Re: [linus:master] [mm] c0bff412e6: stress-ng.clone.ops_per_sec -2.9% regression

[Mateusz Guzik <mjguzik@gmail.com>]

On Mon, Aug 12, 2024 at 10:12 AM David Hildenbrand <david@redhat.com> wrote:
>
> On 12.08.24 06:49, Mateusz Guzik wrote:
> > On Mon, Aug 12, 2024 at 12:43:08PM +0800, Yin Fengwei wrote:
> >> Hi David,
> >>
> >> On 8/1/24 09:44, David Hildenbrand wrote:
> >>> On 01.08.24 15:37, Mateusz Guzik wrote:
> >>>> On Thu, Aug 1, 2024 at 3:34 PM David Hildenbrand <david@redhat.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:
> >> Do you want us to test this change? Or you have further optimization
> >> ongoing? Thanks.
> >
> > I verified the thing below boots, I have no idea about performance. If
> > it helps it can be massaged later from style perspective.
>
> As quite a lot of setups already run with the vmemmap optimization enabled, I
> wonder how effective this would be (might need more fine tuning, did not look
> at the generated code):
>
>
> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
> index 085dd8dcbea2..7ddcdbd712ec 100644
> --- a/include/linux/page-flags.h
> +++ b/include/linux/page-flags.h
> @@ -233,7 +233,7 @@ static __always_inline int page_is_fake_head(const struct page *page)
>          return page_fixed_fake_head(page) != page;
>   }
>
> -static inline unsigned long _compound_head(const struct page *page)
> +static __always_inline unsigned long _compound_head(const struct page *page)
>   {
>          unsigned long head = READ_ONCE(page->compound_head);
>
>

Well one may need to justify it with bloat-o-meter which is why I did
not just straight up inline the entire thing.

But if you are down to fight opposition of the sort I agree this is
the patch to benchmark. :)
-- 
Mateusz Guzik <mjguzik gmail.com>