Re: [PATCH v3 1/4] x86/clear_page: extend clear_page*() for multi-page clearing

[Mateusz Guzik <mjguzik@gmail.com>]

On Tue, Apr 15, 2025 at 10:02 PM Ankur Arora <ankur.a.arora@oracle.com> wrote:
>
>
> Mateusz Guzik <mjguzik@gmail.com> writes:
>
> > On Tue, Apr 15, 2025 at 8:14 AM Ankur Arora <ankur.a.arora@oracle.com> wrote:
> >>
> >>
> >> Mateusz Guzik <mjguzik@gmail.com> writes:
> >> > With that sucker out of the way, an optional quest is to figure out if
> >> > rep stosq vs rep stosb makes any difference for pages -- for all I know
> >> > rep stosq is the way. This would require testing on quite a few uarchs
> >> > and I'm not going to blame anyone for not being interested.
> >>
> >> IIRC some recent AMD models (Rome?) did expose REP_GOOD but not ERMS.
> >>
> >
> > The uarch does not have it or the bit magically fails to show up?
> > Worst case, should rep stosb be faster on that uarch, the kernel can
> > pretend the bit is set.
>
> It's a synthetic bit so the uarch has both. I think REP STOSB is optimized
> post FSRS (AIUI Zen3)
>
>         if (c->x86 >= 0x10)
>                 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
>
>         /* AMD FSRM also implies FSRS */
>         if (cpu_has(c, X86_FEATURE_FSRM))
>                 set_cpu_cap(c, X86_FEATURE_FSRS);
>
>
> >> > Let's say nobody bothered OR rep stosb provides a win. In that case this
> >> > can trivially ALTERNATIVE between rep stosb and rep stosq based on ERMS,
> >> > no func calls necessary.
> >>
> >> We shouldn't need any function calls for ERMS and REP_GOOD.
> >>
> >> I think something like this untested code should work:
> >>
> >>         asm volatile(
> >>             ALTERNATIVE_2("call clear_pages_orig",
> >>                           "rep stosb", X86_FEATURE_REP_GOOD,
> >>                           "shrl $3,%ecx; rep stosq", X86_FEATURE_ERMS,
> >>                           : "+c" (size), "+D" (addr), ASM_CALL_CONSTRAINT
> >>                           : "a" (0)))
> >>
> >
> > That's what I'm suggesting, with one difference: whack
> > clear_pages_orig altogether.
>
> What do we gain by getting rid of it? Maybe there's old hardware with
> unoptimized rep; stos*.
>

The string routines (memset, memcpy et al) need a lot of love and
preferably nobody would bother spending time placating non-rep users
while sorting them out.

According to wiki the AMD CPUs started with REP_GOOD in 2007, meaning
you would need something even older than that to not have it. Intel is
presumably in a similar boat.

So happens gcc spent several years emitting inlined rep stosq and rep
movsq, so either users don't care or there are no users (well
realistically someone somewhere has a machine like that in the garage,
but fringe cases are not an argument).

rep_movs_alternative already punts to rep mov ignoring the issue of
REP_GOOD for some time now (admittedly, I removed the non-rep support
:P) and again there are no pitchforks (that I had seen).

So I think it would be best for everyone in the long run to completely
reap out the REP_GOOD thing. For all I know the kernel stopped booting
on machines with such uarchs long time ago for unrelated reasons.

As far as this specific patchset goes, it's just a waste of testing to
make sure it still works, but I can't *insist* on removing the
routine. I guess it is x86 maintainers call whether to whack this.
-- 
Mateusz Guzik <mjguzik gmail.com>