig@amd.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Jani Nikula <jani.nikula@linux.intel.com>
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Felix Kuehling <Felix.Kuehling@amd.com>
Cc: Jason Gunthorpe <jgg@mellanox.com>
Cc: Ross Zwisler <zwisler@kernel.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krcmar <rkrcmar@redhat.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Christian Koenig <christian.koenig@amd.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 03:20:49 +03:00
address & HPAGE_PMD_MASK ,
2018-12-28 11:38:09 +03:00
( address & HPAGE_PMD_MASK ) + HPAGE_PMD_SIZE ) ;
mmu_notifier_invalidate_range_start ( & range ) ;
ptl = pmd_lock ( vma - > vm_mm , pmd ) ;
2016-07-14 22:07:32 +03:00
/*
2022-01-21 18:44:52 +03:00
* If caller asks to setup a migration entry , we need a folio to check
* pmd against . Otherwise we can end up replacing wrong folio .
2016-07-14 22:07:32 +03:00
*/
2022-01-21 18:44:52 +03:00
VM_BUG_ON ( freeze & & ! folio ) ;
2022-04-06 15:41:39 +03:00
VM_WARN_ON_ONCE ( folio & & ! folio_test_locked ( folio ) ) ;
2016-07-14 22:07:32 +03:00
2022-03-25 04:13:56 +03:00
if ( pmd_trans_huge ( * pmd ) | | pmd_devmap ( * pmd ) | |
2022-04-06 15:41:39 +03:00
is_pmd_migration_entry ( * pmd ) ) {
2022-07-04 16:22:00 +03:00
/*
* It ' s safe to call pmd_page when folio is set because it ' s
* guaranteed that pmd is present .
*/
2022-04-06 15:41:39 +03:00
if ( folio & & folio ! = page_folio ( pmd_page ( * pmd ) ) )
goto out ;
2022-03-25 04:13:56 +03:00
__split_huge_pmd_locked ( vma , pmd , range . start , freeze ) ;
2022-04-06 15:41:39 +03:00
}
2022-03-25 04:13:56 +03:00
2016-01-16 03:54:33 +03:00
out :
2016-01-16 03:53:53 +03:00
spin_unlock ( ptl ) ;
2017-11-16 04:34:11 +03:00
/*
* No need to double call mmu_notifier - > invalidate_range ( ) callback .
* They are 3 cases to consider inside __split_huge_pmd_locked ( ) :
* 1 ) pmdp_huge_clear_flush_notify ( ) call invalidate_range ( ) obvious
* 2 ) __split_huge_zero_page_pmd ( ) read only zero page and any write
* fault will trigger a flush_notify before pointing to a new page
* ( it is fine if the secondary mmu keeps pointing to the old zero
* page in the meantime )
* 3 ) Split a huge pmd into pte pointing to the same page . No need
* to invalidate secondary tlb entry they are all still valid .
* any further changes to individual pte will notify . So no need
* to call mmu_notifier - > invalidate_range ( )
*/
2018-12-28 11:38:09 +03:00
mmu_notifier_invalidate_range_only_end ( & range ) ;
2016-01-16 03:53:53 +03:00
}
2016-03-18 00:20:10 +03:00
void split_huge_pmd_address ( struct vm_area_struct * vma , unsigned long address ,
2022-01-21 18:44:52 +03:00
bool freeze , struct folio * folio )
2011-01-14 02:47:08 +03:00
{
2022-07-07 02:59:26 +03:00
pmd_t * pmd = mm_find_pmd ( vma - > vm_mm , address ) ;
2011-01-14 02:47:08 +03:00
2022-07-07 02:59:26 +03:00
if ( ! pmd )
mm: let mm_find_pmd fix buggy race with THP fault
Trinity has reported:
BUG: unable to handle kernel NULL pointer dereference at 0000000000000018
IP: __lock_acquire (kernel/locking/lockdep.c:3070 (discriminator 1))
CPU: 6 PID: 16173 Comm: trinity-c364 Tainted: G W
3.15.0-rc1-next-20140415-sasha-00020-gaa90d09 #398
lock_acquire (arch/x86/include/asm/current.h:14
kernel/locking/lockdep.c:3602)
_raw_spin_lock (include/linux/spinlock_api_smp.h:143
kernel/locking/spinlock.c:151)
remove_migration_pte (mm/migrate.c:137)
rmap_walk (mm/rmap.c:1628 mm/rmap.c:1699)
remove_migration_ptes (mm/migrate.c:224)
migrate_pages (mm/migrate.c:922 mm/migrate.c:960 mm/migrate.c:1126)
migrate_misplaced_page (mm/migrate.c:1733)
__handle_mm_fault (mm/memory.c:3762 mm/memory.c:3812 mm/memory.c:3925)
handle_mm_fault (mm/memory.c:3948)
__get_user_pages (mm/memory.c:1851)
__mlock_vma_pages_range (mm/mlock.c:255)
__mm_populate (mm/mlock.c:711)
SyS_mlockall (include/linux/mm.h:1799 mm/mlock.c:817 mm/mlock.c:791)
I believe this comes about because, whereas collapsing and splitting THP
functions take anon_vma lock in write mode (which excludes concurrent
rmap walks), faulting THP functions (write protection and misplaced
NUMA) do not - and mostly they do not need to.
But they do use a pmdp_clear_flush(), set_pmd_at() sequence which, for
an instant (indeed, for a long instant, given the inter-CPU TLB flush in
there), leaves *pmd neither present not trans_huge.
Which can confuse a concurrent rmap walk, as when removing migration
ptes, seen in the dumped trace. Although that rmap walk has a 4k page
to insert, anon_vmas containing THPs are in no way segregated from
4k-page anon_vmas, so the 4k-intent mm_find_pmd() does need to cope with
that instant when a trans_huge pmd is temporarily absent.
I don't think we need strengthen the locking at the THP end: it's easily
handled with an ACCESS_ONCE() before testing both conditions.
And since mm_find_pmd() had only one caller who wanted a THP rather than
a pmd, let's slightly repurpose it to fail when it hits a THP or
non-present pmd, and open code split_huge_page_address() again.
Signed-off-by: Hugh Dickins <hughd@google.com>
Reported-by: Sasha Levin <sasha.levin@oracle.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Konstantin Khlebnikov <koct9i@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Lameter <cl@gentwo.org>
Cc: Dave Jones <davej@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-24 00:22:05 +04:00
return ;
2022-01-21 18:44:52 +03:00
__split_huge_pmd ( vma , pmd , address , freeze , folio ) ;
2011-01-14 02:47:08 +03:00
}
2021-05-05 04:33:52 +03:00
static inline void split_huge_pmd_if_needed ( struct vm_area_struct * vma , unsigned long address )
{
/*
* If the new address isn ' t hpage aligned and it could previously
* contain an hugepage : check if we need to split an huge pmd .
*/
if ( ! IS_ALIGNED ( address , HPAGE_PMD_SIZE ) & &
range_in_vma ( vma , ALIGN_DOWN ( address , HPAGE_PMD_SIZE ) ,
ALIGN ( address , HPAGE_PMD_SIZE ) ) )
split_huge_pmd_address ( vma , address , false , NULL ) ;
}
2015-09-09 00:58:37 +03:00
void vma_adjust_trans_huge ( struct vm_area_struct * vma ,
2011-01-14 02:47:08 +03:00
unsigned long start ,
unsigned long end ,
long adjust_next )
{
2021-05-05 04:33:52 +03:00
/* Check if we need to split start first. */
split_huge_pmd_if_needed ( vma , start ) ;
2011-01-14 02:47:08 +03:00
2021-05-05 04:33:52 +03:00
/* Check if we need to split end next. */
split_huge_pmd_if_needed ( vma , end ) ;
2011-01-14 02:47:08 +03:00
/*
2022-09-06 22:49:00 +03:00
* If we ' re also updating the next vma vm_start ,
2021-05-05 04:33:52 +03:00
* check if we need to split it .
2011-01-14 02:47:08 +03:00
*/
if ( adjust_next > 0 ) {
2022-09-06 22:49:00 +03:00
struct vm_area_struct * next = find_vma ( vma - > vm_mm , vma - > vm_end ) ;
2011-01-14 02:47:08 +03:00
unsigned long nstart = next - > vm_start ;
2020-10-14 02:53:57 +03:00
nstart + = adjust_next ;
2021-05-05 04:33:52 +03:00
split_huge_pmd_if_needed ( next , nstart ) ;
2011-01-14 02:47:08 +03:00
}
}
2016-01-16 03:54:10 +03:00
2022-09-02 22:46:49 +03:00
static void unmap_folio ( struct folio * folio )
2016-01-16 03:54:10 +03:00
{
2021-07-01 04:54:16 +03:00
enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
TTU_SYNC ;
2016-01-16 03:54:10 +03:00
2022-09-02 22:46:49 +03:00
VM_BUG_ON_FOLIO ( ! folio_test_large ( folio ) , folio ) ;
2016-01-16 03:54:10 +03:00
2021-07-01 04:54:16 +03:00
/*
* Anon pages need migration entries to preserve them , but file
* pages can simply be left unmapped , then faulted back on demand .
* If that is ever changed ( perhaps for mlock ) , update remap_page ( ) .
*/
2022-01-28 22:29:43 +03:00
if ( folio_test_anon ( folio ) )
try_to_migrate ( folio , ttu_flags ) ;
2021-07-01 04:54:16 +03:00
else
2022-02-15 17:28:49 +03:00
try_to_unmap ( folio , ttu_flags | TTU_IGNORE_MLOCK ) ;
2016-01-16 03:54:10 +03:00
}
2022-01-29 07:32:59 +03:00
static void remap_page ( struct folio * folio , unsigned long nr )
2016-01-16 03:54:10 +03:00
{
2022-01-29 07:32:59 +03:00
int i = 0 ;
2021-07-01 04:52:04 +03:00
2022-09-02 22:46:49 +03:00
/* If unmap_folio() uses try_to_migrate() on file, remove this check */
2022-01-29 07:32:59 +03:00
if ( ! folio_test_anon ( folio ) )
2021-07-01 04:52:04 +03:00
return ;
2022-01-29 07:32:59 +03:00
for ( ; ; ) {
remove_migration_ptes ( folio , folio , true ) ;
i + = folio_nr_pages ( folio ) ;
if ( i > = nr )
break ;
folio = folio_next ( folio ) ;
2017-02-25 01:57:45 +03:00
}
2016-01-16 03:54:10 +03:00
}
2020-12-15 23:33:24 +03:00
static void lru_add_page_tail ( struct page * head , struct page * tail ,
mm/thp: move lru_add_page_tail() to huge_memory.c
Patch series "per memcg lru lock", v21.
This patchset includes 3 parts:
1) some code cleanup and minimum optimization as preparation
2) use TestCleanPageLRU as page isolation's precondition
3) replace per node lru_lock with per memcg per node lru_lock
Current lru_lock is one for each of node, pgdat->lru_lock, that guard
for lru lists, but now we had moved the lru lists into memcg for long
time. Still using per node lru_lock is clearly unscalable, pages on
each of memcgs have to compete each others for a whole lru_lock. This
patchset try to use per lruvec/memcg lru_lock to repleace per node lru
lock to guard lru lists, make it scalable for memcgs and get performance
gain.
Currently lru_lock still guards both lru list and page's lru bit, that's
ok. but if we want to use specific lruvec lock on the page, we need to
pin down the page's lruvec/memcg during locking. Just taking lruvec
lock first may be undermined by the page's memcg charge/migration. To
fix this problem, we could take out the page's lru bit clear and use it
as pin down action to block the memcg changes. That's the reason for
new atomic func TestClearPageLRU. So now isolating a page need both
actions: TestClearPageLRU and hold the lru_lock.
The typical usage of this is isolate_migratepages_block() in
compaction.c we have to take lru bit before lru lock, that serialized
the page isolation in memcg page charge/migration which will change
page's lruvec and new lru_lock in it.
The above solution suggested by Johannes Weiner, and based on his new
memcg charge path, then have this patchset. (Hugh Dickins tested and
contributed much code from compaction fix to general code polish, thanks
a lot!).
Daniel Jordan's testing show 62% improvement on modified readtwice case
on his 2P * 10 core * 2 HT broadwell box on v18, which has no much
different with this v20.
https://lore.kernel.org/lkml/20200915165807.kpp7uhiw7l3loofu@ca-dmjordan1.us.oracle.com/
Thanks to Hugh Dickins and Konstantin Khlebnikov, they both brought this
idea 8 years ago, and others who gave comments as well: Daniel Jordan,
Mel Gorman, Shakeel Butt, Matthew Wilcox, Alexander Duyck etc.
Thanks for Testing support from Intel 0day and Rong Chen, Fengguang Wu,
and Yun Wang. Hugh Dickins also shared his kbuild-swap case.
This patch (of 19):
lru_add_page_tail() is only used in huge_memory.c, defining it in other
file with a CONFIG_TRANSPARENT_HUGEPAGE macro restrict just looks weird.
Let's move it THP. And make it static as Hugh Dickins suggested.
Link: https://lkml.kernel.org/r/1604566549-62481-1-git-send-email-alex.shi@linux.alibaba.com
Link: https://lkml.kernel.org/r/1604566549-62481-2-git-send-email-alex.shi@linux.alibaba.com
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Tejun Heo <tj@kernel.org>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: "Chen, Rong A" <rong.a.chen@intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jann Horn <jannh@google.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 23:33:20 +03:00
struct lruvec * lruvec , struct list_head * list )
{
2020-12-15 23:33:24 +03:00
VM_BUG_ON_PAGE ( ! PageHead ( head ) , head ) ;
VM_BUG_ON_PAGE ( PageCompound ( tail ) , head ) ;
VM_BUG_ON_PAGE ( PageLRU ( tail ) , head ) ;
2020-12-15 23:34:29 +03:00
lockdep_assert_held ( & lruvec - > lru_lock ) ;
mm/thp: move lru_add_page_tail() to huge_memory.c
Patch series "per memcg lru lock", v21.
This patchset includes 3 parts:
1) some code cleanup and minimum optimization as preparation
2) use TestCleanPageLRU as page isolation's precondition
3) replace per node lru_lock with per memcg per node lru_lock
Current lru_lock is one for each of node, pgdat->lru_lock, that guard
for lru lists, but now we had moved the lru lists into memcg for long
time. Still using per node lru_lock is clearly unscalable, pages on
each of memcgs have to compete each others for a whole lru_lock. This
patchset try to use per lruvec/memcg lru_lock to repleace per node lru
lock to guard lru lists, make it scalable for memcgs and get performance
gain.
Currently lru_lock still guards both lru list and page's lru bit, that's
ok. but if we want to use specific lruvec lock on the page, we need to
pin down the page's lruvec/memcg during locking. Just taking lruvec
lock first may be undermined by the page's memcg charge/migration. To
fix this problem, we could take out the page's lru bit clear and use it
as pin down action to block the memcg changes. That's the reason for
new atomic func TestClearPageLRU. So now isolating a page need both
actions: TestClearPageLRU and hold the lru_lock.
The typical usage of this is isolate_migratepages_block() in
compaction.c we have to take lru bit before lru lock, that serialized
the page isolation in memcg page charge/migration which will change
page's lruvec and new lru_lock in it.
The above solution suggested by Johannes Weiner, and based on his new
memcg charge path, then have this patchset. (Hugh Dickins tested and
contributed much code from compaction fix to general code polish, thanks
a lot!).
Daniel Jordan's testing show 62% improvement on modified readtwice case
on his 2P * 10 core * 2 HT broadwell box on v18, which has no much
different with this v20.
https://lore.kernel.org/lkml/20200915165807.kpp7uhiw7l3loofu@ca-dmjordan1.us.oracle.com/
Thanks to Hugh Dickins and Konstantin Khlebnikov, they both brought this
idea 8 years ago, and others who gave comments as well: Daniel Jordan,
Mel Gorman, Shakeel Butt, Matthew Wilcox, Alexander Duyck etc.
Thanks for Testing support from Intel 0day and Rong Chen, Fengguang Wu,
and Yun Wang. Hugh Dickins also shared his kbuild-swap case.
This patch (of 19):
lru_add_page_tail() is only used in huge_memory.c, defining it in other
file with a CONFIG_TRANSPARENT_HUGEPAGE macro restrict just looks weird.
Let's move it THP. And make it static as Hugh Dickins suggested.
Link: https://lkml.kernel.org/r/1604566549-62481-1-git-send-email-alex.shi@linux.alibaba.com
Link: https://lkml.kernel.org/r/1604566549-62481-2-git-send-email-alex.shi@linux.alibaba.com
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Tejun Heo <tj@kernel.org>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: "Chen, Rong A" <rong.a.chen@intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jann Horn <jannh@google.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 23:33:20 +03:00
2020-12-15 23:33:29 +03:00
if ( list ) {
mm/thp: move lru_add_page_tail() to huge_memory.c
Patch series "per memcg lru lock", v21.
This patchset includes 3 parts:
1) some code cleanup and minimum optimization as preparation
2) use TestCleanPageLRU as page isolation's precondition
3) replace per node lru_lock with per memcg per node lru_lock
Current lru_lock is one for each of node, pgdat->lru_lock, that guard
for lru lists, but now we had moved the lru lists into memcg for long
time. Still using per node lru_lock is clearly unscalable, pages on
each of memcgs have to compete each others for a whole lru_lock. This
patchset try to use per lruvec/memcg lru_lock to repleace per node lru
lock to guard lru lists, make it scalable for memcgs and get performance
gain.
Currently lru_lock still guards both lru list and page's lru bit, that's
ok. but if we want to use specific lruvec lock on the page, we need to
pin down the page's lruvec/memcg during locking. Just taking lruvec
lock first may be undermined by the page's memcg charge/migration. To
fix this problem, we could take out the page's lru bit clear and use it
as pin down action to block the memcg changes. That's the reason for
new atomic func TestClearPageLRU. So now isolating a page need both
actions: TestClearPageLRU and hold the lru_lock.
The typical usage of this is isolate_migratepages_block() in
compaction.c we have to take lru bit before lru lock, that serialized
the page isolation in memcg page charge/migration which will change
page's lruvec and new lru_lock in it.
The above solution suggested by Johannes Weiner, and based on his new
memcg charge path, then have this patchset. (Hugh Dickins tested and
contributed much code from compaction fix to general code polish, thanks
a lot!).
Daniel Jordan's testing show 62% improvement on modified readtwice case
on his 2P * 10 core * 2 HT broadwell box on v18, which has no much
different with this v20.
https://lore.kernel.org/lkml/20200915165807.kpp7uhiw7l3loofu@ca-dmjordan1.us.oracle.com/
Thanks to Hugh Dickins and Konstantin Khlebnikov, they both brought this
idea 8 years ago, and others who gave comments as well: Daniel Jordan,
Mel Gorman, Shakeel Butt, Matthew Wilcox, Alexander Duyck etc.
Thanks for Testing support from Intel 0day and Rong Chen, Fengguang Wu,
and Yun Wang. Hugh Dickins also shared his kbuild-swap case.
This patch (of 19):
lru_add_page_tail() is only used in huge_memory.c, defining it in other
file with a CONFIG_TRANSPARENT_HUGEPAGE macro restrict just looks weird.
Let's move it THP. And make it static as Hugh Dickins suggested.
Link: https://lkml.kernel.org/r/1604566549-62481-1-git-send-email-alex.shi@linux.alibaba.com
Link: https://lkml.kernel.org/r/1604566549-62481-2-git-send-email-alex.shi@linux.alibaba.com
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Tejun Heo <tj@kernel.org>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: "Chen, Rong A" <rong.a.chen@intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jann Horn <jannh@google.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 23:33:20 +03:00
/* page reclaim is reclaiming a huge page */
2020-12-15 23:33:29 +03:00
VM_WARN_ON ( PageLRU ( head ) ) ;
2020-12-15 23:33:24 +03:00
get_page ( tail ) ;
list_add_tail ( & tail - > lru , list ) ;
mm/thp: move lru_add_page_tail() to huge_memory.c
Patch series "per memcg lru lock", v21.
This patchset includes 3 parts:
1) some code cleanup and minimum optimization as preparation
2) use TestCleanPageLRU as page isolation's precondition
3) replace per node lru_lock with per memcg per node lru_lock
Current lru_lock is one for each of node, pgdat->lru_lock, that guard
for lru lists, but now we had moved the lru lists into memcg for long
time. Still using per node lru_lock is clearly unscalable, pages on
each of memcgs have to compete each others for a whole lru_lock. This
patchset try to use per lruvec/memcg lru_lock to repleace per node lru
lock to guard lru lists, make it scalable for memcgs and get performance
gain.
Currently lru_lock still guards both lru list and page's lru bit, that's
ok. but if we want to use specific lruvec lock on the page, we need to
pin down the page's lruvec/memcg during locking. Just taking lruvec
lock first may be undermined by the page's memcg charge/migration. To
fix this problem, we could take out the page's lru bit clear and use it
as pin down action to block the memcg changes. That's the reason for
new atomic func TestClearPageLRU. So now isolating a page need both
actions: TestClearPageLRU and hold the lru_lock.
The typical usage of this is isolate_migratepages_block() in
compaction.c we have to take lru bit before lru lock, that serialized
the page isolation in memcg page charge/migration which will change
page's lruvec and new lru_lock in it.
The above solution suggested by Johannes Weiner, and based on his new
memcg charge path, then have this patchset. (Hugh Dickins tested and
contributed much code from compaction fix to general code polish, thanks
a lot!).
Daniel Jordan's testing show 62% improvement on modified readtwice case
on his 2P * 10 core * 2 HT broadwell box on v18, which has no much
different with this v20.
https://lore.kernel.org/lkml/20200915165807.kpp7uhiw7l3loofu@ca-dmjordan1.us.oracle.com/
Thanks to Hugh Dickins and Konstantin Khlebnikov, they both brought this
idea 8 years ago, and others who gave comments as well: Daniel Jordan,
Mel Gorman, Shakeel Butt, Matthew Wilcox, Alexander Duyck etc.
Thanks for Testing support from Intel 0day and Rong Chen, Fengguang Wu,
and Yun Wang. Hugh Dickins also shared his kbuild-swap case.
This patch (of 19):
lru_add_page_tail() is only used in huge_memory.c, defining it in other
file with a CONFIG_TRANSPARENT_HUGEPAGE macro restrict just looks weird.
Let's move it THP. And make it static as Hugh Dickins suggested.
Link: https://lkml.kernel.org/r/1604566549-62481-1-git-send-email-alex.shi@linux.alibaba.com
Link: https://lkml.kernel.org/r/1604566549-62481-2-git-send-email-alex.shi@linux.alibaba.com
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Tejun Heo <tj@kernel.org>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: "Chen, Rong A" <rong.a.chen@intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jann Horn <jannh@google.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 23:33:20 +03:00
} else {
2020-12-15 23:33:29 +03:00
/* head is still on lru (and we have it frozen) */
VM_WARN_ON ( ! PageLRU ( head ) ) ;
2022-02-15 05:29:54 +03:00
if ( PageUnevictable ( tail ) )
tail - > mlock_count = 0 ;
else
list_add_tail ( & tail - > lru , & head - > lru ) ;
2020-12-15 23:33:29 +03:00
SetPageLRU ( tail ) ;
mm/thp: move lru_add_page_tail() to huge_memory.c
Patch series "per memcg lru lock", v21.
This patchset includes 3 parts:
1) some code cleanup and minimum optimization as preparation
2) use TestCleanPageLRU as page isolation's precondition
3) replace per node lru_lock with per memcg per node lru_lock
Current lru_lock is one for each of node, pgdat->lru_lock, that guard
for lru lists, but now we had moved the lru lists into memcg for long
time. Still using per node lru_lock is clearly unscalable, pages on
each of memcgs have to compete each others for a whole lru_lock. This
patchset try to use per lruvec/memcg lru_lock to repleace per node lru
lock to guard lru lists, make it scalable for memcgs and get performance
gain.
Currently lru_lock still guards both lru list and page's lru bit, that's
ok. but if we want to use specific lruvec lock on the page, we need to
pin down the page's lruvec/memcg during locking. Just taking lruvec
lock first may be undermined by the page's memcg charge/migration. To
fix this problem, we could take out the page's lru bit clear and use it
as pin down action to block the memcg changes. That's the reason for
new atomic func TestClearPageLRU. So now isolating a page need both
actions: TestClearPageLRU and hold the lru_lock.
The typical usage of this is isolate_migratepages_block() in
compaction.c we have to take lru bit before lru lock, that serialized
the page isolation in memcg page charge/migration which will change
page's lruvec and new lru_lock in it.
The above solution suggested by Johannes Weiner, and based on his new
memcg charge path, then have this patchset. (Hugh Dickins tested and
contributed much code from compaction fix to general code polish, thanks
a lot!).
Daniel Jordan's testing show 62% improvement on modified readtwice case
on his 2P * 10 core * 2 HT broadwell box on v18, which has no much
different with this v20.
https://lore.kernel.org/lkml/20200915165807.kpp7uhiw7l3loofu@ca-dmjordan1.us.oracle.com/
Thanks to Hugh Dickins and Konstantin Khlebnikov, they both brought this
idea 8 years ago, and others who gave comments as well: Daniel Jordan,
Mel Gorman, Shakeel Butt, Matthew Wilcox, Alexander Duyck etc.
Thanks for Testing support from Intel 0day and Rong Chen, Fengguang Wu,
and Yun Wang. Hugh Dickins also shared his kbuild-swap case.
This patch (of 19):
lru_add_page_tail() is only used in huge_memory.c, defining it in other
file with a CONFIG_TRANSPARENT_HUGEPAGE macro restrict just looks weird.
Let's move it THP. And make it static as Hugh Dickins suggested.
Link: https://lkml.kernel.org/r/1604566549-62481-1-git-send-email-alex.shi@linux.alibaba.com
Link: https://lkml.kernel.org/r/1604566549-62481-2-git-send-email-alex.shi@linux.alibaba.com
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Tejun Heo <tj@kernel.org>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: "Chen, Rong A" <rong.a.chen@intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jann Horn <jannh@google.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 23:33:20 +03:00
}
}
2016-03-16 00:57:30 +03:00
static void __split_huge_page_tail ( struct page * head , int tail ,
2016-01-16 03:54:10 +03:00
struct lruvec * lruvec , struct list_head * list )
{
struct page * page_tail = head + tail ;
2016-03-16 00:57:30 +03:00
VM_BUG_ON_PAGE ( atomic_read ( & page_tail - > _mapcount ) ! = - 1 , page_tail ) ;
2016-01-16 03:54:10 +03:00
/*
2018-04-06 02:23:28 +03:00
* Clone page flags before unfreezing refcount .
*
* After successful get_page_unless_zero ( ) might follow flags change ,
2020-12-16 07:47:26 +03:00
* for example lock_page ( ) which set PG_waiters .
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
*
* Note that for mapped sub - pages of an anonymous THP ,
2022-09-02 22:46:49 +03:00
* PG_anon_exclusive has been cleared in unmap_folio ( ) and is stored in
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
* the migration entry instead from where remap_page ( ) will restore it .
* We can still have PG_anon_exclusive set on effectively unmapped and
* unreferenced sub - pages of an anonymous THP : we can simply drop
* PG_anon_exclusive ( - > PG_mappedtodisk ) for these here .
2016-01-16 03:54:10 +03:00
*/
page_tail - > flags & = ~ PAGE_FLAGS_CHECK_AT_PREP ;
page_tail - > flags | = ( head - > flags &
( ( 1L < < PG_referenced ) |
( 1L < < PG_swapbacked ) |
mm, THP, swap: delay splitting THP during swap out
Patch series "THP swap: Delay splitting THP during swapping out", v11.
This patchset is to optimize the performance of Transparent Huge Page
(THP) swap.
Recently, the performance of the storage devices improved so fast that
we cannot saturate the disk bandwidth with single logical CPU when do
page swap out even on a high-end server machine. Because the
performance of the storage device improved faster than that of single
logical CPU. And it seems that the trend will not change in the near
future. On the other hand, the THP becomes more and more popular
because of increased memory size. So it becomes necessary to optimize
THP swap performance.
The advantages of the THP swap support include:
- Batch the swap operations for the THP to reduce lock
acquiring/releasing, including allocating/freeing the swap space,
adding/deleting to/from the swap cache, and writing/reading the swap
space, etc. This will help improve the performance of the THP swap.
- The THP swap space read/write will be 2M sequential IO. It is
particularly helpful for the swap read, which are usually 4k random
IO. This will improve the performance of the THP swap too.
- It will help the memory fragmentation, especially when the THP is
heavily used by the applications. The 2M continuous pages will be
free up after THP swapping out.
- It will improve the THP utilization on the system with the swap
turned on. Because the speed for khugepaged to collapse the normal
pages into the THP is quite slow. After the THP is split during the
swapping out, it will take quite long time for the normal pages to
collapse back into the THP after being swapped in. The high THP
utilization helps the efficiency of the page based memory management
too.
There are some concerns regarding THP swap in, mainly because possible
enlarged read/write IO size (for swap in/out) may put more overhead on
the storage device. To deal with that, the THP swap in should be turned
on only when necessary. For example, it can be selected via
"always/never/madvise" logic, to be turned on globally, turned off
globally, or turned on only for VMA with MADV_HUGEPAGE, etc.
This patchset is the first step for the THP swap support. The plan is
to delay splitting THP step by step, finally avoid splitting THP during
the THP swapping out and swap out/in the THP as a whole.
As the first step, in this patchset, the splitting huge page is delayed
from almost the first step of swapping out to after allocating the swap
space for the THP and adding the THP into the swap cache. This will
reduce lock acquiring/releasing for the locks used for the swap cache
management.
With the patchset, the swap out throughput improves 15.5% (from about
3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
This patch (of 5):
In this patch, splitting huge page is delayed from almost the first step
of swapping out to after allocating the swap space for the THP
(Transparent Huge Page) and adding the THP into the swap cache. This
will batch the corresponding operation, thus improve THP swap out
throughput.
This is the first step for the THP swap optimization. The plan is to
delay splitting the THP step by step and avoid splitting the THP
finally.
In this patch, one swap cluster is used to hold the contents of each THP
swapped out. So, the size of the swap cluster is changed to that of the
THP (Transparent Huge Page) on x86_64 architecture (512). For other
architectures which want such THP swap optimization,
ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for
the architecture. In effect, this will enlarge swap cluster size by 2
times on x86_64. Which may make it harder to find a free cluster when
the swap space becomes fragmented. So that, this may reduce the
continuous swap space allocation and sequential write in theory. The
performance test in 0day shows no regressions caused by this.
In the future of THP swap optimization, some information of the swapped
out THP (such as compound map count) will be recorded in the
swap_cluster_info data structure.
The mem cgroup swap accounting functions are enhanced to support charge
or uncharge a swap cluster backing a THP as a whole.
The swap cluster allocate/free functions are added to allocate/free a
swap cluster for a THP. A fair simple algorithm is used for swap
cluster allocation, that is, only the first swap device in priority list
will be tried to allocate the swap cluster. The function will fail if
the trying is not successful, and the caller will fallback to allocate a
single swap slot instead. This works good enough for normal cases. If
the difference of the number of the free swap clusters among multiple
swap devices is significant, it is possible that some THPs are split
earlier than necessary. For example, this could be caused by big size
difference among multiple swap devices.
The swap cache functions is enhanced to support add/delete THP to/from
the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be
enhanced in the future with multi-order radix tree. But because we will
split the THP soon during swapping out, that optimization doesn't make
much sense for this first step.
The THP splitting functions are enhanced to support to split THP in swap
cache during swapping out. The page lock will be held during allocating
the swap cluster, adding the THP into the swap cache and splitting the
THP. So in the code path other than swapping out, if the THP need to be
split, the PageSwapCache(THP) will be always false.
The swap cluster is only available for SSD, so the THP swap optimization
in this patchset has no effect for HDD.
[ying.huang@intel.com: fix two issues in THP optimize patch]
Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com
[hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size]
Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option]
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h]
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-07 01:37:18 +03:00
( 1L < < PG_swapcache ) |
2016-01-16 03:54:10 +03:00
( 1L < < PG_mlocked ) |
( 1L < < PG_uptodate ) |
( 1L < < PG_active ) |
mm: workingset: tell cache transitions from workingset thrashing
Refaults happen during transitions between workingsets as well as in-place
thrashing. Knowing the difference between the two has a range of
applications, including measuring the impact of memory shortage on the
system performance, as well as the ability to smarter balance pressure
between the filesystem cache and the swap-backed workingset.
During workingset transitions, inactive cache refaults and pushes out
established active cache. When that active cache isn't stale, however,
and also ends up refaulting, that's bonafide thrashing.
Introduce a new page flag that tells on eviction whether the page has been
active or not in its lifetime. This bit is then stored in the shadow
entry, to classify refaults as transitioning or thrashing.
How many page->flags does this leave us with on 32-bit?
20 bits are always page flags
21 if you have an MMU
23 with the zone bits for DMA, Normal, HighMem, Movable
29 with the sparsemem section bits
30 if PAE is enabled
31 with this patch.
So on 32-bit PAE, that leaves 1 bit for distinguishing two NUMA nodes. If
that's not enough, the system can switch to discontigmem and re-gain the 6
or 7 sparsemem section bits.
Link: http://lkml.kernel.org/r/20180828172258.3185-3-hannes@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Daniel Drake <drake@endlessm.com>
Tested-by: Suren Baghdasaryan <surenb@google.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <jweiner@fb.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Enderborg <peter.enderborg@sony.com>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vinayak Menon <vinmenon@codeaurora.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 01:06:04 +03:00
( 1L < < PG_workingset ) |
2016-01-16 03:54:10 +03:00
( 1L < < PG_locked ) |
mm/huge_memory.c: don't split THP page when MADV_FREE syscall is called
We don't need to split THP page when MADV_FREE syscall is called if
[start, len] is aligned with THP size. The split could be done when VM
decide to free it in reclaim path if memory pressure is heavy. With
that, we could avoid unnecessary THP split.
For the feature, this patch changes pte dirtness marking logic of THP.
Now, it marks every ptes of pages dirty unconditionally in splitting,
which makes MADV_FREE void. So, instead, this patch propagates pmd
dirtiness to all pages via PG_dirty and restores pte dirtiness from
PG_dirty. With this, if pmd is clean(ie, MADV_FREEed) when split
happens(e,g, shrink_page_list), all of pages are clean too so we could
discard them.
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Shaohua Li <shli@kernel.org>
Cc: <yalin.wang2010@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Chen Gang <gang.chen.5i5j@gmail.com>
Cc: Chris Zankel <chris@zankel.net>
Cc: Daniel Micay <danielmicay@gmail.com>
Cc: Darrick J. Wong <darrick.wong@oracle.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Helge Deller <deller@gmx.de>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: Jason Evans <je@fb.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttil <mika.penttila@nextfour.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: Roland Dreier <roland@kernel.org>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Shaohua Li <shli@kernel.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-16 03:55:42 +03:00
( 1L < < PG_unevictable ) |
2022-11-04 04:10:34 +03:00
# ifdef CONFIG_ARCH_USES_PG_ARCH_X
2020-07-02 12:19:30 +03:00
( 1L < < PG_arch_2 ) |
2022-11-04 04:10:37 +03:00
( 1L < < PG_arch_3 ) |
2020-07-02 12:19:30 +03:00
# endif
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 11:00:02 +03:00
( 1L < < PG_dirty ) |
LRU_GEN_MASK | LRU_REFS_MASK ) ) ;
2016-01-16 03:54:10 +03:00
mm,thp,rmap: simplify compound page mapcount handling
Compound page (folio) mapcount calculations have been different for anon
and file (or shmem) THPs, and involved the obscure PageDoubleMap flag.
And each huge mapping and unmapping of a file (or shmem) THP involved
atomically incrementing and decrementing the mapcount of every subpage of
that huge page, dirtying many struct page cachelines.
Add subpages_mapcount field to the struct folio and first tail page, so
that the total of subpage mapcounts is available in one place near the
head: then page_mapcount() and total_mapcount() and page_mapped(), and
their folio equivalents, are so quick that anon and file and hugetlb don't
need to be optimized differently. Delete the unloved PageDoubleMap.
page_add and page_remove rmap functions must now maintain the
subpages_mapcount as well as the subpage _mapcount, when dealing with pte
mappings of huge pages; and correct maintenance of NR_ANON_MAPPED and
NR_FILE_MAPPED statistics still needs reading through the subpages, using
nr_subpages_unmapped() - but only when first or last pmd mapping finds
subpages_mapcount raised (double-map case, not the common case).
But are those counts (used to decide when to split an anon THP, and in
vmscan's pagecache_reclaimable heuristic) correctly maintained? Not
quite: since page_remove_rmap() (and also split_huge_pmd()) is often
called without page lock, there can be races when a subpage pte mapcount
0<->1 while compound pmd mapcount 0<->1 is scanning - races which the
previous implementation had prevented. The statistics might become
inaccurate, and even drift down until they underflow through 0. That is
not good enough, but is better dealt with in a followup patch.
Update a few comments on first and second tail page overlaid fields.
hugepage_add_new_anon_rmap() has to "increment" compound_mapcount, but
subpages_mapcount and compound_pincount are already correctly at 0, so
delete its reinitialization of compound_pincount.
A simple 100 X munmap(mmap(2GB, MAP_SHARED|MAP_POPULATE, tmpfs), 2GB) took
18 seconds on small pages, and used to take 1 second on huge pages, but
now takes 119 milliseconds on huge pages. Mapping by pmds a second time
used to take 860ms and now takes 92ms; mapping by pmds after mapping by
ptes (when the scan is needed) used to take 870ms and now takes 495ms.
But there might be some benchmarks which would show a slowdown, because
tail struct pages now fall out of cache until final freeing checks them.
Link: https://lkml.kernel.org/r/47ad693-717-79c8-e1ba-46c3a6602e48@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: James Houghton <jthoughton@google.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Naoya Horiguchi <naoya.horiguchi@linux.dev>
Cc: Peter Xu <peterx@redhat.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zach O'Keefe <zokeefe@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-03 04:51:38 +03:00
/* ->mapping in first and second tail page is replaced by other uses */
2018-12-01 01:10:16 +03:00
VM_BUG_ON_PAGE ( tail > 2 & & page_tail - > mapping ! = TAIL_MAPPING ,
page_tail ) ;
page_tail - > mapping = head - > mapping ;
page_tail - > index = head - > index + tail ;
2022-10-19 16:41:56 +03:00
/*
* page - > private should not be set in tail pages with the exception
* of swap cache pages that store the swp_entry_t in tail pages .
* Fix up and warn once if private is unexpectedly set .
mm,thp,rmap: simplify compound page mapcount handling
Compound page (folio) mapcount calculations have been different for anon
and file (or shmem) THPs, and involved the obscure PageDoubleMap flag.
And each huge mapping and unmapping of a file (or shmem) THP involved
atomically incrementing and decrementing the mapcount of every subpage of
that huge page, dirtying many struct page cachelines.
Add subpages_mapcount field to the struct folio and first tail page, so
that the total of subpage mapcounts is available in one place near the
head: then page_mapcount() and total_mapcount() and page_mapped(), and
their folio equivalents, are so quick that anon and file and hugetlb don't
need to be optimized differently. Delete the unloved PageDoubleMap.
page_add and page_remove rmap functions must now maintain the
subpages_mapcount as well as the subpage _mapcount, when dealing with pte
mappings of huge pages; and correct maintenance of NR_ANON_MAPPED and
NR_FILE_MAPPED statistics still needs reading through the subpages, using
nr_subpages_unmapped() - but only when first or last pmd mapping finds
subpages_mapcount raised (double-map case, not the common case).
But are those counts (used to decide when to split an anon THP, and in
vmscan's pagecache_reclaimable heuristic) correctly maintained? Not
quite: since page_remove_rmap() (and also split_huge_pmd()) is often
called without page lock, there can be races when a subpage pte mapcount
0<->1 while compound pmd mapcount 0<->1 is scanning - races which the
previous implementation had prevented. The statistics might become
inaccurate, and even drift down until they underflow through 0. That is
not good enough, but is better dealt with in a followup patch.
Update a few comments on first and second tail page overlaid fields.
hugepage_add_new_anon_rmap() has to "increment" compound_mapcount, but
subpages_mapcount and compound_pincount are already correctly at 0, so
delete its reinitialization of compound_pincount.
A simple 100 X munmap(mmap(2GB, MAP_SHARED|MAP_POPULATE, tmpfs), 2GB) took
18 seconds on small pages, and used to take 1 second on huge pages, but
now takes 119 milliseconds on huge pages. Mapping by pmds a second time
used to take 860ms and now takes 92ms; mapping by pmds after mapping by
ptes (when the scan is needed) used to take 870ms and now takes 495ms.
But there might be some benchmarks which would show a slowdown, because
tail struct pages now fall out of cache until final freeing checks them.
Link: https://lkml.kernel.org/r/47ad693-717-79c8-e1ba-46c3a6602e48@google.com
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: James Houghton <jthoughton@google.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Naoya Horiguchi <naoya.horiguchi@linux.dev>
Cc: Peter Xu <peterx@redhat.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zach O'Keefe <zokeefe@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-03 04:51:38 +03:00
*
* What of 32 - bit systems , on which head [ 1 ] . compound_pincount overlays
* head [ 1 ] . private ? No problem : THP_SWAP is not enabled on 32 - bit , and
* compound_pincount must be 0 for folio_ref_freeze ( ) to have succeeded .
2022-10-19 16:41:56 +03:00
*/
if ( ! folio_test_swapcache ( page_folio ( head ) ) ) {
2022-10-22 10:51:06 +03:00
VM_WARN_ON_ONCE_PAGE ( page_tail - > private ! = 0 , page_tail ) ;
2022-10-19 16:41:56 +03:00
page_tail - > private = 0 ;
}
2018-12-01 01:10:16 +03:00
2018-04-06 02:23:28 +03:00
/* Page flags must be visible before we make the page non-compound. */
2016-01-16 03:54:10 +03:00
smp_wmb ( ) ;
2018-04-06 02:23:28 +03:00
/*
* Clear PageTail before unfreezing page refcount .
*
* After successful get_page_unless_zero ( ) might follow put_page ( )
* which needs correct compound_head ( ) .
*/
2016-01-16 03:54:10 +03:00
clear_compound_head ( page_tail ) ;
2018-04-06 02:23:28 +03:00
/* Finally unfreeze refcount. Additional reference from page cache. */
page_ref_unfreeze ( page_tail , 1 + ( ! PageAnon ( head ) | |
PageSwapCache ( head ) ) ) ;
2016-01-16 03:54:10 +03:00
if ( page_is_young ( head ) )
set_page_young ( page_tail ) ;
if ( page_is_idle ( head ) )
set_page_idle ( page_tail ) ;
page_cpupid_xchg_last ( page_tail , page_cpupid_last ( head ) ) ;
2018-04-11 02:30:07 +03:00
/*
* always add to the tail because some iterators expect new
* pages to show after the currently processed elements - e . g .
* migrate_pages
*/
2016-01-16 03:54:10 +03:00
lru_add_page_tail ( head , page_tail , lruvec , list ) ;
}
2016-07-27 01:25:51 +03:00
static void __split_huge_page ( struct page * page , struct list_head * list ,
mm/thp: narrow lru locking
lru_lock and page cache xa_lock have no obvious reason to be taken one
way round or the other: until now, lru_lock has been taken before page
cache xa_lock, when splitting a THP; but nothing else takes them
together. Reverse that ordering: let's narrow the lru locking - but
leave local_irq_disable to block interrupts throughout, like before.
Hugh Dickins point: split_huge_page_to_list() was already silly, to be
using the _irqsave variant: it's just been taking sleeping locks, so
would already be broken if entered with interrupts enabled. So we can
save passing flags argument down to __split_huge_page().
Why change the lock ordering here? That was hard to decide. One reason:
when this series reaches per-memcg lru locking, it relies on the THP's
memcg to be stable when taking the lru_lock: that is now done after the
THP's refcount has been frozen, which ensures page memcg cannot change.
Another reason: previously, lock_page_memcg()'s move_lock was presumed
to nest inside lru_lock; but now lru_lock must nest inside (page cache
lock inside) move_lock, so it becomes possible to use lock_page_memcg()
to stabilize page memcg before taking its lru_lock. That is not the
mechanism used in this series, but it is an option we want to keep open.
[hughd@google.com: rewrite commit log]
Link: https://lkml.kernel.org/r/1604566549-62481-5-git-send-email-alex.shi@linux.alibaba.com
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: "Chen, Rong A" <rong.a.chen@intel.com>
Cc: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jann Horn <jannh@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 23:33:33 +03:00
pgoff_t end )
2016-01-16 03:54:10 +03:00
{
2021-06-29 04:59:47 +03:00
struct folio * folio = page_folio ( page ) ;
struct page * head = & folio - > page ;
2016-01-16 03:54:10 +03:00
struct lruvec * lruvec ;
2019-09-24 01:34:52 +03:00
struct address_space * swap_cache = NULL ;
unsigned long offset = 0 ;
2020-10-16 06:05:36 +03:00
unsigned int nr = thp_nr_pages ( head ) ;
2016-03-16 00:57:30 +03:00
int i ;
2016-01-16 03:54:10 +03:00
/* complete memcg works before add pages to LRU */
2021-03-13 08:08:30 +03:00
split_page_memcg ( head , nr ) ;
2016-01-16 03:54:10 +03:00
2019-09-24 01:34:52 +03:00
if ( PageAnon ( head ) & & PageSwapCache ( head ) ) {
swp_entry_t entry = { . val = page_private ( head ) } ;
offset = swp_offset ( entry ) ;
swap_cache = swap_address_space ( entry ) ;
xa_lock ( & swap_cache - > i_pages ) ;
}
2021-05-07 04:06:47 +03:00
/* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2021-06-29 04:59:47 +03:00
lruvec = folio_lruvec_lock ( folio ) ;
mm/thp: narrow lru locking
lru_lock and page cache xa_lock have no obvious reason to be taken one
way round or the other: until now, lru_lock has been taken before page
cache xa_lock, when splitting a THP; but nothing else takes them
together. Reverse that ordering: let's narrow the lru locking - but
leave local_irq_disable to block interrupts throughout, like before.
Hugh Dickins point: split_huge_page_to_list() was already silly, to be
using the _irqsave variant: it's just been taking sleeping locks, so
would already be broken if entered with interrupts enabled. So we can
save passing flags argument down to __split_huge_page().
Why change the lock ordering here? That was hard to decide. One reason:
when this series reaches per-memcg lru locking, it relies on the THP's
memcg to be stable when taking the lru_lock: that is now done after the
THP's refcount has been frozen, which ensures page memcg cannot change.
Another reason: previously, lock_page_memcg()'s move_lock was presumed
to nest inside lru_lock; but now lru_lock must nest inside (page cache
lock inside) move_lock, so it becomes possible to use lock_page_memcg()
to stabilize page memcg before taking its lru_lock. That is not the
mechanism used in this series, but it is an option we want to keep open.
[hughd@google.com: rewrite commit log]
Link: https://lkml.kernel.org/r/1604566549-62481-5-git-send-email-alex.shi@linux.alibaba.com
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: "Chen, Rong A" <rong.a.chen@intel.com>
Cc: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jann Horn <jannh@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 23:33:33 +03:00
2021-10-29 00:36:11 +03:00
ClearPageHasHWPoisoned ( head ) ;
2020-10-16 06:05:36 +03:00
for ( i = nr - 1 ; i > = 1 ; i - - ) {
2016-03-16 00:57:30 +03:00
__split_huge_page_tail ( head , i , lruvec , list ) ;
2021-09-03 00:54:21 +03:00
/* Some pages can be beyond EOF: drop them from page cache */
2016-07-27 01:25:51 +03:00
if ( head [ i ] . index > = end ) {
2022-06-29 03:15:29 +03:00
struct folio * tail = page_folio ( head + i ) ;
2021-09-03 00:54:21 +03:00
if ( shmem_mapping ( head - > mapping ) )
2016-07-27 01:26:18 +03:00
shmem_uncharge ( head - > mapping - > host , 1 ) ;
2022-06-29 03:15:29 +03:00
else if ( folio_test_clear_dirty ( tail ) )
folio_account_cleaned ( tail ,
inode_to_wb ( folio - > mapping - > host ) ) ;
__filemap_remove_folio ( tail , NULL ) ;
folio_put ( tail ) ;
2019-09-24 01:34:52 +03:00
} else if ( ! PageAnon ( page ) ) {
__xa_store ( & head - > mapping - > i_pages , head [ i ] . index ,
head + i , 0 ) ;
} else if ( swap_cache ) {
__xa_store ( & swap_cache - > i_pages , offset + i ,
head + i , 0 ) ;
2016-07-27 01:25:51 +03:00
}
}
2016-01-16 03:54:10 +03:00
ClearPageCompound ( head ) ;
2020-12-15 23:34:29 +03:00
unlock_page_lruvec ( lruvec ) ;
mm/thp: narrow lru locking
lru_lock and page cache xa_lock have no obvious reason to be taken one
way round or the other: until now, lru_lock has been taken before page
cache xa_lock, when splitting a THP; but nothing else takes them
together. Reverse that ordering: let's narrow the lru locking - but
leave local_irq_disable to block interrupts throughout, like before.
Hugh Dickins point: split_huge_page_to_list() was already silly, to be
using the _irqsave variant: it's just been taking sleeping locks, so
would already be broken if entered with interrupts enabled. So we can
save passing flags argument down to __split_huge_page().
Why change the lock ordering here? That was hard to decide. One reason:
when this series reaches per-memcg lru locking, it relies on the THP's
memcg to be stable when taking the lru_lock: that is now done after the
THP's refcount has been frozen, which ensures page memcg cannot change.
Another reason: previously, lock_page_memcg()'s move_lock was presumed
to nest inside lru_lock; but now lru_lock must nest inside (page cache
lock inside) move_lock, so it becomes possible to use lock_page_memcg()
to stabilize page memcg before taking its lru_lock. That is not the
mechanism used in this series, but it is an option we want to keep open.
[hughd@google.com: rewrite commit log]
Link: https://lkml.kernel.org/r/1604566549-62481-5-git-send-email-alex.shi@linux.alibaba.com
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: "Chen, Rong A" <rong.a.chen@intel.com>
Cc: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jann Horn <jannh@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 23:33:33 +03:00
/* Caller disabled irqs, so they are still disabled here */
2019-08-25 03:54:59 +03:00
2020-10-16 06:05:36 +03:00
split_page_owner ( head , nr ) ;
2019-08-25 03:54:59 +03:00
2016-07-27 01:25:51 +03:00
/* See comment in __split_huge_page_tail() */
if ( PageAnon ( head ) ) {
2017-12-04 18:16:10 +03:00
/* Additional pin to swap cache */
2019-09-24 01:34:52 +03:00
if ( PageSwapCache ( head ) ) {
mm, THP, swap: delay splitting THP during swap out
Patch series "THP swap: Delay splitting THP during swapping out", v11.
This patchset is to optimize the performance of Transparent Huge Page
(THP) swap.
Recently, the performance of the storage devices improved so fast that
we cannot saturate the disk bandwidth with single logical CPU when do
page swap out even on a high-end server machine. Because the
performance of the storage device improved faster than that of single
logical CPU. And it seems that the trend will not change in the near
future. On the other hand, the THP becomes more and more popular
because of increased memory size. So it becomes necessary to optimize
THP swap performance.
The advantages of the THP swap support include:
- Batch the swap operations for the THP to reduce lock
acquiring/releasing, including allocating/freeing the swap space,
adding/deleting to/from the swap cache, and writing/reading the swap
space, etc. This will help improve the performance of the THP swap.
- The THP swap space read/write will be 2M sequential IO. It is
particularly helpful for the swap read, which are usually 4k random
IO. This will improve the performance of the THP swap too.
- It will help the memory fragmentation, especially when the THP is
heavily used by the applications. The 2M continuous pages will be
free up after THP swapping out.
- It will improve the THP utilization on the system with the swap
turned on. Because the speed for khugepaged to collapse the normal
pages into the THP is quite slow. After the THP is split during the
swapping out, it will take quite long time for the normal pages to
collapse back into the THP after being swapped in. The high THP
utilization helps the efficiency of the page based memory management
too.
There are some concerns regarding THP swap in, mainly because possible
enlarged read/write IO size (for swap in/out) may put more overhead on
the storage device. To deal with that, the THP swap in should be turned
on only when necessary. For example, it can be selected via
"always/never/madvise" logic, to be turned on globally, turned off
globally, or turned on only for VMA with MADV_HUGEPAGE, etc.
This patchset is the first step for the THP swap support. The plan is
to delay splitting THP step by step, finally avoid splitting THP during
the THP swapping out and swap out/in the THP as a whole.
As the first step, in this patchset, the splitting huge page is delayed
from almost the first step of swapping out to after allocating the swap
space for the THP and adding the THP into the swap cache. This will
reduce lock acquiring/releasing for the locks used for the swap cache
management.
With the patchset, the swap out throughput improves 15.5% (from about
3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
This patch (of 5):
In this patch, splitting huge page is delayed from almost the first step
of swapping out to after allocating the swap space for the THP
(Transparent Huge Page) and adding the THP into the swap cache. This
will batch the corresponding operation, thus improve THP swap out
throughput.
This is the first step for the THP swap optimization. The plan is to
delay splitting the THP step by step and avoid splitting the THP
finally.
In this patch, one swap cluster is used to hold the contents of each THP
swapped out. So, the size of the swap cluster is changed to that of the
THP (Transparent Huge Page) on x86_64 architecture (512). For other
architectures which want such THP swap optimization,
ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for
the architecture. In effect, this will enlarge swap cluster size by 2
times on x86_64. Which may make it harder to find a free cluster when
the swap space becomes fragmented. So that, this may reduce the
continuous swap space allocation and sequential write in theory. The
performance test in 0day shows no regressions caused by this.
In the future of THP swap optimization, some information of the swapped
out THP (such as compound map count) will be recorded in the
swap_cluster_info data structure.
The mem cgroup swap accounting functions are enhanced to support charge
or uncharge a swap cluster backing a THP as a whole.
The swap cluster allocate/free functions are added to allocate/free a
swap cluster for a THP. A fair simple algorithm is used for swap
cluster allocation, that is, only the first swap device in priority list
will be tried to allocate the swap cluster. The function will fail if
the trying is not successful, and the caller will fallback to allocate a
single swap slot instead. This works good enough for normal cases. If
the difference of the number of the free swap clusters among multiple
swap devices is significant, it is possible that some THPs are split
earlier than necessary. For example, this could be caused by big size
difference among multiple swap devices.
The swap cache functions is enhanced to support add/delete THP to/from
the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be
enhanced in the future with multi-order radix tree. But because we will
split the THP soon during swapping out, that optimization doesn't make
much sense for this first step.
The THP splitting functions are enhanced to support to split THP in swap
cache during swapping out. The page lock will be held during allocating
the swap cluster, adding the THP into the swap cache and splitting the
THP. So in the code path other than swapping out, if the THP need to be
split, the PageSwapCache(THP) will be always false.
The swap cluster is only available for SSD, so the THP swap optimization
in this patchset has no effect for HDD.
[ying.huang@intel.com: fix two issues in THP optimize patch]
Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com
[hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size]
Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option]
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h]
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-07 01:37:18 +03:00
page_ref_add ( head , 2 ) ;
2019-09-24 01:34:52 +03:00
xa_unlock ( & swap_cache - > i_pages ) ;
} else {
mm, THP, swap: delay splitting THP during swap out
Patch series "THP swap: Delay splitting THP during swapping out", v11.
This patchset is to optimize the performance of Transparent Huge Page
(THP) swap.
Recently, the performance of the storage devices improved so fast that
we cannot saturate the disk bandwidth with single logical CPU when do
page swap out even on a high-end server machine. Because the
performance of the storage device improved faster than that of single
logical CPU. And it seems that the trend will not change in the near
future. On the other hand, the THP becomes more and more popular
because of increased memory size. So it becomes necessary to optimize
THP swap performance.
The advantages of the THP swap support include:
- Batch the swap operations for the THP to reduce lock
acquiring/releasing, including allocating/freeing the swap space,
adding/deleting to/from the swap cache, and writing/reading the swap
space, etc. This will help improve the performance of the THP swap.
- The THP swap space read/write will be 2M sequential IO. It is
particularly helpful for the swap read, which are usually 4k random
IO. This will improve the performance of the THP swap too.
- It will help the memory fragmentation, especially when the THP is
heavily used by the applications. The 2M continuous pages will be
free up after THP swapping out.
- It will improve the THP utilization on the system with the swap
turned on. Because the speed for khugepaged to collapse the normal
pages into the THP is quite slow. After the THP is split during the
swapping out, it will take quite long time for the normal pages to
collapse back into the THP after being swapped in. The high THP
utilization helps the efficiency of the page based memory management
too.
There are some concerns regarding THP swap in, mainly because possible
enlarged read/write IO size (for swap in/out) may put more overhead on
the storage device. To deal with that, the THP swap in should be turned
on only when necessary. For example, it can be selected via
"always/never/madvise" logic, to be turned on globally, turned off
globally, or turned on only for VMA with MADV_HUGEPAGE, etc.
This patchset is the first step for the THP swap support. The plan is
to delay splitting THP step by step, finally avoid splitting THP during
the THP swapping out and swap out/in the THP as a whole.
As the first step, in this patchset, the splitting huge page is delayed
from almost the first step of swapping out to after allocating the swap
space for the THP and adding the THP into the swap cache. This will
reduce lock acquiring/releasing for the locks used for the swap cache
management.
With the patchset, the swap out throughput improves 15.5% (from about
3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
This patch (of 5):
In this patch, splitting huge page is delayed from almost the first step
of swapping out to after allocating the swap space for the THP
(Transparent Huge Page) and adding the THP into the swap cache. This
will batch the corresponding operation, thus improve THP swap out
throughput.
This is the first step for the THP swap optimization. The plan is to
delay splitting the THP step by step and avoid splitting the THP
finally.
In this patch, one swap cluster is used to hold the contents of each THP
swapped out. So, the size of the swap cluster is changed to that of the
THP (Transparent Huge Page) on x86_64 architecture (512). For other
architectures which want such THP swap optimization,
ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for
the architecture. In effect, this will enlarge swap cluster size by 2
times on x86_64. Which may make it harder to find a free cluster when
the swap space becomes fragmented. So that, this may reduce the
continuous swap space allocation and sequential write in theory. The
performance test in 0day shows no regressions caused by this.
In the future of THP swap optimization, some information of the swapped
out THP (such as compound map count) will be recorded in the
swap_cluster_info data structure.
The mem cgroup swap accounting functions are enhanced to support charge
or uncharge a swap cluster backing a THP as a whole.
The swap cluster allocate/free functions are added to allocate/free a
swap cluster for a THP. A fair simple algorithm is used for swap
cluster allocation, that is, only the first swap device in priority list
will be tried to allocate the swap cluster. The function will fail if
the trying is not successful, and the caller will fallback to allocate a
single swap slot instead. This works good enough for normal cases. If
the difference of the number of the free swap clusters among multiple
swap devices is significant, it is possible that some THPs are split
earlier than necessary. For example, this could be caused by big size
difference among multiple swap devices.
The swap cache functions is enhanced to support add/delete THP to/from
the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be
enhanced in the future with multi-order radix tree. But because we will
split the THP soon during swapping out, that optimization doesn't make
much sense for this first step.
The THP splitting functions are enhanced to support to split THP in swap
cache during swapping out. The page lock will be held during allocating
the swap cluster, adding the THP into the swap cache and splitting the
THP. So in the code path other than swapping out, if the THP need to be
split, the PageSwapCache(THP) will be always false.
The swap cluster is only available for SSD, so the THP swap optimization
in this patchset has no effect for HDD.
[ying.huang@intel.com: fix two issues in THP optimize patch]
Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com
[hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size]
Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option]
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h]
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-07 01:37:18 +03:00
page_ref_inc ( head ) ;
2019-09-24 01:34:52 +03:00
}
2016-07-27 01:25:51 +03:00
} else {
2017-12-04 18:16:10 +03:00
/* Additional pin to page cache */
2016-07-27 01:25:51 +03:00
page_ref_add ( head , 2 ) ;
2018-04-11 02:36:56 +03:00
xa_unlock ( & head - > mapping - > i_pages ) ;
2016-07-27 01:25:51 +03:00
}
mm/thp: narrow lru locking
lru_lock and page cache xa_lock have no obvious reason to be taken one
way round or the other: until now, lru_lock has been taken before page
cache xa_lock, when splitting a THP; but nothing else takes them
together. Reverse that ordering: let's narrow the lru locking - but
leave local_irq_disable to block interrupts throughout, like before.
Hugh Dickins point: split_huge_page_to_list() was already silly, to be
using the _irqsave variant: it's just been taking sleeping locks, so
would already be broken if entered with interrupts enabled. So we can
save passing flags argument down to __split_huge_page().
Why change the lock ordering here? That was hard to decide. One reason:
when this series reaches per-memcg lru locking, it relies on the THP's
memcg to be stable when taking the lru_lock: that is now done after the
THP's refcount has been frozen, which ensures page memcg cannot change.
Another reason: previously, lock_page_memcg()'s move_lock was presumed
to nest inside lru_lock; but now lru_lock must nest inside (page cache
lock inside) move_lock, so it becomes possible to use lock_page_memcg()
to stabilize page memcg before taking its lru_lock. That is not the
mechanism used in this series, but it is an option we want to keep open.
[hughd@google.com: rewrite commit log]
Link: https://lkml.kernel.org/r/1604566549-62481-5-git-send-email-alex.shi@linux.alibaba.com
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: "Chen, Rong A" <rong.a.chen@intel.com>
Cc: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jann Horn <jannh@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 23:33:33 +03:00
local_irq_enable ( ) ;
2016-01-16 03:54:10 +03:00
2022-01-29 07:32:59 +03:00
remap_page ( folio , nr ) ;
2016-01-16 03:54:10 +03:00
mm: fix a race during THP splitting
It is reported that the following bug is triggered if the HDD is used as
swap device,
[ 5758.157556] BUG: kernel NULL pointer dereference, address: 0000000000000007
[ 5758.165331] #PF: supervisor write access in kernel mode
[ 5758.171161] #PF: error_code(0x0002) - not-present page
[ 5758.176894] PGD 0 P4D 0
[ 5758.179721] Oops: 0002 [#1] SMP PTI
[ 5758.183614] CPU: 10 PID: 316 Comm: kswapd1 Kdump: loaded Tainted: G S --------- --- 5.9.0-0.rc3.1.tst.el8.x86_64 #1
[ 5758.196717] Hardware name: Intel Corporation S2600CP/S2600CP, BIOS SE5C600.86B.02.01.0002.082220131453 08/22/2013
[ 5758.208176] RIP: 0010:split_swap_cluster+0x47/0x60
[ 5758.213522] Code: c1 e3 06 48 c1 eb 0f 48 8d 1c d8 48 89 df e8 d0 20 6a 00 80 63 07 fb 48 85 db 74 16 48 89 df c6 07 00 66 66 66 90 31 c0 5b c3 <80> 24 25 07 00 00 00 fb 31 c0 5b c3 b8 f0 ff ff ff 5b c3 66 0f 1f
[ 5758.234478] RSP: 0018:ffffb147442d7af0 EFLAGS: 00010246
[ 5758.240309] RAX: 0000000000000000 RBX: 000000000014b217 RCX: ffffb14779fd9000
[ 5758.248281] RDX: 000000000014b217 RSI: ffff9c52f2ab1400 RDI: 000000000014b217
[ 5758.256246] RBP: ffffe00c51168080 R08: ffffe00c5116fe08 R09: ffff9c52fffd3000
[ 5758.264208] R10: ffffe00c511537c8 R11: ffff9c52fffd3c90 R12: 0000000000000000
[ 5758.272172] R13: ffffe00c51170000 R14: ffffe00c51170000 R15: ffffe00c51168040
[ 5758.280134] FS: 0000000000000000(0000) GS:ffff9c52f2a80000(0000) knlGS:0000000000000000
[ 5758.289163] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 5758.295575] CR2: 0000000000000007 CR3: 0000000022a0e003 CR4: 00000000000606e0
[ 5758.303538] Call Trace:
[ 5758.306273] split_huge_page_to_list+0x88b/0x950
[ 5758.311433] deferred_split_scan+0x1ca/0x310
[ 5758.316202] do_shrink_slab+0x12c/0x2a0
[ 5758.320491] shrink_slab+0x20f/0x2c0
[ 5758.324482] shrink_node+0x240/0x6c0
[ 5758.328469] balance_pgdat+0x2d1/0x550
[ 5758.332652] kswapd+0x201/0x3c0
[ 5758.336157] ? finish_wait+0x80/0x80
[ 5758.340147] ? balance_pgdat+0x550/0x550
[ 5758.344525] kthread+0x114/0x130
[ 5758.348126] ? kthread_park+0x80/0x80
[ 5758.352214] ret_from_fork+0x22/0x30
[ 5758.356203] Modules linked in: fuse zram rfkill sunrpc intel_rapl_msr intel_rapl_common sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp mgag200 iTCO_wdt crct10dif_pclmul iTCO_vendor_support drm_kms_helper crc32_pclmul ghash_clmulni_intel syscopyarea sysfillrect sysimgblt fb_sys_fops cec rapl joydev intel_cstate ipmi_si ipmi_devintf drm intel_uncore i2c_i801 ipmi_msghandler pcspkr lpc_ich mei_me i2c_smbus mei ioatdma ip_tables xfs libcrc32c sr_mod sd_mod cdrom t10_pi sg igb ahci libahci i2c_algo_bit crc32c_intel libata dca wmi dm_mirror dm_region_hash dm_log dm_mod
[ 5758.412673] CR2: 0000000000000007
[ 0.000000] Linux version 5.9.0-0.rc3.1.tst.el8.x86_64 (mockbuild@x86-vm-15.build.eng.bos.redhat.com) (gcc (GCC) 8.3.1 20191121 (Red Hat 8.3.1-5), GNU ld version 2.30-79.el8) #1 SMP Wed Sep 9 16:03:34 EDT 2020
After further digging it's found that the following race condition exists in the
original implementation,
CPU1 CPU2
---- ----
deferred_split_scan()
split_huge_page(page) /* page isn't compound head */
split_huge_page_to_list(page, NULL)
__split_huge_page(page, )
ClearPageCompound(head)
/* unlock all subpages except page (not head) */
add_to_swap(head) /* not THP */
get_swap_page(head)
add_to_swap_cache(head, )
SetPageSwapCache(head)
if PageSwapCache(head)
split_swap_cluster(/* swap entry of head */)
/* Deref sis->cluster_info: NULL accessing! */
So, in split_huge_page_to_list(), PageSwapCache() is called for the already
split and unlocked "head", which may be added to swap cache in another CPU. So
split_swap_cluster() may be called wrongly.
To fix the race, the call to split_swap_cluster() is moved to
__split_huge_page() before all subpages are unlocked. So that the
PageSwapCache() is stable.
Fixes: 59807685a7e77 ("mm, THP, swap: support splitting THP for THP swap out")
Reported-by: Rafael Aquini <aquini@redhat.com>
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Tested-by: Rafael Aquini <aquini@redhat.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Link: https://lkml.kernel.org/r/20201009073647.1531083-1-ying.huang@intel.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-16 06:06:07 +03:00
if ( PageSwapCache ( head ) ) {
swp_entry_t entry = { . val = page_private ( head ) } ;
split_swap_cluster ( entry ) ;
}
2020-10-16 06:05:36 +03:00
for ( i = 0 ; i < nr ; i + + ) {
2016-01-16 03:54:10 +03:00
struct page * subpage = head + i ;
if ( subpage = = page )
continue ;
unlock_page ( subpage ) ;
/*
* Subpages may be freed if there wasn ' t any mapping
* like if add_to_swap ( ) is running on a lru page that
* had its mapping zapped . And freeing these pages
* requires taking the lru_lock so we do the put_page
* of the tail pages after the split is complete .
*/
2022-07-04 16:21:56 +03:00
free_page_and_swap_cache ( subpage ) ;
2016-01-16 03:54:10 +03:00
}
}
mm, THP, swap: check whether THP can be split firstly
To swap out THP (Transparent Huage Page), before splitting the THP, the
swap cluster will be allocated and the THP will be added into the swap
cache. But it is possible that the THP cannot be split, so that we must
delete the THP from the swap cache and free the swap cluster. To avoid
that, in this patch, whether the THP can be split is checked firstly.
The check can only be done racy, but it is good enough for most cases.
With the patch, the swap out throughput improves 3.6% (from about
4.16GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
Link: http://lkml.kernel.org/r/20170515112522.32457-5-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for can_split_huge_page()]
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-07 01:37:28 +03:00
/* Racy check whether the huge page can be split */
2022-02-04 22:13:31 +03:00
bool can_split_folio ( struct folio * folio , int * pextra_pins )
mm, THP, swap: check whether THP can be split firstly
To swap out THP (Transparent Huage Page), before splitting the THP, the
swap cluster will be allocated and the THP will be added into the swap
cache. But it is possible that the THP cannot be split, so that we must
delete the THP from the swap cache and free the swap cluster. To avoid
that, in this patch, whether the THP can be split is checked firstly.
The check can only be done racy, but it is good enough for most cases.
With the patch, the swap out throughput improves 3.6% (from about
4.16GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
Link: http://lkml.kernel.org/r/20170515112522.32457-5-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for can_split_huge_page()]
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-07 01:37:28 +03:00
{
int extra_pins ;
2017-12-04 18:16:10 +03:00
/* Additional pins from page cache */
2022-02-04 22:13:31 +03:00
if ( folio_test_anon ( folio ) )
extra_pins = folio_test_swapcache ( folio ) ?
folio_nr_pages ( folio ) : 0 ;
mm, THP, swap: check whether THP can be split firstly
To swap out THP (Transparent Huage Page), before splitting the THP, the
swap cluster will be allocated and the THP will be added into the swap
cache. But it is possible that the THP cannot be split, so that we must
delete the THP from the swap cache and free the swap cluster. To avoid
that, in this patch, whether the THP can be split is checked firstly.
The check can only be done racy, but it is good enough for most cases.
With the patch, the swap out throughput improves 3.6% (from about
4.16GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
Link: http://lkml.kernel.org/r/20170515112522.32457-5-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for can_split_huge_page()]
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-07 01:37:28 +03:00
else
2022-02-04 22:13:31 +03:00
extra_pins = folio_nr_pages ( folio ) ;
mm, THP, swap: check whether THP can be split firstly
To swap out THP (Transparent Huage Page), before splitting the THP, the
swap cluster will be allocated and the THP will be added into the swap
cache. But it is possible that the THP cannot be split, so that we must
delete the THP from the swap cache and free the swap cluster. To avoid
that, in this patch, whether the THP can be split is checked firstly.
The check can only be done racy, but it is good enough for most cases.
With the patch, the swap out throughput improves 3.6% (from about
4.16GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
Link: http://lkml.kernel.org/r/20170515112522.32457-5-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for can_split_huge_page()]
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-07 01:37:28 +03:00
if ( pextra_pins )
* pextra_pins = extra_pins ;
2022-02-04 22:13:31 +03:00
return folio_mapcount ( folio ) = = folio_ref_count ( folio ) - extra_pins - 1 ;
mm, THP, swap: check whether THP can be split firstly
To swap out THP (Transparent Huage Page), before splitting the THP, the
swap cluster will be allocated and the THP will be added into the swap
cache. But it is possible that the THP cannot be split, so that we must
delete the THP from the swap cache and free the swap cluster. To avoid
that, in this patch, whether the THP can be split is checked firstly.
The check can only be done racy, but it is good enough for most cases.
With the patch, the swap out throughput improves 3.6% (from about
4.16GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
Link: http://lkml.kernel.org/r/20170515112522.32457-5-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for can_split_huge_page()]
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-07 01:37:28 +03:00
}
2016-01-16 03:54:10 +03:00
/*
* This function splits huge page into normal pages . @ page can point to any
* subpage of huge page to split . Split doesn ' t change the position of @ page .
*
* Only caller must hold pin on the @ page , otherwise split fails with - EBUSY .
* The huge page must be locked .
*
* If @ list is null , tail pages will be added to LRU list , otherwise , to @ list .
*
* Both head page and tail pages will inherit mapping , flags , and so on from
* the hugepage .
*
* GUP pin and PG_locked transferred to @ page . Rest subpages can be freed if
* they are not mapped .
*
* Returns 0 if the hugepage is split successfully .
* Returns - EBUSY if the page is pinned or if anon_vma disappeared from under
* us .
*/
int split_huge_page_to_list ( struct page * page , struct list_head * list )
{
2022-01-29 07:32:59 +03:00
struct folio * folio = page_folio ( page ) ;
2022-09-02 22:46:48 +03:00
struct deferred_split * ds_queue = get_deferred_split_queue ( & folio - > page ) ;
XA_STATE ( xas , & folio - > mapping - > i_pages , folio - > index ) ;
2016-07-27 01:25:51 +03:00
struct anon_vma * anon_vma = NULL ;
struct address_space * mapping = NULL ;
2021-06-16 04:24:07 +03:00
int extra_pins , ret ;
2018-12-01 01:10:21 +03:00
pgoff_t end ;
2022-04-29 09:14:43 +03:00
bool is_hzp ;
2016-01-16 03:54:10 +03:00
2022-09-02 22:46:48 +03:00
VM_BUG_ON_FOLIO ( ! folio_test_locked ( folio ) , folio ) ;
VM_BUG_ON_FOLIO ( ! folio_test_large ( folio ) , folio ) ;
2016-01-16 03:54:10 +03:00
2022-09-02 22:46:48 +03:00
is_hzp = is_huge_zero_page ( & folio - > page ) ;
VM_WARN_ON_ONCE_FOLIO ( is_hzp , folio ) ;
2022-04-29 09:14:43 +03:00
if ( is_hzp )
return - EBUSY ;
2022-09-02 22:46:48 +03:00
if ( folio_test_writeback ( folio ) )
2017-09-07 02:22:34 +03:00
return - EBUSY ;
2022-09-02 22:46:48 +03:00
if ( folio_test_anon ( folio ) ) {
2016-07-27 01:25:51 +03:00
/*
2020-06-09 07:33:54 +03:00
* The caller does not necessarily hold an mmap_lock that would
2016-07-27 01:25:51 +03:00
* prevent the anon_vma disappearing so we first we take a
* reference to it and then lock the anon_vma for write . This
2022-01-30 00:06:53 +03:00
* is similar to folio_lock_anon_vma_read except the write lock
2016-07-27 01:25:51 +03:00
* is taken to serialise against parallel split or collapse
* operations .
*/
2022-09-02 22:46:50 +03:00
anon_vma = folio_get_anon_vma ( folio ) ;
2016-07-27 01:25:51 +03:00
if ( ! anon_vma ) {
ret = - EBUSY ;
goto out ;
}
2018-12-01 01:10:21 +03:00
end = - 1 ;
2016-07-27 01:25:51 +03:00
mapping = NULL ;
anon_vma_lock_write ( anon_vma ) ;
} else {
2022-08-10 09:49:07 +03:00
gfp_t gfp ;
2022-09-02 22:46:48 +03:00
mapping = folio - > mapping ;
2016-07-27 01:25:51 +03:00
/* Truncated ? */
if ( ! mapping ) {
ret = - EBUSY ;
goto out ;
}
2022-08-10 09:49:07 +03:00
gfp = current_gfp_context ( mapping_gfp_mask ( mapping ) &
GFP_RECLAIM_MASK ) ;
if ( folio_test_private ( folio ) & &
! filemap_release_folio ( folio , gfp ) ) {
ret = - EBUSY ;
goto out ;
}
2022-09-02 22:46:48 +03:00
xas_split_alloc ( & xas , folio , folio_order ( folio ) , gfp ) ;
2020-06-28 05:19:08 +03:00
if ( xas_error ( & xas ) ) {
ret = xas_error ( & xas ) ;
goto out ;
}
2016-07-27 01:25:51 +03:00
anon_vma = NULL ;
i_mmap_lock_read ( mapping ) ;
2018-12-01 01:10:21 +03:00
/*
* __split_huge_page ( ) may need to trim off pages beyond EOF :
* but on 32 - bit , i_size_read ( ) takes an irq - unsafe seqlock ,
* which cannot be nested inside the page tree lock . So note
* end now : i_size itself may be changed at any moment , but
2022-09-02 22:46:48 +03:00
* folio lock is good enough to serialize the trimming .
2018-12-01 01:10:21 +03:00
*/
end = DIV_ROUND_UP ( i_size_read ( mapping - > host ) , PAGE_SIZE ) ;
2021-09-03 00:54:21 +03:00
if ( shmem_mapping ( mapping ) )
end = shmem_fallocend ( mapping - > host , end ) ;
2016-01-16 03:54:10 +03:00
}
/*
2022-09-02 22:46:49 +03:00
* Racy check if we can split the page , before unmap_folio ( ) will
2016-01-16 03:54:10 +03:00
* split PMDs
*/
2022-02-04 22:13:31 +03:00
if ( ! can_split_folio ( folio , & extra_pins ) ) {
mm: migrate: try again if THP split is failed due to page refcnt
When creating a virtual machine, we will use memfd_create() to get a file
descriptor which can be used to create share memory mappings using the
mmap function, meanwhile the mmap() will set the MAP_POPULATE flag to
allocate physical pages for the virtual machine.
When allocating physical pages for the guest, the host can fallback to
allocate some CMA pages for the guest when over half of the zone's free
memory is in the CMA area.
In guest os, when the application wants to do some data transaction with
DMA, our QEMU will call VFIO_IOMMU_MAP_DMA ioctl to do longterm-pin and
create IOMMU mappings for the DMA pages. However, when calling
VFIO_IOMMU_MAP_DMA ioctl to pin the physical pages, we found it will be
failed to longterm-pin sometimes.
After some invetigation, we found the pages used to do DMA mapping can
contain some CMA pages, and these CMA pages will cause a possible failure
of the longterm-pin, due to failed to migrate the CMA pages. The reason
of migration failure may be temporary reference count or memory allocation
failure. So that will cause the VFIO_IOMMU_MAP_DMA ioctl returns error,
which makes the application failed to start.
I observed one migration failure case (which is not easy to reproduce) is
that, the 'thp_migration_fail' count is 1 and the 'thp_split_page_failed'
count is also 1.
That means when migrating a THP which is in CMA area, but can not allocate
a new THP due to memory fragmentation, so it will split the THP. However
THP split is also failed, probably the reason is temporary reference count
of this THP. And the temporary reference count can be caused by dropping
page caches (I observed the drop caches operation in the system), but we
can not drop the shmem page caches due to they are already dirty at that
time.
Especially for THP split failure, which is caused by temporary reference
count, we can try again to mitigate the failure of migration in this case
according to previous discussion [1].
[1] https://lore.kernel.org/all/470dc638-a300-f261-94b4-e27250e42f96@redhat.com/
Link: https://lkml.kernel.org/r/6784730480a1df82e8f4cba1ed088e4ac767994b.1666599848.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-10-24 11:34:22 +03:00
ret = - EAGAIN ;
2016-01-16 03:54:10 +03:00
goto out_unlock ;
}
2022-09-02 22:46:49 +03:00
unmap_folio ( folio ) ;
2016-01-16 03:54:10 +03:00
mm/thp: narrow lru locking
lru_lock and page cache xa_lock have no obvious reason to be taken one
way round or the other: until now, lru_lock has been taken before page
cache xa_lock, when splitting a THP; but nothing else takes them
together. Reverse that ordering: let's narrow the lru locking - but
leave local_irq_disable to block interrupts throughout, like before.
Hugh Dickins point: split_huge_page_to_list() was already silly, to be
using the _irqsave variant: it's just been taking sleeping locks, so
would already be broken if entered with interrupts enabled. So we can
save passing flags argument down to __split_huge_page().
Why change the lock ordering here? That was hard to decide. One reason:
when this series reaches per-memcg lru locking, it relies on the THP's
memcg to be stable when taking the lru_lock: that is now done after the
THP's refcount has been frozen, which ensures page memcg cannot change.
Another reason: previously, lock_page_memcg()'s move_lock was presumed
to nest inside lru_lock; but now lru_lock must nest inside (page cache
lock inside) move_lock, so it becomes possible to use lock_page_memcg()
to stabilize page memcg before taking its lru_lock. That is not the
mechanism used in this series, but it is an option we want to keep open.
[hughd@google.com: rewrite commit log]
Link: https://lkml.kernel.org/r/1604566549-62481-5-git-send-email-alex.shi@linux.alibaba.com
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: "Chen, Rong A" <rong.a.chen@intel.com>
Cc: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jann Horn <jannh@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 23:33:33 +03:00
/* block interrupt reentry in xa_lock and spinlock */
local_irq_disable ( ) ;
2016-07-27 01:25:51 +03:00
if ( mapping ) {
/*
2022-09-02 22:46:48 +03:00
* Check if the folio is present in page cache .
* We assume all tail are present too , if folio is there .
2016-07-27 01:25:51 +03:00
*/
2020-06-28 05:19:08 +03:00
xas_lock ( & xas ) ;
xas_reset ( & xas ) ;
2022-09-02 22:46:48 +03:00
if ( xas_load ( & xas ) ! = folio )
2016-07-27 01:25:51 +03:00
goto fail ;
}
2016-05-20 03:10:49 +03:00
/* Prevent deferred_split_scan() touching ->_refcount */
2019-09-24 01:38:06 +03:00
spin_lock ( & ds_queue - > split_queue_lock ) ;
2022-09-02 22:46:48 +03:00
if ( folio_ref_freeze ( folio , 1 + extra_pins ) ) {
if ( ! list_empty ( page_deferred_list ( & folio - > page ) ) ) {
2019-09-24 01:38:06 +03:00
ds_queue - > split_queue_len - - ;
2022-09-02 22:46:48 +03:00
list_del ( page_deferred_list ( & folio - > page ) ) ;
2016-01-16 03:54:17 +03:00
}
2020-01-31 09:14:35 +03:00
spin_unlock ( & ds_queue - > split_queue_lock ) ;
2019-10-19 06:20:30 +03:00
if ( mapping ) {
2022-09-02 22:46:48 +03:00
int nr = folio_nr_pages ( folio ) ;
2021-02-24 23:03:27 +03:00
2022-09-02 22:46:48 +03:00
xas_split ( & xas , folio , folio_order ( folio ) ) ;
if ( folio_test_swapbacked ( folio ) ) {
__lruvec_stat_mod_folio ( folio , NR_SHMEM_THPS ,
2021-02-24 23:03:31 +03:00
- nr ) ;
2021-10-19 01:16:19 +03:00
} else {
2022-09-02 22:46:48 +03:00
__lruvec_stat_mod_folio ( folio , NR_FILE_THPS ,
2021-02-24 23:03:27 +03:00
- nr ) ;
2021-10-19 01:16:19 +03:00
filemap_nr_thps_dec ( mapping ) ;
}
2019-10-19 06:20:30 +03:00
}
mm/thp: narrow lru locking
lru_lock and page cache xa_lock have no obvious reason to be taken one
way round or the other: until now, lru_lock has been taken before page
cache xa_lock, when splitting a THP; but nothing else takes them
together. Reverse that ordering: let's narrow the lru locking - but
leave local_irq_disable to block interrupts throughout, like before.
Hugh Dickins point: split_huge_page_to_list() was already silly, to be
using the _irqsave variant: it's just been taking sleeping locks, so
would already be broken if entered with interrupts enabled. So we can
save passing flags argument down to __split_huge_page().
Why change the lock ordering here? That was hard to decide. One reason:
when this series reaches per-memcg lru locking, it relies on the THP's
memcg to be stable when taking the lru_lock: that is now done after the
THP's refcount has been frozen, which ensures page memcg cannot change.
Another reason: previously, lock_page_memcg()'s move_lock was presumed
to nest inside lru_lock; but now lru_lock must nest inside (page cache
lock inside) move_lock, so it becomes possible to use lock_page_memcg()
to stabilize page memcg before taking its lru_lock. That is not the
mechanism used in this series, but it is an option we want to keep open.
[hughd@google.com: rewrite commit log]
Link: https://lkml.kernel.org/r/1604566549-62481-5-git-send-email-alex.shi@linux.alibaba.com
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: "Chen, Rong A" <rong.a.chen@intel.com>
Cc: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jann Horn <jannh@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 23:33:33 +03:00
__split_huge_page ( page , list , end ) ;
mm: fix a race during THP splitting
It is reported that the following bug is triggered if the HDD is used as
swap device,
[ 5758.157556] BUG: kernel NULL pointer dereference, address: 0000000000000007
[ 5758.165331] #PF: supervisor write access in kernel mode
[ 5758.171161] #PF: error_code(0x0002) - not-present page
[ 5758.176894] PGD 0 P4D 0
[ 5758.179721] Oops: 0002 [#1] SMP PTI
[ 5758.183614] CPU: 10 PID: 316 Comm: kswapd1 Kdump: loaded Tainted: G S --------- --- 5.9.0-0.rc3.1.tst.el8.x86_64 #1
[ 5758.196717] Hardware name: Intel Corporation S2600CP/S2600CP, BIOS SE5C600.86B.02.01.0002.082220131453 08/22/2013
[ 5758.208176] RIP: 0010:split_swap_cluster+0x47/0x60
[ 5758.213522] Code: c1 e3 06 48 c1 eb 0f 48 8d 1c d8 48 89 df e8 d0 20 6a 00 80 63 07 fb 48 85 db 74 16 48 89 df c6 07 00 66 66 66 90 31 c0 5b c3 <80> 24 25 07 00 00 00 fb 31 c0 5b c3 b8 f0 ff ff ff 5b c3 66 0f 1f
[ 5758.234478] RSP: 0018:ffffb147442d7af0 EFLAGS: 00010246
[ 5758.240309] RAX: 0000000000000000 RBX: 000000000014b217 RCX: ffffb14779fd9000
[ 5758.248281] RDX: 000000000014b217 RSI: ffff9c52f2ab1400 RDI: 000000000014b217
[ 5758.256246] RBP: ffffe00c51168080 R08: ffffe00c5116fe08 R09: ffff9c52fffd3000
[ 5758.264208] R10: ffffe00c511537c8 R11: ffff9c52fffd3c90 R12: 0000000000000000
[ 5758.272172] R13: ffffe00c51170000 R14: ffffe00c51170000 R15: ffffe00c51168040
[ 5758.280134] FS: 0000000000000000(0000) GS:ffff9c52f2a80000(0000) knlGS:0000000000000000
[ 5758.289163] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 5758.295575] CR2: 0000000000000007 CR3: 0000000022a0e003 CR4: 00000000000606e0
[ 5758.303538] Call Trace:
[ 5758.306273] split_huge_page_to_list+0x88b/0x950
[ 5758.311433] deferred_split_scan+0x1ca/0x310
[ 5758.316202] do_shrink_slab+0x12c/0x2a0
[ 5758.320491] shrink_slab+0x20f/0x2c0
[ 5758.324482] shrink_node+0x240/0x6c0
[ 5758.328469] balance_pgdat+0x2d1/0x550
[ 5758.332652] kswapd+0x201/0x3c0
[ 5758.336157] ? finish_wait+0x80/0x80
[ 5758.340147] ? balance_pgdat+0x550/0x550
[ 5758.344525] kthread+0x114/0x130
[ 5758.348126] ? kthread_park+0x80/0x80
[ 5758.352214] ret_from_fork+0x22/0x30
[ 5758.356203] Modules linked in: fuse zram rfkill sunrpc intel_rapl_msr intel_rapl_common sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp mgag200 iTCO_wdt crct10dif_pclmul iTCO_vendor_support drm_kms_helper crc32_pclmul ghash_clmulni_intel syscopyarea sysfillrect sysimgblt fb_sys_fops cec rapl joydev intel_cstate ipmi_si ipmi_devintf drm intel_uncore i2c_i801 ipmi_msghandler pcspkr lpc_ich mei_me i2c_smbus mei ioatdma ip_tables xfs libcrc32c sr_mod sd_mod cdrom t10_pi sg igb ahci libahci i2c_algo_bit crc32c_intel libata dca wmi dm_mirror dm_region_hash dm_log dm_mod
[ 5758.412673] CR2: 0000000000000007
[ 0.000000] Linux version 5.9.0-0.rc3.1.tst.el8.x86_64 (mockbuild@x86-vm-15.build.eng.bos.redhat.com) (gcc (GCC) 8.3.1 20191121 (Red Hat 8.3.1-5), GNU ld version 2.30-79.el8) #1 SMP Wed Sep 9 16:03:34 EDT 2020
After further digging it's found that the following race condition exists in the
original implementation,
CPU1 CPU2
---- ----
deferred_split_scan()
split_huge_page(page) /* page isn't compound head */
split_huge_page_to_list(page, NULL)
__split_huge_page(page, )
ClearPageCompound(head)
/* unlock all subpages except page (not head) */
add_to_swap(head) /* not THP */
get_swap_page(head)
add_to_swap_cache(head, )
SetPageSwapCache(head)
if PageSwapCache(head)
split_swap_cluster(/* swap entry of head */)
/* Deref sis->cluster_info: NULL accessing! */
So, in split_huge_page_to_list(), PageSwapCache() is called for the already
split and unlocked "head", which may be added to swap cache in another CPU. So
split_swap_cluster() may be called wrongly.
To fix the race, the call to split_swap_cluster() is moved to
__split_huge_page() before all subpages are unlocked. So that the
PageSwapCache() is stable.
Fixes: 59807685a7e77 ("mm, THP, swap: support splitting THP for THP swap out")
Reported-by: Rafael Aquini <aquini@redhat.com>
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Tested-by: Rafael Aquini <aquini@redhat.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Link: https://lkml.kernel.org/r/20201009073647.1531083-1-ying.huang@intel.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-10-16 06:06:07 +03:00
ret = 0 ;
2016-01-16 03:54:10 +03:00
} else {
2019-09-24 01:38:06 +03:00
spin_unlock ( & ds_queue - > split_queue_lock ) ;
2021-06-16 04:24:07 +03:00
fail :
if ( mapping )
2020-06-28 05:19:08 +03:00
xas_unlock ( & xas ) ;
mm/thp: narrow lru locking
lru_lock and page cache xa_lock have no obvious reason to be taken one
way round or the other: until now, lru_lock has been taken before page
cache xa_lock, when splitting a THP; but nothing else takes them
together. Reverse that ordering: let's narrow the lru locking - but
leave local_irq_disable to block interrupts throughout, like before.
Hugh Dickins point: split_huge_page_to_list() was already silly, to be
using the _irqsave variant: it's just been taking sleeping locks, so
would already be broken if entered with interrupts enabled. So we can
save passing flags argument down to __split_huge_page().
Why change the lock ordering here? That was hard to decide. One reason:
when this series reaches per-memcg lru locking, it relies on the THP's
memcg to be stable when taking the lru_lock: that is now done after the
THP's refcount has been frozen, which ensures page memcg cannot change.
Another reason: previously, lock_page_memcg()'s move_lock was presumed
to nest inside lru_lock; but now lru_lock must nest inside (page cache
lock inside) move_lock, so it becomes possible to use lock_page_memcg()
to stabilize page memcg before taking its lru_lock. That is not the
mechanism used in this series, but it is an option we want to keep open.
[hughd@google.com: rewrite commit log]
Link: https://lkml.kernel.org/r/1604566549-62481-5-git-send-email-alex.shi@linux.alibaba.com
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Alexander Duyck <alexander.duyck@gmail.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: "Chen, Rong A" <rong.a.chen@intel.com>
Cc: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jann Horn <jannh@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Wei Yang <richard.weiyang@gmail.com>
Cc: Yang Shi <yang.shi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 23:33:33 +03:00
local_irq_enable ( ) ;
2022-01-29 07:32:59 +03:00
remap_page ( folio , folio_nr_pages ( folio ) ) ;
mm: migrate: try again if THP split is failed due to page refcnt
When creating a virtual machine, we will use memfd_create() to get a file
descriptor which can be used to create share memory mappings using the
mmap function, meanwhile the mmap() will set the MAP_POPULATE flag to
allocate physical pages for the virtual machine.
When allocating physical pages for the guest, the host can fallback to
allocate some CMA pages for the guest when over half of the zone's free
memory is in the CMA area.
In guest os, when the application wants to do some data transaction with
DMA, our QEMU will call VFIO_IOMMU_MAP_DMA ioctl to do longterm-pin and
create IOMMU mappings for the DMA pages. However, when calling
VFIO_IOMMU_MAP_DMA ioctl to pin the physical pages, we found it will be
failed to longterm-pin sometimes.
After some invetigation, we found the pages used to do DMA mapping can
contain some CMA pages, and these CMA pages will cause a possible failure
of the longterm-pin, due to failed to migrate the CMA pages. The reason
of migration failure may be temporary reference count or memory allocation
failure. So that will cause the VFIO_IOMMU_MAP_DMA ioctl returns error,
which makes the application failed to start.
I observed one migration failure case (which is not easy to reproduce) is
that, the 'thp_migration_fail' count is 1 and the 'thp_split_page_failed'
count is also 1.
That means when migrating a THP which is in CMA area, but can not allocate
a new THP due to memory fragmentation, so it will split the THP. However
THP split is also failed, probably the reason is temporary reference count
of this THP. And the temporary reference count can be caused by dropping
page caches (I observed the drop caches operation in the system), but we
can not drop the shmem page caches due to they are already dirty at that
time.
Especially for THP split failure, which is caused by temporary reference
count, we can try again to mitigate the failure of migration in this case
according to previous discussion [1].
[1] https://lore.kernel.org/all/470dc638-a300-f261-94b4-e27250e42f96@redhat.com/
Link: https://lkml.kernel.org/r/6784730480a1df82e8f4cba1ed088e4ac767994b.1666599848.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-10-24 11:34:22 +03:00
ret = - EAGAIN ;
2016-01-16 03:54:10 +03:00
}
out_unlock :
2016-07-27 01:25:51 +03:00
if ( anon_vma ) {
anon_vma_unlock_write ( anon_vma ) ;
put_anon_vma ( anon_vma ) ;
}
if ( mapping )
i_mmap_unlock_read ( mapping ) ;
2016-01-16 03:54:10 +03:00
out :
2022-06-08 22:18:34 +03:00
xas_destroy ( & xas ) ;
2016-01-16 03:54:10 +03:00
count_vm_event ( ! ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED ) ;
return ret ;
}
2016-01-16 03:54:17 +03:00
void free_transhuge_page ( struct page * page )
{
2019-09-24 01:38:15 +03:00
struct deferred_split * ds_queue = get_deferred_split_queue ( page ) ;
2016-01-16 03:54:17 +03:00
unsigned long flags ;
2019-09-24 01:38:06 +03:00
spin_lock_irqsave ( & ds_queue - > split_queue_lock , flags ) ;
2016-01-16 03:54:17 +03:00
if ( ! list_empty ( page_deferred_list ( page ) ) ) {
2019-09-24 01:38:06 +03:00
ds_queue - > split_queue_len - - ;
2016-01-16 03:54:17 +03:00
list_del ( page_deferred_list ( page ) ) ;
}
2019-09-24 01:38:06 +03:00
spin_unlock_irqrestore ( & ds_queue - > split_queue_lock , flags ) ;
2016-01-16 03:54:17 +03:00
free_compound_page ( page ) ;
}
void deferred_split_huge_page ( struct page * page )
{
2019-09-24 01:38:15 +03:00
struct deferred_split * ds_queue = get_deferred_split_queue ( page ) ;
# ifdef CONFIG_MEMCG
mm: memcontrol: Use helpers to read page's memcg data
Patch series "mm: allow mapping accounted kernel pages to userspace", v6.
Currently a non-slab kernel page which has been charged to a memory cgroup
can't be mapped to userspace. The underlying reason is simple: PageKmemcg
flag is defined as a page type (like buddy, offline, etc), so it takes a
bit from a page->mapped counter. Pages with a type set can't be mapped to
userspace.
But in general the kmemcg flag has nothing to do with mapping to
userspace. It only means that the page has been accounted by the page
allocator, so it has to be properly uncharged on release.
Some bpf maps are mapping the vmalloc-based memory to userspace, and their
memory can't be accounted because of this implementation detail.
This patchset removes this limitation by moving the PageKmemcg flag into
one of the free bits of the page->mem_cgroup pointer. Also it formalizes
accesses to the page->mem_cgroup and page->obj_cgroups using new helpers,
adds several checks and removes a couple of obsolete functions. As the
result the code became more robust with fewer open-coded bit tricks.
This patch (of 4):
Currently there are many open-coded reads of the page->mem_cgroup pointer,
as well as a couple of read helpers, which are barely used.
It creates an obstacle on a way to reuse some bits of the pointer for
storing additional bits of information. In fact, we already do this for
slab pages, where the last bit indicates that a pointer has an attached
vector of objcg pointers instead of a regular memcg pointer.
This commits uses 2 existing helpers and introduces a new helper to
converts all read sides to calls of these helpers:
struct mem_cgroup *page_memcg(struct page *page);
struct mem_cgroup *page_memcg_rcu(struct page *page);
struct mem_cgroup *page_memcg_check(struct page *page);
page_memcg_check() is intended to be used in cases when the page can be a
slab page and have a memcg pointer pointing at objcg vector. It does
check the lowest bit, and if set, returns NULL. page_memcg() contains a
VM_BUG_ON_PAGE() check for the page not being a slab page.
To make sure nobody uses a direct access, struct page's
mem_cgroup/obj_cgroups is converted to unsigned long memcg_data.
Signed-off-by: Roman Gushchin <guro@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Link: https://lkml.kernel.org/r/20201027001657.3398190-1-guro@fb.com
Link: https://lkml.kernel.org/r/20201027001657.3398190-2-guro@fb.com
Link: https://lore.kernel.org/bpf/20201201215900.3569844-2-guro@fb.com
2020-12-02 00:58:27 +03:00
struct mem_cgroup * memcg = page_memcg ( compound_head ( page ) ) ;
2019-09-24 01:38:15 +03:00
# endif
2016-01-16 03:54:17 +03:00
unsigned long flags ;
VM_BUG_ON_PAGE ( ! PageTransHuge ( page ) , page ) ;
2019-09-24 01:38:15 +03:00
/*
* The try_to_unmap ( ) in page reclaim path might reach here too ,
* this may cause a race condition to corrupt deferred split queue .
* And , if page reclaim is already handling the same page , it is
* unnecessary to handle it again in shrinker .
*
* Check PageSwapCache to determine if the page is being
* handled by page reclaim since THP swap would add the page into
* swap cache before calling try_to_unmap ( ) .
*/
if ( PageSwapCache ( page ) )
return ;
2019-09-24 01:38:06 +03:00
spin_lock_irqsave ( & ds_queue - > split_queue_lock , flags ) ;
2016-01-16 03:54:17 +03:00
if ( list_empty ( page_deferred_list ( page ) ) ) {
2016-03-18 00:18:45 +03:00
count_vm_event ( THP_DEFERRED_SPLIT_PAGE ) ;
2019-09-24 01:38:06 +03:00
list_add_tail ( page_deferred_list ( page ) , & ds_queue - > split_queue ) ;
ds_queue - > split_queue_len + + ;
2019-09-24 01:38:15 +03:00
# ifdef CONFIG_MEMCG
if ( memcg )
2021-05-05 04:36:11 +03:00
set_shrinker_bit ( memcg , page_to_nid ( page ) ,
deferred_split_shrinker . id ) ;
2019-09-24 01:38:15 +03:00
# endif
2016-01-16 03:54:17 +03:00
}
2019-09-24 01:38:06 +03:00
spin_unlock_irqrestore ( & ds_queue - > split_queue_lock , flags ) ;
2016-01-16 03:54:17 +03:00
}
static unsigned long deferred_split_count ( struct shrinker * shrink ,
struct shrink_control * sc )
{
2016-02-03 03:57:08 +03:00
struct pglist_data * pgdata = NODE_DATA ( sc - > nid ) ;
2019-09-24 01:38:06 +03:00
struct deferred_split * ds_queue = & pgdata - > deferred_split_queue ;
2019-09-24 01:38:15 +03:00
# ifdef CONFIG_MEMCG
if ( sc - > memcg )
ds_queue = & sc - > memcg - > deferred_split_queue ;
# endif
2019-09-24 01:38:06 +03:00
return READ_ONCE ( ds_queue - > split_queue_len ) ;
2016-01-16 03:54:17 +03:00
}
static unsigned long deferred_split_scan ( struct shrinker * shrink ,
struct shrink_control * sc )
{
2016-02-03 03:57:08 +03:00
struct pglist_data * pgdata = NODE_DATA ( sc - > nid ) ;
2019-09-24 01:38:06 +03:00
struct deferred_split * ds_queue = & pgdata - > deferred_split_queue ;
2016-01-16 03:54:17 +03:00
unsigned long flags ;
LIST_HEAD ( list ) , * pos , * next ;
struct page * page ;
int split = 0 ;
2019-09-24 01:38:15 +03:00
# ifdef CONFIG_MEMCG
if ( sc - > memcg )
ds_queue = & sc - > memcg - > deferred_split_queue ;
# endif
2019-09-24 01:38:06 +03:00
spin_lock_irqsave ( & ds_queue - > split_queue_lock , flags ) ;
2016-01-16 03:54:17 +03:00
/* Take pin on all head pages to avoid freeing them under us */
2019-09-24 01:38:06 +03:00
list_for_each_safe ( pos , next , & ds_queue - > split_queue ) {
2021-07-01 04:47:46 +03:00
page = list_entry ( ( void * ) pos , struct page , deferred_list ) ;
2016-01-16 03:54:17 +03:00
page = compound_head ( page ) ;
2016-02-03 03:57:15 +03:00
if ( get_page_unless_zero ( page ) ) {
list_move ( page_deferred_list ( page ) , & list ) ;
} else {
/* We lost race with put_compound_page() */
2016-01-16 03:54:17 +03:00
list_del_init ( page_deferred_list ( page ) ) ;
2019-09-24 01:38:06 +03:00
ds_queue - > split_queue_len - - ;
2016-01-16 03:54:17 +03:00
}
2016-02-03 03:57:15 +03:00
if ( ! - - sc - > nr_to_scan )
break ;
2016-01-16 03:54:17 +03:00
}
2019-09-24 01:38:06 +03:00
spin_unlock_irqrestore ( & ds_queue - > split_queue_lock , flags ) ;
2016-01-16 03:54:17 +03:00
list_for_each_safe ( pos , next , & list ) {
2021-07-01 04:47:46 +03:00
page = list_entry ( ( void * ) pos , struct page , deferred_list ) ;
2018-03-23 02:17:31 +03:00
if ( ! trylock_page ( page ) )
goto next ;
2016-01-16 03:54:17 +03:00
/* split_huge_page() removes page from list on success */
if ( ! split_huge_page ( page ) )
split + + ;
unlock_page ( page ) ;
2018-03-23 02:17:31 +03:00
next :
2016-01-16 03:54:17 +03:00
put_page ( page ) ;
}
2019-09-24 01:38:06 +03:00
spin_lock_irqsave ( & ds_queue - > split_queue_lock , flags ) ;
list_splice_tail ( & list , & ds_queue - > split_queue ) ;
spin_unlock_irqrestore ( & ds_queue - > split_queue_lock , flags ) ;
2016-01-16 03:54:17 +03:00
2016-02-03 03:57:12 +03:00
/*
* Stop shrinker if we didn ' t split any page , but the queue is empty .
* This can happen if pages were freed under us .
*/
2019-09-24 01:38:06 +03:00
if ( ! split & & list_empty ( & ds_queue - > split_queue ) )
2016-02-03 03:57:12 +03:00
return SHRINK_STOP ;
return split ;
2016-01-16 03:54:17 +03:00
}
static struct shrinker deferred_split_shrinker = {
. count_objects = deferred_split_count ,
. scan_objects = deferred_split_scan ,
. seeks = DEFAULT_SEEKS ,
2019-09-24 01:38:15 +03:00
. flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
SHRINKER_NONSLAB ,
2016-01-16 03:54:17 +03:00
} ;
2016-01-16 03:54:40 +03:00
# ifdef CONFIG_DEBUG_FS
2021-05-05 04:34:23 +03:00
static void split_huge_pages_all ( void )
2016-01-16 03:54:40 +03:00
{
struct zone * zone ;
struct page * page ;
unsigned long pfn , max_zone_pfn ;
unsigned long total = 0 , split = 0 ;
2021-05-05 04:34:23 +03:00
pr_debug ( " Split all THPs \n " ) ;
2022-07-04 16:21:57 +03:00
for_each_zone ( zone ) {
if ( ! managed_zone ( zone ) )
continue ;
2016-01-16 03:54:40 +03:00
max_zone_pfn = zone_end_pfn ( zone ) ;
for ( pfn = zone - > zone_start_pfn ; pfn < max_zone_pfn ; pfn + + ) {
2022-07-04 16:21:57 +03:00
int nr_pages ;
2016-01-16 03:54:40 +03:00
2022-09-08 07:11:50 +03:00
page = pfn_to_online_page ( pfn ) ;
if ( ! page | | ! get_page_unless_zero ( page ) )
2016-01-16 03:54:40 +03:00
continue ;
if ( zone ! = page_zone ( page ) )
goto next ;
2016-07-27 01:25:51 +03:00
if ( ! PageHead ( page ) | | PageHuge ( page ) | | ! PageLRU ( page ) )
2016-01-16 03:54:40 +03:00
goto next ;
total + + ;
lock_page ( page ) ;
2022-07-04 16:21:57 +03:00
nr_pages = thp_nr_pages ( page ) ;
2016-01-16 03:54:40 +03:00
if ( ! split_huge_page ( page ) )
split + + ;
2022-07-04 16:21:57 +03:00
pfn + = nr_pages - 1 ;
2016-01-16 03:54:40 +03:00
unlock_page ( page ) ;
next :
put_page ( page ) ;
2021-05-05 04:34:23 +03:00
cond_resched ( ) ;
2016-01-16 03:54:40 +03:00
}
}
2021-05-05 04:34:23 +03:00
pr_debug ( " %lu of %lu THP split \n " , split , total ) ;
}
2016-01-16 03:54:40 +03:00
2021-05-05 04:34:23 +03:00
static inline bool vma_not_suitable_for_thp_split ( struct vm_area_struct * vma )
{
return vma_is_special_huge ( vma ) | | ( vma - > vm_flags & VM_IO ) | |
is_vm_hugetlb_page ( vma ) ;
}
static int split_huge_pages_pid ( int pid , unsigned long vaddr_start ,
unsigned long vaddr_end )
{
int ret = 0 ;
struct task_struct * task ;
struct mm_struct * mm ;
unsigned long total = 0 , split = 0 ;
unsigned long addr ;
vaddr_start & = PAGE_MASK ;
vaddr_end & = PAGE_MASK ;
/* Find the task_struct from pid */
rcu_read_lock ( ) ;
task = find_task_by_vpid ( pid ) ;
if ( ! task ) {
rcu_read_unlock ( ) ;
ret = - ESRCH ;
goto out ;
}
get_task_struct ( task ) ;
rcu_read_unlock ( ) ;
/* Find the mm_struct */
mm = get_task_mm ( task ) ;
put_task_struct ( task ) ;
if ( ! mm ) {
ret = - EINVAL ;
goto out ;
}
pr_debug ( " Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx] \n " ,
pid , vaddr_start , vaddr_end ) ;
mmap_read_lock ( mm ) ;
/*
* always increase addr by PAGE_SIZE , since we could have a PTE page
* table filled with PTE - mapped THPs , each of which is distinct .
*/
for ( addr = vaddr_start ; addr < vaddr_end ; addr + = PAGE_SIZE ) {
2022-07-04 16:21:52 +03:00
struct vm_area_struct * vma = vma_lookup ( mm , addr ) ;
2021-05-05 04:34:23 +03:00
struct page * page ;
2022-07-04 16:21:52 +03:00
if ( ! vma )
2021-05-05 04:34:23 +03:00
break ;
/* skip special VMA and hugetlb VMA */
if ( vma_not_suitable_for_thp_split ( vma ) ) {
addr = vma - > vm_end ;
continue ;
}
/* FOLL_DUMP to ignore special (like zero) pages */
2022-03-23 00:45:29 +03:00
page = follow_page ( vma , addr , FOLL_GET | FOLL_DUMP ) ;
2021-05-05 04:34:23 +03:00
2022-08-23 16:58:41 +03:00
if ( IS_ERR_OR_NULL ( page ) )
2021-05-05 04:34:23 +03:00
continue ;
if ( ! is_transparent_hugepage ( page ) )
goto next ;
total + + ;
2022-02-04 22:13:31 +03:00
if ( ! can_split_folio ( page_folio ( page ) , NULL ) )
2021-05-05 04:34:23 +03:00
goto next ;
if ( ! trylock_page ( page ) )
goto next ;
if ( ! split_huge_page ( page ) )
split + + ;
unlock_page ( page ) ;
next :
put_page ( page ) ;
cond_resched ( ) ;
}
mmap_read_unlock ( mm ) ;
mmput ( mm ) ;
pr_debug ( " %lu of %lu THP split \n " , split , total ) ;
out :
return ret ;
2016-01-16 03:54:40 +03:00
}
2021-05-05 04:34:23 +03:00
2021-05-05 04:34:26 +03:00
static int split_huge_pages_in_file ( const char * file_path , pgoff_t off_start ,
pgoff_t off_end )
{
struct filename * file ;
struct file * candidate ;
struct address_space * mapping ;
int ret = - EINVAL ;
pgoff_t index ;
int nr_pages = 1 ;
unsigned long total = 0 , split = 0 ;
file = getname_kernel ( file_path ) ;
if ( IS_ERR ( file ) )
return ret ;
candidate = file_open_name ( file , O_RDONLY , 0 ) ;
if ( IS_ERR ( candidate ) )
goto out ;
pr_debug ( " split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx] \n " ,
file_path , off_start , off_end ) ;
mapping = candidate - > f_mapping ;
for ( index = off_start ; index < off_end ; index + = nr_pages ) {
2022-10-19 21:33:29 +03:00
struct folio * folio = __filemap_get_folio ( mapping , index ,
FGP_ENTRY , 0 ) ;
2021-05-05 04:34:26 +03:00
nr_pages = 1 ;
2022-10-19 21:33:29 +03:00
if ( xa_is_value ( folio ) | | ! folio )
2021-05-05 04:34:26 +03:00
continue ;
2022-10-19 21:33:29 +03:00
if ( ! folio_test_large ( folio ) )
2021-05-05 04:34:26 +03:00
goto next ;
total + + ;
2022-10-19 21:33:29 +03:00
nr_pages = folio_nr_pages ( folio ) ;
2021-05-05 04:34:26 +03:00
2022-10-19 21:33:29 +03:00
if ( ! folio_trylock ( folio ) )
2021-05-05 04:34:26 +03:00
goto next ;
2022-10-19 21:33:29 +03:00
if ( ! split_folio ( folio ) )
2021-05-05 04:34:26 +03:00
split + + ;
2022-10-19 21:33:29 +03:00
folio_unlock ( folio ) ;
2021-05-05 04:34:26 +03:00
next :
2022-10-19 21:33:29 +03:00
folio_put ( folio ) ;
2021-05-05 04:34:26 +03:00
cond_resched ( ) ;
}
filp_close ( candidate , NULL ) ;
ret = 0 ;
pr_debug ( " %lu of %lu file-backed THP split \n " , split , total ) ;
out :
putname ( file ) ;
return ret ;
}
2021-05-05 04:34:23 +03:00
# define MAX_INPUT_BUF_SZ 255
static ssize_t split_huge_pages_write ( struct file * file , const char __user * buf ,
size_t count , loff_t * ppops )
{
static DEFINE_MUTEX ( split_debug_mutex ) ;
ssize_t ret ;
2021-05-05 04:34:26 +03:00
/* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
char input_buf [ MAX_INPUT_BUF_SZ ] ;
2021-05-05 04:34:23 +03:00
int pid ;
unsigned long vaddr_start , vaddr_end ;
ret = mutex_lock_interruptible ( & split_debug_mutex ) ;
if ( ret )
return ret ;
ret = - EFAULT ;
memset ( input_buf , 0 , MAX_INPUT_BUF_SZ ) ;
if ( copy_from_user ( input_buf , buf , min_t ( size_t , count , MAX_INPUT_BUF_SZ ) ) )
goto out ;
input_buf [ MAX_INPUT_BUF_SZ - 1 ] = ' \0 ' ;
2021-05-05 04:34:26 +03:00
if ( input_buf [ 0 ] = = ' / ' ) {
char * tok ;
char * buf = input_buf ;
char file_path [ MAX_INPUT_BUF_SZ ] ;
pgoff_t off_start = 0 , off_end = 0 ;
size_t input_len = strlen ( input_buf ) ;
tok = strsep ( & buf , " , " ) ;
if ( tok ) {
2021-07-01 04:52:11 +03:00
strcpy ( file_path , tok ) ;
2021-05-05 04:34:26 +03:00
} else {
ret = - EINVAL ;
goto out ;
}
ret = sscanf ( buf , " 0x%lx,0x%lx " , & off_start , & off_end ) ;
if ( ret ! = 2 ) {
ret = - EINVAL ;
goto out ;
}
ret = split_huge_pages_in_file ( file_path , off_start , off_end ) ;
if ( ! ret )
ret = input_len ;
goto out ;
}
2021-05-05 04:34:23 +03:00
ret = sscanf ( input_buf , " %d,0x%lx,0x%lx " , & pid , & vaddr_start , & vaddr_end ) ;
if ( ret = = 1 & & pid = = 1 ) {
split_huge_pages_all ( ) ;
ret = strlen ( input_buf ) ;
goto out ;
} else if ( ret ! = 3 ) {
ret = - EINVAL ;
goto out ;
}
ret = split_huge_pages_pid ( pid , vaddr_start , vaddr_end ) ;
if ( ! ret )
ret = strlen ( input_buf ) ;
out :
mutex_unlock ( & split_debug_mutex ) ;
return ret ;
}
static const struct file_operations split_huge_pages_fops = {
. owner = THIS_MODULE ,
. write = split_huge_pages_write ,
. llseek = no_llseek ,
} ;
2016-01-16 03:54:40 +03:00
static int __init split_huge_pages_debugfs ( void )
{
2019-03-06 02:46:09 +03:00
debugfs_create_file ( " split_huge_pages " , 0200 , NULL , NULL ,
& split_huge_pages_fops ) ;
2016-01-16 03:54:40 +03:00
return 0 ;
}
late_initcall ( split_huge_pages_debugfs ) ;
# endif
2017-09-09 02:10:57 +03:00
# ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2022-05-10 04:20:44 +03:00
int set_pmd_migration_entry ( struct page_vma_mapped_walk * pvmw ,
2017-09-09 02:10:57 +03:00
struct page * page )
{
struct vm_area_struct * vma = pvmw - > vma ;
struct mm_struct * mm = vma - > vm_mm ;
unsigned long address = pvmw - > address ;
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
bool anon_exclusive ;
2017-09-09 02:10:57 +03:00
pmd_t pmdval ;
swp_entry_t entry ;
2017-09-09 02:11:04 +03:00
pmd_t pmdswp ;
2017-09-09 02:10:57 +03:00
if ( ! ( pvmw - > pmd & & ! pvmw - > pte ) )
2022-05-10 04:20:44 +03:00
return 0 ;
2017-09-09 02:10:57 +03:00
flush_cache_range ( vma , address , address + HPAGE_PMD_SIZE ) ;
2020-03-06 09:28:29 +03:00
pmdval = pmdp_invalidate ( vma , address , pvmw - > pmd ) ;
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
mm: fix PageAnonExclusive clearing racing with concurrent RCU GUP-fast
commit 6c287605fd56 ("mm: remember exclusively mapped anonymous pages with
PG_anon_exclusive") made sure that when PageAnonExclusive() has to be
cleared during temporary unmapping of a page, that the PTE is
cleared/invalidated and that the TLB is flushed.
What we want to achieve in all cases is that we cannot end up with a pin on
an anonymous page that may be shared, because such pins would be
unreliable and could result in memory corruptions when the mapped page
and the pin go out of sync due to a write fault.
That TLB flush handling was inspired by an outdated comment in
mm/ksm.c:write_protect_page(), which similarly required the TLB flush in
the past to synchronize with GUP-fast. However, ever since general RCU GUP
fast was introduced in commit 2667f50e8b81 ("mm: introduce a general RCU
get_user_pages_fast()"), a TLB flush is no longer sufficient to handle
concurrent GUP-fast in all cases -- it only handles traditional IPI-based
GUP-fast correctly.
Peter Xu (thankfully) questioned whether that TLB flush is really
required. On architectures that send an IPI broadcast on TLB flush,
it works as expected. To synchronize with RCU GUP-fast properly, we're
conceptually fine, however, we have to enforce a certain memory order and
are missing memory barriers.
Let's document that, avoid the TLB flush where possible and use proper
explicit memory barriers where required. We shouldn't really care about the
additional memory barriers here, as we're not on extremely hot paths --
and we're getting rid of some TLB flushes.
We use a smp_mb() pair for handling concurrent pinning and a
smp_rmb()/smp_wmb() pair for handling the corner case of only temporary
PTE changes but permanent PageAnonExclusive changes.
One extreme example, whereby GUP-fast takes a R/O pin and KSM wants to
convert an exclusive anonymous page to a KSM page, and that page is already
mapped write-protected (-> no PTE change) would be:
Thread 0 (KSM) Thread 1 (GUP-fast)
(B1) Read the PTE
# (B2) skipped without FOLL_WRITE
(A1) Clear PTE
smp_mb()
(A2) Check pinned
(B3) Pin the mapped page
smp_mb()
(A3) Clear PageAnonExclusive
smp_wmb()
(A4) Restore PTE
(B4) Check if the PTE changed
smp_rmb()
(B5) Check PageAnonExclusive
Thread 1 will properly detect that PageAnonExclusive was cleared and
back off.
Note that we don't need a memory barrier between checking if the page is
pinned and clearing PageAnonExclusive, because stores are not
speculated.
The possible issues due to reordering are of theoretical nature so far
and attempts to reproduce the race failed.
Especially the "no PTE change" case isn't the common case, because we'd
need an exclusive anonymous page that's mapped R/O and the PTE is clean
in KSM code -- and using KSM with page pinning isn't extremely common.
Further, the clear+TLB flush we used for now implies a memory barrier.
So the problematic missing part should be the missing memory barrier
after pinning but before checking if the PTE changed.
Link: https://lkml.kernel.org/r/20220901083559.67446-1-david@redhat.com
Fixes: 6c287605fd56 ("mm: remember exclusively mapped anonymous pages with PG_anon_exclusive")
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Yang Shi <shy828301@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Andrea Parri <parri.andrea@gmail.com>
Cc: Will Deacon <will@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: "Paul E. McKenney" <paulmck@kernel.org>
Cc: Christoph von Recklinghausen <crecklin@redhat.com>
Cc: Don Dutile <ddutile@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-01 11:35:59 +03:00
/* See page_try_share_anon_rmap(): invalidate PMD first. */
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
anon_exclusive = PageAnon ( page ) & & PageAnonExclusive ( page ) ;
if ( anon_exclusive & & page_try_share_anon_rmap ( page ) ) {
set_pmd_at ( mm , address , pvmw - > pmd , pmdval ) ;
2022-05-10 04:20:44 +03:00
return - EBUSY ;
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
}
2017-09-09 02:10:57 +03:00
if ( pmd_dirty ( pmdval ) )
set_page_dirty ( page ) ;
2021-07-01 04:54:09 +03:00
if ( pmd_write ( pmdval ) )
entry = make_writable_migration_entry ( page_to_pfn ( page ) ) ;
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
else if ( anon_exclusive )
entry = make_readable_exclusive_migration_entry ( page_to_pfn ( page ) ) ;
2021-07-01 04:54:09 +03:00
else
entry = make_readable_migration_entry ( page_to_pfn ( page ) ) ;
2022-08-11 19:13:29 +03:00
if ( pmd_young ( pmdval ) )
entry = make_migration_entry_young ( entry ) ;
if ( pmd_dirty ( pmdval ) )
entry = make_migration_entry_dirty ( entry ) ;
2017-09-09 02:11:04 +03:00
pmdswp = swp_entry_to_pmd ( entry ) ;
if ( pmd_soft_dirty ( pmdval ) )
pmdswp = pmd_swp_mksoft_dirty ( pmdswp ) ;
set_pmd_at ( mm , address , pvmw - > pmd , pmdswp ) ;
mm/munlock: rmap call mlock_vma_page() munlock_vma_page()
Add vma argument to mlock_vma_page() and munlock_vma_page(), make them
inline functions which check (vma->vm_flags & VM_LOCKED) before calling
mlock_page() and munlock_page() in mm/mlock.c.
Add bool compound to mlock_vma_page() and munlock_vma_page(): this is
because we have understandable difficulty in accounting pte maps of THPs,
and if passed a PageHead page, mlock_page() and munlock_page() cannot
tell whether it's a pmd map to be counted or a pte map to be ignored.
Add vma arg to page_add_file_rmap() and page_remove_rmap(), like the
others, and use that to call mlock_vma_page() at the end of the page
adds, and munlock_vma_page() at the end of page_remove_rmap() (end or
beginning? unimportant, but end was easier for assertions in testing).
No page lock is required (although almost all adds happen to hold it):
delete the "Serialize with page migration" BUG_ON(!PageLocked(page))s.
Certainly page lock did serialize with page migration, but I'm having
difficulty explaining why that was ever important.
Mlock accounting on THPs has been hard to define, differed between anon
and file, involved PageDoubleMap in some places and not others, required
clear_page_mlock() at some points. Keep it simple now: just count the
pmds and ignore the ptes, there is no reason for ptes to undo pmd mlocks.
page_add_new_anon_rmap() callers unchanged: they have long been calling
lru_cache_add_inactive_or_unevictable(), which does its own VM_LOCKED
handling (it also checks for not VM_SPECIAL: I think that's overcautious,
and inconsistent with other checks, that mmap_region() already prevents
VM_LOCKED on VM_SPECIAL; but haven't quite convinced myself to change it).
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
2022-02-15 05:26:39 +03:00
page_remove_rmap ( page , vma , true ) ;
2017-09-09 02:10:57 +03:00
put_page ( page ) ;
2022-03-25 04:09:58 +03:00
trace_set_migration_pmd ( address , pmd_val ( pmdswp ) ) ;
2022-05-10 04:20:44 +03:00
return 0 ;
2017-09-09 02:10:57 +03:00
}
void remove_migration_pmd ( struct page_vma_mapped_walk * pvmw , struct page * new )
{
struct vm_area_struct * vma = pvmw - > vma ;
struct mm_struct * mm = vma - > vm_mm ;
unsigned long address = pvmw - > address ;
2022-07-04 16:21:51 +03:00
unsigned long haddr = address & HPAGE_PMD_MASK ;
2017-09-09 02:10:57 +03:00
pmd_t pmde ;
swp_entry_t entry ;
if ( ! ( pvmw - > pmd & & ! pvmw - > pte ) )
return ;
entry = pmd_to_swp_entry ( * pvmw - > pmd ) ;
get_page ( new ) ;
2022-08-11 19:13:29 +03:00
pmde = mk_huge_pmd ( new , READ_ONCE ( vma - > vm_page_prot ) ) ;
2017-09-09 02:11:04 +03:00
if ( pmd_swp_soft_dirty ( * pvmw - > pmd ) )
pmde = pmd_mksoft_dirty ( pmde ) ;
2021-07-01 04:54:09 +03:00
if ( is_writable_migration_entry ( entry ) )
2017-11-29 20:01:01 +03:00
pmde = maybe_pmd_mkwrite ( pmde , vma ) ;
mm/userfaultfd: fix uffd-wp special cases for fork()
We tried to do something similar in b569a1760782 ("userfaultfd: wp: drop
_PAGE_UFFD_WP properly when fork") previously, but it's not doing it all
right.. A few fixes around the code path:
1. We were referencing VM_UFFD_WP vm_flags on the _old_ vma rather
than the new vma. That's overlooked in b569a1760782, so it won't work
as expected. Thanks to the recent rework on fork code
(7a4830c380f3a8b3), we can easily get the new vma now, so switch the
checks to that.
2. Dropping the uffd-wp bit in copy_huge_pmd() could be wrong if the
huge pmd is a migration huge pmd. When it happens, instead of using
pmd_uffd_wp(), we should use pmd_swp_uffd_wp(). The fix is simply to
handle them separately.
3. Forget to carry over uffd-wp bit for a write migration huge pmd
entry. This also happens in copy_huge_pmd(), where we converted a
write huge migration entry into a read one.
4. In copy_nonpresent_pte(), drop uffd-wp if necessary for swap ptes.
5. In copy_present_page() when COW is enforced when fork(), we also
need to pass over the uffd-wp bit if VM_UFFD_WP is armed on the new
vma, and when the pte to be copied has uffd-wp bit set.
Remove the comment in copy_present_pte() about this. It won't help a huge
lot to only comment there, but comment everywhere would be an overkill.
Let's assume the commit messages would help.
[peterx@redhat.com: fix a few thp pmd missing uffd-wp bit]
Link: https://lkml.kernel.org/r/20210428225030.9708-4-peterx@redhat.com
Link: https://lkml.kernel.org/r/20210428225030.9708-3-peterx@redhat.com
Fixes: b569a1760782f ("userfaultfd: wp: drop _PAGE_UFFD_WP properly when fork")
Signed-off-by: Peter Xu <peterx@redhat.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Brian Geffon <bgeffon@google.com>
Cc: "Dr . David Alan Gilbert" <dgilbert@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Joe Perches <joe@perches.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Lokesh Gidra <lokeshgidra@google.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Oliver Upton <oupton@google.com>
Cc: Shaohua Li <shli@fb.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Wang Qing <wangqing@vivo.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-01 04:49:02 +03:00
if ( pmd_swp_uffd_wp ( * pvmw - > pmd ) )
pmde = pmd_wrprotect ( pmd_mkuffd_wp ( pmde ) ) ;
2022-08-11 19:13:29 +03:00
if ( ! is_migration_entry_young ( entry ) )
pmde = pmd_mkold ( pmde ) ;
/* NOTE: this may contain setting soft-dirty on some archs */
if ( PageDirty ( new ) & & is_migration_entry_dirty ( entry ) )
pmde = pmd_mkdirty ( pmde ) ;
2017-09-09 02:10:57 +03:00
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
if ( PageAnon ( new ) ) {
rmap_t rmap_flags = RMAP_COMPOUND ;
if ( ! is_readable_migration_entry ( entry ) )
rmap_flags | = RMAP_EXCLUSIVE ;
2022-07-04 16:21:51 +03:00
page_add_anon_rmap ( new , vma , haddr , rmap_flags ) ;
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
} else {
mm/munlock: rmap call mlock_vma_page() munlock_vma_page()
Add vma argument to mlock_vma_page() and munlock_vma_page(), make them
inline functions which check (vma->vm_flags & VM_LOCKED) before calling
mlock_page() and munlock_page() in mm/mlock.c.
Add bool compound to mlock_vma_page() and munlock_vma_page(): this is
because we have understandable difficulty in accounting pte maps of THPs,
and if passed a PageHead page, mlock_page() and munlock_page() cannot
tell whether it's a pmd map to be counted or a pte map to be ignored.
Add vma arg to page_add_file_rmap() and page_remove_rmap(), like the
others, and use that to call mlock_vma_page() at the end of the page
adds, and munlock_vma_page() at the end of page_remove_rmap() (end or
beginning? unimportant, but end was easier for assertions in testing).
No page lock is required (although almost all adds happen to hold it):
delete the "Serialize with page migration" BUG_ON(!PageLocked(page))s.
Certainly page lock did serialize with page migration, but I'm having
difficulty explaining why that was ever important.
Mlock accounting on THPs has been hard to define, differed between anon
and file, involved PageDoubleMap in some places and not others, required
clear_page_mlock() at some points. Keep it simple now: just count the
pmds and ignore the ptes, there is no reason for ptes to undo pmd mlocks.
page_add_new_anon_rmap() callers unchanged: they have long been calling
lru_cache_add_inactive_or_unevictable(), which does its own VM_LOCKED
handling (it also checks for not VM_SPECIAL: I think that's overcautious,
and inconsistent with other checks, that mmap_region() already prevents
VM_LOCKED on VM_SPECIAL; but haven't quite convinced myself to change it).
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
2022-02-15 05:26:39 +03:00
page_add_file_rmap ( new , vma , true ) ;
mm: remember exclusively mapped anonymous pages with PG_anon_exclusive
Let's mark exclusively mapped anonymous pages with PG_anon_exclusive as
exclusive, and use that information to make GUP pins reliable and stay
consistent with the page mapped into the page table even if the page table
entry gets write-protected.
With that information at hand, we can extend our COW logic to always reuse
anonymous pages that are exclusive. For anonymous pages that might be
shared, the existing logic applies.
As already documented, PG_anon_exclusive is usually only expressive in
combination with a page table entry. Especially PTE vs. PMD-mapped
anonymous pages require more thought, some examples: due to mremap() we
can easily have a single compound page PTE-mapped into multiple page
tables exclusively in a single process -- multiple page table locks apply.
Further, due to MADV_WIPEONFORK we might not necessarily write-protect
all PTEs, and only some subpages might be pinned. Long story short: once
PTE-mapped, we have to track information about exclusivity per sub-page,
but until then, we can just track it for the compound page in the head
page and not having to update a whole bunch of subpages all of the time
for a simple PMD mapping of a THP.
For simplicity, this commit mostly talks about "anonymous pages", while
it's for THP actually "the part of an anonymous folio referenced via a
page table entry".
To not spill PG_anon_exclusive code all over the mm code-base, we let the
anon rmap code to handle all PG_anon_exclusive logic it can easily handle.
If a writable, present page table entry points at an anonymous (sub)page,
that (sub)page must be PG_anon_exclusive. If GUP wants to take a reliably
pin (FOLL_PIN) on an anonymous page references via a present page table
entry, it must only pin if PG_anon_exclusive is set for the mapped
(sub)page.
This commit doesn't adjust GUP, so this is only implicitly handled for
FOLL_WRITE, follow-up commits will teach GUP to also respect it for
FOLL_PIN without FOLL_WRITE, to make all GUP pins of anonymous pages fully
reliable.
Whenever an anonymous page is to be shared (fork(), KSM), or when
temporarily unmapping an anonymous page (swap, migration), the relevant
PG_anon_exclusive bit has to be cleared to mark the anonymous page
possibly shared. Clearing will fail if there are GUP pins on the page:
* For fork(), this means having to copy the page and not being able to
share it. fork() protects against concurrent GUP using the PT lock and
the src_mm->write_protect_seq.
* For KSM, this means sharing will fail. For swap this means, unmapping
will fail, For migration this means, migration will fail early. All
three cases protect against concurrent GUP using the PT lock and a
proper clear/invalidate+flush of the relevant page table entry.
This fixes memory corruptions reported for FOLL_PIN | FOLL_WRITE, when a
pinned page gets mapped R/O and the successive write fault ends up
replacing the page instead of reusing it. It improves the situation for
O_DIRECT/vmsplice/... that still use FOLL_GET instead of FOLL_PIN, if
fork() is *not* involved, however swapout and fork() are still
problematic. Properly using FOLL_PIN instead of FOLL_GET for these GUP
users will fix the issue for them.
I. Details about basic handling
I.1. Fresh anonymous pages
page_add_new_anon_rmap() and hugepage_add_new_anon_rmap() will mark the
given page exclusive via __page_set_anon_rmap(exclusive=1). As that is
the mechanism fresh anonymous pages come into life (besides migration code
where we copy the page->mapping), all fresh anonymous pages will start out
as exclusive.
I.2. COW reuse handling of anonymous pages
When a COW handler stumbles over a (sub)page that's marked exclusive, it
simply reuses it. Otherwise, the handler tries harder under page lock to
detect if the (sub)page is exclusive and can be reused. If exclusive,
page_move_anon_rmap() will mark the given (sub)page exclusive.
Note that hugetlb code does not yet check for PageAnonExclusive(), as it
still uses the old COW logic that is prone to the COW security issue
because hugetlb code cannot really tolerate unnecessary/wrong COW as huge
pages are a scarce resource.
I.3. Migration handling
try_to_migrate() has to try marking an exclusive anonymous page shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. migrate_vma_collect_pmd() and
__split_huge_pmd_locked() are handled similarly.
Writable migration entries implicitly point at shared anonymous pages.
For readable migration entries that information is stored via a new
"readable-exclusive" migration entry, specific to anonymous pages.
When restoring a migration entry in remove_migration_pte(), information
about exlusivity is detected via the migration entry type, and
RMAP_EXCLUSIVE is set accordingly for
page_add_anon_rmap()/hugepage_add_anon_rmap() to restore that information.
I.4. Swapout handling
try_to_unmap() has to try marking the mapped page possibly shared via
page_try_share_anon_rmap(). If it fails because there are GUP pins on the
page, unmap fails. For now, information about exclusivity is lost. In
the future, we might want to remember that information in the swap entry
in some cases, however, it requires more thought, care, and a way to store
that information in swap entries.
I.5. Swapin handling
do_swap_page() will never stumble over exclusive anonymous pages in the
swap cache, as try_to_migrate() prohibits that. do_swap_page() always has
to detect manually if an anonymous page is exclusive and has to set
RMAP_EXCLUSIVE for page_add_anon_rmap() accordingly.
I.6. THP handling
__split_huge_pmd_locked() has to move the information about exclusivity
from the PMD to the PTEs.
a) In case we have a readable-exclusive PMD migration entry, simply
insert readable-exclusive PTE migration entries.
b) In case we have a present PMD entry and we don't want to freeze
("convert to migration entries"), simply forward PG_anon_exclusive to
all sub-pages, no need to temporarily clear the bit.
c) In case we have a present PMD entry and want to freeze, handle it
similar to try_to_migrate(): try marking the page shared first. In
case we fail, we ignore the "freeze" instruction and simply split
ordinarily. try_to_migrate() will properly fail because the THP is
still mapped via PTEs.
When splitting a compound anonymous folio (THP), the information about
exclusivity is implicitly handled via the migration entries: no need to
replicate PG_anon_exclusive manually.
I.7. fork() handling fork() handling is relatively easy, because
PG_anon_exclusive is only expressive for some page table entry types.
a) Present anonymous pages
page_try_dup_anon_rmap() will mark the given subpage shared -- which will
fail if the page is pinned. If it failed, we have to copy (or PTE-map a
PMD to handle it on the PTE level).
Note that device exclusive entries are just a pointer at a PageAnon()
page. fork() will first convert a device exclusive entry to a present
page table and handle it just like present anonymous pages.
b) Device private entry
Device private entries point at PageAnon() pages that cannot be mapped
directly and, therefore, cannot get pinned.
page_try_dup_anon_rmap() will mark the given subpage shared, which cannot
fail because they cannot get pinned.
c) HW poison entries
PG_anon_exclusive will remain untouched and is stale -- the page table
entry is just a placeholder after all.
d) Migration entries
Writable and readable-exclusive entries are converted to readable entries:
possibly shared.
I.8. mprotect() handling
mprotect() only has to properly handle the new readable-exclusive
migration entry:
When write-protecting a migration entry that points at an anonymous page,
remember the information about exclusivity via the "readable-exclusive"
migration entry type.
II. Migration and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a migration entry, we have to mark the page possibly
shared and synchronize against GUP-fast by a proper clear/invalidate+flush
to make the following scenario impossible:
1. try_to_migrate() places a migration entry after checking for GUP pins
and marks the page possibly shared.
2. GUP-fast pins the page due to lack of synchronization
3. fork() converts the "writable/readable-exclusive" migration entry into a
readable migration entry
4. Migration fails due to the GUP pin (failing to freeze the refcount)
5. Migration entries are restored. PG_anon_exclusive is lost
-> We have a pinned page that is not marked exclusive anymore.
Note that we move information about exclusivity from the page to the
migration entry as it otherwise highly overcomplicates fork() and
PTE-mapping a THP.
III. Swapout and GUP-fast
Whenever replacing a present page table entry that maps an exclusive
anonymous page by a swap entry, we have to mark the page possibly shared
and synchronize against GUP-fast by a proper clear/invalidate+flush to
make the following scenario impossible:
1. try_to_unmap() places a swap entry after checking for GUP pins and
clears exclusivity information on the page.
2. GUP-fast pins the page due to lack of synchronization.
-> We have a pinned page that is not marked exclusive anymore.
If we'd ever store information about exclusivity in the swap entry,
similar to migration handling, the same considerations as in II would
apply. This is future work.
Link: https://lkml.kernel.org/r/20220428083441.37290-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Khalid Aziz <khalid.aziz@oracle.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Liang Zhang <zhangliang5@huawei.com>
Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Nadav Amit <namit@vmware.com>
Cc: Oded Gabbay <oded.gabbay@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Rik van Riel <riel@surriel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 04:20:44 +03:00
}
VM_BUG_ON ( pmd_write ( pmde ) & & PageAnon ( new ) & & ! PageAnonExclusive ( new ) ) ;
2022-07-04 16:21:51 +03:00
set_pmd_at ( mm , haddr , pvmw - > pmd , pmde ) ;
2022-03-23 00:41:53 +03:00
/* No need to invalidate - it was non-present before */
2017-09-09 02:10:57 +03:00
update_mmu_cache_pmd ( vma , address , pvmw - > pmd ) ;
2022-03-25 04:09:58 +03:00
trace_remove_migration_pmd ( address , pmd_val ( pmde ) ) ;
2017-09-09 02:10:57 +03:00
}
# endif