Re: [PATCH] mm,page_alloc: PF_WQ_WORKER threads must sleep at should_reclaim_retry().

[Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>]

On 2018/07/30 23:54, Tejun Heo wrote:
> Hello,
> 
> On Mon, Jul 30, 2018 at 04:46:47PM +0200, Michal Hocko wrote:
>> On Mon 30-07-18 23:34:23, Tetsuo Handa wrote:
>>> On 2018/07/30 18:32, Michal Hocko wrote:
>> [...]
>>>> This one is waiting for draining and we are in mm_percpu_wq WQ context
>>>> which has its rescuer so no other activity can block us for ever. So
>>>> this certainly shouldn't deadlock. It can be dead slow but well, this is
>>>> what you will get when your shoot your system to death.
>>>
>>> We need schedule_timeout_*() to allow such WQ_MEM_RECLAIM workqueues to wake up. (Tejun,
>>> is my understanding correct?) Lack of schedule_timeout_*() does block WQ_MEM_RECLAIM
>>> workqueues forever.
>>
>> Hmm. This doesn't match my understanding of what WQ_MEM_RECLAIM actually
>> guarantees. If you are right then the whole thing sounds quite fragile
>> to me TBH.
> 
> Workqueue doesn't think the cpu is stalled as long as one of the
> per-cpu kworkers is running.  The assumption is that kernel threads
> are not supposed to be busy-looping indefinitely (and they really
> shouldn't).

WQ_MEM_RECLAIM guarantees that "struct task_struct" is preallocated. But
WQ_MEM_RECLAIM does not guarantee that the pending work is started as soon
as an item was queued. Same rule applies to both WQ_MEM_RECLAIM workqueues 
and !WQ_MEM_RECLAIM workqueues regarding when to start a pending work (i.e.
when schedule_timeout_*() is called).

Is this correct?

>              We can add timeout mechanism to workqueue so that it
> kicks off other kworkers if one of them is in running state for too
> long, but idk, if there's an indefinite busy loop condition in kernel
> threads, we really should get rid of them and hung task watchdog is
> pretty effective at finding these cases (at least with preemption
> disabled).

Currently the page allocator has a path which can loop forever with
only cond_resched().

> 
> Thanks.
>