lockref scalability on x86-64 vs cpu_relax

[Nicholas Piggin]

On Fri Jan 13, 2023 at 2:15 PM AEST, Linus Torvalds wrote:
> On Thu, Jan 12, 2023 at 9:20 PM Nicholas Piggin <npiggin at gmail.com> wrote:
> >
> > Actually what we'd really want is an arch specific implementation of
> > lockref.
>
> The problem is mainly that then you need to generate the asm versions
> of all those different CMPXCHG_LOOP()  variants.
>
> They are all fairly simple, though, and it woudln't be hard to make
> the current lib/lockref.c just be the generic fallback if you don't
> have an arch-specific one.

Yeah, it doesn't look too onerous so it's probably worth seeing what
the code and some numbers look like here.

> And even if you do have the arch-specific LL/SC version, you'd still
> want the generic fallback for the case where a spinlock isn't a single
> word any more (which happens when the spinlock debugging options are
> on).

You're right, good point.

Thanks,
Nick