Re: [PATCH v3 1/3] RISC-V: Issue a local tlbflush if possible.

From: Atish Patra
Date: Thu Aug 22 2019 - 09:44:00 EST


On Thu, 2019-08-22 at 06:27 +0200, hch@xxxxxx wrote:
> On Thu, Aug 22, 2019 at 04:01:24AM +0000, Atish Patra wrote:
> > The downside of this is that for every !cmask case in true SMP
> > (more
> > common probably) it will execute 2 extra cpumask instructions. As
> > tlbflush path is in performance critical path, I think we should
> > favor
> > more common case (SMP with more than 1 core).
>
> Actually, looking at both the current mainline code, and the code
> from my
> cleanups tree I don't think remote_sfence_vma / __sbi_tlb_flush_range
> can ever be called with NULL cpumask, as we always have a valid mm.
>

Yes. You are correct.

As both cpumask functions here will crash if cpumask is null, we should
probably leave a harmless comment to warn the consequeunce of cpumask
being null.

> So this is a bit of a moot point, and we can drop andling that case
> entirely. With that we can also use a simple if / else for the local
> cpu only vs remote case.

Done.

> Btw, what was the reason you didn't like
> using cpumask_any_but like x86, which should be more efficient than
> cpumask_test_cpu + hweigt?

I had it in v2 patch but removed as it can potentially return garbage
value if cpumask is empty.

However, we are already checking empty cpumask before the local cpu
check. I will replace cpumask_test_cpu + hweight with
cpumask_any_but().

--
Regards,
Atish