2005-04-17 02:20:36 +04:00
/*
* Copyright ( C ) 2001 Momchil Velikov
* Portions Copyright ( C ) 2001 Christoph Hellwig
2008-07-04 20:59:22 +04:00
* Copyright ( C ) 2005 SGI , Christoph Lameter
2006-12-07 07:33:44 +03:00
* Copyright ( C ) 2006 Nick Piggin
2012-03-29 01:42:53 +04:00
* Copyright ( C ) 2012 Konstantin Khlebnikov
2005-04-17 02:20:36 +04:00
*
* This program is free software ; you can redistribute it and / or
* modify it under the terms of the GNU General Public License as
* published by the Free Software Foundation ; either version 2 , or ( at
* your option ) any later version .
*
* 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 .
*
* You should have received a copy of the GNU General Public License
* along with this program ; if not , write to the Free Software
* Foundation , Inc . , 675 Mass Ave , Cambridge , MA 0213 9 , USA .
*/
# include <linux/errno.h>
# include <linux/init.h>
# include <linux/kernel.h>
2011-11-17 06:29:17 +04:00
# include <linux/export.h>
2005-04-17 02:20:36 +04:00
# include <linux/radix-tree.h>
# include <linux/percpu.h>
# include <linux/slab.h>
2014-06-07 01:38:18 +04:00
# include <linux/kmemleak.h>
2005-04-17 02:20:36 +04:00
# include <linux/notifier.h>
# include <linux/cpu.h>
# include <linux/string.h>
# include <linux/bitops.h>
2006-12-07 07:33:44 +03:00
# include <linux/rcupdate.h>
sched/preempt: Merge preempt_mask.h into preempt.h
preempt_mask.h defines all the preempt_count semantics and related
symbols: preempt, softirq, hardirq, nmi, preempt active, need resched,
etc...
preempt.h defines the accessors and mutators of preempt_count.
But there is a messy dependency game around those two header files:
* preempt_mask.h includes preempt.h in order to access preempt_count()
* preempt_mask.h defines all preempt_count semantic and symbols
except PREEMPT_NEED_RESCHED that is needed by asm/preempt.h
Thus we need to define it from preempt.h, right before including
asm/preempt.h, instead of defining it to preempt_mask.h with the
other preempt_count symbols. Therefore the preempt_count semantics
happen to be spread out.
* We plan to introduce preempt_active_[enter,exit]() to consolidate
preempt_schedule*() code. But we'll need to access both preempt_count
mutators (preempt_count_add()) and preempt_count symbols
(PREEMPT_ACTIVE, PREEMPT_OFFSET). The usual place to define preempt
operations is in preempt.h but then we'll need symbols in
preempt_mask.h which already includes preempt.h. So we end up with
a ressource circle dependency.
Lets merge preempt_mask.h into preempt.h to solve these dependency issues.
This way we gather semantic symbols and operation definition of
preempt_count in a single file.
This is a dumb copy-paste merge. Further merge re-arrangments are
performed in a subsequent patch to ease review.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1431441711-29753-2-git-send-email-fweisbec@gmail.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-12 17:41:46 +03:00
# include <linux/preempt.h> /* in_interrupt() */
2005-04-17 02:20:36 +04:00
2007-10-16 12:24:49 +04:00
/*
* The height_to_maxindex array needs to be one deeper than the maximum
* path as height 0 holds only 1 entry .
*/
static unsigned long height_to_maxindex [ RADIX_TREE_MAX_PATH + 1 ] __read_mostly ;
2005-04-17 02:20:36 +04:00
/*
* Radix tree node cache .
*/
2006-12-07 07:33:20 +03:00
static struct kmem_cache * radix_tree_node_cachep ;
2005-04-17 02:20:36 +04:00
2012-05-30 02:07:34 +04:00
/*
* The radix tree is variable - height , so an insert operation not only has
* to build the branch to its corresponding item , it also has to build the
* branch to existing items if the size has to be increased ( by
* radix_tree_extend ) .
*
* The worst case is a zero height tree with just a single item at index 0 ,
* and then inserting an item at index ULONG_MAX . This requires 2 new branches
* of RADIX_TREE_MAX_PATH size to be created , with only the root node shared .
* Hence :
*/
# define RADIX_TREE_PRELOAD_SIZE (RADIX_TREE_MAX_PATH * 2 - 1)
2005-04-17 02:20:36 +04:00
/*
* Per - cpu pool of preloaded nodes
*/
struct radix_tree_preload {
int nr ;
2015-06-26 01:02:19 +03:00
/* nodes->private_data points to next preallocated node */
struct radix_tree_node * nodes ;
2005-04-17 02:20:36 +04:00
} ;
2009-01-07 01:40:50 +03:00
static DEFINE_PER_CPU ( struct radix_tree_preload , radix_tree_preloads ) = { 0 , } ;
2005-04-17 02:20:36 +04:00
2010-11-12 01:05:19 +03:00
static inline void * ptr_to_indirect ( void * ptr )
{
return ( void * ) ( ( unsigned long ) ptr | RADIX_TREE_INDIRECT_PTR ) ;
}
static inline void * indirect_to_ptr ( void * ptr )
{
return ( void * ) ( ( unsigned long ) ptr & ~ RADIX_TREE_INDIRECT_PTR ) ;
}
2006-06-23 13:03:22 +04:00
static inline gfp_t root_gfp_mask ( struct radix_tree_root * root )
{
return root - > gfp_mask & __GFP_BITS_MASK ;
}
radix-tree: fix small lockless radix-tree bug
We shrink a radix tree when its root node has only one child, in the left
most slot. The child becomes the new root node. To perform this
operation in a manner compatible with concurrent lockless lookups, we
atomically switch the root pointer from the parent to its child.
However a concurrent lockless lookup may now have loaded a pointer to the
parent (and is presently deciding what to do next). For this reason, we
also have to keep the parent node in a valid state after shrinking the
tree, until the next RCU grace period -- otherwise this lookup with the
parent pointer may not do the right thing. Notably, we need to keep the
child in the left most slot there in case that is requested by the lookup.
This is all pretty standard RCU stuff. It is worth repeating because in
my eagerness to obey the radix tree node constructor scheme, I had broken
it by zeroing the radix tree node before the grace period.
What could happen is that a lookup can load the parent pointer, then
decide it wants to follow the left most child slot, only to find the slot
contained NULL due to the concurrent shrinker having zeroed the parent
node before waiting for a grace period. The lookup would return a false
negative as a result.
Fix it by doing that clearing in the RCU callback. I would normally want
to rip out the constructor entirely, but radix tree nodes are one of those
places where they make sense (only few cachelines will be touched soon
after allocation).
This was never actually found in any lockless pagecache testing or by the
test harness, but by seeing the odd problem with my scalable vmap rewrite.
I have not tickled the test harness into reproducing it yet, but I'll
keep working at it.
Fortunately, it is not a problem anywhere lockless pagecache is used in
mainline kernels (pagecache probe is not a guarantee, and brd does not
have concurrent lookups and deletes).
Signed-off-by: Nick Piggin <npiggin@suse.de>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: "Paul E. McKenney" <paulmck@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-06-13 02:21:52 +04:00
static inline void tag_set ( struct radix_tree_node * node , unsigned int tag ,
int offset )
{
__set_bit ( offset , node - > tags [ tag ] ) ;
}
static inline void tag_clear ( struct radix_tree_node * node , unsigned int tag ,
int offset )
{
__clear_bit ( offset , node - > tags [ tag ] ) ;
}
static inline int tag_get ( struct radix_tree_node * node , unsigned int tag ,
int offset )
{
return test_bit ( offset , node - > tags [ tag ] ) ;
}
static inline void root_tag_set ( struct radix_tree_root * root , unsigned int tag )
{
root - > gfp_mask | = ( __force gfp_t ) ( 1 < < ( tag + __GFP_BITS_SHIFT ) ) ;
}
static inline void root_tag_clear ( struct radix_tree_root * root , unsigned int tag )
{
root - > gfp_mask & = ( __force gfp_t ) ~ ( 1 < < ( tag + __GFP_BITS_SHIFT ) ) ;
}
static inline void root_tag_clear_all ( struct radix_tree_root * root )
{
root - > gfp_mask & = __GFP_BITS_MASK ;
}
static inline int root_tag_get ( struct radix_tree_root * root , unsigned int tag )
{
return ( __force unsigned ) root - > gfp_mask & ( 1 < < ( tag + __GFP_BITS_SHIFT ) ) ;
}
/*
* Returns 1 if any slot in the node has this tag set .
* Otherwise returns 0.
*/
static inline int any_tag_set ( struct radix_tree_node * node , unsigned int tag )
{
int idx ;
for ( idx = 0 ; idx < RADIX_TREE_TAG_LONGS ; idx + + ) {
if ( node - > tags [ tag ] [ idx ] )
return 1 ;
}
return 0 ;
}
2012-03-29 01:42:53 +04:00
/**
* radix_tree_find_next_bit - find the next set bit in a memory region
*
* @ addr : The address to base the search on
* @ size : The bitmap size in bits
* @ offset : The bitnumber to start searching at
*
* Unrollable variant of find_next_bit ( ) for constant size arrays .
* Tail bits starting from size to roundup ( size , BITS_PER_LONG ) must be zero .
* Returns next bit offset , or size if nothing found .
*/
static __always_inline unsigned long
radix_tree_find_next_bit ( const unsigned long * addr ,
unsigned long size , unsigned long offset )
{
if ( ! __builtin_constant_p ( size ) )
return find_next_bit ( addr , size , offset ) ;
if ( offset < size ) {
unsigned long tmp ;
addr + = offset / BITS_PER_LONG ;
tmp = * addr > > ( offset % BITS_PER_LONG ) ;
if ( tmp )
return __ffs ( tmp ) + offset ;
offset = ( offset + BITS_PER_LONG ) & ~ ( BITS_PER_LONG - 1 ) ;
while ( offset < size ) {
tmp = * + + addr ;
if ( tmp )
return __ffs ( tmp ) + offset ;
offset + = BITS_PER_LONG ;
}
}
return size ;
}
2005-04-17 02:20:36 +04:00
/*
* This assumes that the caller has performed appropriate preallocation , and
* that the caller has pinned this thread of control to the current CPU .
*/
static struct radix_tree_node *
radix_tree_node_alloc ( struct radix_tree_root * root )
{
2008-02-05 09:29:10 +03:00
struct radix_tree_node * ret = NULL ;
2006-06-23 13:03:22 +04:00
gfp_t gfp_mask = root_gfp_mask ( root ) ;
2005-04-17 02:20:36 +04:00
2013-09-12 01:26:05 +04:00
/*
* Preload code isn ' t irq safe and it doesn ' t make sence to use
* preloading in the interrupt anyway as all the allocations have to
* be atomic . So just do normal allocation when in interrupt .
*/
if ( ! ( gfp_mask & __GFP_WAIT ) & & ! in_interrupt ( ) ) {
2005-04-17 02:20:36 +04:00
struct radix_tree_preload * rtp ;
2008-02-05 09:29:10 +03:00
/*
* Provided the caller has preloaded here , we will always
* succeed in getting a node here ( and never reach
* kmem_cache_alloc )
*/
2014-06-05 03:07:56 +04:00
rtp = this_cpu_ptr ( & radix_tree_preloads ) ;
2005-04-17 02:20:36 +04:00
if ( rtp - > nr ) {
2015-06-26 01:02:19 +03:00
ret = rtp - > nodes ;
rtp - > nodes = ret - > private_data ;
ret - > private_data = NULL ;
2005-04-17 02:20:36 +04:00
rtp - > nr - - ;
}
2014-06-07 01:38:18 +04:00
/*
* Update the allocation stack trace as this is more useful
* for debugging .
*/
kmemleak_update_trace ( ret ) ;
2005-04-17 02:20:36 +04:00
}
2008-02-05 09:29:10 +03:00
if ( ret = = NULL )
2008-04-28 13:12:05 +04:00
ret = kmem_cache_alloc ( radix_tree_node_cachep , gfp_mask ) ;
2008-02-05 09:29:10 +03:00
2007-10-16 12:24:42 +04:00
BUG_ON ( radix_tree_is_indirect_ptr ( ret ) ) ;
2005-04-17 02:20:36 +04:00
return ret ;
}
2006-12-07 07:33:44 +03:00
static void radix_tree_node_rcu_free ( struct rcu_head * head )
{
struct radix_tree_node * node =
container_of ( head , struct radix_tree_node , rcu_head ) ;
2010-08-23 04:33:19 +04:00
int i ;
radix-tree: fix small lockless radix-tree bug
We shrink a radix tree when its root node has only one child, in the left
most slot. The child becomes the new root node. To perform this
operation in a manner compatible with concurrent lockless lookups, we
atomically switch the root pointer from the parent to its child.
However a concurrent lockless lookup may now have loaded a pointer to the
parent (and is presently deciding what to do next). For this reason, we
also have to keep the parent node in a valid state after shrinking the
tree, until the next RCU grace period -- otherwise this lookup with the
parent pointer may not do the right thing. Notably, we need to keep the
child in the left most slot there in case that is requested by the lookup.
This is all pretty standard RCU stuff. It is worth repeating because in
my eagerness to obey the radix tree node constructor scheme, I had broken
it by zeroing the radix tree node before the grace period.
What could happen is that a lookup can load the parent pointer, then
decide it wants to follow the left most child slot, only to find the slot
contained NULL due to the concurrent shrinker having zeroed the parent
node before waiting for a grace period. The lookup would return a false
negative as a result.
Fix it by doing that clearing in the RCU callback. I would normally want
to rip out the constructor entirely, but radix tree nodes are one of those
places where they make sense (only few cachelines will be touched soon
after allocation).
This was never actually found in any lockless pagecache testing or by the
test harness, but by seeing the odd problem with my scalable vmap rewrite.
I have not tickled the test harness into reproducing it yet, but I'll
keep working at it.
Fortunately, it is not a problem anywhere lockless pagecache is used in
mainline kernels (pagecache probe is not a guarantee, and brd does not
have concurrent lookups and deletes).
Signed-off-by: Nick Piggin <npiggin@suse.de>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: "Paul E. McKenney" <paulmck@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-06-13 02:21:52 +04:00
/*
* must only free zeroed nodes into the slab . radix_tree_shrink
* can leave us with a non - NULL entry in the first slot , so clear
* that here to make sure .
*/
2010-08-23 04:33:19 +04:00
for ( i = 0 ; i < RADIX_TREE_MAX_TAGS ; i + + )
tag_clear ( node , i , 0 ) ;
radix-tree: fix small lockless radix-tree bug
We shrink a radix tree when its root node has only one child, in the left
most slot. The child becomes the new root node. To perform this
operation in a manner compatible with concurrent lockless lookups, we
atomically switch the root pointer from the parent to its child.
However a concurrent lockless lookup may now have loaded a pointer to the
parent (and is presently deciding what to do next). For this reason, we
also have to keep the parent node in a valid state after shrinking the
tree, until the next RCU grace period -- otherwise this lookup with the
parent pointer may not do the right thing. Notably, we need to keep the
child in the left most slot there in case that is requested by the lookup.
This is all pretty standard RCU stuff. It is worth repeating because in
my eagerness to obey the radix tree node constructor scheme, I had broken
it by zeroing the radix tree node before the grace period.
What could happen is that a lookup can load the parent pointer, then
decide it wants to follow the left most child slot, only to find the slot
contained NULL due to the concurrent shrinker having zeroed the parent
node before waiting for a grace period. The lookup would return a false
negative as a result.
Fix it by doing that clearing in the RCU callback. I would normally want
to rip out the constructor entirely, but radix tree nodes are one of those
places where they make sense (only few cachelines will be touched soon
after allocation).
This was never actually found in any lockless pagecache testing or by the
test harness, but by seeing the odd problem with my scalable vmap rewrite.
I have not tickled the test harness into reproducing it yet, but I'll
keep working at it.
Fortunately, it is not a problem anywhere lockless pagecache is used in
mainline kernels (pagecache probe is not a guarantee, and brd does not
have concurrent lookups and deletes).
Signed-off-by: Nick Piggin <npiggin@suse.de>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: "Paul E. McKenney" <paulmck@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-06-13 02:21:52 +04:00
node - > slots [ 0 ] = NULL ;
node - > count = 0 ;
2006-12-07 07:33:44 +03:00
kmem_cache_free ( radix_tree_node_cachep , node ) ;
}
2005-04-17 02:20:36 +04:00
static inline void
radix_tree_node_free ( struct radix_tree_node * node )
{
2006-12-07 07:33:44 +03:00
call_rcu ( & node - > rcu_head , radix_tree_node_rcu_free ) ;
2005-04-17 02:20:36 +04:00
}
/*
* Load up this CPU ' s radix_tree_node buffer with sufficient objects to
* ensure that the addition of a single element in the tree cannot fail . On
* success , return zero , with preemption disabled . On error , return - ENOMEM
* with preemption not disabled .
FS-Cache: Use radix tree preload correctly in tracking of pages to be stored
__fscache_write_page() attempts to load the radix tree preallocation pool for
the CPU it is on before calling radix_tree_insert(), as the insertion must be
done inside a pair of spinlocks.
Use of the preallocation pool, however, is contingent on the radix tree being
initialised without __GFP_WAIT specified. __fscache_acquire_cookie() was
passing GFP_NOFS to INIT_RADIX_TREE() - but that includes __GFP_WAIT.
The solution is to AND out __GFP_WAIT.
Additionally, the banner comment to radix_tree_preload() is altered to make
note of this prerequisite. Possibly there should be a WARN_ON() too.
Without this fix, I have seen the following recursive deadlock caused by
radix_tree_insert() attempting to allocate memory inside the spinlocked
region, which resulted in FS-Cache being called back into to release memory -
which required the spinlock already held.
=============================================
[ INFO: possible recursive locking detected ]
2.6.32-rc6-cachefs #24
---------------------------------------------
nfsiod/7916 is trying to acquire lock:
(&cookie->lock){+.+.-.}, at: [<ffffffffa0076872>] __fscache_uncache_page+0xdb/0x160 [fscache]
but task is already holding lock:
(&cookie->lock){+.+.-.}, at: [<ffffffffa0076acc>] __fscache_write_page+0x15c/0x3f3 [fscache]
other info that might help us debug this:
5 locks held by nfsiod/7916:
#0: (nfsiod){+.+.+.}, at: [<ffffffff81048290>] worker_thread+0x19a/0x2e2
#1: (&task->u.tk_work#2){+.+.+.}, at: [<ffffffff81048290>] worker_thread+0x19a/0x2e2
#2: (&cookie->lock){+.+.-.}, at: [<ffffffffa0076acc>] __fscache_write_page+0x15c/0x3f3 [fscache]
#3: (&object->lock#2){+.+.-.}, at: [<ffffffffa0076b07>] __fscache_write_page+0x197/0x3f3 [fscache]
#4: (&cookie->stores_lock){+.+...}, at: [<ffffffffa0076b0f>] __fscache_write_page+0x19f/0x3f3 [fscache]
stack backtrace:
Pid: 7916, comm: nfsiod Not tainted 2.6.32-rc6-cachefs #24
Call Trace:
[<ffffffff8105ac7f>] __lock_acquire+0x1649/0x16e3
[<ffffffff81059ded>] ? __lock_acquire+0x7b7/0x16e3
[<ffffffff8100e27d>] ? dump_trace+0x248/0x257
[<ffffffff8105ad70>] lock_acquire+0x57/0x6d
[<ffffffffa0076872>] ? __fscache_uncache_page+0xdb/0x160 [fscache]
[<ffffffff8135467c>] _spin_lock+0x2c/0x3b
[<ffffffffa0076872>] ? __fscache_uncache_page+0xdb/0x160 [fscache]
[<ffffffffa0076872>] __fscache_uncache_page+0xdb/0x160 [fscache]
[<ffffffffa0077eb7>] ? __fscache_check_page_write+0x0/0x71 [fscache]
[<ffffffffa00b4755>] nfs_fscache_release_page+0x86/0xc4 [nfs]
[<ffffffffa00907f0>] nfs_release_page+0x3c/0x41 [nfs]
[<ffffffff81087ffb>] try_to_release_page+0x32/0x3b
[<ffffffff81092c2b>] shrink_page_list+0x316/0x4ac
[<ffffffff81058a9b>] ? mark_held_locks+0x52/0x70
[<ffffffff8135451b>] ? _spin_unlock_irq+0x2b/0x31
[<ffffffff81093153>] shrink_inactive_list+0x392/0x67c
[<ffffffff81058a9b>] ? mark_held_locks+0x52/0x70
[<ffffffff810934ca>] shrink_list+0x8d/0x8f
[<ffffffff81093744>] shrink_zone+0x278/0x33c
[<ffffffff81052c70>] ? ktime_get_ts+0xad/0xba
[<ffffffff8109453b>] try_to_free_pages+0x22e/0x392
[<ffffffff8109184c>] ? isolate_pages_global+0x0/0x212
[<ffffffff8108e16b>] __alloc_pages_nodemask+0x3dc/0x5cf
[<ffffffff810ae24a>] cache_alloc_refill+0x34d/0x6c1
[<ffffffff811bcf74>] ? radix_tree_node_alloc+0x52/0x5c
[<ffffffff810ae929>] kmem_cache_alloc+0xb2/0x118
[<ffffffff811bcf74>] radix_tree_node_alloc+0x52/0x5c
[<ffffffff811bcfd5>] radix_tree_insert+0x57/0x19c
[<ffffffffa0076b53>] __fscache_write_page+0x1e3/0x3f3 [fscache]
[<ffffffffa00b4248>] __nfs_readpage_to_fscache+0x58/0x11e [nfs]
[<ffffffffa009bb77>] nfs_readpage_release+0x34/0x9b [nfs]
[<ffffffffa009c0d9>] nfs_readpage_release_full+0x32/0x4b [nfs]
[<ffffffffa0006cff>] rpc_release_calldata+0x12/0x14 [sunrpc]
[<ffffffffa0006e2d>] rpc_free_task+0x59/0x61 [sunrpc]
[<ffffffffa0006f03>] rpc_async_release+0x10/0x12 [sunrpc]
[<ffffffff810482e5>] worker_thread+0x1ef/0x2e2
[<ffffffff81048290>] ? worker_thread+0x19a/0x2e2
[<ffffffff81352433>] ? thread_return+0x3e/0x101
[<ffffffffa0006ef3>] ? rpc_async_release+0x0/0x12 [sunrpc]
[<ffffffff8104bff5>] ? autoremove_wake_function+0x0/0x34
[<ffffffff81058d25>] ? trace_hardirqs_on+0xd/0xf
[<ffffffff810480f6>] ? worker_thread+0x0/0x2e2
[<ffffffff8104bd21>] kthread+0x7a/0x82
[<ffffffff8100beda>] child_rip+0xa/0x20
[<ffffffff8100b87c>] ? restore_args+0x0/0x30
[<ffffffff8104c2b9>] ? add_wait_queue+0x15/0x44
[<ffffffff8104bca7>] ? kthread+0x0/0x82
[<ffffffff8100bed0>] ? child_rip+0x0/0x20
Signed-off-by: David Howells <dhowells@redhat.com>
2009-11-19 21:11:14 +03:00
*
* To make use of this facility , the radix tree must be initialised without
* __GFP_WAIT being passed to INIT_RADIX_TREE ( ) .
2005-04-17 02:20:36 +04:00
*/
2013-09-12 01:26:05 +04:00
static int __radix_tree_preload ( gfp_t gfp_mask )
2005-04-17 02:20:36 +04:00
{
struct radix_tree_preload * rtp ;
struct radix_tree_node * node ;
int ret = - ENOMEM ;
preempt_disable ( ) ;
2014-06-05 03:07:56 +04:00
rtp = this_cpu_ptr ( & radix_tree_preloads ) ;
2015-06-26 01:02:19 +03:00
while ( rtp - > nr < RADIX_TREE_PRELOAD_SIZE ) {
2005-04-17 02:20:36 +04:00
preempt_enable ( ) ;
2008-04-28 13:12:05 +04:00
node = kmem_cache_alloc ( radix_tree_node_cachep , gfp_mask ) ;
2005-04-17 02:20:36 +04:00
if ( node = = NULL )
goto out ;
preempt_disable ( ) ;
2014-06-05 03:07:56 +04:00
rtp = this_cpu_ptr ( & radix_tree_preloads ) ;
2015-06-26 01:02:19 +03:00
if ( rtp - > nr < RADIX_TREE_PRELOAD_SIZE ) {
node - > private_data = rtp - > nodes ;
rtp - > nodes = node ;
rtp - > nr + + ;
} else {
2005-04-17 02:20:36 +04:00
kmem_cache_free ( radix_tree_node_cachep , node ) ;
2015-06-26 01:02:19 +03:00
}
2005-04-17 02:20:36 +04:00
}
ret = 0 ;
out :
return ret ;
}
2013-09-12 01:26:05 +04:00
/*
* Load up this CPU ' s radix_tree_node buffer with sufficient objects to
* ensure that the addition of a single element in the tree cannot fail . On
* success , return zero , with preemption disabled . On error , return - ENOMEM
* with preemption not disabled .
*
* To make use of this facility , the radix tree must be initialised without
* __GFP_WAIT being passed to INIT_RADIX_TREE ( ) .
*/
int radix_tree_preload ( gfp_t gfp_mask )
{
/* Warn on non-sensical use... */
WARN_ON_ONCE ( ! ( gfp_mask & __GFP_WAIT ) ) ;
return __radix_tree_preload ( gfp_mask ) ;
}
2007-07-14 10:05:04 +04:00
EXPORT_SYMBOL ( radix_tree_preload ) ;
2005-04-17 02:20:36 +04:00
2013-09-12 01:26:05 +04:00
/*
* The same as above function , except we don ' t guarantee preloading happens .
* We do it , if we decide it helps . On success , return zero with preemption
* disabled . On error , return - ENOMEM with preemption not disabled .
*/
int radix_tree_maybe_preload ( gfp_t gfp_mask )
{
if ( gfp_mask & __GFP_WAIT )
return __radix_tree_preload ( gfp_mask ) ;
/* Preloading doesn't help anything with this gfp mask, skip it */
preempt_disable ( ) ;
return 0 ;
}
EXPORT_SYMBOL ( radix_tree_maybe_preload ) ;
2005-04-17 02:20:36 +04:00
/*
* Return the maximum key which can be store into a
* radix tree with height HEIGHT .
*/
static inline unsigned long radix_tree_maxindex ( unsigned int height )
{
return height_to_maxindex [ height ] ;
}
/*
* Extend a radix tree so it can store key @ index .
*/
static int radix_tree_extend ( struct radix_tree_root * root , unsigned long index )
{
struct radix_tree_node * node ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
struct radix_tree_node * slot ;
2005-04-17 02:20:36 +04:00
unsigned int height ;
int tag ;
/* Figure out what the height should be. */
height = root - > height + 1 ;
while ( index > radix_tree_maxindex ( height ) )
height + + ;
if ( root - > rnode = = NULL ) {
root - > height = height ;
goto out ;
}
do {
2006-12-07 07:33:44 +03:00
unsigned int newheight ;
2005-04-17 02:20:36 +04:00
if ( ! ( node = radix_tree_node_alloc ( root ) ) )
return - ENOMEM ;
/* Propagate the aggregated tag info into the new root */
2006-03-25 14:08:05 +03:00
for ( tag = 0 ; tag < RADIX_TREE_MAX_TAGS ; tag + + ) {
2006-06-23 13:03:22 +04:00
if ( root_tag_get ( root , tag ) )
2005-04-17 02:20:36 +04:00
tag_set ( node , tag , 0 ) ;
}
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
/* Increase the height. */
2006-12-07 07:33:44 +03:00
newheight = root - > height + 1 ;
mm: keep page cache radix tree nodes in check
Previously, page cache radix tree nodes were freed after reclaim emptied
out their page pointers. But now reclaim stores shadow entries in their
place, which are only reclaimed when the inodes themselves are
reclaimed. This is problematic for bigger files that are still in use
after they have a significant amount of their cache reclaimed, without
any of those pages actually refaulting. The shadow entries will just
sit there and waste memory. In the worst case, the shadow entries will
accumulate until the machine runs out of memory.
To get this under control, the VM will track radix tree nodes
exclusively containing shadow entries on a per-NUMA node list. Per-NUMA
rather than global because we expect the radix tree nodes themselves to
be allocated node-locally and we want to reduce cross-node references of
otherwise independent cache workloads. A simple shrinker will then
reclaim these nodes on memory pressure.
A few things need to be stored in the radix tree node to implement the
shadow node LRU and allow tree deletions coming from the list:
1. There is no index available that would describe the reverse path
from the node up to the tree root, which is needed to perform a
deletion. To solve this, encode in each node its offset inside the
parent. This can be stored in the unused upper bits of the same
member that stores the node's height at no extra space cost.
2. The number of shadow entries needs to be counted in addition to the
regular entries, to quickly detect when the node is ready to go to
the shadow node LRU list. The current entry count is an unsigned
int but the maximum number of entries is 64, so a shadow counter
can easily be stored in the unused upper bits.
3. Tree modification needs tree lock and tree root, which are located
in the address space, so store an address_space backpointer in the
node. The parent pointer of the node is in a union with the 2-word
rcu_head, so the backpointer comes at no extra cost as well.
4. The node needs to be linked to an LRU list, which requires a list
head inside the node. This does increase the size of the node, but
it does not change the number of objects that fit into a slab page.
[akpm@linux-foundation.org: export the right function]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-04 01:47:56 +04:00
BUG_ON ( newheight & ~ RADIX_TREE_HEIGHT_MASK ) ;
node - > path = newheight ;
2005-04-17 02:20:36 +04:00
node - > count = 1 ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
node - > parent = NULL ;
slot = root - > rnode ;
if ( newheight > 1 ) {
slot = indirect_to_ptr ( slot ) ;
slot - > parent = node ;
}
node - > slots [ 0 ] = slot ;
2010-11-12 01:05:19 +03:00
node = ptr_to_indirect ( node ) ;
2006-12-07 07:33:44 +03:00
rcu_assign_pointer ( root - > rnode , node ) ;
root - > height = newheight ;
2005-04-17 02:20:36 +04:00
} while ( height > root - > height ) ;
out :
return 0 ;
}
/**
2014-04-04 01:47:54 +04:00
* __radix_tree_create - create a slot in a radix tree
2005-04-17 02:20:36 +04:00
* @ root : radix tree root
* @ index : index key
2014-04-04 01:47:54 +04:00
* @ nodep : returns node
* @ slotp : returns slot
2005-04-17 02:20:36 +04:00
*
2014-04-04 01:47:54 +04:00
* Create , if necessary , and return the node and slot for an item
* at position @ index in the radix tree @ root .
*
* Until there is more than one item in the tree , no nodes are
* allocated and @ root - > rnode is used as a direct slot instead of
* pointing to a node , in which case * @ nodep will be NULL .
*
* Returns - ENOMEM , or 0 for success .
2005-04-17 02:20:36 +04:00
*/
2014-04-04 01:47:54 +04:00
int __radix_tree_create ( struct radix_tree_root * root , unsigned long index ,
struct radix_tree_node * * nodep , void * * * slotp )
2005-04-17 02:20:36 +04:00
{
2005-09-07 02:16:46 +04:00
struct radix_tree_node * node = NULL , * slot ;
2014-04-04 01:47:54 +04:00
unsigned int height , shift , offset ;
2005-04-17 02:20:36 +04:00
int error ;
/* Make sure the tree is high enough. */
2006-06-23 13:03:22 +04:00
if ( index > radix_tree_maxindex ( root - > height ) ) {
2005-04-17 02:20:36 +04:00
error = radix_tree_extend ( root , index ) ;
if ( error )
return error ;
}
2010-11-12 01:05:19 +03:00
slot = indirect_to_ptr ( root - > rnode ) ;
2007-10-16 12:24:42 +04:00
2005-04-17 02:20:36 +04:00
height = root - > height ;
shift = ( height - 1 ) * RADIX_TREE_MAP_SHIFT ;
offset = 0 ; /* uninitialised var warning */
2006-06-23 13:03:22 +04:00
while ( height > 0 ) {
2005-09-07 02:16:46 +04:00
if ( slot = = NULL ) {
2005-04-17 02:20:36 +04:00
/* Have to add a child node. */
2005-09-07 02:16:46 +04:00
if ( ! ( slot = radix_tree_node_alloc ( root ) ) )
2005-04-17 02:20:36 +04:00
return - ENOMEM ;
mm: keep page cache radix tree nodes in check
Previously, page cache radix tree nodes were freed after reclaim emptied
out their page pointers. But now reclaim stores shadow entries in their
place, which are only reclaimed when the inodes themselves are
reclaimed. This is problematic for bigger files that are still in use
after they have a significant amount of their cache reclaimed, without
any of those pages actually refaulting. The shadow entries will just
sit there and waste memory. In the worst case, the shadow entries will
accumulate until the machine runs out of memory.
To get this under control, the VM will track radix tree nodes
exclusively containing shadow entries on a per-NUMA node list. Per-NUMA
rather than global because we expect the radix tree nodes themselves to
be allocated node-locally and we want to reduce cross-node references of
otherwise independent cache workloads. A simple shrinker will then
reclaim these nodes on memory pressure.
A few things need to be stored in the radix tree node to implement the
shadow node LRU and allow tree deletions coming from the list:
1. There is no index available that would describe the reverse path
from the node up to the tree root, which is needed to perform a
deletion. To solve this, encode in each node its offset inside the
parent. This can be stored in the unused upper bits of the same
member that stores the node's height at no extra space cost.
2. The number of shadow entries needs to be counted in addition to the
regular entries, to quickly detect when the node is ready to go to
the shadow node LRU list. The current entry count is an unsigned
int but the maximum number of entries is 64, so a shadow counter
can easily be stored in the unused upper bits.
3. Tree modification needs tree lock and tree root, which are located
in the address space, so store an address_space backpointer in the
node. The parent pointer of the node is in a union with the 2-word
rcu_head, so the backpointer comes at no extra cost as well.
4. The node needs to be linked to an LRU list, which requires a list
head inside the node. This does increase the size of the node, but
it does not change the number of objects that fit into a slab page.
[akpm@linux-foundation.org: export the right function]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-04 01:47:56 +04:00
slot - > path = height ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
slot - > parent = node ;
2005-09-07 02:16:46 +04:00
if ( node ) {
2006-12-07 07:33:44 +03:00
rcu_assign_pointer ( node - > slots [ offset ] , slot ) ;
2005-04-17 02:20:36 +04:00
node - > count + + ;
mm: keep page cache radix tree nodes in check
Previously, page cache radix tree nodes were freed after reclaim emptied
out their page pointers. But now reclaim stores shadow entries in their
place, which are only reclaimed when the inodes themselves are
reclaimed. This is problematic for bigger files that are still in use
after they have a significant amount of their cache reclaimed, without
any of those pages actually refaulting. The shadow entries will just
sit there and waste memory. In the worst case, the shadow entries will
accumulate until the machine runs out of memory.
To get this under control, the VM will track radix tree nodes
exclusively containing shadow entries on a per-NUMA node list. Per-NUMA
rather than global because we expect the radix tree nodes themselves to
be allocated node-locally and we want to reduce cross-node references of
otherwise independent cache workloads. A simple shrinker will then
reclaim these nodes on memory pressure.
A few things need to be stored in the radix tree node to implement the
shadow node LRU and allow tree deletions coming from the list:
1. There is no index available that would describe the reverse path
from the node up to the tree root, which is needed to perform a
deletion. To solve this, encode in each node its offset inside the
parent. This can be stored in the unused upper bits of the same
member that stores the node's height at no extra space cost.
2. The number of shadow entries needs to be counted in addition to the
regular entries, to quickly detect when the node is ready to go to
the shadow node LRU list. The current entry count is an unsigned
int but the maximum number of entries is 64, so a shadow counter
can easily be stored in the unused upper bits.
3. Tree modification needs tree lock and tree root, which are located
in the address space, so store an address_space backpointer in the
node. The parent pointer of the node is in a union with the 2-word
rcu_head, so the backpointer comes at no extra cost as well.
4. The node needs to be linked to an LRU list, which requires a list
head inside the node. This does increase the size of the node, but
it does not change the number of objects that fit into a slab page.
[akpm@linux-foundation.org: export the right function]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-04 01:47:56 +04:00
slot - > path | = offset < < RADIX_TREE_HEIGHT_SHIFT ;
2005-09-07 02:16:46 +04:00
} else
2010-11-12 01:05:19 +03:00
rcu_assign_pointer ( root - > rnode , ptr_to_indirect ( slot ) ) ;
2005-04-17 02:20:36 +04:00
}
/* Go a level down */
offset = ( index > > shift ) & RADIX_TREE_MAP_MASK ;
2005-09-07 02:16:46 +04:00
node = slot ;
slot = node - > slots [ offset ] ;
2005-04-17 02:20:36 +04:00
shift - = RADIX_TREE_MAP_SHIFT ;
height - - ;
2006-06-23 13:03:22 +04:00
}
2005-04-17 02:20:36 +04:00
2014-04-04 01:47:54 +04:00
if ( nodep )
* nodep = node ;
if ( slotp )
* slotp = node ? node - > slots + offset : ( void * * ) & root - > rnode ;
return 0 ;
}
/**
* radix_tree_insert - insert into a radix tree
* @ root : radix tree root
* @ index : index key
* @ item : item to insert
*
* Insert an item into the radix tree at position @ index .
*/
int radix_tree_insert ( struct radix_tree_root * root ,
unsigned long index , void * item )
{
struct radix_tree_node * node ;
void * * slot ;
int error ;
BUG_ON ( radix_tree_is_indirect_ptr ( item ) ) ;
error = __radix_tree_create ( root , index , & node , & slot ) ;
if ( error )
return error ;
if ( * slot ! = NULL )
2005-04-17 02:20:36 +04:00
return - EEXIST ;
2014-04-04 01:47:54 +04:00
rcu_assign_pointer ( * slot , item ) ;
2005-09-07 02:16:46 +04:00
2006-06-23 13:03:22 +04:00
if ( node ) {
node - > count + + ;
2014-04-04 01:47:54 +04:00
BUG_ON ( tag_get ( node , 0 , index & RADIX_TREE_MAP_MASK ) ) ;
BUG_ON ( tag_get ( node , 1 , index & RADIX_TREE_MAP_MASK ) ) ;
2006-06-23 13:03:22 +04:00
} else {
BUG_ON ( root_tag_get ( root , 0 ) ) ;
BUG_ON ( root_tag_get ( root , 1 ) ) ;
}
2005-04-17 02:20:36 +04:00
return 0 ;
}
EXPORT_SYMBOL ( radix_tree_insert ) ;
2014-04-04 01:47:54 +04:00
/**
* __radix_tree_lookup - lookup an item in a radix tree
* @ root : radix tree root
* @ index : index key
* @ nodep : returns node
* @ slotp : returns slot
*
* Lookup and return the item at position @ index in the radix
* tree @ root .
*
* Until there is more than one item in the tree , no nodes are
* allocated and @ root - > rnode is used as a direct slot instead of
* pointing to a node , in which case * @ nodep will be NULL .
2006-12-07 07:33:44 +03:00
*/
2014-04-04 01:47:54 +04:00
void * __radix_tree_lookup ( struct radix_tree_root * root , unsigned long index ,
struct radix_tree_node * * nodep , void * * * slotp )
2005-04-17 02:20:36 +04:00
{
2014-04-04 01:47:54 +04:00
struct radix_tree_node * node , * parent ;
2005-04-17 02:20:36 +04:00
unsigned int height , shift ;
2014-04-04 01:47:54 +04:00
void * * slot ;
2006-06-23 13:03:22 +04:00
2010-02-23 04:04:54 +03:00
node = rcu_dereference_raw ( root - > rnode ) ;
2006-12-07 07:33:44 +03:00
if ( node = = NULL )
2005-04-17 02:20:36 +04:00
return NULL ;
2007-10-16 12:24:42 +04:00
if ( ! radix_tree_is_indirect_ptr ( node ) ) {
2006-12-07 07:33:44 +03:00
if ( index > 0 )
return NULL ;
2014-04-04 01:47:54 +04:00
if ( nodep )
* nodep = NULL ;
if ( slotp )
* slotp = ( void * * ) & root - > rnode ;
return node ;
2006-12-07 07:33:44 +03:00
}
2010-11-12 01:05:19 +03:00
node = indirect_to_ptr ( node ) ;
2006-12-07 07:33:44 +03:00
mm: keep page cache radix tree nodes in check
Previously, page cache radix tree nodes were freed after reclaim emptied
out their page pointers. But now reclaim stores shadow entries in their
place, which are only reclaimed when the inodes themselves are
reclaimed. This is problematic for bigger files that are still in use
after they have a significant amount of their cache reclaimed, without
any of those pages actually refaulting. The shadow entries will just
sit there and waste memory. In the worst case, the shadow entries will
accumulate until the machine runs out of memory.
To get this under control, the VM will track radix tree nodes
exclusively containing shadow entries on a per-NUMA node list. Per-NUMA
rather than global because we expect the radix tree nodes themselves to
be allocated node-locally and we want to reduce cross-node references of
otherwise independent cache workloads. A simple shrinker will then
reclaim these nodes on memory pressure.
A few things need to be stored in the radix tree node to implement the
shadow node LRU and allow tree deletions coming from the list:
1. There is no index available that would describe the reverse path
from the node up to the tree root, which is needed to perform a
deletion. To solve this, encode in each node its offset inside the
parent. This can be stored in the unused upper bits of the same
member that stores the node's height at no extra space cost.
2. The number of shadow entries needs to be counted in addition to the
regular entries, to quickly detect when the node is ready to go to
the shadow node LRU list. The current entry count is an unsigned
int but the maximum number of entries is 64, so a shadow counter
can easily be stored in the unused upper bits.
3. Tree modification needs tree lock and tree root, which are located
in the address space, so store an address_space backpointer in the
node. The parent pointer of the node is in a union with the 2-word
rcu_head, so the backpointer comes at no extra cost as well.
4. The node needs to be linked to an LRU list, which requires a list
head inside the node. This does increase the size of the node, but
it does not change the number of objects that fit into a slab page.
[akpm@linux-foundation.org: export the right function]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-04 01:47:56 +04:00
height = node - > path & RADIX_TREE_HEIGHT_MASK ;
2006-12-07 07:33:44 +03:00
if ( index > radix_tree_maxindex ( height ) )
return NULL ;
2006-06-23 13:03:22 +04:00
2005-04-17 02:20:36 +04:00
shift = ( height - 1 ) * RADIX_TREE_MAP_SHIFT ;
2006-12-07 07:33:44 +03:00
do {
2014-04-04 01:47:54 +04:00
parent = node ;
slot = node - > slots + ( ( index > > shift ) & RADIX_TREE_MAP_MASK ) ;
2010-02-23 04:04:54 +03:00
node = rcu_dereference_raw ( * slot ) ;
2006-12-07 07:33:44 +03:00
if ( node = = NULL )
2005-04-17 02:20:36 +04:00
return NULL ;
shift - = RADIX_TREE_MAP_SHIFT ;
height - - ;
2006-12-07 07:33:44 +03:00
} while ( height > 0 ) ;
2005-04-17 02:20:36 +04:00
2014-04-04 01:47:54 +04:00
if ( nodep )
* nodep = parent ;
if ( slotp )
* slotp = slot ;
return node ;
2009-06-17 02:33:42 +04:00
}
/**
* radix_tree_lookup_slot - lookup a slot in a radix tree
* @ root : radix tree root
* @ index : index key
*
* Returns : the slot corresponding to the position @ index in the
* radix tree @ root . This is useful for update - if - exists operations .
*
* This function can be called under rcu_read_lock iff the slot is not
* modified by radix_tree_replace_slot , otherwise it must be called
* exclusive from other writers . Any dereference of the slot must be done
* using radix_tree_deref_slot .
*/
void * * radix_tree_lookup_slot ( struct radix_tree_root * root , unsigned long index )
{
2014-04-04 01:47:54 +04:00
void * * slot ;
if ( ! __radix_tree_lookup ( root , index , NULL , & slot ) )
return NULL ;
return slot ;
2005-11-07 11:59:29 +03:00
}
EXPORT_SYMBOL ( radix_tree_lookup_slot ) ;
/**
* radix_tree_lookup - perform lookup operation on a radix tree
* @ root : radix tree root
* @ index : index key
*
* Lookup the item at the position @ index in the radix tree @ root .
2006-12-07 07:33:44 +03:00
*
* This function can be called under rcu_read_lock , however the caller
* must manage lifetimes of leaf nodes ( eg . RCU may also be used to free
* them safely ) . No RCU barriers are required to access or modify the
* returned item , however .
2005-11-07 11:59:29 +03:00
*/
void * radix_tree_lookup ( struct radix_tree_root * root , unsigned long index )
{
2014-04-04 01:47:54 +04:00
return __radix_tree_lookup ( root , index , NULL , NULL ) ;
2005-04-17 02:20:36 +04:00
}
EXPORT_SYMBOL ( radix_tree_lookup ) ;
/**
* radix_tree_tag_set - set a tag on a radix tree node
* @ root : radix tree root
* @ index : index key
* @ tag : tag index
*
2006-03-25 14:08:05 +03:00
* Set the search tag ( which must be < RADIX_TREE_MAX_TAGS )
* corresponding to @ index in the radix tree . From
2005-04-17 02:20:36 +04:00
* the root all the way down to the leaf node .
*
* Returns the address of the tagged item . Setting a tag on a not - present
* item is a bug .
*/
void * radix_tree_tag_set ( struct radix_tree_root * root ,
2006-03-25 14:08:05 +03:00
unsigned long index , unsigned int tag )
2005-04-17 02:20:36 +04:00
{
unsigned int height , shift ;
2005-09-07 02:16:46 +04:00
struct radix_tree_node * slot ;
2005-04-17 02:20:36 +04:00
height = root - > height ;
2006-06-23 13:03:25 +04:00
BUG_ON ( index > radix_tree_maxindex ( height ) ) ;
2005-04-17 02:20:36 +04:00
2010-11-12 01:05:19 +03:00
slot = indirect_to_ptr ( root - > rnode ) ;
2006-06-23 13:03:22 +04:00
shift = ( height - 1 ) * RADIX_TREE_MAP_SHIFT ;
2005-04-17 02:20:36 +04:00
while ( height > 0 ) {
int offset ;
offset = ( index > > shift ) & RADIX_TREE_MAP_MASK ;
2006-01-08 12:01:41 +03:00
if ( ! tag_get ( slot , tag , offset ) )
tag_set ( slot , tag , offset ) ;
2005-09-07 02:16:46 +04:00
slot = slot - > slots [ offset ] ;
BUG_ON ( slot = = NULL ) ;
2005-04-17 02:20:36 +04:00
shift - = RADIX_TREE_MAP_SHIFT ;
height - - ;
}
2006-06-23 13:03:22 +04:00
/* set the root's tag bit */
if ( slot & & ! root_tag_get ( root , tag ) )
root_tag_set ( root , tag ) ;
2005-09-07 02:16:46 +04:00
return slot ;
2005-04-17 02:20:36 +04:00
}
EXPORT_SYMBOL ( radix_tree_tag_set ) ;
/**
* radix_tree_tag_clear - clear a tag on a radix tree node
* @ root : radix tree root
* @ index : index key
* @ tag : tag index
*
2006-03-25 14:08:05 +03:00
* Clear the search tag ( which must be < RADIX_TREE_MAX_TAGS )
* corresponding to @ index in the radix tree . If
2005-04-17 02:20:36 +04:00
* this causes the leaf node to have no tags set then clear the tag in the
* next - to - leaf node , etc .
*
* Returns the address of the tagged item on success , else NULL . ie :
* has the same return value and semantics as radix_tree_lookup ( ) .
*/
void * radix_tree_tag_clear ( struct radix_tree_root * root ,
2006-03-25 14:08:05 +03:00
unsigned long index , unsigned int tag )
2005-04-17 02:20:36 +04:00
{
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
struct radix_tree_node * node = NULL ;
2006-06-23 13:03:22 +04:00
struct radix_tree_node * slot = NULL ;
2005-04-17 02:20:36 +04:00
unsigned int height , shift ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
int uninitialized_var ( offset ) ;
2005-04-17 02:20:36 +04:00
height = root - > height ;
if ( index > radix_tree_maxindex ( height ) )
goto out ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
shift = height * RADIX_TREE_MAP_SHIFT ;
2010-11-12 01:05:19 +03:00
slot = indirect_to_ptr ( root - > rnode ) ;
2005-04-17 02:20:36 +04:00
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
while ( shift ) {
2005-09-07 02:16:46 +04:00
if ( slot = = NULL )
2005-04-17 02:20:36 +04:00
goto out ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
shift - = RADIX_TREE_MAP_SHIFT ;
2005-04-17 02:20:36 +04:00
offset = ( index > > shift ) & RADIX_TREE_MAP_MASK ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
node = slot ;
2005-09-07 02:16:46 +04:00
slot = slot - > slots [ offset ] ;
2005-04-17 02:20:36 +04:00
}
2006-06-23 13:03:22 +04:00
if ( slot = = NULL )
2005-04-17 02:20:36 +04:00
goto out ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
while ( node ) {
if ( ! tag_get ( node , tag , offset ) )
2006-01-08 12:01:41 +03:00
goto out ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
tag_clear ( node , tag , offset ) ;
if ( any_tag_set ( node , tag ) )
2006-01-08 12:01:40 +03:00
goto out ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
index > > = RADIX_TREE_MAP_SHIFT ;
offset = index & RADIX_TREE_MAP_MASK ;
node = node - > parent ;
2006-06-23 13:03:22 +04:00
}
/* clear the root's tag bit */
if ( root_tag_get ( root , tag ) )
root_tag_clear ( root , tag ) ;
2005-04-17 02:20:36 +04:00
out :
2006-06-23 13:03:22 +04:00
return slot ;
2005-04-17 02:20:36 +04:00
}
EXPORT_SYMBOL ( radix_tree_tag_clear ) ;
/**
2005-09-07 02:16:48 +04:00
* radix_tree_tag_get - get a tag on a radix tree node
* @ root : radix tree root
* @ index : index key
2006-03-25 14:08:05 +03:00
* @ tag : tag index ( < RADIX_TREE_MAX_TAGS )
2005-04-17 02:20:36 +04:00
*
2005-09-07 02:16:48 +04:00
* Return values :
2005-04-17 02:20:36 +04:00
*
2006-06-23 13:03:22 +04:00
* 0 : tag not present or not set
* 1 : tag set
radix_tree_tag_get() is not as safe as the docs make out [ver #2]
radix_tree_tag_get() is not safe to use concurrently with radix_tree_tag_set()
or radix_tree_tag_clear(). The problem is that the double tag_get() in
radix_tree_tag_get():
if (!tag_get(node, tag, offset))
saw_unset_tag = 1;
if (height == 1) {
int ret = tag_get(node, tag, offset);
may see the value change due to the action of set/clear. RCU is no protection
against this as no pointers are being changed, no nodes are being replaced
according to a COW protocol - set/clear alter the node directly.
The documentation in linux/radix-tree.h, however, says that
radix_tree_tag_get() is an exception to the rule that "any function modifying
the tree or tags (...) must exclude other modifications, and exclude any
functions reading the tree".
The problem is that the next statement in radix_tree_tag_get() checks that the
tag doesn't vary over time:
BUG_ON(ret && saw_unset_tag);
This has been seen happening in FS-Cache:
https://www.redhat.com/archives/linux-cachefs/2010-April/msg00013.html
To this end, remove the BUG_ON() from radix_tree_tag_get() and note in various
comments that the value of the tag may change whilst the RCU read lock is held,
and thus that the return value of radix_tree_tag_get() may not be relied upon
unless radix_tree_tag_set/clear() and radix_tree_delete() are excluded from
running concurrently with it.
Reported-by: Romain DEGEZ <romain.degez@smartjog.com>
Signed-off-by: David Howells <dhowells@redhat.com>
Acked-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-04-07 01:36:20 +04:00
*
* Note that the return value of this function may not be relied on , even if
* the RCU lock is held , unless tag modification and node deletion are excluded
* from concurrency .
2005-04-17 02:20:36 +04:00
*/
int radix_tree_tag_get ( struct radix_tree_root * root ,
2006-03-25 14:08:05 +03:00
unsigned long index , unsigned int tag )
2005-04-17 02:20:36 +04:00
{
unsigned int height , shift ;
2006-12-07 07:33:44 +03:00
struct radix_tree_node * node ;
2005-04-17 02:20:36 +04:00
2006-06-23 13:03:22 +04:00
/* check the root's tag bit */
if ( ! root_tag_get ( root , tag ) )
return 0 ;
2010-02-23 04:04:54 +03:00
node = rcu_dereference_raw ( root - > rnode ) ;
2006-12-07 07:33:44 +03:00
if ( node = = NULL )
return 0 ;
2007-10-16 12:24:42 +04:00
if ( ! radix_tree_is_indirect_ptr ( node ) )
2006-12-07 07:33:44 +03:00
return ( index = = 0 ) ;
2010-11-12 01:05:19 +03:00
node = indirect_to_ptr ( node ) ;
2006-12-07 07:33:44 +03:00
mm: keep page cache radix tree nodes in check
Previously, page cache radix tree nodes were freed after reclaim emptied
out their page pointers. But now reclaim stores shadow entries in their
place, which are only reclaimed when the inodes themselves are
reclaimed. This is problematic for bigger files that are still in use
after they have a significant amount of their cache reclaimed, without
any of those pages actually refaulting. The shadow entries will just
sit there and waste memory. In the worst case, the shadow entries will
accumulate until the machine runs out of memory.
To get this under control, the VM will track radix tree nodes
exclusively containing shadow entries on a per-NUMA node list. Per-NUMA
rather than global because we expect the radix tree nodes themselves to
be allocated node-locally and we want to reduce cross-node references of
otherwise independent cache workloads. A simple shrinker will then
reclaim these nodes on memory pressure.
A few things need to be stored in the radix tree node to implement the
shadow node LRU and allow tree deletions coming from the list:
1. There is no index available that would describe the reverse path
from the node up to the tree root, which is needed to perform a
deletion. To solve this, encode in each node its offset inside the
parent. This can be stored in the unused upper bits of the same
member that stores the node's height at no extra space cost.
2. The number of shadow entries needs to be counted in addition to the
regular entries, to quickly detect when the node is ready to go to
the shadow node LRU list. The current entry count is an unsigned
int but the maximum number of entries is 64, so a shadow counter
can easily be stored in the unused upper bits.
3. Tree modification needs tree lock and tree root, which are located
in the address space, so store an address_space backpointer in the
node. The parent pointer of the node is in a union with the 2-word
rcu_head, so the backpointer comes at no extra cost as well.
4. The node needs to be linked to an LRU list, which requires a list
head inside the node. This does increase the size of the node, but
it does not change the number of objects that fit into a slab page.
[akpm@linux-foundation.org: export the right function]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-04 01:47:56 +04:00
height = node - > path & RADIX_TREE_HEIGHT_MASK ;
2006-12-07 07:33:44 +03:00
if ( index > radix_tree_maxindex ( height ) )
return 0 ;
2006-06-23 13:03:22 +04:00
2005-04-17 02:20:36 +04:00
shift = ( height - 1 ) * RADIX_TREE_MAP_SHIFT ;
for ( ; ; ) {
int offset ;
2006-12-07 07:33:44 +03:00
if ( node = = NULL )
2005-04-17 02:20:36 +04:00
return 0 ;
offset = ( index > > shift ) & RADIX_TREE_MAP_MASK ;
2006-12-07 07:33:44 +03:00
if ( ! tag_get ( node , tag , offset ) )
2011-11-01 04:07:02 +04:00
return 0 ;
radix_tree_tag_get() is not as safe as the docs make out [ver #2]
radix_tree_tag_get() is not safe to use concurrently with radix_tree_tag_set()
or radix_tree_tag_clear(). The problem is that the double tag_get() in
radix_tree_tag_get():
if (!tag_get(node, tag, offset))
saw_unset_tag = 1;
if (height == 1) {
int ret = tag_get(node, tag, offset);
may see the value change due to the action of set/clear. RCU is no protection
against this as no pointers are being changed, no nodes are being replaced
according to a COW protocol - set/clear alter the node directly.
The documentation in linux/radix-tree.h, however, says that
radix_tree_tag_get() is an exception to the rule that "any function modifying
the tree or tags (...) must exclude other modifications, and exclude any
functions reading the tree".
The problem is that the next statement in radix_tree_tag_get() checks that the
tag doesn't vary over time:
BUG_ON(ret && saw_unset_tag);
This has been seen happening in FS-Cache:
https://www.redhat.com/archives/linux-cachefs/2010-April/msg00013.html
To this end, remove the BUG_ON() from radix_tree_tag_get() and note in various
comments that the value of the tag may change whilst the RCU read lock is held,
and thus that the return value of radix_tree_tag_get() may not be relied upon
unless radix_tree_tag_set/clear() and radix_tree_delete() are excluded from
running concurrently with it.
Reported-by: Romain DEGEZ <romain.degez@smartjog.com>
Signed-off-by: David Howells <dhowells@redhat.com>
Acked-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-04-07 01:36:20 +04:00
if ( height = = 1 )
2011-11-01 04:07:02 +04:00
return 1 ;
2010-02-23 04:04:54 +03:00
node = rcu_dereference_raw ( node - > slots [ offset ] ) ;
2005-04-17 02:20:36 +04:00
shift - = RADIX_TREE_MAP_SHIFT ;
height - - ;
}
}
EXPORT_SYMBOL ( radix_tree_tag_get ) ;
2012-03-29 01:42:53 +04:00
/**
* radix_tree_next_chunk - find next chunk of slots for iteration
*
* @ root : radix tree root
* @ iter : iterator state
* @ flags : RADIX_TREE_ITER_ * flags and tag index
* Returns : pointer to chunk first slot , or NULL if iteration is over
*/
void * * radix_tree_next_chunk ( struct radix_tree_root * root ,
struct radix_tree_iter * iter , unsigned flags )
{
unsigned shift , tag = flags & RADIX_TREE_ITER_TAG_MASK ;
struct radix_tree_node * rnode , * node ;
mm: keep page cache radix tree nodes in check
Previously, page cache radix tree nodes were freed after reclaim emptied
out their page pointers. But now reclaim stores shadow entries in their
place, which are only reclaimed when the inodes themselves are
reclaimed. This is problematic for bigger files that are still in use
after they have a significant amount of their cache reclaimed, without
any of those pages actually refaulting. The shadow entries will just
sit there and waste memory. In the worst case, the shadow entries will
accumulate until the machine runs out of memory.
To get this under control, the VM will track radix tree nodes
exclusively containing shadow entries on a per-NUMA node list. Per-NUMA
rather than global because we expect the radix tree nodes themselves to
be allocated node-locally and we want to reduce cross-node references of
otherwise independent cache workloads. A simple shrinker will then
reclaim these nodes on memory pressure.
A few things need to be stored in the radix tree node to implement the
shadow node LRU and allow tree deletions coming from the list:
1. There is no index available that would describe the reverse path
from the node up to the tree root, which is needed to perform a
deletion. To solve this, encode in each node its offset inside the
parent. This can be stored in the unused upper bits of the same
member that stores the node's height at no extra space cost.
2. The number of shadow entries needs to be counted in addition to the
regular entries, to quickly detect when the node is ready to go to
the shadow node LRU list. The current entry count is an unsigned
int but the maximum number of entries is 64, so a shadow counter
can easily be stored in the unused upper bits.
3. Tree modification needs tree lock and tree root, which are located
in the address space, so store an address_space backpointer in the
node. The parent pointer of the node is in a union with the 2-word
rcu_head, so the backpointer comes at no extra cost as well.
4. The node needs to be linked to an LRU list, which requires a list
head inside the node. This does increase the size of the node, but
it does not change the number of objects that fit into a slab page.
[akpm@linux-foundation.org: export the right function]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-04 01:47:56 +04:00
unsigned long index , offset , height ;
2012-03-29 01:42:53 +04:00
if ( ( flags & RADIX_TREE_ITER_TAGGED ) & & ! root_tag_get ( root , tag ) )
return NULL ;
/*
* Catch next_index overflow after ~ 0UL . iter - > index never overflows
* during iterating ; it can be zero only at the beginning .
* And we cannot overflow iter - > next_index in a single step ,
* because RADIX_TREE_MAP_SHIFT < BITS_PER_LONG .
2012-06-05 21:36:33 +04:00
*
* This condition also used by radix_tree_next_slot ( ) to stop
* contiguous iterating , and forbid swithing to the next chunk .
2012-03-29 01:42:53 +04:00
*/
index = iter - > next_index ;
if ( ! index & & iter - > index )
return NULL ;
rnode = rcu_dereference_raw ( root - > rnode ) ;
if ( radix_tree_is_indirect_ptr ( rnode ) ) {
rnode = indirect_to_ptr ( rnode ) ;
} else if ( rnode & & ! index ) {
/* Single-slot tree */
iter - > index = 0 ;
iter - > next_index = 1 ;
iter - > tags = 1 ;
return ( void * * ) & root - > rnode ;
} else
return NULL ;
restart :
mm: keep page cache radix tree nodes in check
Previously, page cache radix tree nodes were freed after reclaim emptied
out their page pointers. But now reclaim stores shadow entries in their
place, which are only reclaimed when the inodes themselves are
reclaimed. This is problematic for bigger files that are still in use
after they have a significant amount of their cache reclaimed, without
any of those pages actually refaulting. The shadow entries will just
sit there and waste memory. In the worst case, the shadow entries will
accumulate until the machine runs out of memory.
To get this under control, the VM will track radix tree nodes
exclusively containing shadow entries on a per-NUMA node list. Per-NUMA
rather than global because we expect the radix tree nodes themselves to
be allocated node-locally and we want to reduce cross-node references of
otherwise independent cache workloads. A simple shrinker will then
reclaim these nodes on memory pressure.
A few things need to be stored in the radix tree node to implement the
shadow node LRU and allow tree deletions coming from the list:
1. There is no index available that would describe the reverse path
from the node up to the tree root, which is needed to perform a
deletion. To solve this, encode in each node its offset inside the
parent. This can be stored in the unused upper bits of the same
member that stores the node's height at no extra space cost.
2. The number of shadow entries needs to be counted in addition to the
regular entries, to quickly detect when the node is ready to go to
the shadow node LRU list. The current entry count is an unsigned
int but the maximum number of entries is 64, so a shadow counter
can easily be stored in the unused upper bits.
3. Tree modification needs tree lock and tree root, which are located
in the address space, so store an address_space backpointer in the
node. The parent pointer of the node is in a union with the 2-word
rcu_head, so the backpointer comes at no extra cost as well.
4. The node needs to be linked to an LRU list, which requires a list
head inside the node. This does increase the size of the node, but
it does not change the number of objects that fit into a slab page.
[akpm@linux-foundation.org: export the right function]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-04 01:47:56 +04:00
height = rnode - > path & RADIX_TREE_HEIGHT_MASK ;
shift = ( height - 1 ) * RADIX_TREE_MAP_SHIFT ;
2012-03-29 01:42:53 +04:00
offset = index > > shift ;
/* Index outside of the tree */
if ( offset > = RADIX_TREE_MAP_SIZE )
return NULL ;
node = rnode ;
while ( 1 ) {
if ( ( flags & RADIX_TREE_ITER_TAGGED ) ?
! test_bit ( offset , node - > tags [ tag ] ) :
! node - > slots [ offset ] ) {
/* Hole detected */
if ( flags & RADIX_TREE_ITER_CONTIG )
return NULL ;
if ( flags & RADIX_TREE_ITER_TAGGED )
offset = radix_tree_find_next_bit (
node - > tags [ tag ] ,
RADIX_TREE_MAP_SIZE ,
offset + 1 ) ;
else
while ( + + offset < RADIX_TREE_MAP_SIZE ) {
if ( node - > slots [ offset ] )
break ;
}
index & = ~ ( ( RADIX_TREE_MAP_SIZE < < shift ) - 1 ) ;
index + = offset < < shift ;
/* Overflow after ~0UL */
if ( ! index )
return NULL ;
if ( offset = = RADIX_TREE_MAP_SIZE )
goto restart ;
}
/* This is leaf-node */
if ( ! shift )
break ;
node = rcu_dereference_raw ( node - > slots [ offset ] ) ;
if ( node = = NULL )
goto restart ;
shift - = RADIX_TREE_MAP_SHIFT ;
offset = ( index > > shift ) & RADIX_TREE_MAP_MASK ;
}
/* Update the iterator state */
iter - > index = index ;
iter - > next_index = ( index | RADIX_TREE_MAP_MASK ) + 1 ;
/* Construct iter->tags bit-mask from node->tags[tag] array */
if ( flags & RADIX_TREE_ITER_TAGGED ) {
unsigned tag_long , tag_bit ;
tag_long = offset / BITS_PER_LONG ;
tag_bit = offset % BITS_PER_LONG ;
iter - > tags = node - > tags [ tag ] [ tag_long ] > > tag_bit ;
/* This never happens if RADIX_TREE_TAG_LONGS == 1 */
if ( tag_long < RADIX_TREE_TAG_LONGS - 1 ) {
/* Pick tags from next element */
if ( tag_bit )
iter - > tags | = node - > tags [ tag ] [ tag_long + 1 ] < <
( BITS_PER_LONG - tag_bit ) ;
/* Clip chunk size, here only BITS_PER_LONG tags */
iter - > next_index = index + BITS_PER_LONG ;
}
}
return node - > slots + offset ;
}
EXPORT_SYMBOL ( radix_tree_next_chunk ) ;
2010-08-10 04:19:11 +04:00
/**
* radix_tree_range_tag_if_tagged - for each item in given range set given
* tag if item has another tag set
* @ root : radix tree root
* @ first_indexp : pointer to a starting index of a range to scan
* @ last_index : last index of a range to scan
* @ nr_to_tag : maximum number items to tag
* @ iftag : tag index to test
* @ settag : tag index to set if tested tag is set
*
* This function scans range of radix tree from first_index to last_index
* ( inclusive ) . For each item in the range if iftag is set , the function sets
* also settag . The function stops either after tagging nr_to_tag items or
* after reaching last_index .
*
2010-08-23 04:33:53 +04:00
* The tags must be set from the leaf level only and propagated back up the
* path to the root . We must do this so that we resolve the full path before
* setting any tags on intermediate nodes . If we set tags as we descend , then
* we can get to the leaf node and find that the index that has the iftag
* set is outside the range we are scanning . This reults in dangling tags and
* can lead to problems with later tag operations ( e . g . livelocks on lookups ) .
*
2010-08-10 04:19:11 +04:00
* The function returns number of leaves where the tag was set and sets
* * first_indexp to the first unscanned index .
2010-08-20 01:13:33 +04:00
* WARNING ! * first_indexp can wrap if last_index is ULONG_MAX . Caller must
* be prepared to handle that .
2010-08-10 04:19:11 +04:00
*/
unsigned long radix_tree_range_tag_if_tagged ( struct radix_tree_root * root ,
unsigned long * first_indexp , unsigned long last_index ,
unsigned long nr_to_tag ,
unsigned int iftag , unsigned int settag )
{
2010-08-23 04:33:53 +04:00
unsigned int height = root - > height ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
struct radix_tree_node * node = NULL ;
2010-08-23 04:33:53 +04:00
struct radix_tree_node * slot ;
unsigned int shift ;
unsigned long tagged = 0 ;
unsigned long index = * first_indexp ;
2010-08-10 04:19:11 +04:00
last_index = min ( last_index , radix_tree_maxindex ( height ) ) ;
if ( index > last_index )
return 0 ;
if ( ! nr_to_tag )
return 0 ;
if ( ! root_tag_get ( root , iftag ) ) {
* first_indexp = last_index + 1 ;
return 0 ;
}
if ( height = = 0 ) {
* first_indexp = last_index + 1 ;
root_tag_set ( root , settag ) ;
return 1 ;
}
shift = ( height - 1 ) * RADIX_TREE_MAP_SHIFT ;
2010-11-12 01:05:19 +03:00
slot = indirect_to_ptr ( root - > rnode ) ;
2010-08-10 04:19:11 +04:00
for ( ; ; ) {
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
unsigned long upindex ;
2010-08-10 04:19:11 +04:00
int offset ;
offset = ( index > > shift ) & RADIX_TREE_MAP_MASK ;
if ( ! slot - > slots [ offset ] )
goto next ;
if ( ! tag_get ( slot , iftag , offset ) )
goto next ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
if ( shift ) {
2010-08-23 04:33:53 +04:00
/* Go down one level */
shift - = RADIX_TREE_MAP_SHIFT ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
node = slot ;
2010-08-23 04:33:53 +04:00
slot = slot - > slots [ offset ] ;
continue ;
}
/* tag the leaf */
tagged + + ;
2010-08-10 04:19:11 +04:00
tag_set ( slot , settag , offset ) ;
2010-08-23 04:33:53 +04:00
/* walk back up the path tagging interior nodes */
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
upindex = index ;
while ( node ) {
upindex > > = RADIX_TREE_MAP_SHIFT ;
offset = upindex & RADIX_TREE_MAP_MASK ;
2010-08-23 04:33:53 +04:00
/* stop if we find a node with the tag already set */
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
if ( tag_get ( node , settag , offset ) )
2010-08-23 04:33:53 +04:00
break ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
tag_set ( node , settag , offset ) ;
node = node - > parent ;
2010-08-10 04:19:11 +04:00
}
2010-08-23 04:33:53 +04:00
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
/*
* Small optimization : now clear that node pointer .
* Since all of this slot ' s ancestors now have the tag set
* from setting it above , we have no further need to walk
* back up the tree setting tags , until we update slot to
* point to another radix_tree_node .
*/
node = NULL ;
2010-08-10 04:19:11 +04:00
next :
/* Go to next item at level determined by 'shift' */
index = ( ( index > > shift ) + 1 ) < < shift ;
2010-08-20 01:13:33 +04:00
/* Overflow can happen when last_index is ~0UL... */
if ( index > last_index | | ! index )
2010-08-10 04:19:11 +04:00
break ;
if ( tagged > = nr_to_tag )
break ;
while ( ( ( index > > shift ) & RADIX_TREE_MAP_MASK ) = = 0 ) {
/*
* We ' ve fully scanned this node . Go up . Because
* last_index is guaranteed to be in the tree , what
* we do below cannot wander astray .
*/
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
slot = slot - > parent ;
2010-08-10 04:19:11 +04:00
shift + = RADIX_TREE_MAP_SHIFT ;
}
}
/*
radix_tree: radix_tree_gang_lookup_tag_slot() may never return
Executed command: fsstress -d /mnt -n 600 -p 850
crash> bt
PID: 7947 TASK: ffff880160546a70 CPU: 0 COMMAND: "fsstress"
#0 [ffff8800dfc07d00] machine_kexec at ffffffff81030db9
#1 [ffff8800dfc07d70] crash_kexec at ffffffff810a7952
#2 [ffff8800dfc07e40] oops_end at ffffffff814aa7c8
#3 [ffff8800dfc07e70] die_nmi at ffffffff814aa969
#4 [ffff8800dfc07ea0] do_nmi_callback at ffffffff8102b07b
#5 [ffff8800dfc07f10] do_nmi at ffffffff814aa514
#6 [ffff8800dfc07f50] nmi at ffffffff814a9d60
[exception RIP: __lookup_tag+100]
RIP: ffffffff812274b4 RSP: ffff88016056b998 RFLAGS: 00000287
RAX: 0000000000000000 RBX: 0000000000000002 RCX: 0000000000000006
RDX: 000000000000001d RSI: ffff88016056bb18 RDI: ffff8800c85366e0
RBP: ffff88016056b9c8 R8: ffff88016056b9e8 R9: 0000000000000000
R10: 000000000000000e R11: ffff8800c8536908 R12: 0000000000000010
R13: 0000000000000040 R14: ffffffffffffffc0 R15: ffff8800c85366e0
ORIG_RAX: ffffffffffffffff CS: 0010 SS: 0018
<NMI exception stack>
#7 [ffff88016056b998] __lookup_tag at ffffffff812274b4
#8 [ffff88016056b9d0] radix_tree_gang_lookup_tag_slot at ffffffff81227605
#9 [ffff88016056ba20] find_get_pages_tag at ffffffff810fc110
#10 [ffff88016056ba80] pagevec_lookup_tag at ffffffff81105e85
#11 [ffff88016056baa0] write_cache_pages at ffffffff81104c47
#12 [ffff88016056bbd0] generic_writepages at ffffffff81105014
#13 [ffff88016056bbe0] do_writepages at ffffffff81105055
#14 [ffff88016056bbf0] __filemap_fdatawrite_range at ffffffff810fb2cb
#15 [ffff88016056bc40] filemap_write_and_wait_range at ffffffff810fb32a
#16 [ffff88016056bc70] generic_file_direct_write at ffffffff810fb3dc
#17 [ffff88016056bce0] __generic_file_aio_write at ffffffff810fcee5
#18 [ffff88016056bda0] generic_file_aio_write at ffffffff810fd085
#19 [ffff88016056bdf0] do_sync_write at ffffffff8114f9ea
#20 [ffff88016056bf00] vfs_write at ffffffff8114fcf8
#21 [ffff88016056bf30] sys_write at ffffffff81150691
#22 [ffff88016056bf80] system_call_fastpath at ffffffff8100c0b2
I think this root cause is the following:
radix_tree_range_tag_if_tagged() always tags the root tag with settag
if the root tag is set with iftag even if there are no iftag tags
in the specified range (Of course, there are some iftag tags
outside the specified range).
===============================================================================
[[[Detailed description]]]
(1) Why cannot radix_tree_gang_lookup_tag_slot() return forever?
__lookup_tag():
- Return with 0.
- Return with the index which is not bigger than the old one as the
input parameter.
Therefore the following "while" repeats forever because the above
conditions cause "ret" not to be updated and the cur_index cannot be
changed into the bigger one.
(So, radix_tree_gang_lookup_tag_slot() cannot return forever.)
radix_tree_gang_lookup_tag_slot():
1178 while (ret < max_items) {
1179 unsigned int slots_found;
1180 unsigned long next_index; /* Index of next search */
1181
1182 if (cur_index > max_index)
1183 break;
1184 slots_found = __lookup_tag(node, results + ret,
1185 cur_index, max_items - ret, &next_index,
tag);
1186 ret += slots_found;
// cannot update ret because slots_found == 0.
// so, this while loops forever.
1187 if (next_index == 0)
1188 break;
1189 cur_index = next_index;
1190 }
(2) Why does __lookup_tag() return with 0 and doesn't update the index?
Assuming the following:
- the one of the slot in radix_tree_node is NULL.
- the one of the tag which corresponds to the slot sets with
PAGECACHE_TAG_TOWRITE or other.
- In a certain height(!=0), the corresponding index is 0.
a) __lookup_tag() notices that the tag is set.
1005 static unsigned int
1006 __lookup_tag(struct radix_tree_node *slot, void ***results, unsigned long index,
1007 unsigned int max_items, unsigned long *next_index, unsigned int tag)
1008 {
1009 unsigned int nr_found = 0;
1010 unsigned int shift, height;
1011
1012 height = slot->height;
1013 if (height == 0)
1014 goto out;
1015 shift = (height-1) * RADIX_TREE_MAP_SHIFT;
1016
1017 while (height > 0) {
1018 unsigned long i = (index >> shift) & RADIX_TREE_MAP_MASK ;
1019
1020 for (;;) {
1021 if (tag_get(slot, tag, i))
1022 break;
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
* the index is not updated yet.
b) __lookup_tag() notices that the slot is NULL.
1023 index &= ~((1UL << shift) - 1);
1024 index += 1UL << shift;
1025 if (index == 0)
1026 goto out; /* 32-bit wraparound */
1027 i++;
1028 if (i == RADIX_TREE_MAP_SIZE)
1029 goto out;
1030 }
1031 height--;
1032 if (height == 0) { /* Bottom level: grab some items */
...
1055 }
1056 shift -= RADIX_TREE_MAP_SHIFT;
1057 slot = rcu_dereference_raw(slot->slots[i]);
1058 if (slot == NULL)
1059 break;
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
c) __lookup_tag() doesn't update the index and return with 0.
1060 }
1061 out:
1062 *next_index = index;
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1063 return nr_found;
1064 }
(3) Why is the slot NULL even if the tag is set?
Because radix_tree_range_tag_if_tagged() always sets the root tag with
PAGECACHE_TAG_TOWRITE if the root tag is set with PAGECACHE_TAG_DIRTY,
even if there is no tag which can be set with PAGECACHE_TAG_TOWRITE
in the specified range (from *first_indexp to last_index). Of course,
some PAGECACHE_TAG_DIRTY nodes must exist outside the specified range.
(radix_tree_range_tag_if_tagged() is called only from tag_pages_for_writeback())
640 unsigned long radix_tree_range_tag_if_tagged(struct radix_tree_root
*root,
641 unsigned long *first_indexp, unsigned long last_index,
642 unsigned long nr_to_tag,
643 unsigned int iftag, unsigned int settag)
644 {
645 unsigned int height = root->height;
646 struct radix_tree_path path[height];
647 struct radix_tree_path *pathp = path;
648 struct radix_tree_node *slot;
649 unsigned int shift;
650 unsigned long tagged = 0;
651 unsigned long index = *first_indexp;
652
653 last_index = min(last_index, radix_tree_maxindex(height));
654 if (index > last_index)
655 return 0;
656 if (!nr_to_tag)
657 return 0;
658 if (!root_tag_get(root, iftag)) {
659 *first_indexp = last_index + 1;
660 return 0;
661 }
662 if (height == 0) {
663 *first_indexp = last_index + 1;
664 root_tag_set(root, settag);
665 return 1;
666 }
...
733 root_tag_set(root, settag);
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
734 *first_indexp = index;
735
736 return tagged;
737 }
As the result, there is no radix_tree_node which is set with
PAGECACHE_TAG_TOWRITE but the root tag(radix_tree_root) is set with
PAGECACHE_TAG_TOWRITE.
[figure: inside radix_tree]
(Please see the figure with typewriter font)
===========================================
[roottag = DIRTY]
| tag=0:NOTHING
tag[0 0 0 1] 1:DIRTY
[x x x +] 2:WRITEBACK
| 3:DIRTY,WRITEBACK
p 4:TOWRITE
<---> 5:DIRTY,TOWRITE ...
specified range (index: 0 to 2)
* There is no DIRTY tag within the specified range.
(But there is a DIRTY tag outside that range.)
| | | | | | | | |
after calling tag_pages_for_writeback()
| | | | | | | | |
v v v v v v v v v
[roottag = DIRTY,TOWRITE]
| p is "page".
tag[0 0 0 1] x is NULL.
[x x x +] +- is a pointer to "page".
|
p
* But TOWRITE tag is set on the root tag.
============================================
After that, radix_tree_extend() via radix_tree_insert() is called
when the page is added.
This function sets the new radix_tree_node with PAGECACHE_TAG_TOWRITE
to succeed the status of the root tag.
246 static int radix_tree_extend(struct radix_tree_root *root, unsigned long
index)
247 {
248 struct radix_tree_node *node;
249 unsigned int height;
250 int tag;
251
252 /* Figure out what the height should be. */
253 height = root->height + 1;
254 while (index > radix_tree_maxindex(height))
255 height++;
256
257 if (root->rnode == NULL) {
258 root->height = height;
259 goto out;
260 }
261
262 do {
263 unsigned int newheight;
264 if (!(node = radix_tree_node_alloc(root)))
265 return -ENOMEM;
266
267 /* Increase the height. */
268 node->slots[0] = radix_tree_indirect_to_ptr(root->rnode);
269
270 /* Propagate the aggregated tag info into the new root */
271 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
272 if (root_tag_get(root, tag))
273 tag_set(node, tag, 0);
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
274 }
===========================================
[roottag = DIRTY,TOWRITE]
| :
tag[0 0 0 1] [0 0 0 0]
[x x x +] [+ x x x]
| |
p p (new page)
| | | | | | | | |
after calling radix_tree_insert
| | | | | | | | |
v v v v v v v v v
[roottag = DIRTY,TOWRITE]
|
tag [5 0 0 0] * DIRTY and TOWRITE tags are
[+ + x x] succeeded to the new node.
| |
tag [0 0 0 1] [0 0 0 0]
[x x x +] [+ x x x]
| |
p p
============================================
After that, the index 3 page is released by remove_from_page_cache().
Then we can make the situation that the tag is set with PAGECACHE_TAG_TOWRITE
and that the slot which corresponds to the tag is NULL.
===========================================
[roottag = DIRTY,TOWRITE]
|
tag [5 0 0 0]
[+ + x x]
| |
tag [0 0 0 1] [0 0 0 0]
[x x x +] [+ x x x]
| |
p p
(remove)
| | | | | | | | |
after calling remove_page_cache
| | | | | | | | |
v v v v v v v v v
[roottag = DIRTY,TOWRITE]
|
tag [4 0 0 0] * Only DIRTY tag is cleared
[x + x x] because no TOWRITE tag is existed
| in the bottom node.
[0 0 0 0]
[+ x x x]
|
p
============================================
To solve this problem
Change to that radix_tree_tag_if_tagged() doesn't tag the root tag
if it doesn't set any tags within the specified range.
Like this.
============================================
640 unsigned long radix_tree_range_tag_if_tagged(struct radix_tree_root
*root,
641 unsigned long *first_indexp, unsigned long last_index,
642 unsigned long nr_to_tag,
643 unsigned int iftag, unsigned int settag)
644 {
650 unsigned long tagged = 0;
...
733 if (tagged)
^^^^^^^^^^^^^^^^^^^^^^^^
734 root_tag_set(root, settag);
735 *first_indexp = index;
736
737 return tagged;
738 }
============================================
Signed-off-by: Toshiyuki Okajima <toshi.okajima@jp.fujitsu.com>
Acked-by: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-26 02:07:32 +03:00
* We need not to tag the root tag if there is no tag which is set with
* settag within the range from * first_indexp to last_index .
2010-08-10 04:19:11 +04:00
*/
radix_tree: radix_tree_gang_lookup_tag_slot() may never return
Executed command: fsstress -d /mnt -n 600 -p 850
crash> bt
PID: 7947 TASK: ffff880160546a70 CPU: 0 COMMAND: "fsstress"
#0 [ffff8800dfc07d00] machine_kexec at ffffffff81030db9
#1 [ffff8800dfc07d70] crash_kexec at ffffffff810a7952
#2 [ffff8800dfc07e40] oops_end at ffffffff814aa7c8
#3 [ffff8800dfc07e70] die_nmi at ffffffff814aa969
#4 [ffff8800dfc07ea0] do_nmi_callback at ffffffff8102b07b
#5 [ffff8800dfc07f10] do_nmi at ffffffff814aa514
#6 [ffff8800dfc07f50] nmi at ffffffff814a9d60
[exception RIP: __lookup_tag+100]
RIP: ffffffff812274b4 RSP: ffff88016056b998 RFLAGS: 00000287
RAX: 0000000000000000 RBX: 0000000000000002 RCX: 0000000000000006
RDX: 000000000000001d RSI: ffff88016056bb18 RDI: ffff8800c85366e0
RBP: ffff88016056b9c8 R8: ffff88016056b9e8 R9: 0000000000000000
R10: 000000000000000e R11: ffff8800c8536908 R12: 0000000000000010
R13: 0000000000000040 R14: ffffffffffffffc0 R15: ffff8800c85366e0
ORIG_RAX: ffffffffffffffff CS: 0010 SS: 0018
<NMI exception stack>
#7 [ffff88016056b998] __lookup_tag at ffffffff812274b4
#8 [ffff88016056b9d0] radix_tree_gang_lookup_tag_slot at ffffffff81227605
#9 [ffff88016056ba20] find_get_pages_tag at ffffffff810fc110
#10 [ffff88016056ba80] pagevec_lookup_tag at ffffffff81105e85
#11 [ffff88016056baa0] write_cache_pages at ffffffff81104c47
#12 [ffff88016056bbd0] generic_writepages at ffffffff81105014
#13 [ffff88016056bbe0] do_writepages at ffffffff81105055
#14 [ffff88016056bbf0] __filemap_fdatawrite_range at ffffffff810fb2cb
#15 [ffff88016056bc40] filemap_write_and_wait_range at ffffffff810fb32a
#16 [ffff88016056bc70] generic_file_direct_write at ffffffff810fb3dc
#17 [ffff88016056bce0] __generic_file_aio_write at ffffffff810fcee5
#18 [ffff88016056bda0] generic_file_aio_write at ffffffff810fd085
#19 [ffff88016056bdf0] do_sync_write at ffffffff8114f9ea
#20 [ffff88016056bf00] vfs_write at ffffffff8114fcf8
#21 [ffff88016056bf30] sys_write at ffffffff81150691
#22 [ffff88016056bf80] system_call_fastpath at ffffffff8100c0b2
I think this root cause is the following:
radix_tree_range_tag_if_tagged() always tags the root tag with settag
if the root tag is set with iftag even if there are no iftag tags
in the specified range (Of course, there are some iftag tags
outside the specified range).
===============================================================================
[[[Detailed description]]]
(1) Why cannot radix_tree_gang_lookup_tag_slot() return forever?
__lookup_tag():
- Return with 0.
- Return with the index which is not bigger than the old one as the
input parameter.
Therefore the following "while" repeats forever because the above
conditions cause "ret" not to be updated and the cur_index cannot be
changed into the bigger one.
(So, radix_tree_gang_lookup_tag_slot() cannot return forever.)
radix_tree_gang_lookup_tag_slot():
1178 while (ret < max_items) {
1179 unsigned int slots_found;
1180 unsigned long next_index; /* Index of next search */
1181
1182 if (cur_index > max_index)
1183 break;
1184 slots_found = __lookup_tag(node, results + ret,
1185 cur_index, max_items - ret, &next_index,
tag);
1186 ret += slots_found;
// cannot update ret because slots_found == 0.
// so, this while loops forever.
1187 if (next_index == 0)
1188 break;
1189 cur_index = next_index;
1190 }
(2) Why does __lookup_tag() return with 0 and doesn't update the index?
Assuming the following:
- the one of the slot in radix_tree_node is NULL.
- the one of the tag which corresponds to the slot sets with
PAGECACHE_TAG_TOWRITE or other.
- In a certain height(!=0), the corresponding index is 0.
a) __lookup_tag() notices that the tag is set.
1005 static unsigned int
1006 __lookup_tag(struct radix_tree_node *slot, void ***results, unsigned long index,
1007 unsigned int max_items, unsigned long *next_index, unsigned int tag)
1008 {
1009 unsigned int nr_found = 0;
1010 unsigned int shift, height;
1011
1012 height = slot->height;
1013 if (height == 0)
1014 goto out;
1015 shift = (height-1) * RADIX_TREE_MAP_SHIFT;
1016
1017 while (height > 0) {
1018 unsigned long i = (index >> shift) & RADIX_TREE_MAP_MASK ;
1019
1020 for (;;) {
1021 if (tag_get(slot, tag, i))
1022 break;
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
* the index is not updated yet.
b) __lookup_tag() notices that the slot is NULL.
1023 index &= ~((1UL << shift) - 1);
1024 index += 1UL << shift;
1025 if (index == 0)
1026 goto out; /* 32-bit wraparound */
1027 i++;
1028 if (i == RADIX_TREE_MAP_SIZE)
1029 goto out;
1030 }
1031 height--;
1032 if (height == 0) { /* Bottom level: grab some items */
...
1055 }
1056 shift -= RADIX_TREE_MAP_SHIFT;
1057 slot = rcu_dereference_raw(slot->slots[i]);
1058 if (slot == NULL)
1059 break;
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
c) __lookup_tag() doesn't update the index and return with 0.
1060 }
1061 out:
1062 *next_index = index;
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1063 return nr_found;
1064 }
(3) Why is the slot NULL even if the tag is set?
Because radix_tree_range_tag_if_tagged() always sets the root tag with
PAGECACHE_TAG_TOWRITE if the root tag is set with PAGECACHE_TAG_DIRTY,
even if there is no tag which can be set with PAGECACHE_TAG_TOWRITE
in the specified range (from *first_indexp to last_index). Of course,
some PAGECACHE_TAG_DIRTY nodes must exist outside the specified range.
(radix_tree_range_tag_if_tagged() is called only from tag_pages_for_writeback())
640 unsigned long radix_tree_range_tag_if_tagged(struct radix_tree_root
*root,
641 unsigned long *first_indexp, unsigned long last_index,
642 unsigned long nr_to_tag,
643 unsigned int iftag, unsigned int settag)
644 {
645 unsigned int height = root->height;
646 struct radix_tree_path path[height];
647 struct radix_tree_path *pathp = path;
648 struct radix_tree_node *slot;
649 unsigned int shift;
650 unsigned long tagged = 0;
651 unsigned long index = *first_indexp;
652
653 last_index = min(last_index, radix_tree_maxindex(height));
654 if (index > last_index)
655 return 0;
656 if (!nr_to_tag)
657 return 0;
658 if (!root_tag_get(root, iftag)) {
659 *first_indexp = last_index + 1;
660 return 0;
661 }
662 if (height == 0) {
663 *first_indexp = last_index + 1;
664 root_tag_set(root, settag);
665 return 1;
666 }
...
733 root_tag_set(root, settag);
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
734 *first_indexp = index;
735
736 return tagged;
737 }
As the result, there is no radix_tree_node which is set with
PAGECACHE_TAG_TOWRITE but the root tag(radix_tree_root) is set with
PAGECACHE_TAG_TOWRITE.
[figure: inside radix_tree]
(Please see the figure with typewriter font)
===========================================
[roottag = DIRTY]
| tag=0:NOTHING
tag[0 0 0 1] 1:DIRTY
[x x x +] 2:WRITEBACK
| 3:DIRTY,WRITEBACK
p 4:TOWRITE
<---> 5:DIRTY,TOWRITE ...
specified range (index: 0 to 2)
* There is no DIRTY tag within the specified range.
(But there is a DIRTY tag outside that range.)
| | | | | | | | |
after calling tag_pages_for_writeback()
| | | | | | | | |
v v v v v v v v v
[roottag = DIRTY,TOWRITE]
| p is "page".
tag[0 0 0 1] x is NULL.
[x x x +] +- is a pointer to "page".
|
p
* But TOWRITE tag is set on the root tag.
============================================
After that, radix_tree_extend() via radix_tree_insert() is called
when the page is added.
This function sets the new radix_tree_node with PAGECACHE_TAG_TOWRITE
to succeed the status of the root tag.
246 static int radix_tree_extend(struct radix_tree_root *root, unsigned long
index)
247 {
248 struct radix_tree_node *node;
249 unsigned int height;
250 int tag;
251
252 /* Figure out what the height should be. */
253 height = root->height + 1;
254 while (index > radix_tree_maxindex(height))
255 height++;
256
257 if (root->rnode == NULL) {
258 root->height = height;
259 goto out;
260 }
261
262 do {
263 unsigned int newheight;
264 if (!(node = radix_tree_node_alloc(root)))
265 return -ENOMEM;
266
267 /* Increase the height. */
268 node->slots[0] = radix_tree_indirect_to_ptr(root->rnode);
269
270 /* Propagate the aggregated tag info into the new root */
271 for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
272 if (root_tag_get(root, tag))
273 tag_set(node, tag, 0);
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
274 }
===========================================
[roottag = DIRTY,TOWRITE]
| :
tag[0 0 0 1] [0 0 0 0]
[x x x +] [+ x x x]
| |
p p (new page)
| | | | | | | | |
after calling radix_tree_insert
| | | | | | | | |
v v v v v v v v v
[roottag = DIRTY,TOWRITE]
|
tag [5 0 0 0] * DIRTY and TOWRITE tags are
[+ + x x] succeeded to the new node.
| |
tag [0 0 0 1] [0 0 0 0]
[x x x +] [+ x x x]
| |
p p
============================================
After that, the index 3 page is released by remove_from_page_cache().
Then we can make the situation that the tag is set with PAGECACHE_TAG_TOWRITE
and that the slot which corresponds to the tag is NULL.
===========================================
[roottag = DIRTY,TOWRITE]
|
tag [5 0 0 0]
[+ + x x]
| |
tag [0 0 0 1] [0 0 0 0]
[x x x +] [+ x x x]
| |
p p
(remove)
| | | | | | | | |
after calling remove_page_cache
| | | | | | | | |
v v v v v v v v v
[roottag = DIRTY,TOWRITE]
|
tag [4 0 0 0] * Only DIRTY tag is cleared
[x + x x] because no TOWRITE tag is existed
| in the bottom node.
[0 0 0 0]
[+ x x x]
|
p
============================================
To solve this problem
Change to that radix_tree_tag_if_tagged() doesn't tag the root tag
if it doesn't set any tags within the specified range.
Like this.
============================================
640 unsigned long radix_tree_range_tag_if_tagged(struct radix_tree_root
*root,
641 unsigned long *first_indexp, unsigned long last_index,
642 unsigned long nr_to_tag,
643 unsigned int iftag, unsigned int settag)
644 {
650 unsigned long tagged = 0;
...
733 if (tagged)
^^^^^^^^^^^^^^^^^^^^^^^^
734 root_tag_set(root, settag);
735 *first_indexp = index;
736
737 return tagged;
738 }
============================================
Signed-off-by: Toshiyuki Okajima <toshi.okajima@jp.fujitsu.com>
Acked-by: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-01-26 02:07:32 +03:00
if ( tagged > 0 )
root_tag_set ( root , settag ) ;
2010-08-10 04:19:11 +04:00
* first_indexp = index ;
return tagged ;
}
EXPORT_SYMBOL ( radix_tree_range_tag_if_tagged ) ;
2005-04-17 02:20:36 +04:00
/**
* radix_tree_gang_lookup - perform multiple lookup on a radix tree
* @ root : radix tree root
* @ results : where the results of the lookup are placed
* @ first_index : start the lookup from this key
* @ max_items : place up to this many items at * results
*
* Performs an index - ascending scan of the tree for present items . Places
* them at * @ results and returns the number of items which were placed at
* * @ results .
*
* The implementation is naive .
2006-12-07 07:33:44 +03:00
*
* Like radix_tree_lookup , radix_tree_gang_lookup may be called under
* rcu_read_lock . In this case , rather than the returned results being
* an atomic snapshot of the tree at a single point in time , the semantics
* of an RCU protected gang lookup are as though multiple radix_tree_lookups
* have been issued in individual locks , and results stored in ' results ' .
2005-04-17 02:20:36 +04:00
*/
unsigned int
radix_tree_gang_lookup ( struct radix_tree_root * root , void * * results ,
unsigned long first_index , unsigned int max_items )
{
2012-03-29 01:42:53 +04:00
struct radix_tree_iter iter ;
void * * slot ;
unsigned int ret = 0 ;
2006-12-07 07:33:44 +03:00
2012-03-29 01:42:53 +04:00
if ( unlikely ( ! max_items ) )
2006-12-07 07:33:44 +03:00
return 0 ;
2005-04-17 02:20:36 +04:00
2012-03-29 01:42:53 +04:00
radix_tree_for_each_slot ( slot , root , & iter , first_index ) {
results [ ret ] = indirect_to_ptr ( rcu_dereference_raw ( * slot ) ) ;
if ( ! results [ ret ] )
continue ;
if ( + + ret = = max_items )
2005-04-17 02:20:36 +04:00
break ;
}
2006-12-07 07:33:44 +03:00
2005-04-17 02:20:36 +04:00
return ret ;
}
EXPORT_SYMBOL ( radix_tree_gang_lookup ) ;
2008-07-26 06:45:29 +04:00
/**
* radix_tree_gang_lookup_slot - perform multiple slot lookup on radix tree
* @ root : radix tree root
* @ results : where the results of the lookup are placed
radix_tree: exceptional entries and indices
A patchset to extend tmpfs to MAX_LFS_FILESIZE by abandoning its
peculiar swap vector, instead keeping a file's swap entries in the same
radix tree as its struct page pointers: thus saving memory, and
simplifying its code and locking.
This patch:
The radix_tree is used by several subsystems for different purposes. A
major use is to store the struct page pointers of a file's pagecache for
memory management. But what if mm wanted to store something other than
page pointers there too?
The low bit of a radix_tree entry is already used to denote an indirect
pointer, for internal use, and the unlikely radix_tree_deref_retry()
case.
Define the next bit as denoting an exceptional entry, and supply inline
functions radix_tree_exception() to return non-0 in either unlikely
case, and radix_tree_exceptional_entry() to return non-0 in the second
case.
If a subsystem already uses radix_tree with that bit set, no problem: it
does not affect internal workings at all, but is defined for the
convenience of those storing well-aligned pointers in the radix_tree.
The radix_tree_gang_lookups have an implicit assumption that the caller
can deduce the offset of each entry returned e.g. by the page->index of
a struct page. But that may not be feasible for some kinds of item to
be stored there.
radix_tree_gang_lookup_slot() allow for an optional indices argument,
output array in which to return those offsets. The same could be added
to other radix_tree_gang_lookups, but for now keep it to the only one
for which we need it.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-08-04 03:21:18 +04:00
* @ indices : where their indices should be placed ( but usually NULL )
2008-07-26 06:45:29 +04:00
* @ first_index : start the lookup from this key
* @ max_items : place up to this many items at * results
*
* Performs an index - ascending scan of the tree for present items . Places
* their slots at * @ results and returns the number of items which were
* placed at * @ results .
*
* The implementation is naive .
*
* Like radix_tree_gang_lookup as far as RCU and locking goes . Slots must
* be dereferenced with radix_tree_deref_slot , and if using only RCU
* protection , radix_tree_deref_slot may fail requiring a retry .
*/
unsigned int
radix_tree: exceptional entries and indices
A patchset to extend tmpfs to MAX_LFS_FILESIZE by abandoning its
peculiar swap vector, instead keeping a file's swap entries in the same
radix tree as its struct page pointers: thus saving memory, and
simplifying its code and locking.
This patch:
The radix_tree is used by several subsystems for different purposes. A
major use is to store the struct page pointers of a file's pagecache for
memory management. But what if mm wanted to store something other than
page pointers there too?
The low bit of a radix_tree entry is already used to denote an indirect
pointer, for internal use, and the unlikely radix_tree_deref_retry()
case.
Define the next bit as denoting an exceptional entry, and supply inline
functions radix_tree_exception() to return non-0 in either unlikely
case, and radix_tree_exceptional_entry() to return non-0 in the second
case.
If a subsystem already uses radix_tree with that bit set, no problem: it
does not affect internal workings at all, but is defined for the
convenience of those storing well-aligned pointers in the radix_tree.
The radix_tree_gang_lookups have an implicit assumption that the caller
can deduce the offset of each entry returned e.g. by the page->index of
a struct page. But that may not be feasible for some kinds of item to
be stored there.
radix_tree_gang_lookup_slot() allow for an optional indices argument,
output array in which to return those offsets. The same could be added
to other radix_tree_gang_lookups, but for now keep it to the only one
for which we need it.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-08-04 03:21:18 +04:00
radix_tree_gang_lookup_slot ( struct radix_tree_root * root ,
void * * * results , unsigned long * indices ,
2008-07-26 06:45:29 +04:00
unsigned long first_index , unsigned int max_items )
{
2012-03-29 01:42:53 +04:00
struct radix_tree_iter iter ;
void * * slot ;
unsigned int ret = 0 ;
2008-07-26 06:45:29 +04:00
2012-03-29 01:42:53 +04:00
if ( unlikely ( ! max_items ) )
2008-07-26 06:45:29 +04:00
return 0 ;
2012-03-29 01:42:53 +04:00
radix_tree_for_each_slot ( slot , root , & iter , first_index ) {
results [ ret ] = slot ;
radix_tree: exceptional entries and indices
A patchset to extend tmpfs to MAX_LFS_FILESIZE by abandoning its
peculiar swap vector, instead keeping a file's swap entries in the same
radix tree as its struct page pointers: thus saving memory, and
simplifying its code and locking.
This patch:
The radix_tree is used by several subsystems for different purposes. A
major use is to store the struct page pointers of a file's pagecache for
memory management. But what if mm wanted to store something other than
page pointers there too?
The low bit of a radix_tree entry is already used to denote an indirect
pointer, for internal use, and the unlikely radix_tree_deref_retry()
case.
Define the next bit as denoting an exceptional entry, and supply inline
functions radix_tree_exception() to return non-0 in either unlikely
case, and radix_tree_exceptional_entry() to return non-0 in the second
case.
If a subsystem already uses radix_tree with that bit set, no problem: it
does not affect internal workings at all, but is defined for the
convenience of those storing well-aligned pointers in the radix_tree.
The radix_tree_gang_lookups have an implicit assumption that the caller
can deduce the offset of each entry returned e.g. by the page->index of
a struct page. But that may not be feasible for some kinds of item to
be stored there.
radix_tree_gang_lookup_slot() allow for an optional indices argument,
output array in which to return those offsets. The same could be added
to other radix_tree_gang_lookups, but for now keep it to the only one
for which we need it.
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-08-04 03:21:18 +04:00
if ( indices )
2012-03-29 01:42:53 +04:00
indices [ ret ] = iter . index ;
if ( + + ret = = max_items )
2008-07-26 06:45:29 +04:00
break ;
}
return ret ;
}
EXPORT_SYMBOL ( radix_tree_gang_lookup_slot ) ;
2005-04-17 02:20:36 +04:00
/**
* radix_tree_gang_lookup_tag - perform multiple lookup on a radix tree
* based on a tag
* @ root : radix tree root
* @ results : where the results of the lookup are placed
* @ first_index : start the lookup from this key
* @ max_items : place up to this many items at * results
2006-03-25 14:08:05 +03:00
* @ tag : the tag index ( < RADIX_TREE_MAX_TAGS )
2005-04-17 02:20:36 +04:00
*
* Performs an index - ascending scan of the tree for present items which
* have the tag indexed by @ tag set . Places the items at * @ results and
* returns the number of items which were placed at * @ results .
*/
unsigned int
radix_tree_gang_lookup_tag ( struct radix_tree_root * root , void * * results ,
2006-03-25 14:08:05 +03:00
unsigned long first_index , unsigned int max_items ,
unsigned int tag )
2005-04-17 02:20:36 +04:00
{
2012-03-29 01:42:53 +04:00
struct radix_tree_iter iter ;
void * * slot ;
unsigned int ret = 0 ;
2006-06-23 13:03:22 +04:00
2012-03-29 01:42:53 +04:00
if ( unlikely ( ! max_items ) )
2006-12-07 07:33:44 +03:00
return 0 ;
2012-03-29 01:42:53 +04:00
radix_tree_for_each_tagged ( slot , root , & iter , first_index , tag ) {
results [ ret ] = indirect_to_ptr ( rcu_dereference_raw ( * slot ) ) ;
if ( ! results [ ret ] )
continue ;
if ( + + ret = = max_items )
2005-04-17 02:20:36 +04:00
break ;
}
2006-12-07 07:33:44 +03:00
2005-04-17 02:20:36 +04:00
return ret ;
}
EXPORT_SYMBOL ( radix_tree_gang_lookup_tag ) ;
2008-07-26 06:45:29 +04:00
/**
* radix_tree_gang_lookup_tag_slot - perform multiple slot lookup on a
* radix tree based on a tag
* @ root : radix tree root
* @ results : where the results of the lookup are placed
* @ first_index : start the lookup from this key
* @ max_items : place up to this many items at * results
* @ tag : the tag index ( < RADIX_TREE_MAX_TAGS )
*
* Performs an index - ascending scan of the tree for present items which
* have the tag indexed by @ tag set . Places the slots at * @ results and
* returns the number of slots which were placed at * @ results .
*/
unsigned int
radix_tree_gang_lookup_tag_slot ( struct radix_tree_root * root , void * * * results ,
unsigned long first_index , unsigned int max_items ,
unsigned int tag )
{
2012-03-29 01:42:53 +04:00
struct radix_tree_iter iter ;
void * * slot ;
unsigned int ret = 0 ;
2008-07-26 06:45:29 +04:00
2012-03-29 01:42:53 +04:00
if ( unlikely ( ! max_items ) )
2008-07-26 06:45:29 +04:00
return 0 ;
2012-03-29 01:42:53 +04:00
radix_tree_for_each_tagged ( slot , root , & iter , first_index , tag ) {
results [ ret ] = slot ;
if ( + + ret = = max_items )
2008-07-26 06:45:29 +04:00
break ;
}
return ret ;
}
EXPORT_SYMBOL ( radix_tree_gang_lookup_tag_slot ) ;
tmpfs radix_tree: locate_item to speed up swapoff
We have already acknowledged that swapoff of a tmpfs file is slower than
it was before conversion to the generic radix_tree: a little slower
there will be acceptable, if the hotter paths are faster.
But it was a shock to find swapoff of a 500MB file 20 times slower on my
laptop, taking 10 minutes; and at that rate it significantly slows down
my testing.
Now, most of that turned out to be overhead from PROVE_LOCKING and
PROVE_RCU: without those it was only 4 times slower than before; and
more realistic tests on other machines don't fare as badly.
I've tried a number of things to improve it, including tagging the swap
entries, then doing lookup by tag: I'd expected that to halve the time,
but in practice it's erratic, and often counter-productive.
The only change I've so far found to make a consistent improvement, is
to short-circuit the way we go back and forth, gang lookup packing
entries into the array supplied, then shmem scanning that array for the
target entry. Scanning in place doubles the speed, so it's now only
twice as slow as before (or three times slower when the PROVEs are on).
So, add radix_tree_locate_item() as an expedient, once-off,
single-caller hack to do the lookup directly in place. #ifdef it on
CONFIG_SHMEM and CONFIG_SWAP, as much to document its limited
applicability as save space in other configurations. And, sadly,
#include sched.h for cond_resched().
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-08-04 03:21:27 +04:00
# if defined(CONFIG_SHMEM) && defined(CONFIG_SWAP)
# include <linux/sched.h> /* for cond_resched() */
/*
* This linear search is at present only useful to shmem_unuse_inode ( ) .
*/
static unsigned long __locate ( struct radix_tree_node * slot , void * item ,
unsigned long index , unsigned long * found_index )
{
unsigned int shift , height ;
unsigned long i ;
mm: keep page cache radix tree nodes in check
Previously, page cache radix tree nodes were freed after reclaim emptied
out their page pointers. But now reclaim stores shadow entries in their
place, which are only reclaimed when the inodes themselves are
reclaimed. This is problematic for bigger files that are still in use
after they have a significant amount of their cache reclaimed, without
any of those pages actually refaulting. The shadow entries will just
sit there and waste memory. In the worst case, the shadow entries will
accumulate until the machine runs out of memory.
To get this under control, the VM will track radix tree nodes
exclusively containing shadow entries on a per-NUMA node list. Per-NUMA
rather than global because we expect the radix tree nodes themselves to
be allocated node-locally and we want to reduce cross-node references of
otherwise independent cache workloads. A simple shrinker will then
reclaim these nodes on memory pressure.
A few things need to be stored in the radix tree node to implement the
shadow node LRU and allow tree deletions coming from the list:
1. There is no index available that would describe the reverse path
from the node up to the tree root, which is needed to perform a
deletion. To solve this, encode in each node its offset inside the
parent. This can be stored in the unused upper bits of the same
member that stores the node's height at no extra space cost.
2. The number of shadow entries needs to be counted in addition to the
regular entries, to quickly detect when the node is ready to go to
the shadow node LRU list. The current entry count is an unsigned
int but the maximum number of entries is 64, so a shadow counter
can easily be stored in the unused upper bits.
3. Tree modification needs tree lock and tree root, which are located
in the address space, so store an address_space backpointer in the
node. The parent pointer of the node is in a union with the 2-word
rcu_head, so the backpointer comes at no extra cost as well.
4. The node needs to be linked to an LRU list, which requires a list
head inside the node. This does increase the size of the node, but
it does not change the number of objects that fit into a slab page.
[akpm@linux-foundation.org: export the right function]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-04 01:47:56 +04:00
height = slot - > path & RADIX_TREE_HEIGHT_MASK ;
tmpfs radix_tree: locate_item to speed up swapoff
We have already acknowledged that swapoff of a tmpfs file is slower than
it was before conversion to the generic radix_tree: a little slower
there will be acceptable, if the hotter paths are faster.
But it was a shock to find swapoff of a 500MB file 20 times slower on my
laptop, taking 10 minutes; and at that rate it significantly slows down
my testing.
Now, most of that turned out to be overhead from PROVE_LOCKING and
PROVE_RCU: without those it was only 4 times slower than before; and
more realistic tests on other machines don't fare as badly.
I've tried a number of things to improve it, including tagging the swap
entries, then doing lookup by tag: I'd expected that to halve the time,
but in practice it's erratic, and often counter-productive.
The only change I've so far found to make a consistent improvement, is
to short-circuit the way we go back and forth, gang lookup packing
entries into the array supplied, then shmem scanning that array for the
target entry. Scanning in place doubles the speed, so it's now only
twice as slow as before (or three times slower when the PROVEs are on).
So, add radix_tree_locate_item() as an expedient, once-off,
single-caller hack to do the lookup directly in place. #ifdef it on
CONFIG_SHMEM and CONFIG_SWAP, as much to document its limited
applicability as save space in other configurations. And, sadly,
#include sched.h for cond_resched().
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-08-04 03:21:27 +04:00
shift = ( height - 1 ) * RADIX_TREE_MAP_SHIFT ;
for ( ; height > 1 ; height - - ) {
i = ( index > > shift ) & RADIX_TREE_MAP_MASK ;
for ( ; ; ) {
if ( slot - > slots [ i ] ! = NULL )
break ;
index & = ~ ( ( 1UL < < shift ) - 1 ) ;
index + = 1UL < < shift ;
if ( index = = 0 )
goto out ; /* 32-bit wraparound */
i + + ;
if ( i = = RADIX_TREE_MAP_SIZE )
goto out ;
}
shift - = RADIX_TREE_MAP_SHIFT ;
slot = rcu_dereference_raw ( slot - > slots [ i ] ) ;
if ( slot = = NULL )
goto out ;
}
/* Bottom level: check items */
for ( i = 0 ; i < RADIX_TREE_MAP_SIZE ; i + + ) {
if ( slot - > slots [ i ] = = item ) {
* found_index = index + i ;
index = 0 ;
goto out ;
}
}
index + = RADIX_TREE_MAP_SIZE ;
out :
return index ;
}
/**
* radix_tree_locate_item - search through radix tree for item
* @ root : radix tree root
* @ item : item to be found
*
* Returns index where item was found , or - 1 if not found .
* Caller must hold no lock ( since this time - consuming function needs
* to be preemptible ) , and must check afterwards if item is still there .
*/
unsigned long radix_tree_locate_item ( struct radix_tree_root * root , void * item )
{
struct radix_tree_node * node ;
unsigned long max_index ;
unsigned long cur_index = 0 ;
unsigned long found_index = - 1 ;
do {
rcu_read_lock ( ) ;
node = rcu_dereference_raw ( root - > rnode ) ;
if ( ! radix_tree_is_indirect_ptr ( node ) ) {
rcu_read_unlock ( ) ;
if ( node = = item )
found_index = 0 ;
break ;
}
node = indirect_to_ptr ( node ) ;
mm: keep page cache radix tree nodes in check
Previously, page cache radix tree nodes were freed after reclaim emptied
out their page pointers. But now reclaim stores shadow entries in their
place, which are only reclaimed when the inodes themselves are
reclaimed. This is problematic for bigger files that are still in use
after they have a significant amount of their cache reclaimed, without
any of those pages actually refaulting. The shadow entries will just
sit there and waste memory. In the worst case, the shadow entries will
accumulate until the machine runs out of memory.
To get this under control, the VM will track radix tree nodes
exclusively containing shadow entries on a per-NUMA node list. Per-NUMA
rather than global because we expect the radix tree nodes themselves to
be allocated node-locally and we want to reduce cross-node references of
otherwise independent cache workloads. A simple shrinker will then
reclaim these nodes on memory pressure.
A few things need to be stored in the radix tree node to implement the
shadow node LRU and allow tree deletions coming from the list:
1. There is no index available that would describe the reverse path
from the node up to the tree root, which is needed to perform a
deletion. To solve this, encode in each node its offset inside the
parent. This can be stored in the unused upper bits of the same
member that stores the node's height at no extra space cost.
2. The number of shadow entries needs to be counted in addition to the
regular entries, to quickly detect when the node is ready to go to
the shadow node LRU list. The current entry count is an unsigned
int but the maximum number of entries is 64, so a shadow counter
can easily be stored in the unused upper bits.
3. Tree modification needs tree lock and tree root, which are located
in the address space, so store an address_space backpointer in the
node. The parent pointer of the node is in a union with the 2-word
rcu_head, so the backpointer comes at no extra cost as well.
4. The node needs to be linked to an LRU list, which requires a list
head inside the node. This does increase the size of the node, but
it does not change the number of objects that fit into a slab page.
[akpm@linux-foundation.org: export the right function]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-04 01:47:56 +04:00
max_index = radix_tree_maxindex ( node - > path &
RADIX_TREE_HEIGHT_MASK ) ;
2014-03-04 03:38:23 +04:00
if ( cur_index > max_index ) {
rcu_read_unlock ( ) ;
tmpfs radix_tree: locate_item to speed up swapoff
We have already acknowledged that swapoff of a tmpfs file is slower than
it was before conversion to the generic radix_tree: a little slower
there will be acceptable, if the hotter paths are faster.
But it was a shock to find swapoff of a 500MB file 20 times slower on my
laptop, taking 10 minutes; and at that rate it significantly slows down
my testing.
Now, most of that turned out to be overhead from PROVE_LOCKING and
PROVE_RCU: without those it was only 4 times slower than before; and
more realistic tests on other machines don't fare as badly.
I've tried a number of things to improve it, including tagging the swap
entries, then doing lookup by tag: I'd expected that to halve the time,
but in practice it's erratic, and often counter-productive.
The only change I've so far found to make a consistent improvement, is
to short-circuit the way we go back and forth, gang lookup packing
entries into the array supplied, then shmem scanning that array for the
target entry. Scanning in place doubles the speed, so it's now only
twice as slow as before (or three times slower when the PROVEs are on).
So, add radix_tree_locate_item() as an expedient, once-off,
single-caller hack to do the lookup directly in place. #ifdef it on
CONFIG_SHMEM and CONFIG_SWAP, as much to document its limited
applicability as save space in other configurations. And, sadly,
#include sched.h for cond_resched().
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-08-04 03:21:27 +04:00
break ;
2014-03-04 03:38:23 +04:00
}
tmpfs radix_tree: locate_item to speed up swapoff
We have already acknowledged that swapoff of a tmpfs file is slower than
it was before conversion to the generic radix_tree: a little slower
there will be acceptable, if the hotter paths are faster.
But it was a shock to find swapoff of a 500MB file 20 times slower on my
laptop, taking 10 minutes; and at that rate it significantly slows down
my testing.
Now, most of that turned out to be overhead from PROVE_LOCKING and
PROVE_RCU: without those it was only 4 times slower than before; and
more realistic tests on other machines don't fare as badly.
I've tried a number of things to improve it, including tagging the swap
entries, then doing lookup by tag: I'd expected that to halve the time,
but in practice it's erratic, and often counter-productive.
The only change I've so far found to make a consistent improvement, is
to short-circuit the way we go back and forth, gang lookup packing
entries into the array supplied, then shmem scanning that array for the
target entry. Scanning in place doubles the speed, so it's now only
twice as slow as before (or three times slower when the PROVEs are on).
So, add radix_tree_locate_item() as an expedient, once-off,
single-caller hack to do the lookup directly in place. #ifdef it on
CONFIG_SHMEM and CONFIG_SWAP, as much to document its limited
applicability as save space in other configurations. And, sadly,
#include sched.h for cond_resched().
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-08-04 03:21:27 +04:00
cur_index = __locate ( node , item , cur_index , & found_index ) ;
rcu_read_unlock ( ) ;
cond_resched ( ) ;
} while ( cur_index ! = 0 & & cur_index < = max_index ) ;
return found_index ;
}
# else
unsigned long radix_tree_locate_item ( struct radix_tree_root * root , void * item )
{
return - 1 ;
}
# endif /* CONFIG_SHMEM && CONFIG_SWAP */
2008-07-26 06:45:29 +04:00
2006-01-08 12:01:41 +03:00
/**
* radix_tree_shrink - shrink height of a radix tree to minimal
* @ root radix tree root
*/
static inline void radix_tree_shrink ( struct radix_tree_root * root )
{
/* try to shrink tree height */
2007-10-16 12:24:42 +04:00
while ( root - > height > 0 ) {
2006-01-08 12:01:41 +03:00
struct radix_tree_node * to_free = root - > rnode ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
struct radix_tree_node * slot ;
2006-01-08 12:01:41 +03:00
2007-10-16 12:24:42 +04:00
BUG_ON ( ! radix_tree_is_indirect_ptr ( to_free ) ) ;
2010-11-12 01:05:19 +03:00
to_free = indirect_to_ptr ( to_free ) ;
2007-10-16 12:24:42 +04:00
/*
* The candidate node has more than one child , or its child
* is not at the leftmost slot , we cannot shrink .
*/
if ( to_free - > count ! = 1 )
break ;
if ( ! to_free - > slots [ 0 ] )
break ;
2006-12-07 07:33:44 +03:00
/*
* We don ' t need rcu_assign_pointer ( ) , since we are simply
2010-11-12 01:05:19 +03:00
* moving the node from one part of the tree to another : if it
* was safe to dereference the old pointer to it
2006-12-07 07:33:44 +03:00
* ( to_free - > slots [ 0 ] ) , it will be safe to dereference the new
2010-11-12 01:05:19 +03:00
* one ( root - > rnode ) as far as dependent read barriers go .
2006-12-07 07:33:44 +03:00
*/
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
slot = to_free - > slots [ 0 ] ;
if ( root - > height > 1 ) {
slot - > parent = NULL ;
slot = ptr_to_indirect ( slot ) ;
}
root - > rnode = slot ;
2006-01-08 12:01:41 +03:00
root - > height - - ;
2010-11-12 01:05:19 +03:00
/*
* We have a dilemma here . The node ' s slot [ 0 ] must not be
* NULLed in case there are concurrent lookups expecting to
* find the item . However if this was a bottom - level node ,
* then it may be subject to the slot pointer being visible
* to callers dereferencing it . If item corresponding to
* slot [ 0 ] is subsequently deleted , these callers would expect
* their slot to become empty sooner or later .
*
* For example , lockless pagecache will look up a slot , deref
* the page pointer , and if the page is 0 refcount it means it
* was concurrently deleted from pagecache so try the deref
* again . Fortunately there is already a requirement for logic
* to retry the entire slot lookup - - the indirect pointer
* problem ( replacing direct root node with an indirect pointer
* also results in a stale slot ) . So tag the slot as indirect
* to force callers to retry .
*/
if ( root - > height = = 0 )
* ( ( unsigned long * ) & to_free - > slots [ 0 ] ) | =
RADIX_TREE_INDIRECT_PTR ;
2006-01-08 12:01:41 +03:00
radix_tree_node_free ( to_free ) ;
}
}
2014-04-04 01:47:54 +04:00
/**
* __radix_tree_delete_node - try to free node after clearing a slot
* @ root : radix tree root
* @ node : node containing @ index
*
* After clearing the slot at @ index in @ node from radix tree
* rooted at @ root , call this function to attempt freeing the
* node and shrinking the tree .
*
* Returns % true if @ node was freed , % false otherwise .
*/
mm: keep page cache radix tree nodes in check
Previously, page cache radix tree nodes were freed after reclaim emptied
out their page pointers. But now reclaim stores shadow entries in their
place, which are only reclaimed when the inodes themselves are
reclaimed. This is problematic for bigger files that are still in use
after they have a significant amount of their cache reclaimed, without
any of those pages actually refaulting. The shadow entries will just
sit there and waste memory. In the worst case, the shadow entries will
accumulate until the machine runs out of memory.
To get this under control, the VM will track radix tree nodes
exclusively containing shadow entries on a per-NUMA node list. Per-NUMA
rather than global because we expect the radix tree nodes themselves to
be allocated node-locally and we want to reduce cross-node references of
otherwise independent cache workloads. A simple shrinker will then
reclaim these nodes on memory pressure.
A few things need to be stored in the radix tree node to implement the
shadow node LRU and allow tree deletions coming from the list:
1. There is no index available that would describe the reverse path
from the node up to the tree root, which is needed to perform a
deletion. To solve this, encode in each node its offset inside the
parent. This can be stored in the unused upper bits of the same
member that stores the node's height at no extra space cost.
2. The number of shadow entries needs to be counted in addition to the
regular entries, to quickly detect when the node is ready to go to
the shadow node LRU list. The current entry count is an unsigned
int but the maximum number of entries is 64, so a shadow counter
can easily be stored in the unused upper bits.
3. Tree modification needs tree lock and tree root, which are located
in the address space, so store an address_space backpointer in the
node. The parent pointer of the node is in a union with the 2-word
rcu_head, so the backpointer comes at no extra cost as well.
4. The node needs to be linked to an LRU list, which requires a list
head inside the node. This does increase the size of the node, but
it does not change the number of objects that fit into a slab page.
[akpm@linux-foundation.org: export the right function]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-04 01:47:56 +04:00
bool __radix_tree_delete_node ( struct radix_tree_root * root ,
2014-04-04 01:47:54 +04:00
struct radix_tree_node * node )
{
bool deleted = false ;
do {
struct radix_tree_node * parent ;
if ( node - > count ) {
if ( node = = indirect_to_ptr ( root - > rnode ) ) {
radix_tree_shrink ( root ) ;
if ( root - > height = = 0 )
deleted = true ;
}
return deleted ;
}
parent = node - > parent ;
if ( parent ) {
mm: keep page cache radix tree nodes in check
Previously, page cache radix tree nodes were freed after reclaim emptied
out their page pointers. But now reclaim stores shadow entries in their
place, which are only reclaimed when the inodes themselves are
reclaimed. This is problematic for bigger files that are still in use
after they have a significant amount of their cache reclaimed, without
any of those pages actually refaulting. The shadow entries will just
sit there and waste memory. In the worst case, the shadow entries will
accumulate until the machine runs out of memory.
To get this under control, the VM will track radix tree nodes
exclusively containing shadow entries on a per-NUMA node list. Per-NUMA
rather than global because we expect the radix tree nodes themselves to
be allocated node-locally and we want to reduce cross-node references of
otherwise independent cache workloads. A simple shrinker will then
reclaim these nodes on memory pressure.
A few things need to be stored in the radix tree node to implement the
shadow node LRU and allow tree deletions coming from the list:
1. There is no index available that would describe the reverse path
from the node up to the tree root, which is needed to perform a
deletion. To solve this, encode in each node its offset inside the
parent. This can be stored in the unused upper bits of the same
member that stores the node's height at no extra space cost.
2. The number of shadow entries needs to be counted in addition to the
regular entries, to quickly detect when the node is ready to go to
the shadow node LRU list. The current entry count is an unsigned
int but the maximum number of entries is 64, so a shadow counter
can easily be stored in the unused upper bits.
3. Tree modification needs tree lock and tree root, which are located
in the address space, so store an address_space backpointer in the
node. The parent pointer of the node is in a union with the 2-word
rcu_head, so the backpointer comes at no extra cost as well.
4. The node needs to be linked to an LRU list, which requires a list
head inside the node. This does increase the size of the node, but
it does not change the number of objects that fit into a slab page.
[akpm@linux-foundation.org: export the right function]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-04 01:47:56 +04:00
unsigned int offset ;
2014-04-04 01:47:54 +04:00
mm: keep page cache radix tree nodes in check
Previously, page cache radix tree nodes were freed after reclaim emptied
out their page pointers. But now reclaim stores shadow entries in their
place, which are only reclaimed when the inodes themselves are
reclaimed. This is problematic for bigger files that are still in use
after they have a significant amount of their cache reclaimed, without
any of those pages actually refaulting. The shadow entries will just
sit there and waste memory. In the worst case, the shadow entries will
accumulate until the machine runs out of memory.
To get this under control, the VM will track radix tree nodes
exclusively containing shadow entries on a per-NUMA node list. Per-NUMA
rather than global because we expect the radix tree nodes themselves to
be allocated node-locally and we want to reduce cross-node references of
otherwise independent cache workloads. A simple shrinker will then
reclaim these nodes on memory pressure.
A few things need to be stored in the radix tree node to implement the
shadow node LRU and allow tree deletions coming from the list:
1. There is no index available that would describe the reverse path
from the node up to the tree root, which is needed to perform a
deletion. To solve this, encode in each node its offset inside the
parent. This can be stored in the unused upper bits of the same
member that stores the node's height at no extra space cost.
2. The number of shadow entries needs to be counted in addition to the
regular entries, to quickly detect when the node is ready to go to
the shadow node LRU list. The current entry count is an unsigned
int but the maximum number of entries is 64, so a shadow counter
can easily be stored in the unused upper bits.
3. Tree modification needs tree lock and tree root, which are located
in the address space, so store an address_space backpointer in the
node. The parent pointer of the node is in a union with the 2-word
rcu_head, so the backpointer comes at no extra cost as well.
4. The node needs to be linked to an LRU list, which requires a list
head inside the node. This does increase the size of the node, but
it does not change the number of objects that fit into a slab page.
[akpm@linux-foundation.org: export the right function]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-04 01:47:56 +04:00
offset = node - > path > > RADIX_TREE_HEIGHT_SHIFT ;
parent - > slots [ offset ] = NULL ;
2014-04-04 01:47:54 +04:00
parent - > count - - ;
} else {
root_tag_clear_all ( root ) ;
root - > height = 0 ;
root - > rnode = NULL ;
}
radix_tree_node_free ( node ) ;
deleted = true ;
node = parent ;
} while ( node ) ;
return deleted ;
}
2005-04-17 02:20:36 +04:00
/**
2014-04-04 01:47:39 +04:00
* radix_tree_delete_item - delete an item from a radix tree
2005-04-17 02:20:36 +04:00
* @ root : radix tree root
* @ index : index key
2014-04-04 01:47:39 +04:00
* @ item : expected item
2005-04-17 02:20:36 +04:00
*
2014-04-04 01:47:39 +04:00
* Remove @ item at @ index from the radix tree rooted at @ root .
2005-04-17 02:20:36 +04:00
*
2014-04-04 01:47:39 +04:00
* Returns the address of the deleted item , or NULL if it was not present
* or the entry at the given @ index was not @ item .
2005-04-17 02:20:36 +04:00
*/
2014-04-04 01:47:39 +04:00
void * radix_tree_delete_item ( struct radix_tree_root * root ,
unsigned long index , void * item )
2005-04-17 02:20:36 +04:00
{
2014-04-04 01:47:54 +04:00
struct radix_tree_node * node ;
unsigned int offset ;
void * * slot ;
void * entry ;
2006-01-08 12:01:41 +03:00
int tag ;
2005-04-17 02:20:36 +04:00
2014-04-04 01:47:54 +04:00
entry = __radix_tree_lookup ( root , index , & node , & slot ) ;
if ( ! entry )
return NULL ;
2005-04-17 02:20:36 +04:00
2014-04-04 01:47:54 +04:00
if ( item & & entry ! = item )
return NULL ;
if ( ! node ) {
2006-06-23 13:03:22 +04:00
root_tag_clear_all ( root ) ;
root - > rnode = NULL ;
2014-04-04 01:47:54 +04:00
return entry ;
2006-06-23 13:03:22 +04:00
}
2005-04-17 02:20:36 +04:00
2014-04-04 01:47:54 +04:00
offset = index & RADIX_TREE_MAP_MASK ;
2014-04-04 01:47:39 +04:00
2005-04-17 02:20:36 +04:00
/*
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
* Clear all tags associated with the item to be deleted .
* This way of doing it would be inefficient , but seldom is any set .
2005-04-17 02:20:36 +04:00
*/
2006-03-25 14:08:05 +03:00
for ( tag = 0 ; tag < RADIX_TREE_MAX_TAGS ; tag + + ) {
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
if ( tag_get ( node , tag , offset ) )
2006-06-23 13:03:22 +04:00
radix_tree_tag_clear ( root , index , tag ) ;
2006-01-08 12:01:41 +03:00
}
2005-04-17 02:20:36 +04:00
2014-04-04 01:47:54 +04:00
node - > slots [ offset ] = NULL ;
node - > count - - ;
radix_tree: take radix_tree_path off stack
Down, down in the deepest depths of GFP_NOIO page reclaim, we have
shrink_page_list() calling __remove_mapping() calling __delete_from_
swap_cache() or __delete_from_page_cache().
You would not expect those to need much stack, but in fact they call
radix_tree_delete(): which declares a 192-byte radix_tree_path array on
its stack (to record the node,offsets it visits when descending, in case
it needs to ascend to update them). And if any tag is still set [1],
that calls radix_tree_tag_clear(), which declares a further such
192-byte radix_tree_path array on the stack. (At least we have
interrupts disabled here, so won't then be pushing registers too.)
That was probably a good choice when most users were 32-bit (array of
half the size), and adding fields to radix_tree_node would have bloated
it unnecessarily. But nowadays many are 64-bit, and each
radix_tree_node contains a struct rcu_head, which is only used when
freeing; whereas the radix_tree_path info is only used for updating the
tree (deleting, clearing tags or setting tags if tagged) when a lock
must be held, of no interest when accessing the tree locklessly.
So add a parent pointer to the radix_tree_node, in union with the
rcu_head, and remove all uses of the radix_tree_path. There would be
space in that union to save the offset when descending as before (we can
argue that a lock must already be held to exclude other users), but
recalculating it when ascending is both easy (a constant shift and a
constant mask) and uncommon, so it seems better just to do that.
Two little optimizations: no need to decrement height when descending,
adjusting shift is enough; and once radix_tree_tag_if_tagged() has set
tag on a node and its ancestors, it need not ascend from that node
again.
perf on the radix tree test harness reports radix_tree_insert() as 2%
slower (now having to set parent), but radix_tree_delete() 24% faster.
Surely that's an exaggeration from rtth's artificially low map shift 3,
but forcing it back to 6 still rates radix_tree_delete() 8% faster.
[1] Can a pagecache tag (dirty, writeback or towrite) actually still be
set at the time of radix_tree_delete()? Perhaps not if the filesystem is
well-behaved. But although I've not tracked any stack overflow down to
this cause, I have observed a curious case in which a dirty tag is set
and left set on tmpfs: page migration's migrate_page_copy() happens to
use __set_page_dirty_nobuffers() to set PageDirty on the newpage, and
that sets PAGECACHE_TAG_DIRTY as a side-effect - harmless to a
filesystem which doesn't use tags, except for this stack depth issue.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Nai Xia <nai.xia@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-13 05:20:41 +04:00
mm: keep page cache radix tree nodes in check
Previously, page cache radix tree nodes were freed after reclaim emptied
out their page pointers. But now reclaim stores shadow entries in their
place, which are only reclaimed when the inodes themselves are
reclaimed. This is problematic for bigger files that are still in use
after they have a significant amount of their cache reclaimed, without
any of those pages actually refaulting. The shadow entries will just
sit there and waste memory. In the worst case, the shadow entries will
accumulate until the machine runs out of memory.
To get this under control, the VM will track radix tree nodes
exclusively containing shadow entries on a per-NUMA node list. Per-NUMA
rather than global because we expect the radix tree nodes themselves to
be allocated node-locally and we want to reduce cross-node references of
otherwise independent cache workloads. A simple shrinker will then
reclaim these nodes on memory pressure.
A few things need to be stored in the radix tree node to implement the
shadow node LRU and allow tree deletions coming from the list:
1. There is no index available that would describe the reverse path
from the node up to the tree root, which is needed to perform a
deletion. To solve this, encode in each node its offset inside the
parent. This can be stored in the unused upper bits of the same
member that stores the node's height at no extra space cost.
2. The number of shadow entries needs to be counted in addition to the
regular entries, to quickly detect when the node is ready to go to
the shadow node LRU list. The current entry count is an unsigned
int but the maximum number of entries is 64, so a shadow counter
can easily be stored in the unused upper bits.
3. Tree modification needs tree lock and tree root, which are located
in the address space, so store an address_space backpointer in the
node. The parent pointer of the node is in a union with the 2-word
rcu_head, so the backpointer comes at no extra cost as well.
4. The node needs to be linked to an LRU list, which requires a list
head inside the node. This does increase the size of the node, but
it does not change the number of objects that fit into a slab page.
[akpm@linux-foundation.org: export the right function]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-04 01:47:56 +04:00
__radix_tree_delete_node ( root , node ) ;
2006-06-23 13:03:22 +04:00
2014-04-04 01:47:54 +04:00
return entry ;
2005-04-17 02:20:36 +04:00
}
2014-04-04 01:47:39 +04:00
EXPORT_SYMBOL ( radix_tree_delete_item ) ;
/**
* radix_tree_delete - delete an item from a radix tree
* @ root : radix tree root
* @ index : index key
*
* Remove the item at @ index from the radix tree rooted at @ root .
*
* Returns the address of the deleted item , or NULL if it was not present .
*/
void * radix_tree_delete ( struct radix_tree_root * root , unsigned long index )
{
return radix_tree_delete_item ( root , index , NULL ) ;
}
2005-04-17 02:20:36 +04:00
EXPORT_SYMBOL ( radix_tree_delete ) ;
/**
* radix_tree_tagged - test whether any items in the tree are tagged
* @ root : radix tree root
* @ tag : tag to test
*/
2006-03-25 14:08:05 +03:00
int radix_tree_tagged ( struct radix_tree_root * root , unsigned int tag )
2005-04-17 02:20:36 +04:00
{
2006-06-23 13:03:22 +04:00
return root_tag_get ( root , tag ) ;
2005-04-17 02:20:36 +04:00
}
EXPORT_SYMBOL ( radix_tree_tagged ) ;
static void
mm: keep page cache radix tree nodes in check
Previously, page cache radix tree nodes were freed after reclaim emptied
out their page pointers. But now reclaim stores shadow entries in their
place, which are only reclaimed when the inodes themselves are
reclaimed. This is problematic for bigger files that are still in use
after they have a significant amount of their cache reclaimed, without
any of those pages actually refaulting. The shadow entries will just
sit there and waste memory. In the worst case, the shadow entries will
accumulate until the machine runs out of memory.
To get this under control, the VM will track radix tree nodes
exclusively containing shadow entries on a per-NUMA node list. Per-NUMA
rather than global because we expect the radix tree nodes themselves to
be allocated node-locally and we want to reduce cross-node references of
otherwise independent cache workloads. A simple shrinker will then
reclaim these nodes on memory pressure.
A few things need to be stored in the radix tree node to implement the
shadow node LRU and allow tree deletions coming from the list:
1. There is no index available that would describe the reverse path
from the node up to the tree root, which is needed to perform a
deletion. To solve this, encode in each node its offset inside the
parent. This can be stored in the unused upper bits of the same
member that stores the node's height at no extra space cost.
2. The number of shadow entries needs to be counted in addition to the
regular entries, to quickly detect when the node is ready to go to
the shadow node LRU list. The current entry count is an unsigned
int but the maximum number of entries is 64, so a shadow counter
can easily be stored in the unused upper bits.
3. Tree modification needs tree lock and tree root, which are located
in the address space, so store an address_space backpointer in the
node. The parent pointer of the node is in a union with the 2-word
rcu_head, so the backpointer comes at no extra cost as well.
4. The node needs to be linked to an LRU list, which requires a list
head inside the node. This does increase the size of the node, but
it does not change the number of objects that fit into a slab page.
[akpm@linux-foundation.org: export the right function]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-04 01:47:56 +04:00
radix_tree_node_ctor ( void * arg )
2005-04-17 02:20:36 +04:00
{
mm: keep page cache radix tree nodes in check
Previously, page cache radix tree nodes were freed after reclaim emptied
out their page pointers. But now reclaim stores shadow entries in their
place, which are only reclaimed when the inodes themselves are
reclaimed. This is problematic for bigger files that are still in use
after they have a significant amount of their cache reclaimed, without
any of those pages actually refaulting. The shadow entries will just
sit there and waste memory. In the worst case, the shadow entries will
accumulate until the machine runs out of memory.
To get this under control, the VM will track radix tree nodes
exclusively containing shadow entries on a per-NUMA node list. Per-NUMA
rather than global because we expect the radix tree nodes themselves to
be allocated node-locally and we want to reduce cross-node references of
otherwise independent cache workloads. A simple shrinker will then
reclaim these nodes on memory pressure.
A few things need to be stored in the radix tree node to implement the
shadow node LRU and allow tree deletions coming from the list:
1. There is no index available that would describe the reverse path
from the node up to the tree root, which is needed to perform a
deletion. To solve this, encode in each node its offset inside the
parent. This can be stored in the unused upper bits of the same
member that stores the node's height at no extra space cost.
2. The number of shadow entries needs to be counted in addition to the
regular entries, to quickly detect when the node is ready to go to
the shadow node LRU list. The current entry count is an unsigned
int but the maximum number of entries is 64, so a shadow counter
can easily be stored in the unused upper bits.
3. Tree modification needs tree lock and tree root, which are located
in the address space, so store an address_space backpointer in the
node. The parent pointer of the node is in a union with the 2-word
rcu_head, so the backpointer comes at no extra cost as well.
4. The node needs to be linked to an LRU list, which requires a list
head inside the node. This does increase the size of the node, but
it does not change the number of objects that fit into a slab page.
[akpm@linux-foundation.org: export the right function]
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Rik van Riel <riel@redhat.com>
Reviewed-by: Minchan Kim <minchan@kernel.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Bob Liu <bob.liu@oracle.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jan Kara <jack@suse.cz>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Luigi Semenzato <semenzato@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Metin Doslu <metin@citusdata.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: Ozgun Erdogan <ozgun@citusdata.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin <klamm@yandex-team.ru>
Cc: Ryan Mallon <rmallon@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-04 01:47:56 +04:00
struct radix_tree_node * node = arg ;
memset ( node , 0 , sizeof ( * node ) ) ;
INIT_LIST_HEAD ( & node - > private_list ) ;
2005-04-17 02:20:36 +04:00
}
static __init unsigned long __maxindex ( unsigned int height )
{
2007-10-17 10:29:35 +04:00
unsigned int width = height * RADIX_TREE_MAP_SHIFT ;
int shift = RADIX_TREE_INDEX_BITS - width ;
if ( shift < 0 )
return ~ 0UL ;
if ( shift > = BITS_PER_LONG )
return 0UL ;
return ~ 0UL > > shift ;
2005-04-17 02:20:36 +04:00
}
static __init void radix_tree_init_maxindex ( void )
{
unsigned int i ;
for ( i = 0 ; i < ARRAY_SIZE ( height_to_maxindex ) ; i + + )
height_to_maxindex [ i ] = __maxindex ( i ) ;
}
static int radix_tree_callback ( struct notifier_block * nfb ,
unsigned long action ,
void * hcpu )
{
int cpu = ( long ) hcpu ;
struct radix_tree_preload * rtp ;
2015-06-26 01:02:19 +03:00
struct radix_tree_node * node ;
2005-04-17 02:20:36 +04:00
/* Free per-cpu pool of perloaded nodes */
2007-05-09 13:35:10 +04:00
if ( action = = CPU_DEAD | | action = = CPU_DEAD_FROZEN ) {
2005-04-17 02:20:36 +04:00
rtp = & per_cpu ( radix_tree_preloads , cpu ) ;
while ( rtp - > nr ) {
2015-06-26 01:02:19 +03:00
node = rtp - > nodes ;
rtp - > nodes = node - > private_data ;
kmem_cache_free ( radix_tree_node_cachep , node ) ;
rtp - > nr - - ;
2005-04-17 02:20:36 +04:00
}
}
return NOTIFY_OK ;
}
void __init radix_tree_init ( void )
{
radix_tree_node_cachep = kmem_cache_create ( " radix_tree_node " ,
sizeof ( struct radix_tree_node ) , 0 ,
2008-04-28 13:12:05 +04:00
SLAB_PANIC | SLAB_RECLAIM_ACCOUNT ,
radix_tree_node_ctor ) ;
2005-04-17 02:20:36 +04:00
radix_tree_init_maxindex ( ) ;
hotcpu_notifier ( radix_tree_callback , 0 ) ;
}