Paolo Bonzini <pbonzini@xxxxxxxxxx> writes: > On 7/21/22 23:43, Sean Christopherson wrote: >> diff --git a/arch/x86/kvm/hyperv.c b/arch/x86/kvm/hyperv.c >> index c284a605e453..ca91547034e4 100644 >> --- a/arch/x86/kvm/hyperv.c >> +++ b/arch/x86/kvm/hyperv.c >> @@ -1282,7 +1282,7 @@ static bool hv_check_msr_access(struct kvm_vcpu_hv *hv_vcpu, u32 msr) >> case HV_X64_MSR_SYNDBG_OPTIONS: >> case HV_X64_MSR_SYNDBG_CONTROL ... HV_X64_MSR_SYNDBG_PENDING_BUFFER: >> return hv_vcpu->cpuid_cache.features_edx & >> - HV_FEATURE_DEBUG_MSRS_AVAILABLE; >> + HV_ACCESS_DEBUG_MSRS; >> default: >> break; >> } >> > > Yes, and this will need some kind of hack in QEMU to expose both CPUID > bits. Fortunately hv-syndbg shouldn't be in much use in the wild, so I > think we can avoid quirks etc. Properly behaving VMM should always expose both bits. I'm not sure what would it mean if only 'access' bit is present: you can try accessing the missing feature but you get #GP anyway most likely. When the feature is available but 'access' bit is not set -- the result is also #GP. In case we really want to support this behavior in KVM we should probably check *both* bits in hv_check_msr_access() but I don't really see a use-case. I've lazily kept HV_FEATURE_DEBUG_MSRS_AVAILABLE here just to be QEMU compatible. -- Vitaly
