On Mon, May 17, 2021, Reiji Watanabe wrote:
> > void kvm_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)
> > {
> > + unsigned long old_cr0 = kvm_read_cr0(vcpu);
> > + unsigned long old_cr4 = kvm_read_cr4(vcpu);
> > +
> > kvm_lapic_reset(vcpu, init_event);
> >
> > vcpu->arch.hflags = 0;
> > @@ -10483,6 +10485,10 @@ void kvm_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)
> > vcpu->arch.ia32_xss = 0;
> >
> > static_call(kvm_x86_vcpu_reset)(vcpu, init_event);
> > +
> > + if (kvm_cr0_mmu_role_changed(old_cr0, kvm_read_cr0(vcpu)) ||
> > + kvm_cr4_mmu_role_changed(old_cr4, kvm_read_cr4(vcpu)))
> > + kvm_mmu_reset_context(vcpu);
> > }
>
> I'm wondering if kvm_vcpu_reset() should call kvm_mmu_reset_context()
> for a change in EFER.NX as well.
Oooh. So there _should_ be no need. Paging has to be enabled for EFER.NX to
be relevant, and INIT toggles CR0.PG 1=>0 if paging was enabled and so is
guaranteed to trigger a context reset. And we do want to skip the context reset,
e.g. INIT-SIPI-SIPI when the vCPU has paging disabled should continue using the
same MMU.
But, kvm_calc_mmu_role_common() neglects to ignore NX if CR0.PG=0, and so the
MMU role will be stale if INIT clears EFER.NX without forcing a context reset.
However, that's benign from a functionality perspective because the context
itself correctly incorporates CR0.PG, it's only the role that's borked. I.e.
KVM will fail to reuse a page/context due to the spurious role.nxe, but the
permission checks are always be correct.
I'll add a comment here and send a patch to fix the role calculation.
