kvm: vmx: Limit guest PMCs to those supported on the host
authorJim Mattson <jmattson@google.com>
Mon, 30 Sep 2019 23:38:54 +0000 (16:38 -0700)
committerPaolo Bonzini <pbonzini@redhat.com>
Tue, 1 Oct 2019 13:15:06 +0000 (15:15 +0200)
KVM can only virtualize as many PMCs as the host supports.

Limit the number of generic counters and fixed counters to the number
of corresponding counters supported on the host, rather than to
INTEL_PMC_MAX_GENERIC and INTEL_PMC_MAX_FIXED, respectively.

Note that INTEL_PMC_MAX_GENERIC is currently 32, which exceeds the 18
contiguous MSR indices reserved by Intel for event selectors. Since
the existing code relies on a contiguous range of MSR indices for
event selectors, it can't possibly work for more than 18 general
purpose counters.

Fixes: f5132b01386b5a ("KVM: Expose a version 2 architectural PMU to a guests")
Signed-off-by: Jim Mattson <jmattson@google.com>
Reviewed-by: Marc Orr <marcorr@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
arch/x86/kvm/vmx/pmu_intel.c

index 4dea0e0..3e9c059 100644 (file)
@@ -262,6 +262,7 @@ static int intel_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
 static void intel_pmu_refresh(struct kvm_vcpu *vcpu)
 {
        struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
+       struct x86_pmu_capability x86_pmu;
        struct kvm_cpuid_entry2 *entry;
        union cpuid10_eax eax;
        union cpuid10_edx edx;
@@ -283,8 +284,10 @@ static void intel_pmu_refresh(struct kvm_vcpu *vcpu)
        if (!pmu->version)
                return;
 
+       perf_get_x86_pmu_capability(&x86_pmu);
+
        pmu->nr_arch_gp_counters = min_t(int, eax.split.num_counters,
-                                       INTEL_PMC_MAX_GENERIC);
+                                        x86_pmu.num_counters_gp);
        pmu->counter_bitmask[KVM_PMC_GP] = ((u64)1 << eax.split.bit_width) - 1;
        pmu->available_event_types = ~entry->ebx &
                                        ((1ull << eax.split.mask_length) - 1);
@@ -294,7 +297,7 @@ static void intel_pmu_refresh(struct kvm_vcpu *vcpu)
        } else {
                pmu->nr_arch_fixed_counters =
                        min_t(int, edx.split.num_counters_fixed,
-                               INTEL_PMC_MAX_FIXED);
+                             x86_pmu.num_counters_fixed);
                pmu->counter_bitmask[KVM_PMC_FIXED] =
                        ((u64)1 << edx.split.bit_width_fixed) - 1;
        }