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

[Tetsuo Handa]

On 2018/08/08 5:38, Tetsuo Handa wrote:
> 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().

Also, before the OOM reaper was introduced, we waited until TIF_MEMDIE is
cleared from the OOM victim thread. Compared to pre OOM reaper era, giving up
so early is certainly a regression.