Re: [PATCH v3 07/16] KVM: SVM: Move core EFER.SVME enablement to kernel

[Sean Christopherson]

On Thu, Feb 26, 2026, Chao Gao wrote:
> >-static inline void kvm_cpu_svm_disable(void)
> >-{
> >-	uint64_t efer;
> >-
> >-	wrmsrq(MSR_VM_HSAVE_PA, 0);
> >-	rdmsrq(MSR_EFER, efer);
> >-	if (efer & EFER_SVME) {
> >-		/*
> >-		 * Force GIF=1 prior to disabling SVM, e.g. to ensure INIT and
> >-		 * NMI aren't blocked.
> >-		 */
> >-		stgi();
> >-		wrmsrq(MSR_EFER, efer & ~EFER_SVME);
> >-	}
> >-}
> >-
> > static void svm_emergency_disable_virtualization_cpu(void)
> > {
> >-	virt_rebooting = true;
> >-
> >-	kvm_cpu_svm_disable();
> >+	wrmsrq(MSR_VM_HSAVE_PA, 0);
> > }
> > 
> > static void svm_disable_virtualization_cpu(void)
> >@@ -507,7 +489,7 @@ static void svm_disable_virtualization_cpu(void)
> > 	if (tsc_scaling)
> > 		__svm_write_tsc_multiplier(SVM_TSC_RATIO_DEFAULT);
> > 
> >-	kvm_cpu_svm_disable();
> >+	x86_svm_disable_virtualization_cpu();
> 
> There's a functional change here. The new x86_svm_disable_virtualization_cpu()
> doesn't reset MSR_VM_HSAVE_PA, but the old kvm_cpu_svm_disable() does.

Doh.  I'll squash this as fixup, assuming there are no other goofs:

diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index 5f033bf3ba83..fc08450cb4b7 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -490,6 +490,7 @@ static void svm_disable_virtualization_cpu(void)
                __svm_write_tsc_multiplier(SVM_TSC_RATIO_DEFAULT);
 
        x86_svm_disable_virtualization_cpu();
+       wrmsrq(MSR_VM_HSAVE_PA, 0);
 
        amd_pmu_disable_virt();
 }

Very nice catch!

P.S. This reminded me that there's a lurking wart with __sev_snp_init_locked()
where it forces MSR_VM_HSAVE_PA to '0' on all CPUs.  That's firmly a "hypervisor"
thing so it doesn't really fit here (and code wise it's also kludgy), just thought
I'd mention it in case someone has a brilliant idea and/or runs into problems with
it.  IIRC, we ran into a problem where __sev_snp_init_locked() clobbered KVM's
value, but I think the underlying problem was effectively fixed by commit
6f1d5a3513c2 ("KVM: SVM: Add support to initialize SEV/SNP functionality in KVM").