lockref scalability on x86-64 vs cpu_relax

[Linus Torvalds]

On Thu, Jan 12, 2023 at 6:31 PM Luck, Tony <tony.luck at intel.com> wrote:
>
> There's not much "simultaneous" in the SMT on ia64.

Oh, I forgot about the whole SoEMT fiasco.

Yeah, that might make ia64 act a bit differently here.

But I don't think anybody cares any more, so I don't think that merits
making this a per-architecture choice.

The s390 people hated cpu_relax() here, but for them it was really
because it was bad *everywhere*, and they just made it a no-op (see
commit 22b6430d3665 "locking/core, s390: Make cpu_relax() a barrier
again"). There had been a (failed) attempt at "cpu_relax_lowlatency()"
for the s390 issues.

                  Linus