On Tue, Dec 22, 2020, Babu Moger wrote: > > On 12/9/20 5:11 PM, Jim Mattson wrote: > > On Wed, Dec 9, 2020 at 2:39 PM Babu Moger <babu.moger@xxxxxxx> wrote: > >> > >> On 12/7/20 5:22 PM, Jim Mattson wrote: > >>> On Mon, Dec 7, 2020 at 2:38 PM Babu Moger <babu.moger@xxxxxxx> wrote: > >>>> diff --git a/arch/x86/include/asm/cpufeatures.h b/arch/x86/include/asm/cpufeatures.h > >>>> index dad350d42ecf..d649ac5ed7c7 100644 > >>>> --- a/arch/x86/include/asm/cpufeatures.h > >>>> +++ b/arch/x86/include/asm/cpufeatures.h > >>>> @@ -335,6 +335,7 @@ > >>>> #define X86_FEATURE_AVIC (15*32+13) /* Virtual Interrupt Controller */ > >>>> #define X86_FEATURE_V_VMSAVE_VMLOAD (15*32+15) /* Virtual VMSAVE VMLOAD */ > >>>> #define X86_FEATURE_VGIF (15*32+16) /* Virtual GIF */ > >>>> +#define X86_FEATURE_V_SPEC_CTRL (15*32+20) /* Virtual SPEC_CTRL */ > >>> > >>> Shouldn't this bit be reported by KVM_GET_SUPPORTED_CPUID when it's > >>> enumerated on the host? > >> > >> Jim, I am not sure if this needs to be reported by > >> KVM_GET_SUPPORTED_CPUID. I dont see V_VMSAVE_VMLOAD or VGIF being reported > >> via KVM_GET_SUPPORTED_CPUID. Do you see the need for that? > > > > Every little bit helps. No, it isn't *needed*. But then again, this > > entire patchset isn't *needed*, is it? > > > > Working on v2 of these patches. Saw this code comment(in > arch/x86/kvm/cpuid.c) on about exposing SVM features to the guest. > > > /* > * Hide all SVM features by default, SVM will set the cap bits for > * features it emulates and/or exposes for L1. > */ > kvm_cpu_cap_mask(CPUID_8000_000A_EDX, 0); > > > Should we go ahead with the changes here? Probably not, as the current SVM implementation aligns with the intended use of KVM_GET_SUPPORTED_CPUID. The current approach is to enumerate what SVM features KVM can virtualize or emulate for a nested VM, i.e. what SVM features an L1 VMM can use and thus can be set in a vCPU's CPUID model. For V_SPEC_CTRL, I'm pretty sure Jim was providing feedback for the non-nested case of reporting host/KVM support of the feature itself. There is the question of whether or not KVM should have an ioctl() to report what virtualization features are supported/enabled. AFAIK, it's not truly required as userspace can glean the information via /proc/cpuinfo (especially now that vmx_features exists), raw CPUID, and KVM module params. Providing an ioctl() would likely be a bit cleaner for userspace, but I'm guessing that ship has already sailed for most VMMs.
