[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <5c0979a2-9a56-4284-82d2-42da62bda4a5@redhat.com>
Date: Thu, 1 Aug 2024 15:44:12 +0200
From: David Hildenbrand <david@...hat.com>
To: Mateusz Guzik <mjguzik@...il.com>
Cc: "Yin, Fengwei" <fengwei.yin@...el.com>,
kernel test robot <oliver.sang@...el.com>, Peter Xu <peterx@...hat.com>,
oe-lkp@...ts.linux.dev, lkp@...el.com, linux-kernel@...r.kernel.org,
Andrew Morton <akpm@...ux-foundation.org>,
Huacai Chen <chenhuacai@...nel.org>, Jason Gunthorpe <jgg@...dia.com>,
Matthew Wilcox <willy@...radead.org>, Nathan Chancellor <nathan@...nel.org>,
Ryan Roberts <ryan.roberts@....com>, WANG Xuerui <kernel@...0n.name>,
linux-mm@...ck.org, ying.huang@...el.com, feng.tang@...el.com
Subject: Re: [linus:master] [mm] c0bff412e6: stress-ng.clone.ops_per_sec -2.9%
regression
On 01.08.24 15:37, Mateusz Guzik wrote:
> On Thu, Aug 1, 2024 at 3:34 PM David Hildenbrand <david@...hat.com> wrote:
>>
>> On 01.08.24 15:30, Mateusz Guzik wrote:
>>> On Thu, Aug 01, 2024 at 08:49:27AM +0200, David Hildenbrand wrote:
>>>> Yes indeed. fork() can be extremely sensitive to each added instruction.
>>>>
>>>> I even pointed out to Peter why I didn't add the PageHuge check in there
>>>> originally [1].
>>>>
>>>> "Well, and I didn't want to have runtime-hugetlb checks in
>>>> PageAnonExclusive code called on certainly-not-hugetlb code paths."
>>>>
>>>>
>>>> We now have to do a page_folio(page) and then test for hugetlb.
>>>>
>>>> return folio_test_hugetlb(page_folio(page));
>>>>
>>>> Nowadays, folio_test_hugetlb() will be faster than at c0bff412e6 times, so
>>>> maybe at least part of the overhead is gone.
>>>>
>>>
>>> I'll note page_folio expands to a call to _compound_head.
>>>
>>> While _compound_head is declared as an inline, it ends up being big
>>> enough that the compiler decides to emit a real function instead and
>>> real func calls are not particularly cheap.
>>>
>>> I had a brief look with a profiler myself and for single-threaded usage
>>> the func is quite high up there, while it manages to get out with the
>>> first branch -- that is to say there is definitely performance lost for
>>> having a func call instead of an inlined branch.
>>>
>>> The routine is deinlined because of a call to page_fixed_fake_head,
>>> which itself is annotated with always_inline.
>>>
>>> This is of course patchable with minor shoveling.
>>>
>>> I did not go for it because stress-ng results were too unstable for me
>>> to confidently state win/loss.
>>>
>>> But should you want to whack the regression, this is what I would look
>>> into.
>>>
>>
>> This might improve it, at least for small folios I guess:
>>
>> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
>> index 5769fe6e4950..7796ae116018 100644
>> --- a/include/linux/page-flags.h
>> +++ b/include/linux/page-flags.h
>> @@ -1086,7 +1086,7 @@ PAGE_TYPE_OPS(Zsmalloc, zsmalloc, zsmalloc)
>> */
>> static inline bool PageHuge(const struct page *page)
>> {
>> - return folio_test_hugetlb(page_folio(page));
>> + return PageCompound(page) && folio_test_hugetlb(page_folio(page));
>> }
>>
>> /*
>>
>>
>> We would avoid the function call for small folios.
>>
>
> why not massage _compound_head back to an inlineable form instead? for
> all i know you may even register a small win in total
Agreed, likely it will increase code size a bit which is why the
compiler decides to not inline. We could force it with __always_inline.
Finding ways to shrink page_fixed_fake_head() might be even better.
--
Cheers,
David / dhildenb
Powered by blists - more mailing lists