Re: [PATCH] KVM: x86/pmu: Return KVM_MSR_RET_INVALID for invalid PMU MSR access

["Mi, Dapeng" <dapeng1.mi@linux.intel.com>]

On 7/10/2024 2:45 AM, Sean Christopherson wrote:
> On Tue, Jul 09, 2024, Dapeng Mi wrote:
>> Return KVM_MSR_RET_INVALID instead of 0 to inject #GP to guest for all
>> invalid PMU MSRs access
>>
>> Currently KVM silently drops the access and doesn't inject #GP for some
>> invalid PMU MSRs like MSR_P6_PERFCTR0/MSR_P6_PERFCTR1,
>> MSR_P6_EVNTSEL0/MSR_P6_EVNTSEL1, but KVM still injects #GP for all other
>> invalid PMU MSRs. This leads to guest see different behavior on invalid
>> PMU access and may confuse guest.
> This is by design.  I'm not saying it's _good_ design, but it is very much
> intended.  More importantly, it's established behavior, i.e. having KVM inject
> #GP could break existing setups.
>
>> This behavior is introduced by the
>> 'commit 5753785fa977 ("KVM: do not #GP on perf MSR writes when vPMU is disabled")'
>> in 2012. This commit seems to want to keep back compatible with weird
>> behavior of some guests in vPMU disabled case,
> Ya, because at the time, guest kernels hadn't been taught to play nice with
> unexpected virtualization setups, i.e. VMs without PMUs.
>
>> but strongly suspect if it's still available nowadays.
> I don't follow this comment.
>
>> Since Perfmon v6 starts, the GP counters could become discontinuous on
>> HW, It's possible that HW doesn't support GP counters 0 and 1.
>> Considering this situation KVM should inject #GP for all invalid PMU MSRs
>> access.
> IIUC, the behavior you want is inject a #GP if the vCPU has a PMU and the MSR is
> not valid.  We can do that and still maintain backwards compatibility, hopefully
> without too much ugliness (maybe even an improvement!).
>
> This? (completely untested)

Seems no better method. Would adopt this. Thanks.

>
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 5aa7581802f7..b5e95e5f1f32 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -4063,9 +4063,8 @@ int kvm_set_msr_common(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
>         case MSR_P6_PERFCTR0 ... MSR_P6_PERFCTR1:
>         case MSR_K7_EVNTSEL0 ... MSR_K7_EVNTSEL3:
>         case MSR_P6_EVNTSEL0 ... MSR_P6_EVNTSEL1:
> -               if (kvm_pmu_is_valid_msr(vcpu, msr))
> -                       return kvm_pmu_set_msr(vcpu, msr_info);
> -
> +               if (vcpu_to_pmu(vcpu)->version)
> +                       goto default_handler;
>                 if (data)
>                         kvm_pr_unimpl_wrmsr(vcpu, msr, data);
>                 break;
> @@ -4146,6 +4145,7 @@ int kvm_set_msr_common(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
>                 break;
>  #endif
>         default:
> +default_handler:
>                 if (kvm_pmu_is_valid_msr(vcpu, msr))
>                         return kvm_pmu_set_msr(vcpu, msr_info);
>  
> @@ -4251,8 +4251,8 @@ int kvm_get_msr_common(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
>         case MSR_K7_PERFCTR0 ... MSR_K7_PERFCTR3:
>         case MSR_P6_PERFCTR0 ... MSR_P6_PERFCTR1:
>         case MSR_P6_EVNTSEL0 ... MSR_P6_EVNTSEL1:
> -               if (kvm_pmu_is_valid_msr(vcpu, msr_info->index))
> -                       return kvm_pmu_get_msr(vcpu, msr_info);
> +               if (vcpu_to_pmu(vcpu)->version)
> +                       goto default_handler;
>                 msr_info->data = 0;
>                 break;
>         case MSR_IA32_UCODE_REV:
> @@ -4505,6 +4505,7 @@ int kvm_get_msr_common(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
>                 break;
>  #endif
>         default:
> +default_handler:
>                 if (kvm_pmu_is_valid_msr(vcpu, msr_info->index))
>                         return kvm_pmu_get_msr(vcpu, msr_info);
>  
> diff --git a/tools/testing/selftests/kvm/x86_64/pmu_counters_test.c b/tools/testing/selftests/kvm/x86_64/pmu_counters_test.c
> index 96446134c00b..0de606b542ac 100644
> --- a/tools/testing/selftests/kvm/x86_64/pmu_counters_test.c
> +++ b/tools/testing/selftests/kvm/x86_64/pmu_counters_test.c
> @@ -344,7 +344,8 @@ static void guest_test_rdpmc(uint32_t rdpmc_idx, bool expect_success,
>  static void guest_rd_wr_counters(uint32_t base_msr, uint8_t nr_possible_counters,
>                                  uint8_t nr_counters, uint32_t or_mask)
>  {
> -       const bool pmu_has_fast_mode = !guest_get_pmu_version();
> +       const u8 pmu_version = guest_get_pmu_version();
> +       const bool pmu_has_fast_mode = !pmu_version;
>         uint8_t i;
>  
>         for (i = 0; i < nr_possible_counters; i++) {
> @@ -363,12 +364,13 @@ static void guest_rd_wr_counters(uint32_t base_msr, uint8_t nr_possible_counters
>                 const bool expect_success = i < nr_counters || (or_mask & BIT(i));
>  
>                 /*
> -                * KVM drops writes to MSR_P6_PERFCTR[0|1] if the counters are
> -                * unsupported, i.e. doesn't #GP and reads back '0'.
> +                * KVM drops writes to MSR_P6_PERFCTR[0|1] if the vCPU doesn't
> +                * have a PMU, i.e. doesn't #GP and reads back '0'.
>                  */
>                 const uint64_t expected_val = expect_success ? test_val : 0;
> -               const bool expect_gp = !expect_success && msr != MSR_P6_PERFCTR0 &&
> -                                      msr != MSR_P6_PERFCTR1;
> +               const bool expect_gp = !expect_success &&
> +                                      (pmu_version ||
> +                                       (msr != MSR_P6_PERFCTR0 && msr != MSR_P6_PERFCTR1));
>                 uint32_t rdpmc_idx;
>                 uint8_t vector;
>                 uint64_t val;
>