[PATCH] locking/atomic: Make test_and_*_bit() ordered on failure

[Jon Nettleton]

On Tue, Aug 16, 2022 at 8:02 PM Linus Torvalds
<torvalds at linux-foundation.org> wrote:
>
> On Tue, Aug 16, 2022 at 10:49 AM Jon Nettleton <jon at solid-run.com> wrote:
> >
> > It is moot if Linus has already taken the patch, but with a stock
> > kernel config I am
> > still seeing a slight performance dip but only ~1-2% in the specific
> > tests I was running.
>
> It would be interesting to hear if you can pinpoint in the profiles
> where the time is spent.
>
> It might be some random place that really doesn't care about ordering
> at all, and then we could easily rewrite _that_ particular case to do
> the unordered test explicitly, ie something like
>
> -        if (test_and_set_bit()) ...
> +       if (test_bit() || test_and_set_bit()) ...
>
> or even introduce an explicitly unordered "test_and_set_bit_relaxed()" thing.
>
>                  Linus

This is very interesting, the additional performance overhead doesn't seem
to be coming from within the kernel but from userspace. Comparing patched
and unpatched kernels I am seeing more cycles being taken up by glibc
atomics like __aarch64_cas4_acq  and __aarch64_ldadd4_acq_rel.

I need to test further to see if there is less effect on a system with
less cores,
This is a 16-core Cortex-A72, it is possible this is less of an issue on 4 core
A72's and A53's.

-Jon