On Fri, Jun 16, 2023, Xiong Zhang wrote:
> Perf defines four types of perf event: per cpu pinned event, per process
> pinned event, per cpu event, per process event, their prioirity are from
> high to low. vLBR event is per process pinned event. So durng vm exit
> handler, if vLBR event preempts perf low priority LBR event, perf will
> disable LBR and let guest control LBR, or if vLBR event is preempted by
> perf high priority LBR event, perf will enable LBR. In a word LBR status
> may be changed during vm exit handler.
>
> MSR_IA32_DEBUGCTLMSR[0] controls LBR enabling, kvm saves its value into
> vmx->host_debugctlmsr in vcpu_load(), and kvm restores its value from
> vmx->host_debugctlmsr after vm exit immediately. Since
> MSR_IA32_DEBUGCTLMSR[0] could be changed during vm exit handler, the
> saved value vmx->host_debugctlmsr could be wrong. So this commit saves
> MSR_IA32_DEBUGCTLMSR into vmx->host_debugctlmsr just before vm entry to
> reflect the real hardware value.
>
> Signed-off-by: Xiong Zhang <xiong.y.zhang@xxxxxxxxx>
> ---
> arch/x86/kvm/vmx/vmx.c | 5 +----
> 1 file changed, 1 insertion(+), 4 deletions(-)
>
> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> index 44fb619803b8..5ca61a26d0d7 100644
> --- a/arch/x86/kvm/vmx/vmx.c
> +++ b/arch/x86/kvm/vmx/vmx.c
> @@ -1459,13 +1459,9 @@ void vmx_vcpu_load_vmcs(struct kvm_vcpu *vcpu, int cpu,
> */
> static void vmx_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
> {
> - struct vcpu_vmx *vmx = to_vmx(vcpu);
> -
> vmx_vcpu_load_vmcs(vcpu, cpu, NULL);
>
> vmx_vcpu_pi_load(vcpu, cpu);
> -
> - vmx->host_debugctlmsr = get_debugctlmsr();
> }
>
> static void vmx_vcpu_put(struct kvm_vcpu *vcpu)
> @@ -7273,6 +7269,7 @@ static fastpath_t vmx_vcpu_run(struct kvm_vcpu *vcpu)
> atomic_switch_perf_msrs(vmx);
> if (intel_pmu_lbr_is_enabled(vcpu))
> vmx_passthrough_lbr_msrs(vcpu);
> + vmx->host_debugctlmsr = get_debugctlmsr();
Reading DEBUG_CTL on every VM-Entry is either unnecessary or insufficient. If
the DEBUG_CTL value is being changed synchronously, then just fix whatever KVM
path leads to a change in the host avlue. If DEBUG_CTL is being changed
asynchronously, then I'm guessing the change is coming from NMI context, which
means that KVM is buggy no matter how close we put this to VM-Enter.
