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

[Limonciello, Mario]

[Public]



> -----Original Message-----
> From: Dave Hansen <dave.hansen@xxxxxxxxx>
> Sent: Tuesday, April 5, 2022 10:10
> To: Peter Zijlstra <peterz@xxxxxxxxxxxxx>; Karny, Wyes
> <Wyes.Karny@xxxxxxx>
> Cc: linux-kernel@xxxxxxxxxxxxxxx; Carroll, Lewis <Lewis.Carroll@xxxxxxx>;
> Limonciello, Mario <Mario.Limonciello@xxxxxxx>; Shenoy, Gautham Ranjal
> <gautham.shenoy@xxxxxxx>; Narayan, Ananth
> <Ananth.Narayan@xxxxxxx>; Rao, Bharata Bhasker <bharata@xxxxxxx>;
> len.brown@xxxxxxxxx; x86@xxxxxxxxxx; tglx@xxxxxxxxxxxxx;
> 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
> 
> On 4/5/22 07:07, Peter Zijlstra wrote:
> > On Tue, Apr 05, 2022 at 06:30:21PM +0530, Wyes Karny wrote:
> >> +static inline bool early_mwait_supported(const struct cpuinfo_x86 *c)
> >> +{
> >> +	if (c->x86_vendor == X86_VENDOR_INTEL)
> >> +		return true;
> >> +
> >> +	if (c->x86_vendor == X86_VENDOR_AMD && cpu_has(c,
> X86_FEATURE_ZEN))
> > What about Hygon? For some reason you guys don't co-ordinate and we
> end
> > up getting endless 'make-same' patches, sometimes separated by years :/
> 
> Believe it or not Hygon seems to work OK with this because:
> 
> > static void init_hygon(struct cpuinfo_x86 *c)
> > {
> ...
> >         set_cpu_cap(c, X86_FEATURE_ZEN);
> 
> I do worry a bit though that using X86_FEATURE_ZEN is going to bite us
> long-term.  It currently claims to be set for "family 0x17 or above":
> 
> > #define X86_FEATURE_ZEN                 ( 7*32+28) /* "" CPU is AMD family
> 0x17 or above (Zen) */
> 
> But then it goes and gets used in side-channel defense:
> 
> >         if (!static_cpu_has(X86_FEATURE_ZEN)) {
> >                 msr |= ssbd_tif_to_amd_ls_cfg(tifn);
> >                 wrmsrl(MSR_AMD64_LS_CFG, msr);
> >                 return;
> >         }
> 
> This seem _bit_ at odds with the commit message (and the AMD SSBD
> whitepaper):
> 
> >     Add the necessary synchronization logic for AMD family 17H.
> 
> So, is X86_FEATURE_ZEN for family==0x17, or family>=0x17?

There are Zen family CPUs and APUs from family 0x19.  Perhaps at the
time of the whitepaper there weren't yet though.