Re: [PATCH 02/15] KVM: x86: Emulate RDPID only if RDTSCP is supported

[Sean Christopherson]

On Mon, May 10, 2021, Maxim Levitsky wrote:
> On Tue, 2021-05-04 at 10:17 -0700, Sean Christopherson wrote:
> > Do not advertise emulation support for RDPID if RDTSCP is unsupported.
> > RDPID emulation subtly relies on MSR_TSC_AUX to exist in hardware, as
> > both vmx_get_msr() and svm_get_msr() will return an error if the MSR is
> > unsupported, i.e. ctxt->ops->get_msr() will fail and the emulator will
> > inject a #UD.
> > 
> > Note, RDPID emulation also relies on RDTSCP being enabled in the guest,
> > but this is a KVM bug and will eventually be fixed.
> > 
> > Fixes: fb6d4d340e05 ("KVM: x86: emulate RDPID")
> > Cc: stable@xxxxxxxxxxxxxxx
> > Signed-off-by: Sean Christopherson <seanjc@xxxxxxxxxx>
> > ---
> >  arch/x86/kvm/cpuid.c | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> > 
> > diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c
> > index f765bf7a529c..c96f79c9fff2 100644
> > --- a/arch/x86/kvm/cpuid.c
> > +++ b/arch/x86/kvm/cpuid.c
> > @@ -637,7 +637,8 @@ static int __do_cpuid_func_emulated(struct kvm_cpuid_array *array, u32 func)
> >  	case 7:
> >  		entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
> >  		entry->eax = 0;
> > -		entry->ecx = F(RDPID);
> > +		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
> > +			entry->ecx = F(RDPID);
> >  		++array->nent;
> >  	default:
> >  		break;
> 
> Just to make sure that I understand this correctly:
> 
> This is what I know:
> 
> Both RDTSCP and RDPID are instructions that read IA32_TSC_AUX
> (and RDTSCP also reads the TSC).
> 
> Both instructions have their own CPUID bits (X86_FEATURE_RDPID, X86_FEATURE_RDTSCP)
> If either of these CPUID bits are present, IA32_TSC_AUX should be supported.

Yep.

> RDPID is a newer feature, thus I can at least for the sanity sake assume that
> usually a CPU will either have neither of the features, have only RDTSCP,
> and IA32_AUX, or have both RDSCP and RDPID.

Yep.

> If not supported in hardware KVM only emulates RDPID as I see.

Yep.

> Why btw? Performance wise guest that only wants the IA32_AUX in userspace,
> is better to use RDTSCP and pay the penalty of saving/restoring of the
> unwanted registers, than use RDPID with a vmexit.

FWIW, Linux doesn't even fall back to RDTSCP.  If RDPID isn't supported, Linux
throws the info into the limit of a dummy segment in the GDT and uses LSL to get
at the data.  Turns out that RDTSCP is too slow for its intended use case :-)

> My own guess for an answer to this question is that RDPID emulation is there
> to aid migration from a host that does support RDPID to a host that doesn't.

That's my assumption as well.  Paolo's commit is a bit light on why emulation
was added in the first place, but emulating to allow migrating to old hardware
is the only motivation I can come up with.

commit fb6d4d340e0532032c808a9933eaaa7b8de435ab
Author: Paolo Bonzini <pbonzini@xxxxxxxxxx>
Date:   Tue Jul 12 11:04:26 2016 +0200

    KVM: x86: emulate RDPID

    This is encoded as F3 0F C7 /7 with a register argument.  The register
    argument is the second array in the group9 GroupDual, while F3 is the
    fourth element of a Prefix.

> Having said all that, assuming that we don't want to emulate the RDTSCP too,
> when it is not supported, then this patch does make sense.
> 
> Reviewed-by: Maxim Levitsky <mlevitsk@xxxxxxxxxx>
> 
> Best regards,
> 	Maxim Levitsky
> 
>