On Wed, Aug 16, 2023, Kai Huang wrote: > > > > > --- a/arch/x86/kvm/vmx/vmx.c > > > > +++ b/arch/x86/kvm/vmx/vmx.c > > > > @@ -7783,6 +7783,9 @@ static void vmx_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu) > > > > vmx->msr_ia32_feature_control_valid_bits &= > > > > ~FEAT_CTL_SGX_LC_ENABLED; > > > > > > > > + if (boot_cpu_has(X86_FEATURE_LAM)) > > > > + kvm_governed_feature_check_and_set(vcpu, X86_FEATURE_LAM); > > > > + > > > If you want to use boot_cpu_has(), it's better to be done at your last patch to > > > only set the cap bit when boot_cpu_has() is true, I suppose. > > Yes, but new version of kvm_governed_feature_check_and_set() of > > KVM-governed feature framework will check against kvm_cpu_cap_has() as well. > > I will remove the if statement and call > > kvm_governed_feature_check_and_set() directly. > > https://lore.kernel.org/kvm/20230815203653.519297-2-seanjc@xxxxxxxxxx/ > > > > I mean kvm_cpu_cap_has() checks against the host CPUID directly while here you > are using boot_cpu_has(). They are not the same. > > If LAM should be only supported when boot_cpu_has() is true then it seems you > can just only set the LAM cap bit when boot_cpu_has() is true. As you also > mentioned above the kvm_governed_feature_check_and_set() here internally does > kvm_cpu_cap_has(). That's covered by the last patch: diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c index e961e9a05847..06061c11d74d 100644 --- a/arch/x86/kvm/cpuid.c +++ b/arch/x86/kvm/cpuid.c @@ -677,7 +677,7 @@ void kvm_set_cpu_caps(void) kvm_cpu_cap_mask(CPUID_7_1_EAX, F(AVX_VNNI) | F(AVX512_BF16) | F(CMPCCXADD) | F(FZRM) | F(FSRS) | F(FSRC) | - F(AMX_FP16) | F(AVX_IFMA) + F(AMX_FP16) | F(AVX_IFMA) | F(LAM) ); kvm_cpu_cap_init_kvm_defined(CPUID_7_1_EDX, Which highlights a problem with activating a goverened feature before said feature is actually supported by KVM: it's all kinds of confusing. It'll generate a more churn in git history, but I think we should first enable LAM without a goverened feature, and then activate a goverened feature later on. Using a goverened feature is purely an optimization, i.e. the series needs to be function without using a governed feature. That should yield an easier-to-review series on all fronts: the initial supports won't have any more hidden dependencies than absolutely necessary, switching to a goverened feature should be a very mechanical conversion (if it's not, that's a red flag), and last but not least, it makes it super easy to make a judgment call as to whether using a governed feature flag is justified, because all of the users will be in scope. TL;DR: Do the whole goverened feature thing dead last.
