I'm running a shmem pagefault test case (see attached file) under a 64 CPU
system. Profile shows shmem_inode_info->lock is heavily contented and
100% CPUs time are trying to get the lock. In the pagefault (no swap)
case, shmem_getpage gets the lock twice, the last one is avoidable if we
prealloc a page so we could reduce one time of locking. This is what
below patch does.
The result of the test case:
2.6.35-rc3: ~20s
2.6.35-rc3 + patch: ~12s
so this is 40% improvement.
One might argue if we could have better locking for shmem. But even shmem
is lockless, the pagefault will soon have pagecache lock heavily contented
because shmem must add new page to pagecache. So before we have better
locking for pagecache, improving shmem locking doesn't have too much
improvement. I did a similar pagefault test against a ramfs file, the
test result is ~10.5s.
[akpm@linux-foundation.org: fix comment, clean up code layout, elimintate code duplication]
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: "Zhang, Yanmin" <yanmin.zhang@intel.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The current implementation of tmpfs is not scalable. We found that
stat_lock is contended by multiple threads when we need to get a new page,
leading to useless spinning inside this spin lock.
This patch makes use of the percpu_counter library to maintain local count
of used blocks to speed up getting and returning of pages. So the
acquisition of stat_lock is unnecessary for getting and returning blocks,
improving the performance of tmpfs on system with large number of cpus.
On a 4 socket 32 core NHM-EX system, we saw improvement of 270%.
The implementation below has a slight chance of race between threads
causing a slight overshoot of the maximum configured blocks. However, any
overshoot is small, and is bounded by the number of cpus. This happens
when the number of used blocks is slightly below the maximum configured
blocks when a thread checks the used block count, and another thread
allocates the last block before the current thread does. This should not
be a problem for tmpfs, as the overshoot is most likely to be a few blocks
and bounded. If a strict limit is really desired, then configured the max
blocks to be the limit less the number of cpus in system.
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Add percpu_counter_compare that allows for a quick but accurate comparison
of percpu_counter with a given value.
A rough count is provided by the count field in percpu_counter structure,
without accounting for the other values stored in individual cpu counters.
The actual count is a sum of count and the cpu counters. However, count
field is never different from the actual value by a factor of
batch*num_online_cpu. We do not need to get actual count for comparison
if count is different from the given value by this factor and allows for
quick comparison without summing up all the per cpu counters.
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
As explained in commit 1c0fe6e3bd ("mm: invoke oom-killer from page
fault") , we want to call the architecture independent oom killer when
getting an unexplained OOM from handle_mm_fault, rather than simply
killing current.
Signed-off-by: Nick Piggin <npiggin@suse.de>
Acked-by: Haavard Skinnemoen <hskinnemoen@atmel.com>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
migrate_pages() is using >500 bytes stack. Reduce it.
mm/mempolicy.c: In function 'sys_migrate_pages':
mm/mempolicy.c:1344: warning: the frame size of 528 bytes is larger than 512 bytes
[akpm@linux-foundation.org: don't play with a might-be-NULL pointer]
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Reviewed-by: Christoph Lameter <cl@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Define stubs for the numa_*_id() generic percpu related functions for
non-NUMA configurations in <asm-generic/topology.h> where the other
non-numa stubs live.
Fixes ia64 !NUMA build breakage -- e.g., tiger_defconfig
Back out now unneeded '#ifndef CONFIG_NUMA' guards from ia64 smpboot.c
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Tested-by: Tony Luck <tony.luck@intel.com>
Acked-by: Tony Luck <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We have been used naming try_set_zone_oom and clear_zonelist_oom.
The role of functions is to lock of zonelist for preventing parallel
OOM. So clear_zonelist_oom makes sense but try_set_zone_oome is rather
awkward and unmatched with clear_zonelist_oom.
Let's change it with try_set_zonelist_oom.
Signed-off-by: Minchan Kim <minchan.kim@gmail.com>
Acked-by: David Rientjes <rientjes@google.com>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Remove the redundancy in __oom_kill_task() since:
- init can never be passed to this function: it will never be PF_EXITING
or selectable from select_bad_process(), and
- it will never be passed a task from oom_kill_task() without an ->mm
and we're unconcerned about detachment from exiting tasks, there's no
reason to protect them against SIGKILL or access to memory reserves.
Also moves the kernel log message to a higher level since the verbosity is
not always emitted here; we need not print an error message if an exiting
task is given a longer timeslice.
__oom_kill_task() only has a single caller, so it can be merged into that
function at the same time.
Signed-off-by: David Rientjes <rientjes@google.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
It is possible to remove the special pagefault oom handler by simply oom
locking all system zones and then calling directly into out_of_memory().
All populated zones must have ZONE_OOM_LOCKED set, otherwise there is a
parallel oom killing in progress that will lead to eventual memory freeing
so it's not necessary to needlessly kill another task. The context in
which the pagefault is allocating memory is unknown to the oom killer, so
this is done on a system-wide level.
If a task has already been oom killed and hasn't fully exited yet, this
will be a no-op since select_bad_process() recognizes tasks across the
system with TIF_MEMDIE set.
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Nick Piggin <npiggin@suse.de>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There are various points in the oom killer where the kernel must determine
whether to panic or not. It's better to extract this to a helper function
to remove all the confusion as to its semantics.
Also fix a call to dump_header() where tasklist_lock is not read- locked,
as required.
There's no functional change with this patch.
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If memory has been depleted in lowmem zones even with the protection
afforded to it by /proc/sys/vm/lowmem_reserve_ratio, it is unlikely that
killing current users will help. The memory is either reclaimable (or
migratable) already, in which case we should not invoke the oom killer at
all, or it is pinned by an application for I/O. Killing such an
application may leave the hardware in an unspecified state and there is no
guarantee that it will be able to make a timely exit.
Lowmem allocations are now failed in oom conditions when __GFP_NOFAIL is
not used so that the task can perhaps recover or try again later.
Previously, the heuristic provided some protection for those tasks with
CAP_SYS_RAWIO, but this is no longer necessary since we will not be
killing tasks for the purposes of ISA allocations.
high_zoneidx is gfp_zone(gfp_flags), meaning that ZONE_NORMAL will be the
default for all allocations that are not __GFP_DMA, __GFP_DMA32,
__GFP_HIGHMEM, and __GFP_MOVABLE on kernels configured to support those
flags. Testing for high_zoneidx being less than ZONE_NORMAL will only
return true for allocations that have either __GFP_DMA or __GFP_DMA32.
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The oom killer tasklist dump, enabled with the oom_dump_tasks sysctl, is
very helpful information in diagnosing why a user's task has been killed.
It emits useful information such as each eligible thread's memory usage
that can determine why the system is oom, so it should be enabled by
default.
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The oom killer presently kills current whenever there is no more memory
free or reclaimable on its mempolicy's nodes. There is no guarantee that
current is a memory-hogging task or that killing it will free any
substantial amount of memory, however.
In such situations, it is better to scan the tasklist for nodes that are
allowed to allocate on current's set of nodes and kill the task with the
highest badness() score. This ensures that the most memory-hogging task,
or the one configured by the user with /proc/pid/oom_adj, is always
selected in such scenarios.
Signed-off-by: David Rientjes <rientjes@google.com>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When a task is chosen for oom kill, the oom killer first attempts to
sacrifice a child not sharing its parent's memory instead. Unfortunately,
this often kills in a seemingly random fashion based on the ordering of
the selected task's child list. Additionally, it is not guaranteed at all
to free a large amount of memory that we need to prevent additional oom
killing in the very near future.
Instead, we now only attempt to sacrifice the worst child not sharing its
parent's memory, if one exists. The worst child is indicated with the
highest badness() score. This serves two advantages: we kill a
memory-hogging task more often, and we allow the configurable
/proc/pid/oom_adj value to be considered as a factor in which child to
kill.
Reviewers may observe that the previous implementation would iterate
through the children and attempt to kill each until one was successful and
then the parent if none were found while the new code simply kills the
most memory-hogging task or the parent. Note that the only time
oom_kill_task() fails, however, is when a child does not have an mm or has
a /proc/pid/oom_adj of OOM_DISABLE. badness() returns 0 for both cases,
so the final oom_kill_task() will always succeed.
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Nick Piggin <npiggin@suse.de>
Acked-by: Balbir Singh <balbir@linux.vnet.ibm.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Tasks that do not share the same set of allowed nodes with the task that
triggered the oom should not be considered as candidates for oom kill.
Tasks in other cpusets with a disjoint set of mems would be unfairly
penalized otherwise because of oom conditions elsewhere; an extreme
example could unfairly kill all other applications on the system if a
single task in a user's cpuset sets itself to OOM_DISABLE and then uses
more memory than allowed.
Killing tasks outside of current's cpuset rarely would free memory for
current anyway. To use a sane heuristic, we must ensure that killing a
task would likely free memory for current and avoid needlessly killing
others at all costs just because their potential memory freeing is
unknown. It is better to kill current than another task needlessly.
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Nick Piggin <npiggin@suse.de>
Acked-by: Balbir Singh <balbir@linux.vnet.ibm.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
It's unnecessary to SIGKILL a task that is already PF_EXITING and can
actually cause a NULL pointer dereference of the sighand if it has already
been detached. Instead, simply set TIF_MEMDIE so it has access to memory
reserves and can quickly exit as the comment implies.
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
It's possible to livelock the page allocator if a thread has mm->mmap_sem
and fails to make forward progress because the oom killer selects another
thread sharing the same ->mm to kill that cannot exit until the semaphore
is dropped.
The oom killer will not kill multiple tasks at the same time; each oom
killed task must exit before another task may be killed. Thus, if one
thread is holding mm->mmap_sem and cannot allocate memory, all threads
sharing the same ->mm are blocked from exiting as well. In the oom kill
case, that means the thread holding mm->mmap_sem will never free
additional memory since it cannot get access to memory reserves and the
thread that depends on it with access to memory reserves cannot exit
because it cannot acquire the semaphore. Thus, the page allocators
livelocks.
When the oom killer is called and current happens to have a pending
SIGKILL, this patch automatically gives it access to memory reserves and
returns. Upon returning to the page allocator, its allocation will
hopefully succeed so it can quickly exit and free its memory. If not, the
page allocator will fail the allocation if it is not __GFP_NOFAIL.
Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
