Re: [PATCH v3] mm/compaction: fix set skip in fast_find_migrateblock

[Chuyi Zhou <zhouchuyi@bytedance.com>]

在 2022/8/15 上午11:20, Chuyi Zhou 写道:
> 
> 
> 在 2022/8/15 上午9:24, Andrew Morton 写道:
>> On Tue, 19 Jul 2022 09:28:54 +0100 Mel Gorman 
>> <mgorman@techsingularity.net> wrote:
>>
>>> The return block indeed may not be aligned. It could simply be a
>>> restart. The changelog could still do with a little clarification but
>>> your patch is still fine.
>>
>> I'm not seeing a v4 so I merged the v3.  It would be great if someone
>> could please send along an updated changelog?
> 
> Hi, Morton, Here is the updated changelog:
> The fast_find_migrateblock could return a block aligned with
> pageblock_nr_pages. When we successfully find a block, we use
> pageblock_start_pfn(free_pfn) to get the first pfn of the pageblock,
> the block will be set skip through *set_pageblock_skip*, normally
> the value get from pageblock_start_pfn should be aligned with
> *pageblock_nr_pages*. Then the *first pfn* will be passed to
> isolate_migratepages_block, the whole pageblock will be skipped
> due to the branch
> if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)), because
> !valid_page will be true for the first pageblock, and the low_pfn could
> be aligned with pageblock_nr_pages as mentioned above. Eventually we
> will goto isolate_abort and isolate nothing. That cause
> fast_find_migrateblock useless.
> In this Patch, when we find a suitable pageblock in fast_find_
> migrateblock, we do noting but let isolate_migratepages_block
> to set skip to the pageblock after scan it. Normally,
> we would isolate some pages from the fast-find block.
> I use mmtest/thpscale-madvhugepage test it. Here is the result:
>                               baseline               patch
> Amean     fault-both-1      1331.66 (   0.00%)     1261.04 *   5.30%*
> Amean     fault-both-3      1383.95 (   0.00%)     1191.69 *  13.89%*
> Amean     fault-both-5      1568.13 (   0.00%)     1445.20 *   7.84%*
> Amean     fault-both-7      1819.62 (   0.00%)     1555.13 *  14.54%*
> Amean     fault-both-12     1106.96 (   0.00%)     1149.43 *  -3.84%*
> Amean     fault-both-18     2196.93 (   0.00%)     1875.77 *  14.62%*
> Amean     fault-both-24     2642.69 (   0.00%)     2671.21 *  -1.08%*
> Amean     fault-both-30     2901.89 (   0.00%)     2857.32 *   1.54%*
> Amean     fault-both-32     3747.00 (   0.00%)     3479.23 *   7.15%*
> 
> Fixes: 70b44595eafe9 ("mm, compaction: use free lists to quickly locate 
> a migration source")
> Signed-off-by: zhouchuyi <zhouchuyi@bytedance.com>
> =======================================================================
> Mel think it could be possible that the fast_find_migrateblock can return
> a pageblock not aligned pageblock_nr_pages. In that case,
> isolate_migratepages_block would not set skip flag.
> I want to fix this problem in v4.