mm, oom_reaper: make sure that mmput_async is called only when memory was reaped
authorMichal Hocko <mhocko@suse.com>
Tue, 26 Jul 2016 22:24:50 +0000 (15:24 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 26 Jul 2016 23:19:19 +0000 (16:19 -0700)
Tetsuo is worried that mmput_async might still lead to a premature new
oom victim selection due to the following race:

__oom_reap_task exit_mm
  find_lock_task_mm
  atomic_inc(mm->mm_users) # = 2
  task_unlock
     task_lock
  task->mm = NULL
  up_read(&mm->mmap_sem)
< somebody write locks mmap_sem >
  task_unlock
  mmput
       atomic_dec_and_test # = 1
  exit_oom_victim
  down_read_trylock # failed - no reclaim
  mmput_async # Takes unpredictable amount of time
   < new OOM situation >

the final __mmput will be executed in the delayed context which might
happen far in the future.  Such a race is highly unlikely because the
write holder of mmap_sem would have to be an external task (all direct
holders are already killed or exiting) and it usually have to pin
mm_users in order to do anything reasonable.

We can, however, make sure that the mmput_async is only called when we
do not back off and reap some memory.  That would reduce the impact of
the delayed __mmput because the real content would be already freed.
Pin mm_count to keep it alive after we drop task_lock and before we try
to get mmap_sem.  If the mmap_sem succeeds we can try to grab mm_users
reference and then go on with unmapping the address space.

It is not clear whether this race is possible at all but it is better to
be more robust and do not pin mm_users unless we are sure we are
actually doing some real work during __oom_reap_task.

Link: http://lkml.kernel.org/r/1465306987-30297-1-git-send-email-mhocko@kernel.org
Signed-off-by: Michal Hocko <mhocko@suse.com>
Reported-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/oom_kill.c

index c11f8bdd0c12cea00cf13d57fc53c323439e010c..d4a929d79470aa0c9edc6131854dbf309ca39d47 100644 (file)
@@ -452,7 +452,7 @@ static bool __oom_reap_task(struct task_struct *tsk)
         * We have to make sure to not race with the victim exit path
         * and cause premature new oom victim selection:
         * __oom_reap_task              exit_mm
-        *   atomic_inc_not_zero
+        *   mmget_not_zero
         *                                mmput
         *                                  atomic_dec_and_test
         *                                exit_oom_victim
@@ -474,12 +474,22 @@ static bool __oom_reap_task(struct task_struct *tsk)
        if (!p)
                goto unlock_oom;
        mm = p->mm;
-       atomic_inc(&mm->mm_users);
+       atomic_inc(&mm->mm_count);
        task_unlock(p);
 
        if (!down_read_trylock(&mm->mmap_sem)) {
                ret = false;
-               goto unlock_oom;
+               goto mm_drop;
+       }
+
+       /*
+        * increase mm_users only after we know we will reap something so
+        * that the mmput_async is called only when we have reaped something
+        * and delayed __mmput doesn't matter that much
+        */
+       if (!mmget_not_zero(mm)) {
+               up_read(&mm->mmap_sem);
+               goto mm_drop;
        }
 
        tlb_gather_mmu(&tlb, mm, 0, -1);
@@ -521,15 +531,16 @@ static bool __oom_reap_task(struct task_struct *tsk)
         * to release its memory.
         */
        set_bit(MMF_OOM_REAPED, &mm->flags);
-unlock_oom:
-       mutex_unlock(&oom_lock);
        /*
         * Drop our reference but make sure the mmput slow path is called from a
         * different context because we shouldn't risk we get stuck there and
         * put the oom_reaper out of the way.
         */
-       if (mm)
-               mmput_async(mm);
+       mmput_async(mm);
+mm_drop:
+       mmdrop(mm);
+unlock_oom:
+       mutex_unlock(&oom_lock);
        return ret;
 }