Re: [PATCH 09/14] mm: use compaction feedback for thp backoff conditions

[Michal Hocko]

On Fri 29-04-16 11:16:44, Vlastimil Babka wrote:
> On 04/28/2016 02:35 PM, Michal Hocko wrote:
[...]
> >My main point was to simplify the code and get rid of as much compaction
> >specific hacks as possible. We might very well drop this later on but it
> >would be at least less code to grasp through. I do not have any problem
> >with dropping this but I think this shouldn't collide with other patches
> >much so reducing the number of lines is worth it.

Good point, I have completely missed this part.

> I just realized it also affects khugepaged, and not just THP page faults, so
> it may potentially cripple THP's completely. My main issue is that the
> reasons to bail out includes COMPACT_SKIPPED, and for a wrong reason (see
> the comment above). It also goes against the comment below the noretry
> label:
> 
>  * High-order allocations do not necessarily loop after direct reclaim
>  * and reclaim/compaction depends on compaction being called after
>  * reclaim so call directly if necessary.
> 
> Given that THP's are large, I expect reclaim would indeed be quite often
> necessary before compaction, and the first optimistic async compaction
> attempt will just return SKIPPED. After this patch, there will be no more
> reclaim/compaction attempts for THP's, including khugepaged. And given the
> change of THP page fault defaults, even crippling that path should no longer
> be necessary.
> 
> So I would just drop this for now indeed.

Agreed, thanks for catching this. Andrew, could you drop this patch
please? It was supposed to be a mere clean up without any effect on the
oom detection.

Thanks!
-- 
Michal Hocko
SUSE Labs