Sean Christopherson <sean.j.christopherson@xxxxxxxxx> writes: > On Thu, Jan 16, 2020 at 09:51:47AM +0100, Vitaly Kuznetsov wrote: >> Liran Alon <liran.alon@xxxxxxxxxx> writes: >> >> >> On 16 Jan 2020, at 1:27, Sean Christopherson <sean.j.christopherson@xxxxxxxxx> wrote: >> >> >> >> On Wed, Jan 15, 2020 at 06:10:13PM +0100, Vitaly Kuznetsov wrote: >> >>> With fine grained VMX feature enablement QEMU>=4.2 tries to do KVM_SET_MSRS >> >>> with default (matching CPU model) values and in case eVMCS is also enabled, >> >>> fails. >> >> >> >> As in, Qemu is blindly throwing values at KVM and complains on failure? >> >> That seems like a Qemu bug, especially since Qemu needs to explicitly do >> >> KVM_CAP_HYPERV_ENLIGHTENED_VMCS to enable eVMCS. >> > >> > See: https://patchwork.kernel.org/patch/11316021/ >> > For more context. >> >> Ya, >> >> while it would certainly be possible to require that userspace takes >> into account KVM_CAP_HYPERV_ENLIGHTENED_VMCS (which is an opt-in) when >> doing KVM_SET_MSRS there doesn't seem to be an existing (easy) way to >> figure out which VMX controls were filtered out after enabling >> KVM_CAP_HYPERV_ENLIGHTENED_VMCS: KVM_GET_MSRS returns global >> &vmcs_config.nested values for VMX MSRs (vmx_get_msr_feature()). > > Ah, I was looking at the call to vmx_get_vmx_msr(&vmx->nested.msrs, ...) > in vmx_get_msr(). > > Why not just do this in Qemu? IMO that's not a major ask, e.g. Qemu is > doing a decent amount of manual adjustment anyways. And Qemu isn't even > using the result of KVM_GET_MSRS so I don't think it's fair to say this is > solely KVM's fault. > > diff --git a/target/i386/kvm.c b/target/i386/kvm.c > index 1d10046a6c..6545bb323e 100644 > --- a/target/i386/kvm.c > +++ b/target/i386/kvm.c > @@ -2623,6 +2623,23 @@ static void kvm_msr_entry_add_vmx(X86CPU *cpu, FeatureWordArray f) > MSR_VMX_EPT_UC | MSR_VMX_EPT_WB : 0); > uint64_t fixed_vmx_ept_vpid = kvm_vmx_ept_vpid & fixed_vmx_ept_mask; > > + /* Hyper-V's eVMCS does't support certain features, adjust accordingly. */ > + if (cpu->hyperv_evmcs) { > + f[FEAT_VMX_PINBASED_CTLS] &= ~(VMX_PIN_BASED_VMX_PREEMPTION_TIMER | > + VMX_PIN_BASED_POSTED_INTR); > + f[FEAT_VMX_EXIT_CTLS] &= ~VMX_VM_EXIT_LOAD_IA32_PERF_GLOBAL_CTRL; > + f[FEAT_VMX_ENTRY_CTLS] &= ~VMX_VM_ENTRY_LOAD_IA32_PERF_GLOBAL_CTRL; > + f[FEAT_VMX_SECONDARY_CTLS] &= ~(VMX_SECONDARY_EXEC_VIRTUAL_INTR_DELIVERY | > + VMX_SECONDARY_EXEC_VIRTUALIZE_APIC_ACCESSES | > + VMX_SECONDARY_EXEC_APIC_REGISTER_VIRT | > + VMX_SECONDARY_EXEC_ENABLE_PML | > + VMX_SECONDARY_EXEC_ENABLE_VMFUNC | > + VMX_SECONDARY_EXEC_SHADOW_VMCS | > + /* VMX_SECONDARY_EXEC_TSC_SCALING | */ > + VMX_SECONDARY_EXEC_PAUSE_LOOP_EXITING); > + f[FEAT_VMX_VMFUNC] &= ~MSR_VMX_VMFUNC_EPT_SWITCHING; > + } > + > kvm_msr_entry_add(cpu, MSR_IA32_VMX_TRUE_PROCBASED_CTLS, > make_vmx_msr_value(MSR_IA32_VMX_TRUE_PROCBASED_CTLS, > f[FEAT_VMX_PROCBASED_CTLS])); > I accuse you of not reading my PATCH0 :-) https://lists.nongnu.org/archive/html/qemu-devel/2020-01/msg00123.html does exactly this :-) P.S. I expect Paolo to comment on which hack he hates less :-) -- Vitaly
