Re: [PATCH 2/2] alloc_tag: skip pgalloc_tag_swap if profiling is disabled

[Suren Baghdasaryan]

On Fri, Dec 27, 2024 at 9:28 AM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
>
> On Thu, Dec 26, 2024 at 11:59 PM Andrew Morton
> <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > On Thu, 26 Dec 2024 16:56:00 -0800 Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
> >
> > > On Thu, Dec 26, 2024 at 4:23 PM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> > > >
> > > > On Thu, 26 Dec 2024 15:07:39 -0800 Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
> > > >
> > > > > On Thu, Dec 26, 2024 at 3:01 PM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> > > > > >
> > > > > > On Thu, 26 Dec 2024 13:16:39 -0800 Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
> > > > > >
> > > > > > > When memory allocation profiling is disabled, there is no need to swap
> > > > > > > allocation tags during migration. Skip it to avoid unnecessary overhead.
> > > > > > >
> > > > > > > Fixes: e0a955bf7f61 ("mm/codetag: add pgalloc_tag_copy()")
> > > > > > > Signed-off-by: Suren Baghdasaryan <surenb@xxxxxxxxxx>
> > > > > > > Cc: stable@xxxxxxxxxxxxxxx
> > > > > >
> > > > > > Are these changes worth backporting?  Some indication of how much
> > > > > > difference the patches make would help people understand why we're
> > > > > > proposing a backport.
> > > > >
> > > > > The first patch ("alloc_tag: avoid current->alloc_tag manipulations
> > > > > when profiling is disabled") I think is worth backporting. It
> > > > > eliminates about half of the regression for slab allocations when
> > > > > profiling is disabled.
> > > >
> > > > um, what regression?  The changelog makes no mention of this.  Please
> > > > send along a suitable Reported-by: and Closes: and a summary of the
> > > > benefits so that people can actually see what this patch does, and why.
> > >
> > > Sorry, I should have used "overhead" instead of "regression".
> > > When one sets CONFIG_MEM_ALLOC_PROFILING=y, the code gets instrumented
> > > and even if profiling is turned off, it still has a small performance
> > > cost minimized by the use of mem_alloc_profiling_key static key. I
> > > found a couple of places which were not protected with
> > > mem_alloc_profiling_key, which means that even when profiling is
> > > turned off, the code is still executed. Once I added these checks, the
> > > overhead of the mode when memory profiling is enabled but turned off
> > > went down by about 50%.
> >
> > Well, a 50% reduction in a 0.0000000001% overhead ain't much.
>
> I wish the overhead was that low :)
>
> I ran more comprehensive testing on Pixel 6 on Big, Medium and Little cores:
>
>                  Overhead before fixes            Overhead after fixes
>                  slab alloc      page alloc          slab alloc      page alloc
> Big               6.21%           5.32%                3.31%          4.93%
> Medium       4.51%           5.05%                3.79%          4.39%
> Little            7.62%           1.82%                6.68%          1.02%

Note, this is an allocation microbenchmark doing allocations in a
tight loop. Not a really realistic scenario and useful only to make
performance comparisons.

>
>
> > But I
> > added the final sentence to the changelog.
> >
> > It still doesn't tell us the very simple thing which we're all eager to
> > know: how much faster did the kernel get??