RE: [PATCH] x86: Prefer MWAIT over HALT on AMD processors

[Limonciello, Mario]

[Public]



> -----Original Message-----
> From: Limonciello, Mario
> Sent: Thursday, April 7, 2022 20:24
> To: Dave Hansen <dave.hansen@xxxxxxxxx>; Thomas Gleixner
> <tglx@xxxxxxxxxxxxx>; Peter Zijlstra <peterz@xxxxxxxxxxxxx>; Karny, Wyes
> <Wyes.Karny@xxxxxxx>
> Cc: linux-kernel@xxxxxxxxxxxxxxx; Carroll, Lewis <Lewis.Carroll@xxxxxxx>;
> Shenoy, Gautham Ranjal <gautham.shenoy@xxxxxxx>; Narayan, Ananth
> <Ananth.Narayan@xxxxxxx>; Rao, Bharata Bhasker <bharata@xxxxxxx>;
> len.brown@xxxxxxxxx; x86@xxxxxxxxxx; mingo@xxxxxxxxxx; bp@xxxxxxxxx;
> dave.hansen@xxxxxxxxxxxxxxx; hpa@xxxxxxxxx; chang.seok.bae@xxxxxxxxx;
> keescook@xxxxxxxxxxxx; metze@xxxxxxxxx;
> zhengqi.arch@xxxxxxxxxxxxx; mark.rutland@xxxxxxx
> Subject: RE: [PATCH] x86: Prefer MWAIT over HALT on AMD processors
> 
> [Public]
> 
> >
> > I don't want to give you too hard of a time on this.  But, that's not
> > architecture, that's just telling folks what the implementation is on
> > *one* CPU model.  Don't get me wrong: these model-specific docs are
> > great, and I wish Intel published something like that.
> >
> > But, the code as written depends on behavior for *all* of family 0x19:
> >
> > >         case 0x17: fallthrough;
> > >         case 0x19: init_amd_zn(c); break;
> >
> > So, while the docs for "family 19h model 01h" are *consistent* with this
> > code, there's also nothing preventing the docs for "family 19h model
> > 02h" from breaking this new MWAIT code.
> >
> > Now, AMD is full of smart folks that aren't going out of their way to
> > try and break existing software.  But, actual documentation of the
> > *architecture* is really preferable to what we have now.  It helps
> > establish a shared hardware/software contract that keeps both sides
> honest.
> 
> Point taken.  As you're probably aware releasing documentation is never a
> quick
> process, but Lewis and Wyes can take this to the powers that be internally to
> see
> what can be done for the future.

This thread did splinter a bit as it was between speculation control and MWAIT.
So I just wanted to complete the thoughts related to the speculation as that came
up.  Having gone back over what is in the outward facing documentation I did want
to share to you that APM volume 2 page 67-70 does include the details
for speculation control and proper architectural discovery of the MSR.
https://www.amd.com/system/files/TechDocs/24593.pdf

> 
> >
> > What's missing is something that says:
> >
> > 	All AMD family 0x17 and 0x19 that enumerate support for MWAIT
> > 	also support this "early MWAIT" implementation
> >
> > That represents a promise from AMD that this can't break in the future.
> >  It both gives us something unambiguous to write code with *and*
> > something to help nudge our hardware and microcode colleagues if they
> > start to do something funny in the future.
> >
> > Just having that in the changelog would be fine, and a genuine
> > improvement over what we have now.  Having it one of the
> > model-independent architecture manuals would be even better.
> 
> OK... I think the commit message and/or a comment is doable in the short
> term for v2.