Re: [PATCH v5 02/14] mm, compaction: defer each zone individually instead of preferred zone

[Vlastimil Babka <vbabka@suse.cz>]

On 07/30/2014 06:22 PM, Vlastimil Babka wrote:
> On 07/29/2014 11:12 AM, Vlastimil Babka wrote:
>> On 07/29/2014 08:38 AM, Joonsoo Kim wrote:
>>>
>>> I still don't understand why defer_compaction() is needed here.
>>> defer_compaction() is intended for not struggling doing compaction on
>>> the zone where we already have tried compaction and found that it
>>> isn't suitable for compaction. Allocation failure doesn't tell us
>>> that we have tried compaction for all the zone range so we shouldn't
>>> make a decision here to defer compaction on this zone carelessly.
>>
>> OK I can remove that, it should make the code nicer anyway.
>
> Weird, that removal of this defer_compaction() call seems ho have
> quadrupled compact_stall and compact_fail counts. The scanner pages
> counters however increased by only 10% so that could indicate the
> problem is occuring only in a small zone such as DMA. Could be another
> case of mismatch between watermark checking in compaction and
> allocation? Perhaps the lack of proper classzone_idx in the compaction
> check? Sigh.

Yep so it was the DMA zone returning COMPACT_PARTIAL from the 
compaction_suitable() check done at the very beginning of 
compact_zone(). The meaning of that is "the allocation should succeed 
without compaction", so compaction is not done at all. Yet the 
COMPACT_PARTIAL return value means it counts as a stall, even with the 
patch that doesn't count COMPACT_SKIPPED as stalls.
The watermark check in try_to_compact_pages() also apparently succeeds 
as the compaction is not being deferred. With deferral removed from 
__alloc_pages_direct_compact(), this zone will be attempted uselessly 
each time, and deferred_compaction is practically never reported back.

So for now I think it would be best to leave the defer_compaction() call 
in __alloc_pages_direct_compact() as it is. Fixing this in a better way 
would require more investigation (I guess the lack of classzone_idx in 
compaction makes the difference for the watermark checks here) and 
another patch(es), which I'll attempt, but I don't want to further grow 
this series with new patches right now.

>> I also agree
>> with the argument "for all the zone range" and I also realized that it's
>> not (both before and after this patch) really the case. I planned to fix
>> that in the future, but I can probably do it now.
>> The plan is to call defer_compaction() only when compaction returned
>> COMPACT_COMPLETE (and not COMPACT_PARTIAL) as it means the whole zone
>> was scanned. Otherwise there will be bias towards the beginning of the
>> zone in the migration scanner - compaction will be deferred half-way and
>> then cached pfn's might be reset when it restarts, and the rest of the
>> zone won't be scanned at all.
>
> Hm despite expectations, this didn't seem to make much difference. But
> maybe there will be once I have some idea what happened to those stalls.

Yeah, so it doesn't matter here if I call defer_compaction() with only 
COMPACT_COMPLETE returned from compact_zone(). The whole thing is only 
done when watermarks check fails, and it doesn't for the DMA zone.

>>> Thanks.
>>>
>>
>