On 26.01.2024 19:36, Paolo Bonzini wrote:
On Wed, Jan 24, 2024 at 9:18 PM Maciej S. Szmigiero
<mail@xxxxxxxxxxxxxxxxxxxxx> wrote:
+static void kvm_hv_xsaves_xsavec_maybe_warn_unlocked(struct kvm_vcpu *vcpu)
Calling this function "unlocked" is confusing (others would say
"locked" is confusing instead). The double-underscore convention is
more common.
+{
+ struct kvm *kvm = vcpu->kvm;
+ struct kvm_hv *hv = to_kvm_hv(kvm);
+
+ if (hv->xsaves_xsavec_warned)
+ return;
+
+ if (!vcpu->arch.hyperv_enabled)
+ return;
I think these two should be in kvm_hv_xsaves_xsavec_maybe_warn(),
though the former needs to be checked again under the lock.
+ if ((hv->hv_guest_os_id & KVM_HV_WIN2016_GUEST_ID_MASK) !=
+ KVM_HV_WIN2016_GUEST_ID)
+ return;
At this point there is no need to return. You can set
xsaves_xsavec_warned and save the checks in the future.
+ /* UP configurations aren't affected */
+ if (atomic_read(&kvm->online_vcpus) < 2)
+ return;
+
+ if (boot_cpu_has(X86_FEATURE_XSAVES) ||
+ !guest_cpuid_has(vcpu, X86_FEATURE_XSAVEC))
+ return;
boot_cpu_has can also be done first to cull the whole check.
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 27e23714e960..db0a2c40d749 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -1782,6 +1782,10 @@ static int set_efer
if ((efer ^ old_efer) & KVM_MMU_EFER_ROLE_BITS)
kvm_mmu_reset_context(vcpu);
+ if (guest_cpuid_is_amd_or_hygon(vcpu) &&
+ efer & EFER_SVME)
+ kvm_hv_xsaves_xsavec_maybe_warn(vcpu);
+
return 0;
}
Checking guest_cpuid_is_amd_or_hygon() is relatively expensive, it
should be done after "efer & EFER_SVME" but really the bug can happen
just as well on Intel as far as I understand? It's just less likely
due to the AMD erratum.
Yes, I've checked this guest on an Intel host and it also fails to
boot in !XSAVES && XSAVEC configuration.
Only on Intel it's purely a theoretical problem as AFAIK there's
no corresponding Intel errata that disables just XSAVES.
I'll send a v2.
Paolo
Thanks,
Maciej
