Re: [PATCH RFC 01/14] x86/asm: add iosubmit_cmds512() based on movdir64b CPU instruction

[Dan Williams]

On Fri, Nov 22, 2019 at 10:44 AM Borislav Petkov <bp@xxxxxxxxx> wrote:
>
> On Fri, Nov 22, 2019 at 09:20:39AM -0800, Dan Williams wrote:
> > For those cases the thought would be to have memset512() for case1 and
> > __iowrite512_copy() for case3. Where memset512() writes a
> > non-incrementing source to an incrementing destination, and
> > __iowrite512_copy() copies an incrementing source to an incrementing
> > destination. Those 2 helpers *would* have fallbacks, but with the
> > option to use something like cpu_has_write512() to check in advance
> > whether those routines will fallback, or not.
> >
> > That can be a discussion for a future patchset when those users arrive.
>
> Oh, sure, of course.
>
> My only angle is very simple: if the MOVDIR* et al is only supported on
> upcoming Intel platforms and looking at the use cases:
>
> 1. clear poison/MKTME
> 3. copy iomem in big chunks
>
> I'm going to venture a guess that those two cases are going to be
> happening only on Intel platforms which already support MODVIR*. So
> wouldn't really need to do any generic helpers because those use cases
> are very specific already. Which would make your feature detection a
> one-time, driver-init time thing anyway...
>
> Unless I misunderstand those cases and there really is a use case
> where the thing would fallback and the fallback would really be for an
> "unenlightened" platform without that MOVDIR* hw support...?

At least for something like __iowrite512_copy() it would indeed be
something an unenlightened driver could call. Those drivers would
simply be looking for opportunistic efficiency and could tolerate a
fallback. Just like current __iowrite64_copy() users don't care if
that routine falls back internally.