Re: [PATCH] memcg, oom: be careful about races when warning about no reclaimable task

From: Tetsuo Handa
Date: Tue Aug 07 2018 - 16:38:46 EST


On 2018/08/08 5:19, Johannes Weiner wrote:
> On Tue, Aug 07, 2018 at 07:15:11PM +0900, Tetsuo Handa wrote:
>> On 2018/08/07 16:25, Michal Hocko wrote:
>>> @@ -1703,7 +1703,8 @@ static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int
>>> return OOM_ASYNC;
>>> }
>>>
>>> - if (mem_cgroup_out_of_memory(memcg, mask, order))
>>> + if (mem_cgroup_out_of_memory(memcg, mask, order) ||
>>> + tsk_is_oom_victim(current))
>>> return OOM_SUCCESS;
>>>
>>> WARN(1,"Memory cgroup charge failed because of no reclaimable memory! "
>>>
>>
>> I don't think this patch is appropriate. This patch only avoids hitting WARN(1).
>> This patch does not address the root cause:
>>
>> The task_will_free_mem(current) test in out_of_memory() is returning false
>> because test_bit(MMF_OOM_SKIP, &mm->flags) test in task_will_free_mem() is
>> returning false because MMF_OOM_SKIP was already set by the OOM reaper. The OOM
>> killer does not need to start selecting next OOM victim until "current thread
>> completes __mmput()" or "it fails to complete __mmput() within reasonable
>> period".
>
> I don't see why it matters whether the OOM victim exits or not, unless
> you count the memory consumed by struct task_struct.

We are not counting memory consumed by struct task_struct. But David is
counting memory released between set_bit(MMF_OOM_SKIP, &mm->flags) and
completion of exit_mmap().

>
>> According to https://syzkaller.appspot.com/text?tag=CrashLog&x=15a1c770400000 ,
>> PID=23767 selected PID=23766 as an OOM victim and the OOM reaper set MMF_OOM_SKIP
>> before PID=23766 unnecessarily selects PID=23767 as next OOM victim.
>> At uptime = 366.550949, out_of_memory() should have returned true without selecting
>> next OOM victim because tsk_is_oom_victim(current) == true.
>
> The code works just fine. We have to kill tasks until we a) free
> enough memory or b) run out of tasks or c) kill current. When one of
> these outcomes is reached, we allow the charge and return.
>
> The only problem here is a warning in the wrong place.
>

If forced charge contained a bug, removing this WARN(1) deprives users of chance
to know that something is going wrong.