On Thu, Sep 05, 2024, Xin Li wrote: > On 6/12/2024 2:32 PM, Sean Christopherson wrote: > > On Fri, Apr 19, 2024, Chao Gao wrote: > > > On Wed, Feb 07, 2024 at 09:26:27AM -0800, Xin Li wrote: > > > > Add FRED MSRs to the valid passthrough MSR list and set FRED MSRs intercept > > > > based on FRED enumeration. > > > > This needs a *much* more verbose explanation. It's pretty darn obvious _what_ > > KVM is doing, but it's not at all clear _why_ KVM is passing through FRED MSRs. > > E.g. why is FRED_SSP0 not included in the set of passthrough MSRs? > > > > > > static void vmx_vcpu_config_fred_after_set_cpuid(struct kvm_vcpu *vcpu) > > > > { > > > > struct vcpu_vmx *vmx = to_vmx(vcpu); > > > > + bool fred_enumerated; > > > > > > > > kvm_governed_feature_check_and_set(vcpu, X86_FEATURE_FRED); > > > > + fred_enumerated = guest_can_use(vcpu, X86_FEATURE_FRED); > > > > > > > > - if (guest_can_use(vcpu, X86_FEATURE_FRED)) { > > > > + if (fred_enumerated) { > > > > vm_entry_controls_setbit(vmx, VM_ENTRY_LOAD_IA32_FRED); > > > > secondary_vm_exit_controls_setbit(vmx, > > > > SECONDARY_VM_EXIT_SAVE_IA32_FRED | > > > > @@ -7788,6 +7793,16 @@ static void vmx_vcpu_config_fred_after_set_cpuid(struct kvm_vcpu *vcpu) > > > > SECONDARY_VM_EXIT_SAVE_IA32_FRED | > > > > SECONDARY_VM_EXIT_LOAD_IA32_FRED); > > > > } > > > > + > > > > + vmx_set_intercept_for_msr(vcpu, MSR_IA32_FRED_RSP0, MSR_TYPE_RW, !fred_enumerated); > > > > + vmx_set_intercept_for_msr(vcpu, MSR_IA32_FRED_RSP1, MSR_TYPE_RW, !fred_enumerated); > > > > + vmx_set_intercept_for_msr(vcpu, MSR_IA32_FRED_RSP2, MSR_TYPE_RW, !fred_enumerated); > > > > + vmx_set_intercept_for_msr(vcpu, MSR_IA32_FRED_RSP3, MSR_TYPE_RW, !fred_enumerated); > > > > + vmx_set_intercept_for_msr(vcpu, MSR_IA32_FRED_STKLVLS, MSR_TYPE_RW, !fred_enumerated); > > > > + vmx_set_intercept_for_msr(vcpu, MSR_IA32_FRED_SSP1, MSR_TYPE_RW, !fred_enumerated); > > > > + vmx_set_intercept_for_msr(vcpu, MSR_IA32_FRED_SSP2, MSR_TYPE_RW, !fred_enumerated); > > > > + vmx_set_intercept_for_msr(vcpu, MSR_IA32_FRED_SSP3, MSR_TYPE_RW, !fred_enumerated); > > > > + vmx_set_intercept_for_msr(vcpu, MSR_IA32_FRED_CONFIG, MSR_TYPE_RW, !fred_enumerated); > > > > > > Use a for-loop here? e.g., > > > for (i = MSR_IA32_FRED_RSP0; i <= MSR_IA32_FRED_CONFIG; i++) > > > > Hmm, I'd prefer to keep the open coded version. It's not pretty, but I don't > > expect this to have much, if any, maintenance cost. And using a loop makes it > > harder to both understand _exactly_ what's happening, and to search for relevant > > code. E.g. it's quite difficult to see that FRED_SSP0 is still intercepted (see > > my comment regarding the changelog). > > > I owe you an explanation; I have been thinking about figuring out a way > to include FRED SSP0 in the FRED KVM patch set... > > MSR_IA32_FRED_SSP0 is an alias of the CET MSR_IA32_PL0_SSP and likely to > be used in the same way as FRED RSP0, i.e., host FRED SSP0 _should_ be > restored in arch_exit_to_user_mode_prepare(). However as of today Linux > has no plan to utilize kernel shadow stack thus no one cares host FRED > SSP0 (no?). But lets say anyway it is host's responsibility to manage > host FRED SSP0, then KVM only needs to take care of guest FRED SSP0 > (just like how KVM should handle guest FRED RSP0) even before the > supervisor shadow stack feature is advertised to guest. Heh, I'm not sure what your question is, or if there even is a question. KVM needs to context switch FRED SSP0 if FRED is exposed to the guest, but presumably that will be done through XSAVE state? If that's the long term plan, I would prefer to focus on merging CET virtualization first, and then land FRED virtualization on top so that KVM doesn't have to carry intermediate code to deal with the aliased MSR. Ugh, but what happens if a CPU (or the host kernel) supports FRED but not CET SS? Or is that effectively an illegal combination? > Another question is should KVM handle userspace request to set/get FRED > SSP0? IMO, it should be part of CET state management. Yes, KVM needs to allow userspace to get/set FRED SSP0. In general, KVM needs to allow reads/writes to MSRs even if they can be saved/restored through some other means. In most cases, including this one, it's a moot point, because KVM needs to have the necessary code anyways, e.g. if KVM encounters a RDMSR/WRMSR while emulating.