Re: Lockup in wait_transaction_locked under memory pressure

[Nikolay Borisov <kernel@kyup.com>]

On 06/29/2015 12:38 PM, Michal Hocko wrote:
> On Mon 29-06-15 12:23:16, Nikolay Borisov wrote:
>>
>>
>> On 06/29/2015 12:16 PM, Michal Hocko wrote:
>>> On Mon 29-06-15 12:07:54, Nikolay Borisov wrote:
>>>>
>>>>
>>>> On 06/29/2015 11:32 AM, Michal Hocko wrote:
>>>>> On Thu 25-06-15 18:27:10, Nikolay Borisov wrote:
>>>>>>
>>>>>>
>>>>>> On 06/25/2015 06:18 PM, Michal Hocko wrote:
>>>>>>> On Thu 25-06-15 17:34:22, Nikolay Borisov wrote:
>>>>>>>> On 06/25/2015 05:05 PM, Michal Hocko wrote:
>>>>>>>>> On Thu 25-06-15 16:49:43, Nikolay Borisov wrote:
>>>>>>>>> [...]
>>>>>>>>>> How would you advise to rectify such situation?
>>>>>>>>>
>>>>>>>>> As I've said. Check the oom victim traces and see if it is holding any
>>>>>>>>> of those locks.
>>>>>>>>
>>>>>>>> As mentioned previously all OOM traces are identical to the one I've
>>>>>>>> sent - OOM being called form the page fault path.
>>>>>>>  
>>>>>>> By identical you mean that all of them kill the same task? Or just that
>>>>>>> the path is same (which wouldn't be surprising as this is the only path
>>>>>>> which triggers memcg oom killer)?
>>>>>>
>>>>>> The code path is the same, the tasks being killed are different
>>>>>
>>>>> Is the OOM killer triggered only for a singe memcg or others misbehave
>>>>> as well?
>>>>
>>>> Generally OOM would be triggered for whichever memcg runs out of
>>>> resources but so far I've only observed that the D state issue happens
>>>> in a single containers.
>>>
>>> It is not clear whether it is the OOM memcg which has tasks in the D
>>> state. Anyway I think it all smells like one memcg is throttling others
>>> on another shared resource - journal in your case.
>>
>> Be that as it may, how do I find which cgroup is the culprit?
> 
> Ted has already described that. You have to check all the running tasks
> and try to find which of them is doing the operation which blocks
> others. Transaction commit sounds like the first one to check.

One other, fairly crucial detail - each and every container is on a
separate block device, meaning the journals for different block devices
is not being shared, since the journal is per-block device. I guess this
means that whatever is happening is more or less constrained to the
block device and thus the possibility that different memcg competing for
the journal can be eliminated?

>