Re: [PATCH v3 0/7] Split a folio to any lower order folios

[Zi Yan <ziy@nvidia.com>]

[-- Attachment #1: Type: text/plain, Size: 2952 bytes --]

On 18 Apr 2023, at 6:29, David Hildenbrand wrote:

> On 17.04.23 21:26, Zi Yan wrote:
>> On 17 Apr 2023, at 10:20, David Hildenbrand wrote:
>>
>>> On 16.04.23 20:11, Hugh Dickins wrote:
>>>> On Tue, 4 Apr 2023, Andrew Morton wrote:
>>>>> On Mon,  3 Apr 2023 16:18:32 -0400 Zi Yan <zi.yan@sent.com> wrote:
>>>>>
>>>>>> File folio supports any order and people would like to support flexible orders
>>>>>> for anonymous folio[1] too. Currently, split_huge_page() only splits a huge
>>>>>> page to order-0 pages, but splitting to orders higher than 0 is also useful.
>>>>>> This patchset adds support for splitting a huge page to any lower order pages
>>>>>> and uses it during file folio truncate operations.
>>>>>
>>>>> This series (and its v1 & v2) don't appear to have much in the way of
>>>>> detailed review.  As it's at v3 and has been fairly stable I'll queue
>>>>> it up for some testing now, but I do ask that some reviewers go through
>>>>> it please.
>>>>
>>>> Andrew, please don't let this series drift into 6.4-rc1.
>>>>
>>>> I've seen a bug or two (I'll point out in response to those patches),
>>>> but overall I don't see what the justification for the series is: done
>>>> because it could be done, it seems to me, but liable to add surprises.
>>>>
>>>> The cover letter says "splitting to orders higher than 0 is also useful",
>>>> but it's not clear why; and the infrastructure provided seems unsuited
>>>> to the one use provided - I'll say more on that truncation patch.
>>>
>>> I agree. Maybe this patch set is something we want to have in the future once actual consumers that can benefit are in place, such that we can show actual performance numbers with/without.
>>
>> Ryan is working on large folio for anonymous pages and has shown promising performance
>> results[1]. This patchset would avoid getting base pages during split if possible.
>>
> Yes, I know. And it would be great to get some actual numbers with/without your patches to show that this optimization actually matters in practice.

Sure. Will try to add perf numbers in the next version.

>
> Unrelated, your cover letter mentions "file folio truncate operations.". Would it also apply to FALLOC_FL_PUNCH_HOLE, when partially zapping a THP?

Yes. In the self tests, I have
fallocate(fd, FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE, offset[j], len[j]);
and it uses the truncate operation I updated in patch 6.

>
>>>
>>> Until then, "365 insertions(+), 68 deletions(-)" certainly needs some reasonable motivation.
>>
>> Come on. 225 out of 365 insertions are self tests code. We need motivation to add
>> testing code?
>
> Well, if you add feature X and the tests target feature X, then certainly having the tests require the same motivation as feature X. But yeah, the actual kernel code change is smaller than it looks at first sight.
>
> -- 
> Thanks,
>
> David / dhildenb


--
Best Regards,
Yan, Zi

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 854 bytes --]