2014-11-14 04:36:45 +03:00
/* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
* Copyright ( c ) 2016 Facebook
2014-11-14 04:36:45 +03:00
*
* This program is free software ; you can redistribute it and / or
* modify it under the terms of version 2 of the GNU General Public
* License as published by the Free Software Foundation .
*
* This program is distributed in the hope that it will be useful , but
* WITHOUT ANY WARRANTY ; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE . See the GNU
* General Public License for more details .
*/
# include <linux/bpf.h>
# include <linux/jhash.h>
# include <linux/filter.h>
# include <linux/vmalloc.h>
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
# include "percpu_freelist.h"
2016-11-11 21:55:09 +03:00
# include "bpf_lru_list.h"
2014-11-14 04:36:45 +03:00
2015-12-29 17:40:27 +03:00
struct bucket {
struct hlist_head head ;
raw_spinlock_t lock ;
} ;
2014-11-14 04:36:45 +03:00
struct bpf_htab {
struct bpf_map map ;
2015-12-29 17:40:27 +03:00
struct bucket * buckets ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
void * elems ;
2016-11-11 21:55:09 +03:00
union {
struct pcpu_freelist freelist ;
struct bpf_lru lru ;
} ;
2016-08-06 00:01:27 +03:00
void __percpu * extra_elems ;
2015-12-29 17:40:25 +03:00
atomic_t count ; /* number of elements in this hashtable */
2014-11-14 04:36:45 +03:00
u32 n_buckets ; /* number of hash buckets */
u32 elem_size ; /* size of each element in bytes */
} ;
2016-08-06 00:01:27 +03:00
enum extra_elem_state {
HTAB_NOT_AN_EXTRA_ELEM = 0 ,
HTAB_EXTRA_ELEM_FREE ,
HTAB_EXTRA_ELEM_USED
} ;
2014-11-14 04:36:45 +03:00
/* each htab element is struct htab_elem + key + value */
struct htab_elem {
2016-02-02 09:39:53 +03:00
union {
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
struct hlist_node hash_node ;
struct bpf_htab * htab ;
struct pcpu_freelist_node fnode ;
2016-02-02 09:39:53 +03:00
} ;
2016-08-06 00:01:27 +03:00
union {
struct rcu_head rcu ;
enum extra_elem_state state ;
2016-11-11 21:55:09 +03:00
struct bpf_lru_node lru_node ;
2016-08-06 00:01:27 +03:00
} ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
u32 hash ;
2014-11-14 04:36:45 +03:00
char key [ 0 ] __aligned ( 8 ) ;
} ;
2016-11-11 21:55:09 +03:00
static bool htab_lru_map_delete_node ( void * arg , struct bpf_lru_node * node ) ;
static bool htab_is_lru ( const struct bpf_htab * htab )
{
2016-11-11 21:55:10 +03:00
return htab - > map . map_type = = BPF_MAP_TYPE_LRU_HASH | |
htab - > map . map_type = = BPF_MAP_TYPE_LRU_PERCPU_HASH ;
}
static bool htab_is_percpu ( const struct bpf_htab * htab )
{
return htab - > map . map_type = = BPF_MAP_TYPE_PERCPU_HASH | |
htab - > map . map_type = = BPF_MAP_TYPE_LRU_PERCPU_HASH ;
2016-11-11 21:55:09 +03:00
}
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
static inline void htab_elem_set_ptr ( struct htab_elem * l , u32 key_size ,
void __percpu * pptr )
{
* ( void __percpu * * ) ( l - > key + key_size ) = pptr ;
}
static inline void __percpu * htab_elem_get_ptr ( struct htab_elem * l , u32 key_size )
{
return * ( void __percpu * * ) ( l - > key + key_size ) ;
}
static struct htab_elem * get_htab_elem ( struct bpf_htab * htab , int i )
{
return ( struct htab_elem * ) ( htab - > elems + i * htab - > elem_size ) ;
}
static void htab_free_elems ( struct bpf_htab * htab )
{
int i ;
2016-11-11 21:55:10 +03:00
if ( ! htab_is_percpu ( htab ) )
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
goto free_elems ;
for ( i = 0 ; i < htab - > map . max_entries ; i + + ) {
void __percpu * pptr ;
pptr = htab_elem_get_ptr ( get_htab_elem ( htab , i ) ,
htab - > map . key_size ) ;
free_percpu ( pptr ) ;
}
free_elems :
vfree ( htab - > elems ) ;
}
2016-11-11 21:55:09 +03:00
static struct htab_elem * prealloc_lru_pop ( struct bpf_htab * htab , void * key ,
u32 hash )
{
struct bpf_lru_node * node = bpf_lru_pop_free ( & htab - > lru , hash ) ;
struct htab_elem * l ;
if ( node ) {
l = container_of ( node , struct htab_elem , lru_node ) ;
memcpy ( l - > key , key , htab - > map . key_size ) ;
return l ;
}
return NULL ;
}
static int prealloc_init ( struct bpf_htab * htab )
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
{
int err = - ENOMEM , i ;
htab - > elems = vzalloc ( htab - > elem_size * htab - > map . max_entries ) ;
if ( ! htab - > elems )
return - ENOMEM ;
2016-11-11 21:55:10 +03:00
if ( ! htab_is_percpu ( htab ) )
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
goto skip_percpu_elems ;
for ( i = 0 ; i < htab - > map . max_entries ; i + + ) {
u32 size = round_up ( htab - > map . value_size , 8 ) ;
void __percpu * pptr ;
pptr = __alloc_percpu_gfp ( size , 8 , GFP_USER | __GFP_NOWARN ) ;
if ( ! pptr )
goto free_elems ;
htab_elem_set_ptr ( get_htab_elem ( htab , i ) , htab - > map . key_size ,
pptr ) ;
}
skip_percpu_elems :
2016-11-11 21:55:09 +03:00
if ( htab_is_lru ( htab ) )
err = bpf_lru_init ( & htab - > lru ,
htab - > map . map_flags & BPF_F_NO_COMMON_LRU ,
offsetof ( struct htab_elem , hash ) -
offsetof ( struct htab_elem , lru_node ) ,
htab_lru_map_delete_node ,
htab ) ;
else
err = pcpu_freelist_init ( & htab - > freelist ) ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
if ( err )
goto free_elems ;
2016-11-11 21:55:09 +03:00
if ( htab_is_lru ( htab ) )
bpf_lru_populate ( & htab - > lru , htab - > elems ,
offsetof ( struct htab_elem , lru_node ) ,
htab - > elem_size , htab - > map . max_entries ) ;
else
pcpu_freelist_populate ( & htab - > freelist , htab - > elems ,
htab - > elem_size , htab - > map . max_entries ) ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
return 0 ;
free_elems :
htab_free_elems ( htab ) ;
return err ;
}
2016-11-11 21:55:09 +03:00
static void prealloc_destroy ( struct bpf_htab * htab )
{
htab_free_elems ( htab ) ;
if ( htab_is_lru ( htab ) )
bpf_lru_destroy ( & htab - > lru ) ;
else
pcpu_freelist_destroy ( & htab - > freelist ) ;
}
2016-08-06 00:01:27 +03:00
static int alloc_extra_elems ( struct bpf_htab * htab )
{
void __percpu * pptr ;
int cpu ;
pptr = __alloc_percpu_gfp ( htab - > elem_size , 8 , GFP_USER | __GFP_NOWARN ) ;
if ( ! pptr )
return - ENOMEM ;
for_each_possible_cpu ( cpu ) {
( ( struct htab_elem * ) per_cpu_ptr ( pptr , cpu ) ) - > state =
HTAB_EXTRA_ELEM_FREE ;
}
htab - > extra_elems = pptr ;
return 0 ;
}
2014-11-14 04:36:45 +03:00
/* Called from syscall */
static struct bpf_map * htab_map_alloc ( union bpf_attr * attr )
{
2016-11-11 21:55:10 +03:00
bool percpu = ( attr - > map_type = = BPF_MAP_TYPE_PERCPU_HASH | |
attr - > map_type = = BPF_MAP_TYPE_LRU_PERCPU_HASH ) ;
bool lru = ( attr - > map_type = = BPF_MAP_TYPE_LRU_HASH | |
attr - > map_type = = BPF_MAP_TYPE_LRU_PERCPU_HASH ) ;
2016-11-11 21:55:09 +03:00
/* percpu_lru means each cpu has its own LRU list.
* it is different from BPF_MAP_TYPE_PERCPU_HASH where
* the map ' s value itself is percpu . percpu_lru has
* nothing to do with the map ' s value .
*/
bool percpu_lru = ( attr - > map_flags & BPF_F_NO_COMMON_LRU ) ;
bool prealloc = ! ( attr - > map_flags & BPF_F_NO_PREALLOC ) ;
2014-11-14 04:36:45 +03:00
struct bpf_htab * htab ;
int err , i ;
2016-02-02 09:39:53 +03:00
u64 cost ;
2014-11-14 04:36:45 +03:00
2016-11-11 21:55:09 +03:00
if ( lru & & ! capable ( CAP_SYS_ADMIN ) )
/* LRU implementation is much complicated than other
* maps . Hence , limit to CAP_SYS_ADMIN for now .
*/
return ERR_PTR ( - EPERM ) ;
if ( attr - > map_flags & ~ ( BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU ) )
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
/* reserved bits should not be used */
return ERR_PTR ( - EINVAL ) ;
2016-11-11 21:55:09 +03:00
if ( ! lru & & percpu_lru )
return ERR_PTR ( - EINVAL ) ;
if ( lru & & ! prealloc )
return ERR_PTR ( - ENOTSUPP ) ;
2014-11-14 04:36:45 +03:00
htab = kzalloc ( sizeof ( * htab ) , GFP_USER ) ;
if ( ! htab )
return ERR_PTR ( - ENOMEM ) ;
/* mandatory map attributes */
2016-02-02 09:39:53 +03:00
htab - > map . map_type = attr - > map_type ;
2014-11-14 04:36:45 +03:00
htab - > map . key_size = attr - > key_size ;
htab - > map . value_size = attr - > value_size ;
htab - > map . max_entries = attr - > max_entries ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
htab - > map . map_flags = attr - > map_flags ;
2014-11-14 04:36:45 +03:00
/* check sanity of attributes.
* value_size = = 0 may be allowed in the future to use map as a set
*/
err = - EINVAL ;
if ( htab - > map . max_entries = = 0 | | htab - > map . key_size = = 0 | |
htab - > map . value_size = = 0 )
goto free_htab ;
2016-11-11 21:55:09 +03:00
if ( percpu_lru ) {
/* ensure each CPU's lru list has >=1 elements.
* since we are at it , make each lru list has the same
* number of elements .
*/
htab - > map . max_entries = roundup ( attr - > max_entries ,
num_possible_cpus ( ) ) ;
if ( htab - > map . max_entries < attr - > max_entries )
htab - > map . max_entries = rounddown ( attr - > max_entries ,
num_possible_cpus ( ) ) ;
}
2014-11-14 04:36:45 +03:00
/* hash table size must be power of 2 */
htab - > n_buckets = roundup_pow_of_two ( htab - > map . max_entries ) ;
err = - E2BIG ;
if ( htab - > map . key_size > MAX_BPF_STACK )
/* eBPF programs initialize keys on stack, so they cannot be
* larger than max stack size
*/
goto free_htab ;
2015-11-30 03:59:35 +03:00
if ( htab - > map . value_size > = ( 1 < < ( KMALLOC_SHIFT_MAX - 1 ) ) -
MAX_BPF_STACK - sizeof ( struct htab_elem ) )
/* if value_size is bigger, the user space won't be able to
* access the elements via bpf syscall . This check also makes
* sure that the elem_size doesn ' t overflow and it ' s
* kmalloc - able later in htab_map_update_elem ( )
*/
goto free_htab ;
2016-02-02 09:39:53 +03:00
if ( percpu & & round_up ( htab - > map . value_size , 8 ) > PCPU_MIN_UNIT_SIZE )
/* make sure the size for pcpu_alloc() is reasonable */
goto free_htab ;
2015-11-30 03:59:35 +03:00
htab - > elem_size = sizeof ( struct htab_elem ) +
2016-02-02 09:39:53 +03:00
round_up ( htab - > map . key_size , 8 ) ;
if ( percpu )
htab - > elem_size + = sizeof ( void * ) ;
else
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
htab - > elem_size + = round_up ( htab - > map . value_size , 8 ) ;
2015-11-30 03:59:35 +03:00
2014-11-19 04:32:16 +03:00
/* prevent zero size kmalloc and check for u32 overflow */
if ( htab - > n_buckets = = 0 | |
2015-12-29 17:40:27 +03:00
htab - > n_buckets > U32_MAX / sizeof ( struct bucket ) )
2014-11-19 04:32:16 +03:00
goto free_htab ;
2016-02-02 09:39:53 +03:00
cost = ( u64 ) htab - > n_buckets * sizeof ( struct bucket ) +
( u64 ) htab - > elem_size * htab - > map . max_entries ;
if ( percpu )
cost + = ( u64 ) round_up ( htab - > map . value_size , 8 ) *
num_possible_cpus ( ) * htab - > map . max_entries ;
2016-08-06 00:01:27 +03:00
else
cost + = ( u64 ) htab - > elem_size * num_possible_cpus ( ) ;
2016-02-02 09:39:53 +03:00
if ( cost > = U32_MAX - PAGE_SIZE )
2015-11-30 03:59:35 +03:00
/* make sure page count doesn't overflow */
goto free_htab ;
2016-02-02 09:39:53 +03:00
htab - > map . pages = round_up ( cost , PAGE_SIZE ) > > PAGE_SHIFT ;
2015-11-30 03:59:35 +03:00
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
/* if map size is larger than memlock limit, reject it early */
err = bpf_map_precharge_memlock ( htab - > map . pages ) ;
if ( err )
goto free_htab ;
2015-11-30 03:59:35 +03:00
err = - ENOMEM ;
2015-12-29 17:40:27 +03:00
htab - > buckets = kmalloc_array ( htab - > n_buckets , sizeof ( struct bucket ) ,
2014-11-14 04:36:45 +03:00
GFP_USER | __GFP_NOWARN ) ;
if ( ! htab - > buckets ) {
2015-12-29 17:40:27 +03:00
htab - > buckets = vmalloc ( htab - > n_buckets * sizeof ( struct bucket ) ) ;
2014-11-14 04:36:45 +03:00
if ( ! htab - > buckets )
goto free_htab ;
}
2015-12-29 17:40:27 +03:00
for ( i = 0 ; i < htab - > n_buckets ; i + + ) {
INIT_HLIST_HEAD ( & htab - > buckets [ i ] . head ) ;
raw_spin_lock_init ( & htab - > buckets [ i ] . lock ) ;
}
2014-11-14 04:36:45 +03:00
2016-11-11 21:55:09 +03:00
if ( ! percpu & & ! lru ) {
/* lru itself can remove the least used element, so
* there is no need for an extra elem during map_update .
*/
2016-08-06 00:01:27 +03:00
err = alloc_extra_elems ( htab ) ;
if ( err )
goto free_buckets ;
}
2016-11-11 21:55:09 +03:00
if ( prealloc ) {
err = prealloc_init ( htab ) ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
if ( err )
2016-08-06 00:01:27 +03:00
goto free_extra_elems ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
}
2014-11-14 04:36:45 +03:00
return & htab - > map ;
2016-08-06 00:01:27 +03:00
free_extra_elems :
free_percpu ( htab - > extra_elems ) ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
free_buckets :
kvfree ( htab - > buckets ) ;
2014-11-14 04:36:45 +03:00
free_htab :
kfree ( htab ) ;
return ERR_PTR ( err ) ;
}
static inline u32 htab_map_hash ( const void * key , u32 key_len )
{
return jhash ( key , key_len , 0 ) ;
}
2015-12-29 17:40:27 +03:00
static inline struct bucket * __select_bucket ( struct bpf_htab * htab , u32 hash )
2014-11-14 04:36:45 +03:00
{
return & htab - > buckets [ hash & ( htab - > n_buckets - 1 ) ] ;
}
2015-12-29 17:40:27 +03:00
static inline struct hlist_head * select_bucket ( struct bpf_htab * htab , u32 hash )
{
return & __select_bucket ( htab , hash ) - > head ;
}
2014-11-14 04:36:45 +03:00
static struct htab_elem * lookup_elem_raw ( struct hlist_head * head , u32 hash ,
void * key , u32 key_size )
{
struct htab_elem * l ;
hlist_for_each_entry_rcu ( l , head , hash_node )
if ( l - > hash = = hash & & ! memcmp ( & l - > key , key , key_size ) )
return l ;
return NULL ;
}
/* Called from syscall or from eBPF program */
2016-02-02 09:39:53 +03:00
static void * __htab_map_lookup_elem ( struct bpf_map * map , void * key )
2014-11-14 04:36:45 +03:00
{
struct bpf_htab * htab = container_of ( map , struct bpf_htab , map ) ;
struct hlist_head * head ;
struct htab_elem * l ;
u32 hash , key_size ;
/* Must be called with rcu_read_lock. */
WARN_ON_ONCE ( ! rcu_read_lock_held ( ) ) ;
key_size = map - > key_size ;
hash = htab_map_hash ( key , key_size ) ;
head = select_bucket ( htab , hash ) ;
l = lookup_elem_raw ( head , hash , key , key_size ) ;
2016-02-02 09:39:53 +03:00
return l ;
}
static void * htab_map_lookup_elem ( struct bpf_map * map , void * key )
{
struct htab_elem * l = __htab_map_lookup_elem ( map , key ) ;
2014-11-14 04:36:45 +03:00
if ( l )
return l - > key + round_up ( map - > key_size , 8 ) ;
return NULL ;
}
2016-11-11 21:55:09 +03:00
static void * htab_lru_map_lookup_elem ( struct bpf_map * map , void * key )
{
struct htab_elem * l = __htab_map_lookup_elem ( map , key ) ;
if ( l ) {
bpf_lru_node_set_ref ( & l - > lru_node ) ;
return l - > key + round_up ( map - > key_size , 8 ) ;
}
return NULL ;
}
/* It is called from the bpf_lru_list when the LRU needs to delete
* older elements from the htab .
*/
static bool htab_lru_map_delete_node ( void * arg , struct bpf_lru_node * node )
{
struct bpf_htab * htab = ( struct bpf_htab * ) arg ;
struct htab_elem * l , * tgt_l ;
struct hlist_head * head ;
unsigned long flags ;
struct bucket * b ;
tgt_l = container_of ( node , struct htab_elem , lru_node ) ;
b = __select_bucket ( htab , tgt_l - > hash ) ;
head = & b - > head ;
raw_spin_lock_irqsave ( & b - > lock , flags ) ;
hlist_for_each_entry_rcu ( l , head , hash_node )
if ( l = = tgt_l ) {
hlist_del_rcu ( & l - > hash_node ) ;
break ;
}
raw_spin_unlock_irqrestore ( & b - > lock , flags ) ;
return l = = tgt_l ;
}
2014-11-14 04:36:45 +03:00
/* Called from syscall */
static int htab_map_get_next_key ( struct bpf_map * map , void * key , void * next_key )
{
struct bpf_htab * htab = container_of ( map , struct bpf_htab , map ) ;
struct hlist_head * head ;
struct htab_elem * l , * next_l ;
u32 hash , key_size ;
int i ;
WARN_ON_ONCE ( ! rcu_read_lock_held ( ) ) ;
key_size = map - > key_size ;
hash = htab_map_hash ( key , key_size ) ;
head = select_bucket ( htab , hash ) ;
/* lookup the key */
l = lookup_elem_raw ( head , hash , key , key_size ) ;
if ( ! l ) {
i = 0 ;
goto find_first_elem ;
}
/* key was found, get next key in the same bucket */
next_l = hlist_entry_safe ( rcu_dereference_raw ( hlist_next_rcu ( & l - > hash_node ) ) ,
struct htab_elem , hash_node ) ;
if ( next_l ) {
/* if next elem in this hash list is non-zero, just return it */
memcpy ( next_key , next_l - > key , key_size ) ;
return 0 ;
}
/* no more elements in this hash list, go to the next bucket */
i = hash & ( htab - > n_buckets - 1 ) ;
i + + ;
find_first_elem :
/* iterate over buckets */
for ( ; i < htab - > n_buckets ; i + + ) {
head = select_bucket ( htab , i ) ;
/* pick first element in the bucket */
next_l = hlist_entry_safe ( rcu_dereference_raw ( hlist_first_rcu ( head ) ) ,
struct htab_elem , hash_node ) ;
if ( next_l ) {
/* if it's not empty, just return it */
memcpy ( next_key , next_l - > key , key_size ) ;
return 0 ;
}
}
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
/* iterated over all buckets and all elements */
2014-11-14 04:36:45 +03:00
return - ENOENT ;
}
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
static void htab_elem_free ( struct bpf_htab * htab , struct htab_elem * l )
2016-02-02 09:39:53 +03:00
{
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
if ( htab - > map . map_type = = BPF_MAP_TYPE_PERCPU_HASH )
free_percpu ( htab_elem_get_ptr ( l , htab - > map . key_size ) ) ;
2016-02-02 09:39:53 +03:00
kfree ( l ) ;
}
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
static void htab_elem_free_rcu ( struct rcu_head * head )
2016-02-02 09:39:53 +03:00
{
struct htab_elem * l = container_of ( head , struct htab_elem , rcu ) ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
struct bpf_htab * htab = l - > htab ;
2016-02-02 09:39:53 +03:00
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
/* must increment bpf_prog_active to avoid kprobe+bpf triggering while
* we ' re calling kfree , otherwise deadlock is possible if kprobes
* are placed somewhere inside of slub
*/
preempt_disable ( ) ;
__this_cpu_inc ( bpf_prog_active ) ;
htab_elem_free ( htab , l ) ;
__this_cpu_dec ( bpf_prog_active ) ;
preempt_enable ( ) ;
2016-02-02 09:39:53 +03:00
}
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
static void free_htab_elem ( struct bpf_htab * htab , struct htab_elem * l )
2016-02-02 09:39:53 +03:00
{
2016-08-06 00:01:27 +03:00
if ( l - > state = = HTAB_EXTRA_ELEM_USED ) {
l - > state = HTAB_EXTRA_ELEM_FREE ;
return ;
}
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
if ( ! ( htab - > map . map_flags & BPF_F_NO_PREALLOC ) ) {
pcpu_freelist_push ( & htab - > freelist , & l - > fnode ) ;
2016-02-02 09:39:53 +03:00
} else {
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
atomic_dec ( & htab - > count ) ;
l - > htab = htab ;
call_rcu ( & l - > rcu , htab_elem_free_rcu ) ;
2016-02-02 09:39:53 +03:00
}
}
2016-11-11 21:55:08 +03:00
static void pcpu_copy_value ( struct bpf_htab * htab , void __percpu * pptr ,
void * value , bool onallcpus )
{
if ( ! onallcpus ) {
/* copy true value_size bytes */
memcpy ( this_cpu_ptr ( pptr ) , value , htab - > map . value_size ) ;
} else {
u32 size = round_up ( htab - > map . value_size , 8 ) ;
int off = 0 , cpu ;
for_each_possible_cpu ( cpu ) {
bpf_long_memcpy ( per_cpu_ptr ( pptr , cpu ) ,
value + off , size ) ;
off + = size ;
}
}
}
2016-02-02 09:39:53 +03:00
static struct htab_elem * alloc_htab_elem ( struct bpf_htab * htab , void * key ,
void * value , u32 key_size , u32 hash ,
2016-08-06 00:01:27 +03:00
bool percpu , bool onallcpus ,
bool old_elem_exists )
2016-02-02 09:39:53 +03:00
{
u32 size = htab - > map . value_size ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
bool prealloc = ! ( htab - > map . map_flags & BPF_F_NO_PREALLOC ) ;
2016-02-02 09:39:53 +03:00
struct htab_elem * l_new ;
void __percpu * pptr ;
2016-08-06 00:01:27 +03:00
int err = 0 ;
2016-02-02 09:39:53 +03:00
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
if ( prealloc ) {
l_new = ( struct htab_elem * ) pcpu_freelist_pop ( & htab - > freelist ) ;
if ( ! l_new )
2016-08-06 00:01:27 +03:00
err = - E2BIG ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
} else {
if ( atomic_inc_return ( & htab - > count ) > htab - > map . max_entries ) {
atomic_dec ( & htab - > count ) ;
2016-08-06 00:01:27 +03:00
err = - E2BIG ;
} else {
l_new = kmalloc ( htab - > elem_size ,
GFP_ATOMIC | __GFP_NOWARN ) ;
if ( ! l_new )
return ERR_PTR ( - ENOMEM ) ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
}
2016-08-06 00:01:27 +03:00
}
if ( err ) {
if ( ! old_elem_exists )
return ERR_PTR ( err ) ;
/* if we're updating the existing element and the hash table
* is full , use per - cpu extra elems
*/
l_new = this_cpu_ptr ( htab - > extra_elems ) ;
if ( l_new - > state ! = HTAB_EXTRA_ELEM_FREE )
return ERR_PTR ( - E2BIG ) ;
l_new - > state = HTAB_EXTRA_ELEM_USED ;
} else {
l_new - > state = HTAB_NOT_AN_EXTRA_ELEM ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
}
2016-02-02 09:39:53 +03:00
memcpy ( l_new - > key , key , key_size ) ;
if ( percpu ) {
/* round up value_size to 8 bytes */
size = round_up ( size , 8 ) ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
if ( prealloc ) {
pptr = htab_elem_get_ptr ( l_new , key_size ) ;
} else {
/* alloc_percpu zero-fills */
pptr = __alloc_percpu_gfp ( size , 8 ,
GFP_ATOMIC | __GFP_NOWARN ) ;
if ( ! pptr ) {
kfree ( l_new ) ;
return ERR_PTR ( - ENOMEM ) ;
}
2016-02-02 09:39:53 +03:00
}
2016-11-11 21:55:08 +03:00
pcpu_copy_value ( htab , pptr , value , onallcpus ) ;
bpf: add lookup/update support for per-cpu hash and array maps
The functions bpf_map_lookup_elem(map, key, value) and
bpf_map_update_elem(map, key, value, flags) need to get/set
values from all-cpus for per-cpu hash and array maps,
so that user space can aggregate/update them as necessary.
Example of single counter aggregation in user space:
unsigned int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
long values[nr_cpus];
long value = 0;
bpf_lookup_elem(fd, key, values);
for (i = 0; i < nr_cpus; i++)
value += values[i];
The user space must provide round_up(value_size, 8) * nr_cpus
array to get/set values, since kernel will use 'long' copy
of per-cpu values to try to copy good counters atomically.
It's a best-effort, since bpf programs and user space are racing
to access the same memory.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-02-02 09:39:55 +03:00
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
if ( ! prealloc )
htab_elem_set_ptr ( l_new , key_size , pptr ) ;
2016-02-02 09:39:53 +03:00
} else {
memcpy ( l_new - > key + round_up ( key_size , 8 ) , value , size ) ;
}
l_new - > hash = hash ;
return l_new ;
}
static int check_flags ( struct bpf_htab * htab , struct htab_elem * l_old ,
u64 map_flags )
{
if ( l_old & & map_flags = = BPF_NOEXIST )
/* elem already exists */
return - EEXIST ;
if ( ! l_old & & map_flags = = BPF_EXIST )
/* elem doesn't exist, cannot update it */
return - ENOENT ;
return 0 ;
}
2014-11-14 04:36:45 +03:00
/* Called from syscall or from eBPF program */
static int htab_map_update_elem ( struct bpf_map * map , void * key , void * value ,
u64 map_flags )
{
struct bpf_htab * htab = container_of ( map , struct bpf_htab , map ) ;
2016-02-02 09:39:53 +03:00
struct htab_elem * l_new = NULL , * l_old ;
2014-11-14 04:36:45 +03:00
struct hlist_head * head ;
unsigned long flags ;
2016-02-02 09:39:53 +03:00
struct bucket * b ;
u32 key_size , hash ;
2014-11-14 04:36:45 +03:00
int ret ;
2016-02-02 09:39:53 +03:00
if ( unlikely ( map_flags > BPF_EXIST ) )
2014-11-14 04:36:45 +03:00
/* unknown flags */
return - EINVAL ;
WARN_ON_ONCE ( ! rcu_read_lock_held ( ) ) ;
key_size = map - > key_size ;
2016-02-02 09:39:53 +03:00
hash = htab_map_hash ( key , key_size ) ;
b = __select_bucket ( htab , hash ) ;
2015-12-29 17:40:27 +03:00
head = & b - > head ;
2014-11-14 04:36:45 +03:00
/* bpf_map_update_elem() can be called in_irq() */
2015-12-29 17:40:27 +03:00
raw_spin_lock_irqsave ( & b - > lock , flags ) ;
2014-11-14 04:36:45 +03:00
2016-02-02 09:39:53 +03:00
l_old = lookup_elem_raw ( head , hash , key , key_size ) ;
2014-11-14 04:36:45 +03:00
2016-02-02 09:39:53 +03:00
ret = check_flags ( htab , l_old , map_flags ) ;
if ( ret )
2014-11-14 04:36:45 +03:00
goto err ;
2016-08-06 00:01:27 +03:00
l_new = alloc_htab_elem ( htab , key , value , key_size , hash , false , false ,
! ! l_old ) ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
if ( IS_ERR ( l_new ) ) {
/* all pre-allocated elements are in use or memory exhausted */
ret = PTR_ERR ( l_new ) ;
goto err ;
}
2016-02-02 09:39:53 +03:00
/* add new element to the head of the list, so that
* concurrent search will find it before old elem
2014-11-14 04:36:45 +03:00
*/
hlist_add_head_rcu ( & l_new - > hash_node , head ) ;
if ( l_old ) {
hlist_del_rcu ( & l_old - > hash_node ) ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
free_htab_elem ( htab , l_old ) ;
2014-11-14 04:36:45 +03:00
}
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
ret = 0 ;
2014-11-14 04:36:45 +03:00
err :
2015-12-29 17:40:27 +03:00
raw_spin_unlock_irqrestore ( & b - > lock , flags ) ;
2014-11-14 04:36:45 +03:00
return ret ;
}
2016-11-11 21:55:09 +03:00
static int htab_lru_map_update_elem ( struct bpf_map * map , void * key , void * value ,
u64 map_flags )
{
struct bpf_htab * htab = container_of ( map , struct bpf_htab , map ) ;
struct htab_elem * l_new , * l_old = NULL ;
struct hlist_head * head ;
unsigned long flags ;
struct bucket * b ;
u32 key_size , hash ;
int ret ;
if ( unlikely ( map_flags > BPF_EXIST ) )
/* unknown flags */
return - EINVAL ;
WARN_ON_ONCE ( ! rcu_read_lock_held ( ) ) ;
key_size = map - > key_size ;
hash = htab_map_hash ( key , key_size ) ;
b = __select_bucket ( htab , hash ) ;
head = & b - > head ;
/* For LRU, we need to alloc before taking bucket's
* spinlock because getting free nodes from LRU may need
* to remove older elements from htab and this removal
* operation will need a bucket lock .
*/
l_new = prealloc_lru_pop ( htab , key , hash ) ;
if ( ! l_new )
return - ENOMEM ;
memcpy ( l_new - > key + round_up ( map - > key_size , 8 ) , value , map - > value_size ) ;
/* bpf_map_update_elem() can be called in_irq() */
raw_spin_lock_irqsave ( & b - > lock , flags ) ;
l_old = lookup_elem_raw ( head , hash , key , key_size ) ;
ret = check_flags ( htab , l_old , map_flags ) ;
if ( ret )
goto err ;
/* add new element to the head of the list, so that
* concurrent search will find it before old elem
*/
hlist_add_head_rcu ( & l_new - > hash_node , head ) ;
if ( l_old ) {
bpf_lru_node_set_ref ( & l_new - > lru_node ) ;
hlist_del_rcu ( & l_old - > hash_node ) ;
}
ret = 0 ;
err :
raw_spin_unlock_irqrestore ( & b - > lock , flags ) ;
if ( ret )
bpf_lru_push_free ( & htab - > lru , & l_new - > lru_node ) ;
else if ( l_old )
bpf_lru_push_free ( & htab - > lru , & l_old - > lru_node ) ;
return ret ;
}
bpf: add lookup/update support for per-cpu hash and array maps
The functions bpf_map_lookup_elem(map, key, value) and
bpf_map_update_elem(map, key, value, flags) need to get/set
values from all-cpus for per-cpu hash and array maps,
so that user space can aggregate/update them as necessary.
Example of single counter aggregation in user space:
unsigned int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
long values[nr_cpus];
long value = 0;
bpf_lookup_elem(fd, key, values);
for (i = 0; i < nr_cpus; i++)
value += values[i];
The user space must provide round_up(value_size, 8) * nr_cpus
array to get/set values, since kernel will use 'long' copy
of per-cpu values to try to copy good counters atomically.
It's a best-effort, since bpf programs and user space are racing
to access the same memory.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-02-02 09:39:55 +03:00
static int __htab_percpu_map_update_elem ( struct bpf_map * map , void * key ,
void * value , u64 map_flags ,
bool onallcpus )
2016-02-02 09:39:53 +03:00
{
struct bpf_htab * htab = container_of ( map , struct bpf_htab , map ) ;
struct htab_elem * l_new = NULL , * l_old ;
struct hlist_head * head ;
unsigned long flags ;
struct bucket * b ;
u32 key_size , hash ;
int ret ;
if ( unlikely ( map_flags > BPF_EXIST ) )
/* unknown flags */
return - EINVAL ;
WARN_ON_ONCE ( ! rcu_read_lock_held ( ) ) ;
key_size = map - > key_size ;
hash = htab_map_hash ( key , key_size ) ;
b = __select_bucket ( htab , hash ) ;
head = & b - > head ;
/* bpf_map_update_elem() can be called in_irq() */
raw_spin_lock_irqsave ( & b - > lock , flags ) ;
l_old = lookup_elem_raw ( head , hash , key , key_size ) ;
ret = check_flags ( htab , l_old , map_flags ) ;
if ( ret )
goto err ;
if ( l_old ) {
/* per-cpu hash map can update value in-place */
2016-11-11 21:55:08 +03:00
pcpu_copy_value ( htab , htab_elem_get_ptr ( l_old , key_size ) ,
value , onallcpus ) ;
2016-02-02 09:39:53 +03:00
} else {
l_new = alloc_htab_elem ( htab , key , value , key_size ,
2016-08-06 00:01:27 +03:00
hash , true , onallcpus , false ) ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
if ( IS_ERR ( l_new ) ) {
ret = PTR_ERR ( l_new ) ;
2016-02-02 09:39:53 +03:00
goto err ;
}
hlist_add_head_rcu ( & l_new - > hash_node , head ) ;
}
ret = 0 ;
err :
raw_spin_unlock_irqrestore ( & b - > lock , flags ) ;
return ret ;
}
2016-11-11 21:55:10 +03:00
static int __htab_lru_percpu_map_update_elem ( struct bpf_map * map , void * key ,
void * value , u64 map_flags ,
bool onallcpus )
{
struct bpf_htab * htab = container_of ( map , struct bpf_htab , map ) ;
struct htab_elem * l_new = NULL , * l_old ;
struct hlist_head * head ;
unsigned long flags ;
struct bucket * b ;
u32 key_size , hash ;
int ret ;
if ( unlikely ( map_flags > BPF_EXIST ) )
/* unknown flags */
return - EINVAL ;
WARN_ON_ONCE ( ! rcu_read_lock_held ( ) ) ;
key_size = map - > key_size ;
hash = htab_map_hash ( key , key_size ) ;
b = __select_bucket ( htab , hash ) ;
head = & b - > head ;
/* For LRU, we need to alloc before taking bucket's
* spinlock because LRU ' s elem alloc may need
* to remove older elem from htab and this removal
* operation will need a bucket lock .
*/
if ( map_flags ! = BPF_EXIST ) {
l_new = prealloc_lru_pop ( htab , key , hash ) ;
if ( ! l_new )
return - ENOMEM ;
}
/* bpf_map_update_elem() can be called in_irq() */
raw_spin_lock_irqsave ( & b - > lock , flags ) ;
l_old = lookup_elem_raw ( head , hash , key , key_size ) ;
ret = check_flags ( htab , l_old , map_flags ) ;
if ( ret )
goto err ;
if ( l_old ) {
bpf_lru_node_set_ref ( & l_old - > lru_node ) ;
/* per-cpu hash map can update value in-place */
pcpu_copy_value ( htab , htab_elem_get_ptr ( l_old , key_size ) ,
value , onallcpus ) ;
} else {
pcpu_copy_value ( htab , htab_elem_get_ptr ( l_new , key_size ) ,
value , onallcpus ) ;
hlist_add_head_rcu ( & l_new - > hash_node , head ) ;
l_new = NULL ;
}
ret = 0 ;
err :
raw_spin_unlock_irqrestore ( & b - > lock , flags ) ;
if ( l_new )
bpf_lru_push_free ( & htab - > lru , & l_new - > lru_node ) ;
return ret ;
}
bpf: add lookup/update support for per-cpu hash and array maps
The functions bpf_map_lookup_elem(map, key, value) and
bpf_map_update_elem(map, key, value, flags) need to get/set
values from all-cpus for per-cpu hash and array maps,
so that user space can aggregate/update them as necessary.
Example of single counter aggregation in user space:
unsigned int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
long values[nr_cpus];
long value = 0;
bpf_lookup_elem(fd, key, values);
for (i = 0; i < nr_cpus; i++)
value += values[i];
The user space must provide round_up(value_size, 8) * nr_cpus
array to get/set values, since kernel will use 'long' copy
of per-cpu values to try to copy good counters atomically.
It's a best-effort, since bpf programs and user space are racing
to access the same memory.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-02-02 09:39:55 +03:00
static int htab_percpu_map_update_elem ( struct bpf_map * map , void * key ,
void * value , u64 map_flags )
{
return __htab_percpu_map_update_elem ( map , key , value , map_flags , false ) ;
}
2016-11-11 21:55:10 +03:00
static int htab_lru_percpu_map_update_elem ( struct bpf_map * map , void * key ,
void * value , u64 map_flags )
{
return __htab_lru_percpu_map_update_elem ( map , key , value , map_flags ,
false ) ;
}
2014-11-14 04:36:45 +03:00
/* Called from syscall or from eBPF program */
static int htab_map_delete_elem ( struct bpf_map * map , void * key )
{
struct bpf_htab * htab = container_of ( map , struct bpf_htab , map ) ;
struct hlist_head * head ;
2015-12-29 17:40:27 +03:00
struct bucket * b ;
2014-11-14 04:36:45 +03:00
struct htab_elem * l ;
unsigned long flags ;
u32 hash , key_size ;
int ret = - ENOENT ;
WARN_ON_ONCE ( ! rcu_read_lock_held ( ) ) ;
key_size = map - > key_size ;
hash = htab_map_hash ( key , key_size ) ;
2015-12-29 17:40:27 +03:00
b = __select_bucket ( htab , hash ) ;
head = & b - > head ;
2014-11-14 04:36:45 +03:00
2015-12-29 17:40:27 +03:00
raw_spin_lock_irqsave ( & b - > lock , flags ) ;
2014-11-14 04:36:45 +03:00
l = lookup_elem_raw ( head , hash , key , key_size ) ;
if ( l ) {
hlist_del_rcu ( & l - > hash_node ) ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
free_htab_elem ( htab , l ) ;
2014-11-14 04:36:45 +03:00
ret = 0 ;
}
2015-12-29 17:40:27 +03:00
raw_spin_unlock_irqrestore ( & b - > lock , flags ) ;
2014-11-14 04:36:45 +03:00
return ret ;
}
2016-11-11 21:55:09 +03:00
static int htab_lru_map_delete_elem ( struct bpf_map * map , void * key )
{
struct bpf_htab * htab = container_of ( map , struct bpf_htab , map ) ;
struct hlist_head * head ;
struct bucket * b ;
struct htab_elem * l ;
unsigned long flags ;
u32 hash , key_size ;
int ret = - ENOENT ;
WARN_ON_ONCE ( ! rcu_read_lock_held ( ) ) ;
key_size = map - > key_size ;
hash = htab_map_hash ( key , key_size ) ;
b = __select_bucket ( htab , hash ) ;
head = & b - > head ;
raw_spin_lock_irqsave ( & b - > lock , flags ) ;
l = lookup_elem_raw ( head , hash , key , key_size ) ;
if ( l ) {
hlist_del_rcu ( & l - > hash_node ) ;
ret = 0 ;
}
raw_spin_unlock_irqrestore ( & b - > lock , flags ) ;
if ( l )
bpf_lru_push_free ( & htab - > lru , & l - > lru_node ) ;
return ret ;
}
2014-11-14 04:36:45 +03:00
static void delete_all_elements ( struct bpf_htab * htab )
{
int i ;
for ( i = 0 ; i < htab - > n_buckets ; i + + ) {
struct hlist_head * head = select_bucket ( htab , i ) ;
struct hlist_node * n ;
struct htab_elem * l ;
hlist_for_each_entry_safe ( l , n , head , hash_node ) {
hlist_del_rcu ( & l - > hash_node ) ;
2016-11-04 02:01:19 +03:00
if ( l - > state ! = HTAB_EXTRA_ELEM_USED )
htab_elem_free ( htab , l ) ;
2014-11-14 04:36:45 +03:00
}
}
}
/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
static void htab_map_free ( struct bpf_map * map )
{
struct bpf_htab * htab = container_of ( map , struct bpf_htab , map ) ;
/* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
* so the programs ( can be more than one that used this map ) were
* disconnected from events . Wait for outstanding critical sections in
* these programs to complete
*/
synchronize_rcu ( ) ;
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
/* some of free_htab_elem() callbacks for elements of this map may
* not have executed . Wait for them .
2014-11-14 04:36:45 +03:00
*/
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
rcu_barrier ( ) ;
2016-11-11 21:55:09 +03:00
if ( htab - > map . map_flags & BPF_F_NO_PREALLOC )
bpf: pre-allocate hash map elements
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-03-08 08:57:15 +03:00
delete_all_elements ( htab ) ;
2016-11-11 21:55:09 +03:00
else
prealloc_destroy ( htab ) ;
2016-08-06 00:01:27 +03:00
free_percpu ( htab - > extra_elems ) ;
2014-11-14 04:36:45 +03:00
kvfree ( htab - > buckets ) ;
kfree ( htab ) ;
}
2015-03-01 14:31:42 +03:00
static const struct bpf_map_ops htab_ops = {
2014-11-14 04:36:45 +03:00
. map_alloc = htab_map_alloc ,
. map_free = htab_map_free ,
. map_get_next_key = htab_map_get_next_key ,
. map_lookup_elem = htab_map_lookup_elem ,
. map_update_elem = htab_map_update_elem ,
. map_delete_elem = htab_map_delete_elem ,
} ;
2015-03-01 14:31:42 +03:00
static struct bpf_map_type_list htab_type __read_mostly = {
2014-11-14 04:36:45 +03:00
. ops = & htab_ops ,
. type = BPF_MAP_TYPE_HASH ,
} ;
2016-11-11 21:55:09 +03:00
static const struct bpf_map_ops htab_lru_ops = {
. map_alloc = htab_map_alloc ,
. map_free = htab_map_free ,
. map_get_next_key = htab_map_get_next_key ,
. map_lookup_elem = htab_lru_map_lookup_elem ,
. map_update_elem = htab_lru_map_update_elem ,
. map_delete_elem = htab_lru_map_delete_elem ,
} ;
static struct bpf_map_type_list htab_lru_type __read_mostly = {
. ops = & htab_lru_ops ,
. type = BPF_MAP_TYPE_LRU_HASH ,
} ;
2016-02-02 09:39:53 +03:00
/* Called from eBPF program */
static void * htab_percpu_map_lookup_elem ( struct bpf_map * map , void * key )
{
struct htab_elem * l = __htab_map_lookup_elem ( map , key ) ;
if ( l )
return this_cpu_ptr ( htab_elem_get_ptr ( l , map - > key_size ) ) ;
else
return NULL ;
}
2016-11-11 21:55:10 +03:00
static void * htab_lru_percpu_map_lookup_elem ( struct bpf_map * map , void * key )
{
struct htab_elem * l = __htab_map_lookup_elem ( map , key ) ;
if ( l ) {
bpf_lru_node_set_ref ( & l - > lru_node ) ;
return this_cpu_ptr ( htab_elem_get_ptr ( l , map - > key_size ) ) ;
}
return NULL ;
}
bpf: add lookup/update support for per-cpu hash and array maps
The functions bpf_map_lookup_elem(map, key, value) and
bpf_map_update_elem(map, key, value, flags) need to get/set
values from all-cpus for per-cpu hash and array maps,
so that user space can aggregate/update them as necessary.
Example of single counter aggregation in user space:
unsigned int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
long values[nr_cpus];
long value = 0;
bpf_lookup_elem(fd, key, values);
for (i = 0; i < nr_cpus; i++)
value += values[i];
The user space must provide round_up(value_size, 8) * nr_cpus
array to get/set values, since kernel will use 'long' copy
of per-cpu values to try to copy good counters atomically.
It's a best-effort, since bpf programs and user space are racing
to access the same memory.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-02-02 09:39:55 +03:00
int bpf_percpu_hash_copy ( struct bpf_map * map , void * key , void * value )
{
2016-11-11 21:55:10 +03:00
struct bpf_htab * htab = container_of ( map , struct bpf_htab , map ) ;
bpf: add lookup/update support for per-cpu hash and array maps
The functions bpf_map_lookup_elem(map, key, value) and
bpf_map_update_elem(map, key, value, flags) need to get/set
values from all-cpus for per-cpu hash and array maps,
so that user space can aggregate/update them as necessary.
Example of single counter aggregation in user space:
unsigned int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
long values[nr_cpus];
long value = 0;
bpf_lookup_elem(fd, key, values);
for (i = 0; i < nr_cpus; i++)
value += values[i];
The user space must provide round_up(value_size, 8) * nr_cpus
array to get/set values, since kernel will use 'long' copy
of per-cpu values to try to copy good counters atomically.
It's a best-effort, since bpf programs and user space are racing
to access the same memory.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-02-02 09:39:55 +03:00
struct htab_elem * l ;
void __percpu * pptr ;
int ret = - ENOENT ;
int cpu , off = 0 ;
u32 size ;
/* per_cpu areas are zero-filled and bpf programs can only
* access ' value_size ' of them , so copying rounded areas
* will not leak any kernel data
*/
size = round_up ( map - > value_size , 8 ) ;
rcu_read_lock ( ) ;
l = __htab_map_lookup_elem ( map , key ) ;
if ( ! l )
goto out ;
2016-11-11 21:55:10 +03:00
if ( htab_is_lru ( htab ) )
bpf_lru_node_set_ref ( & l - > lru_node ) ;
bpf: add lookup/update support for per-cpu hash and array maps
The functions bpf_map_lookup_elem(map, key, value) and
bpf_map_update_elem(map, key, value, flags) need to get/set
values from all-cpus for per-cpu hash and array maps,
so that user space can aggregate/update them as necessary.
Example of single counter aggregation in user space:
unsigned int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
long values[nr_cpus];
long value = 0;
bpf_lookup_elem(fd, key, values);
for (i = 0; i < nr_cpus; i++)
value += values[i];
The user space must provide round_up(value_size, 8) * nr_cpus
array to get/set values, since kernel will use 'long' copy
of per-cpu values to try to copy good counters atomically.
It's a best-effort, since bpf programs and user space are racing
to access the same memory.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-02-02 09:39:55 +03:00
pptr = htab_elem_get_ptr ( l , map - > key_size ) ;
for_each_possible_cpu ( cpu ) {
bpf_long_memcpy ( value + off ,
per_cpu_ptr ( pptr , cpu ) , size ) ;
off + = size ;
}
ret = 0 ;
out :
rcu_read_unlock ( ) ;
return ret ;
}
int bpf_percpu_hash_update ( struct bpf_map * map , void * key , void * value ,
u64 map_flags )
{
2016-11-11 21:55:10 +03:00
struct bpf_htab * htab = container_of ( map , struct bpf_htab , map ) ;
2016-02-19 21:53:10 +03:00
int ret ;
rcu_read_lock ( ) ;
2016-11-11 21:55:10 +03:00
if ( htab_is_lru ( htab ) )
ret = __htab_lru_percpu_map_update_elem ( map , key , value ,
map_flags , true ) ;
else
ret = __htab_percpu_map_update_elem ( map , key , value , map_flags ,
true ) ;
2016-02-19 21:53:10 +03:00
rcu_read_unlock ( ) ;
return ret ;
bpf: add lookup/update support for per-cpu hash and array maps
The functions bpf_map_lookup_elem(map, key, value) and
bpf_map_update_elem(map, key, value, flags) need to get/set
values from all-cpus for per-cpu hash and array maps,
so that user space can aggregate/update them as necessary.
Example of single counter aggregation in user space:
unsigned int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
long values[nr_cpus];
long value = 0;
bpf_lookup_elem(fd, key, values);
for (i = 0; i < nr_cpus; i++)
value += values[i];
The user space must provide round_up(value_size, 8) * nr_cpus
array to get/set values, since kernel will use 'long' copy
of per-cpu values to try to copy good counters atomically.
It's a best-effort, since bpf programs and user space are racing
to access the same memory.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2016-02-02 09:39:55 +03:00
}
2016-02-02 09:39:53 +03:00
static const struct bpf_map_ops htab_percpu_ops = {
. map_alloc = htab_map_alloc ,
. map_free = htab_map_free ,
. map_get_next_key = htab_map_get_next_key ,
. map_lookup_elem = htab_percpu_map_lookup_elem ,
. map_update_elem = htab_percpu_map_update_elem ,
. map_delete_elem = htab_map_delete_elem ,
} ;
static struct bpf_map_type_list htab_percpu_type __read_mostly = {
. ops = & htab_percpu_ops ,
. type = BPF_MAP_TYPE_PERCPU_HASH ,
} ;
2016-11-11 21:55:10 +03:00
static const struct bpf_map_ops htab_lru_percpu_ops = {
. map_alloc = htab_map_alloc ,
. map_free = htab_map_free ,
. map_get_next_key = htab_map_get_next_key ,
. map_lookup_elem = htab_lru_percpu_map_lookup_elem ,
. map_update_elem = htab_lru_percpu_map_update_elem ,
. map_delete_elem = htab_lru_map_delete_elem ,
} ;
static struct bpf_map_type_list htab_lru_percpu_type __read_mostly = {
. ops = & htab_lru_percpu_ops ,
. type = BPF_MAP_TYPE_LRU_PERCPU_HASH ,
} ;
2014-11-14 04:36:45 +03:00
static int __init register_htab_map ( void )
{
2015-03-01 14:31:42 +03:00
bpf_register_map_type ( & htab_type ) ;
2016-02-02 09:39:53 +03:00
bpf_register_map_type ( & htab_percpu_type ) ;
2016-11-11 21:55:09 +03:00
bpf_register_map_type ( & htab_lru_type ) ;
2016-11-11 21:55:10 +03:00
bpf_register_map_type ( & htab_lru_percpu_type ) ;
2014-11-14 04:36:45 +03:00
return 0 ;
}
late_initcall ( register_htab_map ) ;