On Thu, Nov 02, 2023, Maxim Levitsky wrote: > On Tue, 2023-10-10 at 20:02 +0000, John Allen wrote: > > @@ -3032,6 +3037,9 @@ static void sev_es_init_vmcb(struct vcpu_svm *svm) > > if (guest_cpuid_has(&svm->vcpu, X86_FEATURE_RDTSCP)) > > svm_clr_intercept(svm, INTERCEPT_RDTSCP); > > } > > + > > + if (kvm_caps.supported_xss) > > + set_msr_interception(vcpu, svm->msrpm, MSR_IA32_XSS, 1, 1); > > This is not just a virtualization hole. This allows the guest to set MSR_IA32_XSS > to whatever value it wants, and thus it might allow XSAVES to access some host msrs > that guest must not be able to access. > > AMD might not yet have such msrs, but on Intel side I do see various components > like 'HDC State', 'HWP state' and such. The approach AMD has taken with SEV-ES+ is to have ucode context switch everything that the guest can access. So, in theory, if/when AMD adds more XCR0/XSS-based features, that state will also be context switched. Don't get me wrong, I hate this with a passion, but it's not *quite* fatally unsafe, just horrific. > I understand that this is needed so that #VC handler could read this msr, and > trying to read it will cause another #VC which is probably not allowed (I > don't know this detail of SEV-ES) > > I guess #VC handler should instead use a kernel cached value of this msr > instead, or at least KVM should only allow reads and not writes to it. Nope, doesn't work. In addition to automatically context switching state, SEV-ES also encrypts the guest state, i.e. KVM *can't* correctly virtualize XSS (or XCR0) for the guest, because KVM *can't* load the guest's desired value into hardware. The guest can do #VMGEXIT (a.k.a. VMMCALL) all it wants to request a certain XSS or XCR0, and there's not a damn thing KVM can do to service the request.
