Re: [PATCH v1 1/2] x86: Implement sched_yield syscall for x86 only.

[Noah Goldstein <goldstein.w.n@gmail.com>]

On Thu, Jun 8, 2023 at 7:08 AM Adhemerval Zanella Netto
<adhemerval.zanella@linaro.org> wrote:
>
>
>
> On 08/06/23 08:43, Florian Weimer wrote:
> > * Noah Goldstein via Libc-alpha:
> >
> >> We slightly optimize it by using `vzeroall` before the actual syscall.
> >> This returns the SSE, AVX, and ZMM_HI256 xsave/xrstor states to the
> >> init-state which allows the imminent context switch to skip
> >> saving/restoring those states.
> >
> > Surely there is a better way to implement this, enabling something
> > similar for all system calls issued by libc on the kernel side?  It
> > changes userspace ABI, so it has to be opt-in.  Maybe it could be an
> > additional flag in the system call number, indicating that it is safe
> > to zap the vector state if it is beneficial.
It seems like a much bigger change than is needed.
>
> Agree, trying to implement it on userland seems really hacky.  It means
> to potentially override and/or add an ifunc variant to any syscall that
> can potentially trigger a context switch; besides adding arch-specific
> implementation for something the kernel already has the information
> (so it can rewrite the syscall entrypoint depending of the ISA).

I don't think we need/want this for every syscall. Only the syscalls
where there is a high probability of a proper ctx switch and the calling
process going back to the schedule loop.
Otherwise the kernel generally just takes care to not touch vector registers
and doesn't bother with the save/restore.