On Tue, Jul 30, 2024 at 11:00:57AM -0700, Sean Christopherson wrote:
> On Tue, Jul 30, 2024, John Allen wrote:
> > Handling deferred, uncorrected MCEs on AMD guests is now possible with
> > additional support in qemu. Ensure that the SUCCOR and OVERFLOW_RECOV
> > bits are advertised to the guest in KVM.
> >
> > Suggested-by: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> > Signed-off-by: John Allen <john.allen@xxxxxxx>
> > ---
> > arch/x86/kvm/cpuid.c | 2 +-
> > arch/x86/kvm/svm/svm.c | 7 +++++++
> > 2 files changed, 8 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c
> > index 2617be544480..4745098416c3 100644
> > --- a/arch/x86/kvm/cpuid.c
> > +++ b/arch/x86/kvm/cpuid.c
> > @@ -1241,7 +1241,7 @@ static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
> >
> > /* mask against host */
> > entry->edx &= boot_cpu_data.x86_power;
> > - entry->eax = entry->ebx = entry->ecx = 0;
> > + entry->eax = entry->ecx = 0;
>
> Needs an override to prevent reporting all of EBX to userspace.
>
> cpuid_entry_override(entry, CPUID_8000_0007_EBX);
Right, I see what you mean. We just want to expose these specific bits
and not all of EBX. I think with the patch as it is along with the
change you suggest below, this should resolve this as the above case
already has the cpuid_entry_override just above where it cuts off. Or is
there another place we need it?
Thanks,
John
>
> > break;
> > case 0x80000008: {
> > /*
> > diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
> > index c115d26844f7..a6820b0915db 100644
> > --- a/arch/x86/kvm/svm/svm.c
> > +++ b/arch/x86/kvm/svm/svm.c
> > @@ -5199,6 +5199,13 @@ static __init void svm_set_cpu_caps(void)
> > kvm_cpu_cap_set(X86_FEATURE_SVME_ADDR_CHK);
> > }
> >
> > + /* CPUID 0x80000007 */
> > + if (boot_cpu_has(X86_FEATURE_SUCCOR))
> > + kvm_cpu_cap_set(X86_FEATURE_SUCCOR);
> > +
> > + if (boot_cpu_has(X86_FEATURE_OVERFLOW_RECOV))
> > + kvm_cpu_cap_set(X86_FEATURE_OVERFLOW_RECOV);
>
> This _could_ use kvm_cpu_cap_check_and_set(), but given that this an AMD specific
> leaf and unlikely to ever be used by Intel, I'm inclined to handle this in cpuid.c,
> with an opporunustic "conversion" to one feature per line[*]:
>
> diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c
> index 2617be544480..ea11a7e45174 100644
> --- a/arch/x86/kvm/cpuid.c
> +++ b/arch/x86/kvm/cpuid.c
> @@ -743,6 +743,11 @@ void kvm_set_cpu_caps(void)
> if (!tdp_enabled && IS_ENABLED(CONFIG_X86_64))
> kvm_cpu_cap_set(X86_FEATURE_GBPAGES);
>
> + kvm_cpu_cap_mask(CPUID_8000_0007_EBX,
> + F(OVERFLOW_RECOV) |
> + F(SUCCOR)
> + );
> +
> kvm_cpu_cap_init_kvm_defined(CPUID_8000_0007_EDX,
> SF(CONSTANT_TSC)
> );
>
>
> [*] https://lore.kernel.org/all/ZoxooTvO5vIEnS5V@xxxxxxxxxx
