On Thu, Jul 04, 2024, Maxim Levitsky wrote: > On Fri, 2024-05-17 at 10:39 -0700, Sean Christopherson wrote: > > +static __always_inline bool guest_cpuid_has(struct kvm_vcpu *vcpu, > > + unsigned int x86_feature) > > { > > const struct cpuid_reg cpuid = x86_feature_cpuid(x86_feature); > > struct kvm_cpuid_entry2 *entry; > > + u32 *reg; > > + > > + /* > > + * XSAVES is a special snowflake. Due to lack of a dedicated intercept > > + * on SVM, KVM must assume that XSAVES (and thus XRSTORS) is usable by > > + * the guest if the host supports XSAVES and *XSAVE* is exposed to the > > + * guest. Although the guest can read/write XSS via XSAVES/XRSTORS, to > > + * minimize the virtualization hole, KVM rejects attempts to read/write > > + * XSS via RDMSR/WRMSR. To make that work, KVM needs to check the raw > > + * guest CPUID, not KVM's view of guest capabilities. > > Hi, > > I think that this comment is wrong: > > The guest can't read/write XSS via XSAVES/XRSTORS. It can only use XSAVES/XRSTORS > to save/restore features that are enabled in XSS, and thus if there are none enabled, > the XSAVES/XRSTORS acts as more or less XSAVEOPTC/XRSTOR except working only when CPL=0) Doh, right you are. > So I don't think that there is a virtualization hole except the fact that VMM can't > really disable XSAVES if it chooses to. There is still a hole. If XSAVES is not supported, KVM runs the guest with the host XSS. See the conditional switching in kvm_load_{guest,host}_xsave_state(). Not treating XSAVES as being available to the guest would allow the guest to read and write host supervisor state. I'll rewrite the comment to call that. > Another "half virtualization hole" is that since we have chosen to not > intercept XSAVES at all, (AMD can't do this at all, and it's slow anyway) we > instead opted to never support some XSS bits (so far all of them, only > upcoming CET will add a few supported bits). > > This creates an unexpected situation for the guest - enabled feature (e.g PT) > but no XSS bit supported to context switch it. x86 arch does allow this > though.