Re: [PATCH v4 3/6] KVM: x86/pmu: Disable counters based on Host-Only/Guest-Only bits in SVM

[Yosry Ahmed <yosry@kernel.org>]

On Mon, Apr 06, 2026 at 06:30:05PM -0700, Sean Christopherson wrote:
> On Thu, Mar 26, 2026, Yosry Ahmed wrote:
> > diff --git a/arch/x86/include/asm/perf_event.h b/arch/x86/include/asm/perf_event.h
> > index ff5acb8b199b0..5961c002b28eb 100644
> > --- a/arch/x86/include/asm/perf_event.h
> > +++ b/arch/x86/include/asm/perf_event.h
> > @@ -60,6 +60,8 @@
> >  #define AMD64_EVENTSEL_INT_CORE_ENABLE			(1ULL << 36)
> >  #define AMD64_EVENTSEL_GUESTONLY			(1ULL << 40)
> >  #define AMD64_EVENTSEL_HOSTONLY				(1ULL << 41)
> > +#define AMD64_EVENTSEL_HOST_GUEST_MASK			\
> > +	(AMD64_EVENTSEL_HOSTONLY | AMD64_EVENTSEL_GUESTONLY)
> >  
> >  #define AMD64_EVENTSEL_INT_CORE_SEL_SHIFT		37
> >  #define AMD64_EVENTSEL_INT_CORE_SEL_MASK		\
> > diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
> > index d6ac3c55fce55..e35d598f809a2 100644
> > --- a/arch/x86/kvm/pmu.c
> > +++ b/arch/x86/kvm/pmu.c
> > @@ -559,6 +559,7 @@ static int reprogram_counter(struct kvm_pmc *pmc)
> >  
> >  	if (kvm_vcpu_has_mediated_pmu(pmu_to_vcpu(pmu))) {
> >  		kvm_mediated_pmu_refresh_event_filter(pmc);
> > +		kvm_pmu_call(mediated_reprogram_counter)(pmc);
> 
> I would rather make a single call from kvm_pmu_handle_event(), and let the vendor
> deal with mediated vs. legacy.  I want to avoid mediated-specific ops as much as
> possible, and I think kvm_x86_ops.reprogram_counters() would be easier to
> understand overall.

I think this doesn't apply anymore now that most nested transitions
won't be handled through kvm_pmu_handle_event(). Also because we need
kvm_mediated_pmu_refresh_event_filter() to still be called before
re-evaluating H/G bits and EFER.SVME.

I think leave this callback as-is and handle everything through
reprogram_counter(). Export reprogram_counter() and rename it to
kvm_pmu_reprogram_counter(), and end up with something like this:

void __kvm_pmu_handle_nested_transition(struct kvm_vcpu *vcpu, bool defer)
{
	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
	DECLARE_BITMAP(bitmap, X86_PMC_IDX_MAX);

	if (bitmap_empty(pmu->reprogram_on_nested_transition, X86_PMC_IDX_MAX))
		return;

	bitmap_copy(bitmap, pmu->reprogram_pmi, X86_PMC_IDX_MAX);
	bitmap_zero(pmu->reprogram_on_nested_transition, X86_PMC_IDX_MAX);

	BUILD_BUG_ON(sizeof(pmu->reprogram_on_nested_transition) != sizeof(atomic64_t));
	if (defer) {
		atomic64_or(*(s64 *)pmu->reprogram_on_nested_transition,
			    &vcpu_to_pmu(vcpu)->__reprogram_pmi);
		kvm_make_request(KVM_REQ_PMU, vcpu);
		return;
	}

	kvm_for_each_pmc(pmu, pmc, bit, bitmap)
		kvm_pmu_reprogram_counter(pmc);
}

void kvm_pmu_handle_nested_transition(struct kvm_vcpu *vcpu)
{
	__kvm_pmu_handle_nested_transition(vcpu, false);
}

Actually, if that's desired, we can move this logic into SVM code now.
We won't be calling kvm_pmu_handle_nested_transition() from inside
enter_guest_mode() and leave_guest_mode() anyway so that we can only
defer for the svm_leave_nested() path.

So we can move:
- kvm_pmu_handle_nested_transition() to
svm_pmu_handle_nested_transition()
- pmu->reprogram_on_nested_transition to
svm->nested.reprogram_on_nested_transition

Not sure if we want to keep SVM-specific logic in SVM code, or if we
want to keep code generic as much as possible. I can see good arguments
for both stances.

> > +
> > +static void amd_mediated_pmu_reprogram_counter(struct kvm_pmc *pmc)
> > +{
> > +	amd_mediated_pmu_handle_host_guest_bits(pmc);
> 
> And then this doesn't need to be such a wonky wrapper, and the "reprogram on
> nested transition" logic can also clear the entire bitmap instead of doing things
> piecemeal, e.g. it can be something like so in the end:

We can still clear the entire bitmap in one go with the above
suggestion, but we'll keep the wonky wrapper :)

> 
> 	if (!kvm_vcpu_has_mediated_pmu(vcpu))
> 		return;
> 
> 	bitmap_zero(pmu->pmc_reprogram_on_nested_transition, X86_PMC_IDX_MAX);
> 
> 	kvm_for_each_pmc(pmu, pmc, bit, bitmap)
> 		amd_mediated_pmu_handle_host_guest_bits(pmc);