Re: [PATCH] KVM: x86/xen: Update Xen CPUID Leaf 4 (tsc info) sub-leaves, if present

[Sean Christopherson]

On Mon, Jun 27, 2022, Durrant, Paul wrote:
> > -----Original Message-----
> [snip]
> > > > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > > > index 00e23dc518e0..8b45f9975e45 100644
> > > > --- a/arch/x86/kvm/x86.c
> > > > +++ b/arch/x86/kvm/x86.c
> > > > @@ -3123,6 +3123,7 @@ static int kvm_guest_time_update(struct kvm_vcpu *v)
> > > >       if (vcpu->xen.vcpu_time_info_cache.active)
> > > >               kvm_setup_guest_pvclock(v, &vcpu->xen.vcpu_time_info_cache, 0);
> > > >       kvm_hv_setup_tsc_page(v->kvm, &vcpu->hv_clock);
> > > > +     kvm_xen_setup_tsc_info(v);
> > >
> > > This can be called inside this if statement, no?
> > >
> > >         if (unlikely(vcpu->hw_tsc_khz != tgt_tsc_khz)) {
> > >
> > >         }
> > >
> 
> I think it ought to be done whenever the shared copy of Xen's vcpu_info is
> updated (it will always match on real Xen) so unconditionally calling it here
> seems reasonable.

But isn't the call pointless if the vCPU's hw_tsc_khz is unchanged?  E.g if the
params were explicitly passed in, then it would look like:

	if (unlikely(vcpu->hw_tsc_khz != tgt_tsc_khz)) {
		kvm_get_time_scale(NSEC_PER_SEC, tgt_tsc_khz * 1000LL,
				   &vcpu->hv_clock.tsc_shift,
				   &vcpu->hv_clock.tsc_to_system_mul);
		vcpu->hw_tsc_khz = tgt_tsc_khz;

		kvm_xen_setup_tsc_info(vcpu, tgt_tsc_khz,
				       vcpu->hv_clock.tsc_shift,
				       vcpu->hv_clock.tsc_to_system_mul);
	}

Explicitly passing in the arguments probably isn't necessary, just use a more
precise name, e.g. kvm_xen_update_tsc_khz(), to make it clear that the update is
limited to TSC frequency changes.

> > > > +{
> > > > +     u32 base = 0;
> > > > +     u32 function;
> > > > +
> > > > +     for_each_possible_hypervisor_cpuid_base(function) {
> > > > +             struct kvm_cpuid_entry2 *entry = kvm_find_cpuid_entry(vcpu, function, 0);
> > > > +
> > > > +             if (entry &&
> > > > +                 entry->ebx == XEN_CPUID_SIGNATURE_EBX &&
> > > > +                 entry->ecx == XEN_CPUID_SIGNATURE_ECX &&
> > > > +                 entry->edx == XEN_CPUID_SIGNATURE_EDX) {
> > > > +                     base = function;
> > > > +                     break;
> > > > +             }
> > > > +     }
> > > > +     if (!base)
> > > > +             return;
> > > > +
> > > > +     function = base | XEN_CPUID_LEAF(3);
> > > > +     vcpu->arch.xen.tsc_info_1 = kvm_find_cpuid_entry(vcpu, function, 1);
> > > > +     vcpu->arch.xen.tsc_info_2 = kvm_find_cpuid_entry(vcpu, function, 2);
> > >
> > > Is it really necessary to cache the leave?  Guest CPUID isn't optimized, but it's
> > > not _that_ slow, and unless I'm missing something updating the TSC frequency and
> > > scaling info should be uncommon, i.e. not performance critical.
> 
> If we're updating the values in the leaves on every entry into the guest (as
> with calls to kvm_setup_guest_pvclock()) then I think the cached pointers are
> worthwhile.

But why would you update on every entry to the guest?   Isn't this a rare operation
if the update is limited to changes in the host CPU's TSC frequency?  Or am I
missing something?