Re: [PATCH v3 1/4] mm/ksm: align smart scan skips with the scan number

[Qiliang Yuan <odys.yuan@gmail.com>]

On 11/7/26 20:51, David Hildenbrand (Arm) wrote:
> On 11/7/26 17:57, Qiliang Yuan wrote:
>>                               before     after      after/before
>>   pairs merged (of 420)       361        420        116%
>>                               334        420        126%
>>   pages looked at per scan    1792       2048       114%
>>
> 
> Oh, I was wondering whether this should actually go to stable? Do you think the
> above is rather an extreme case, and most users won't really run into it that
> severely?

I don't think it is an extreme case, and the test above rather
understates it: it makes the second page of a pair mergeable while the
first one is still young. The age of an rmap_item only grows, and is not
reset when the page content changes, so on a long running system most
unmerged pages are past age 8 and each is looked at every ninth scan, on
a phase of its own.

To check that case, I let 1800 pairs of unique pages age for 32 scans,
with the second page of each pair made mergeable 1 to 9 scans after the
first, and then gave both pages of each pair the same new content:

  kernel           pairs merged after 16 / 32 / 256 scans
  before the fix    112 /  200 /  200
  with the fix     1666 / 1800 / 1800

Before the fix only the pairs whose phases match, one in nine, get
merged, and the others stay unmerged. Three more runs of 64 scans and
one with 128 scans of aging gave 200 and 1800 each time.

It is limited to content that is not in the stable tree yet, since KSM
pages are never skipped and later copies merge with them, and the cost
is memory that stays unmerged. With smart scan on by default since v6.7
and the fix being small, I'd lean towards stable.

I'll send a v4 in the next day or two, which also makes the new selftest
helpers treat a short read or write of smart_scan as an error, and can
add Cc: stable to patch 1 there if you agree.

-- 
Thanks,
Qiliang