Re: [PATCH] Revert "Revert "mm/compaction: fix set skip in fast_find_migrateblock""

[Baolin Wang <baolin.wang@linux.alibaba.com>]

On 4/26/2023 11:33 PM, Mel Gorman wrote:
> On Wed, Apr 26, 2023 at 05:10:14PM +0200, Vlastimil Babka wrote:
>> On 4/26/23 17:03, Baolin Wang wrote:
>>> This reverts commit 95e7a450b8190673675836bfef236262ceff084a.
>>>
>>> When I tested thpscale with v6.3 kernel, I found the compaction efficiency
>>> had a great regression compared to v6.2-rc1 kernel. See below numbers:
>>>                                      v6.2-rc             v6.3
>>> Percentage huge-3        81.35 (   0.00%)       32.97 ( -59.47%)
>>> Percentage huge-5        89.92 (   0.00%)       41.70 ( -53.63%)
>>> Percentage huge-7        92.41 (   0.00%)       34.08 ( -63.12%)
>>> Percentage huge-12       90.29 (   0.00%)       41.10 ( -54.49%)
>>> Percentage huge-18       82.38 (   0.00%)       41.24 ( -49.95%)
>>> Percentage huge-24       80.34 (   0.00%)       35.99 ( -55.20%)
>>> Percentage huge-30       88.90 (   0.00%)       44.20 ( -50.28%)
>>> Percentage huge-32       90.69 (   0.00%)       79.57 ( -12.25%)
>>>
>>> Ops Compaction stalls                 113790.00      207099.00
>>> Ops Compaction success                 33983.00      19488.00
>>> Ops Compaction failures                79807.00      187611.00
>>> Ops Compaction efficiency                 29.86          9.41
>>>
>>> After some investigation, I found the commit 95e7a450b819
>>> ("Revert mm/compaction: fix set skip in fast_find_migrateblock") caused
>>> the regression. This commit revert the commit 7efc3b726103 ("mm/compaction:
>>> fix set skip in fast_find_migrateblock") to fix a CPU stalling issue, which
>>> is caused by compaction stucked in repeating fast_find_migrateblock().
>>>
>>> And now the compaction stalling issue is addressed by commit cfccd2e63e7e
>>> ("mm, compaction: finish pageblocks on complete migration failure"). So
>>
>> IIRC at that time I was pointing out some scenarios that could make the
>> problem appear even after that commit, and we wanted to revisit that
>> when Mel is back.

Ah, I missed that, and will check previous discussion.

> Yes, I've prototyped the fix against 6.3-rc7 and the revert is at the
> end but the revert on its own has the potential for causing problems. The
> series needs to be rebased, retested and posted. What I last tested
> should show up shortly at
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git/ mm-follupfastmigrate-v1r1

Thanks.