lkml.org 
[lkml]   [2018]   [Jul]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
SubjectRe: [patch v3] mm, oom: fix unnecessary killing of additional processes
From
Date
Sigh...

Nacked-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>

because David is not aware what is wrong.

On 2018/07/19 5:22, David Rientjes wrote:
> On Wed, 18 Jul 2018, Tetsuo Handa wrote:
>
>>> diff --git a/mm/mmap.c b/mm/mmap.c
>>> --- a/mm/mmap.c
>>> +++ b/mm/mmap.c
>>> @@ -3059,25 +3059,28 @@ void exit_mmap(struct mm_struct *mm)
>>> if (unlikely(mm_is_oom_victim(mm))) {
>>> /*
>>> * Manually reap the mm to free as much memory as possible.
>>> - * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
>>> - * this mm from further consideration. Taking mm->mmap_sem for
>>> - * write after setting MMF_OOM_SKIP will guarantee that the oom
>>> - * reaper will not run on this mm again after mmap_sem is
>>> - * dropped.
>>> - *
>>> * Nothing can be holding mm->mmap_sem here and the above call
>>> * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
>>> * __oom_reap_task_mm() will not block.
>>> - *
>>> - * This needs to be done before calling munlock_vma_pages_all(),
>>> - * which clears VM_LOCKED, otherwise the oom reaper cannot
>>> - * reliably test it.
>>> */
>>> mutex_lock(&oom_lock);
>>> __oom_reap_task_mm(mm);
>>> mutex_unlock(&oom_lock);
>>>
>>> - set_bit(MMF_OOM_SKIP, &mm->flags);
>>> + /*
>>> + * Now, set MMF_UNSTABLE to avoid racing with the oom reaper.
>>> + * This needs to be done before calling munlock_vma_pages_all(),
>>> + * which clears VM_LOCKED, otherwise the oom reaper cannot
>>> + * reliably test for it. If the oom reaper races with
>>> + * munlock_vma_pages_all(), this can result in a kernel oops if
>>> + * a pmd is zapped, for example, after follow_page_mask() has
>>> + * checked pmd_none().
>>> + *
>>> + * Taking mm->mmap_sem for write after setting MMF_UNSTABLE will
>>> + * guarantee that the oom reaper will not run on this mm again
>>> + * after mmap_sem is dropped.
>>> + */
>>> + set_bit(MMF_UNSTABLE, &mm->flags);
>>
>> Since MMF_UNSTABLE is set by __oom_reap_task_mm() from exit_mmap() before start reaping
>> (because the purpose of MMF_UNSTABLE is to "tell all users of get_user/copy_from_user
>> etc... that the content is no longer stable"), it cannot be used for a flag for indicating
>> that the OOM reaper can't work on the mm anymore.
>>
>
> Why? It should be able to be set by exit_mmap() since nothing else should
> be accessing this mm in the first place. There is no reason to wait for
> the oom reaper and the following down_write();up_write(); cycle will
> guarantee it is not operating on the mm before munlocking.
>

It does not make sense to call set_bit(MMF_UNSTABLE, &mm->flags) again after returning from
__oom_reap_task_mm() because MMF_UNSTABLE is _aready_ set in the beginning of __oom_reap_task_mm().

void __oom_reap_task_mm(struct mm_struct *mm)
{
struct vm_area_struct *vma;

/*
* Tell all users of get_user/copy_from_user etc... that the content
* is no longer stable. No barriers really needed because unmapping
* should imply barriers already and the reader would hit a page fault
* if it stumbled over a reaped memory. If MMF_UNSTABLE is already set,
* reaping as already occurred so nothing left to do.
*/
if (test_and_set_bit(MMF_UNSTABLE, &mm->flags))
return;
(...snipped...)
}

void exit_mmap(struct mm_struct *mm)
{
struct mmu_gather tlb;
struct vm_area_struct *vma;
unsigned long nr_accounted = 0;

/* mm's last user has gone, and its about to be pulled down */
mmu_notifier_release(mm);

if (unlikely(mm_is_oom_victim(mm))) {
/*
* Manually reap the mm to free as much memory as possible.
* Nothing can be holding mm->mmap_sem here and the above call
* to mmu_notifier_release(mm) ensures mmu notifier callbacks in
* __oom_reap_task_mm() will not block.
*/
__oom_reap_task_mm(mm);

/*
* Now, set MMF_UNSTABLE to avoid racing with the oom reaper.
* This needs to be done before calling munlock_vma_pages_all(),
* which clears VM_LOCKED, otherwise the oom reaper cannot
* reliably test for it. If the oom reaper races with
* munlock_vma_pages_all(), this can result in a kernel oops if
* a pmd is zapped, for example, after follow_page_mask() has
* checked pmd_none().
*
* Taking mm->mmap_sem for write after setting MMF_UNSTABLE will
* guarantee that the oom reaper will not run on this mm again
* after mmap_sem is dropped.
*/
set_bit(MMF_UNSTABLE, &mm->flags);
down_write(&mm->mmap_sem);
up_write(&mm->mmap_sem);
}
(...snipped...)
}

>> If the oom_lock serialization is removed, the OOM reaper will give up after (by default)
>> 1 second even if current thread is immediately after set_bit(MMF_UNSTABLE, &mm->flags) from
>> __oom_reap_task_mm() from exit_mmap(). Thus, this patch and the other patch which removes
>> oom_lock serialization should be dropped.
>>
>
> No, it shouldn't, lol. The oom reaper may give up because we have entered
> __oom_reap_task_mm() by way of exit_mmap(), there's no other purpose for
> it acting on the mm. This is very different from giving up by setting
> MMF_OOM_SKIP, which it will wait for oom_free_timeout_ms to do unless the
> thread can make forward progress here in exit_mmap().

Let's call "A" as a thread doing exit_mmap(), and "B" as the OOM reaper kernel thread.

(1) "A" finds that unlikely(mm_is_oom_victim(mm)) == true.
(2) "B" finds that test_bit(MMF_OOM_SKIP, &mm->flags) in oom_reap_task() is false.
(3) "B" finds that !test_bit(MMF_UNSTABLE, &mm->flags) in oom_reap_task() is true.
(4) "B" enters into oom_reap_task_mm(tsk, mm).
(5) "B" finds that !down_read_trylock(&mm->mmap_sem) is false.
(6) "B" finds that mm_has_blockable_invalidate_notifiers(mm) is false.
(7) "B" finds that test_bit(MMF_UNSTABLE, &mm->flags) is false.
(8) "B" enters into __oom_reap_task_mm(mm).
(9) "A" finds that test_and_set_bit(MMF_UNSTABLE, &mm->flags) is false.
(10) "A" is preempted by somebody else.
(11) "B" finds that test_and_set_bit(MMF_UNSTABLE, &mm->flags) is true.
(12) "B" leaves __oom_reap_task_mm(mm).
(13) "B" leaves oom_reap_task_mm().
(14) "B" finds that time_after_eq(jiffies, mm->oom_free_expire) became true.
(15) "B" finds that !test_bit(MMF_OOM_SKIP, &mm->flags) is true.
(16) "B" calls set_bit(MMF_OOM_SKIP, &mm->flags).
(17) "B" finds that test_bit(MMF_OOM_SKIP, &mm->flags) is true.
(18) select_bad_process() finds that MMF_OOM_SKIP is already set.
(19) out_of_memory() kills a new OOM victim.
(20) "A" resumes execution and start reclaiming memory.

because oom_lock serialization was already removed.

>
>>> down_write(&mm->mmap_sem);
>>> up_write(&mm->mmap_sem);
>>> }
>>
>>> @@ -637,25 +649,57 @@ static int oom_reaper(void *unused)
>>> return 0;
>>> }
>>>
>>> +/*
>>> + * Millisecs to wait for an oom mm to free memory before selecting another
>>> + * victim.
>>> + */
>>> +static u64 oom_free_timeout_ms = 1000;
>>> static void wake_oom_reaper(struct task_struct *tsk)
>>> {
>>> - /* tsk is already queued? */
>>> - if (tsk == oom_reaper_list || tsk->oom_reaper_list)
>>> + /*
>>> + * Set the reap timeout; if it's already set, the mm is enqueued and
>>> + * this tsk can be ignored.
>>> + */
>>> + if (cmpxchg(&tsk->signal->oom_mm->oom_free_expire, 0UL,
>>> + jiffies + msecs_to_jiffies(oom_free_timeout_ms)))
>>> return;
>>
>> "expire" must not be 0 in order to avoid double list_add(). See
>> https://lore.kernel.org/lkml/201807130620.w6D6KiAJ093010@www262.sakura.ne.jp/T/#u .
>>
>
> We should not allow oom_free_timeout_ms to be 0 for sure, I assume 1000 is
> the sane minimum since we need to allow time for some memory freeing and
> this will not be radically different from what existed before the patch
> for the various backoffs. Or maybe you meant something else for "expire"
> here?
>

I'm saying that jiffies + msecs_to_jiffies(oom_free_timeout_ms) == 0 will make
tsk->signal->oom_mm->oom_free_expire == 0 and the list will be corrupted by
allowing cmpxchg(&tsk->signal->oom_mm->oom_free_expire) to become true for twice.

\
 
 \ /
  Last update: 2018-07-18 23:21    [W:0.346 / U:0.860 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site