On 2/14/26 11:03 AM, ChaseKnowlden wrote:
> The static variables pmu_version, num_pmu_gp_counters, and
> num_pmu_fixed_counters persist across different VCPU initializations.
> When kvm_init_pmu_info() returns early (e.g., when PMU is disabled or
> there's a vendor mismatch), these variables retain stale values from
> previous initializations.
>
> This causes crashes during guest reboots, particularly with Windows XP,
> when kvm_put_msrs() attempts to write PMU MSRs based on the stale
> pmu_version value for a CPU that doesn't actually support PMU.
Can you provide more details how this may happen? The guest reboots won't change
CPUID. All vCPUs may share the same configuration.
Why "particularly with Windows XP"?
Thank you very much!
Dongli Zhang
>
> Fix this by resetting the PMU state variables to 0 at the start of
> kvm_init_pmu_info() to ensure clean initialization for each VCPU.
>
> Fixes: 4c7f05232c ("target/i386/kvm: reset AMD PMU registers during VM reset")
> Signed-off-by: GitHub Copilot <[email protected]>
>
> Co-authored-by: Copilot <[email protected]>
> ---
> target/i386/kvm/kvm.c | 8 ++++++++
> 1 file changed, 8 insertions(+)
>
> diff --git a/target/i386/kvm/kvm.c b/target/i386/kvm/kvm.c
> index 9f1a4d4cbb..c636f1f487 100644
> --- a/target/i386/kvm/kvm.c
> +++ b/target/i386/kvm/kvm.c
> @@ -2193,6 +2193,14 @@ static void kvm_init_pmu_info(struct kvm_cpuid2
> *cpuid, X86CPU *cpu)
> {
> CPUX86State *env = &cpu->env;
>
> + /*
> + * Reset PMU state to avoid stale values from previous VCPU
> + * initializations affecting subsequent ones.
> + */
> + pmu_version = 0;
> + num_pmu_gp_counters = 0;
> + num_pmu_fixed_counters = 0;
> +
> /*
> * If KVM_CAP_PMU_CAPABILITY is not supported, there is no way to
> * disable the AMD PMU virtualization.
