Re: [PATCH v9 05/12] KVM: x86/pmu: Support PERF_METRICS MSR in mediated vPMU

[Jim Mattson]

On Fri, Sep 18, 2026 at 12:39 PM Zide Chen <zide.chen@xxxxxxxxx> wrote:
>
> +static inline bool kvm_vcpu_has_perf_metrics(struct kvm_vcpu *vcpu)
> +{
> +	return kvm_vcpu_get_perf_caps(vcpu) & PERF_CAP_PERF_METRICS;
> +}

The SDM, volume 3, section 22.3.9.3, says:

> IA32_PERF_GLOBAL_CTRL. EN_PERF_METRICS[48]: If this bit is set *and*
> fixed-function performance-monitoring counter 3 is enabled, built-in
> performance metrics are enabled. [emphasis mine]

and

> The values in MSR_PERF_METRICS are derived from fixed-function
> performance-monitoring counter 3.

Hence, a dependency on fixed counter 3 belongs in this predicate.
Perhaps something like:

	static inline bool kvm_pmu_has_fixed_counter(struct kvm_pmu *pmu,
						     unsigned int idx)
	{
		return idx < pmu->nr_arch_fixed_counters;
	}

	static inline bool kvm_vcpu_has_perf_metrics(struct kvm_vcpu *vcpu)
	{
		return (kvm_vcpu_get_perf_caps(vcpu) & PERF_CAP_PERF_METRICS) &&
		       kvm_pmu_has_fixed_counter(vcpu_to_pmu(vcpu),
						 INTEL_PMC_IDX_FIXED_SLOTS -
						 INTEL_PMC_IDX_FIXED);
	}

with intel_pmu_refresh() calling kvm_vcpu_has_perf_metrics() instead of
testing perf_capabilities directly. (nr_arch_fixed_counters has already
been computed at that point; all_valid_pmc_idx has not.)

KVM only supports contiguous fixed counters today (see patch 2), so a
simple comparison against nr_arch_fixed_counters suffices. However, if KVM
ever honors the CPUID.0AH:ECX fixed counter bitmap, a separate helper
keeps that assumption in one place.

As it stands, if userspace sets PERF_CAP_PERF_METRICS for a version 1
vPMU, MSR_PERF_METRICS is passed through (kvm_vcpu_has_mediated_pmu()
only requires a non-zero version), but intel_pmu_refresh() bails out
before it gets to the PERF_METRICS code. With patch 11, that leaves
perf_metrics_rsvd at ~0ull, so a value the guest writes directly to
hardware can be saved with KVM_GET_MSRS, but not restored with
KVM_SET_MSRS.

Since nr_arch_fixed_counters is 0 for a v1 vPMU, the above handles that
case too. It also takes care of the METRICS_CLEAR_EN unmasking without
fixed counter 3 that Sashiko flagged on patch 12, and it obviates the
NULL pmc check in intel_emulate_rdpmc().

[...]

> +	case MSR_PERF_METRICS:
> +		pmu->perf_metrics = data;
> +		break;

On a part with only four metrics (e.g. Ice Lake), userspace can stuff
bits 63:32 into pmu->perf_metrics here, and the unchecked wrmsrq() in
intel_mediated_pmu_load() will then #GP on every VM-entry. Patch 11
isn't an optional follow-on; it needs to be folded into this patch.