Re: [PATCH] KVM: nSVM: Add missing CET VMRUN consistency checks

From: Yosry Ahmed

Date: Tue Nov 10 2026 - 14:32:35 EST


On Tue, Nov 10, 2026 at 11:18 AM Jim Mattson <jmattson@xxxxxxxxxx> wrote:
>
> On Tue, Nov 10, 2026 at 9:56 AM Yosry Ahmed <yosry@xxxxxxxxxx> wrote:
> >
> > On Mon, Nov 09, 2026 at 04:40:37PM -0800, Jim Mattson wrote:
> > > Per the APM, "Canonicalization and Consistency Checks,"
> > >
> > > [t]he following conditions are considered illegal state combinations:
> > >
> > > - CR4.CET=1 when CR0.WP=0
> > > - Any reserved bit set in S_CET
> > > - CR4.CET=1 and U_CET.SS=1 when EFLAGS.VM=1
> > >
> > > Check these conditions in software before trying to VMRUN vmcb02, so that
> > > the illegal states result in a #VMEXIT to L1 rather than an exit to
> > > userspace with KVM_EXIT_FAIL_ENTRY.
> > >
> > > Fixes: c5ba49458513 ("KVM: nSVM: Save/load CET Shadow Stack state to/from vmcb12/vmcb02")
> > > Signed-off-by: Jim Mattson <jmattson@xxxxxxxxxx>
> > > ---
> > > arch/x86/kvm/svm/nested.c | 15 +++++++++++++++
> > > 1 file changed, 15 insertions(+)
> > >
> > > diff --git a/arch/x86/kvm/svm/nested.c b/arch/x86/kvm/svm/nested.c
> > > index af0ef678433d..20ede403b62d 100644
> > > --- a/arch/x86/kvm/svm/nested.c
> > > +++ b/arch/x86/kvm/svm/nested.c
> > > @@ -469,6 +469,21 @@ static bool nested_vmcb_check_save(struct kvm_vcpu *vcpu,
> > > if (CC(!__kvm_is_valid_cr4(vcpu, save->cr4)))
> > > return false;
> > >
> > > + if (CC((save->cr4 & X86_CR4_CET) && !(save->cr0 & X86_CR0_WP)))
> > > + return false;
> >
> > Why aren't we checking X86_FEATURE_SHSTK here as well? Do we know if
> > older CPUs that didn't support CET had the consistency check?
>
> CR4.CET is reserved if an AMD CPU doesn't have X86_FEATURE_SHSTK.
> Hence, older CPUs without CET support cannot fail the consistency
> check.

Oh I missed this, thanks.

>
> > > +
> > > + if (guest_cpu_cap_has(vcpu, X86_FEATURE_SHSTK) &&
> > > + CC(!kvm_is_valid_u_s_cet(vcpu, save->s_cet)))
> > > + return false;

I wonder if we should clear the cached s_cet based on
X86_FEATURE_SHSTK, similar to how we sanitize some cached control
fields in __nested_copy_vmcb_control_to_cache(). Mostly orthogonal to
this patch tho.

> > > +
> > > + if ((save->cr4 & X86_CR4_CET) && (save->rflags & X86_EFLAGS_VM)) {
> > > + u64 u_cet = 0;
> > > +
> > > + if (!__kvm_emulate_msr_read(vcpu, MSR_IA32_U_CET, &u_cet) &&
> >
> > If X86_FEATURE_SHSTK is not enumerated by the vCPU, then whether or not
> > the consistency check is done depends on ignore_msr. If ignore_msr=0,
> > the MSR read will fail and we won't do the consistency check, but if
> > ignore_msr=1 the MSR read will silently return 0, so the consistency
> > check will be done but will always succeed.
>
> If X86_FEATURE_SHSTK is not enumerated by the vCPU, then CR4.CET == 0
> and this entire block is skipped.

Ack.

> > Instead of this implicit dependency, should we check for
> > X86_FEATURE_SHSTK explicitly?
>
> That seems superfluous.

Given the above, yes.

>
> > Also, if X86_FEATURE_SHSTK is enumberated by the vCPU, and the MSR read
> > fails, the consistency check will silently be skipped. But I don't think
> > this can/should happen. Maybe we should WARN in this case.
>
> You're right; failure to read the MSR should result in a failed VMRUN,
> and a warning would help triage the failure.
>
> > One last question, in the restore path, the MSR may have not been
> > restored yet, right? Can the consistency check mistakenly fail in this
> > case?
>
> Not with a legal snapshot. Note that on the restore path, the 'save'
> area holds L1 state. Hence, RFLAGS.VM must be clear, because VMRUN
> raises #UD in V8086 mode.

Oh that's subtle, but yes, makes sense. I was going to suggest a
comment, but I can't find a good place for it :)