Re: [PATCH v5 07/14] mm, compaction: khugepaged should not give up due to need_resched()

[Vlastimil Babka <vbabka@suse.cz>]

On 07/29/2014 09:31 AM, David Rientjes wrote:
> On Tue, 29 Jul 2014, Joonsoo Kim wrote:
>
>> I have a silly question here.
>> Why need_resched() is criteria to stop async compaction?
>> need_resched() is flagged up when time slice runs out or other reasons.
>> It means that we should stop async compaction at arbitrary timing
>> because process can be on compaction code at arbitrary moment. I think
>> that it isn't reasonable and it doesn't ensure anything. Instead of
>> this approach, how about doing compaction on certain amounts of pageblock
>> for async compaction?
>>
>
> Not a silly question at all, I had the same feeling in
> https://lkml.org/lkml/2014/5/21/730 and proposed it to be a tunable that
> indicates how much work we are willing to do for thp in the pagefault
> path.  It suffers from the fact that past failure to isolate and/or
> migrate memory to free an entire pageblock doesn't indicate that the next
> pageblock will fail as well, but there has to be cutoff at some point or
> async compaction becomes unnecessarily expensive.  We can always rely on
> khugepaged later to do the collapse, assuming we're not faulting memory
> and then immediately pinning it.
>
> I think there's two ways to go about it:
>
>   - allow a single thp fault to be expensive and then rely on deferred
>     compaction to avoid subsequent calls in the near future, or
>
>   - try to make all thp faults be as least expensive as possible so that
>     the cumulative effect of faulting large amounts of memory doesn't end
>     up with lengthy stalls.
>
> Both of these are complex because of the potential for concurrent calls to
> memory compaction when faulting thp on several cpus.
>
> I also think the second point from that email still applies, that we
> should abort isolating pages within a pageblock for migration once it can
> no longer allow a cc->order allocation to succeed.

That was the RFC patch 15, I hope to reintroduce it soon. You could 
still test it meanwhile to see if you see the same extfrag regression as 
me. In my tests, kswapd/khugepaged wasn't doing enough work to 
defragment the pageblocks that the stress-highalloc benchmark 
(configured to behave like thp page fault) was skipping.