Re: [RFC PATCH v2 06/11] KVM: x86: Advertise ARCH_CAP_VIRTUAL_ENUM support

[Xiaoyao Li]

On 5/19/2023 5:57 PM, Chao Gao wrote:
> On Thu, May 18, 2023 at 06:14:40PM +0800, Xiaoyao Li wrote:
> > >    static u32 msr_based_features[ARRAY_SIZE(msr_based_features_all)];
> > > @@ -1591,7 +1593,8 @@ static unsigned int num_msr_based_features;
> > >    	 ARCH_CAP_SKIP_VMENTRY_L1DFLUSH | ARCH_CAP_SSB_NO | ARCH_CAP_MDS_NO | \
> > >    	 ARCH_CAP_PSCHANGE_MC_NO | ARCH_CAP_TSX_CTRL_MSR | ARCH_CAP_TAA_NO | \
> > >    	 ARCH_CAP_SBDR_SSDP_NO | ARCH_CAP_FBSDP_NO | ARCH_CAP_PSDP_NO | \
> > > -	 ARCH_CAP_FB_CLEAR | ARCH_CAP_RRSBA | ARCH_CAP_PBRSB_NO)
> > > +	 ARCH_CAP_FB_CLEAR | ARCH_CAP_RRSBA | ARCH_CAP_PBRSB_NO | \
> > > +	 ARCH_CAP_VIRTUAL_ENUM)
> >
> > We cannot do it.
> >
> > Otherwise, an AMD L1 with X86_FEATURE_ARCH_CAPABILITIES configured is
> > possible to expose MSR_VIRTUAL_ENUMERATION to L2 while no support for it.
>
> How does AMD L1 see the ARCH_CAP_VIRTUAL_ENUM feature in the first
> place? because ...
>
> > >    static u64 kvm_get_arch_capabilities(void)
> > >    {
> > > @@ -1610,6 +1613,17 @@ static u64 kvm_get_arch_capabilities(void)
> > >    	 */
> > >    	data |= ARCH_CAP_PSCHANGE_MC_NO;
> > > +	/*
> > > +	 * Virtual enumeration is a paravirt feature. The only usage for now
> > > +	 * is to bridge the gap caused by microarchitecture changes between
> > > +	 * different Intel processors. And its usage is linked to "virtualize
> > > +	 * IA32_SPEC_CTRL" which is a VMX feature. Whether AMD SVM can benefit
> > > +	 * from the same usage and how to implement it is still unclear. Limit
> > > +	 * virtual enumeration to VMX.
> > > +	 */
> > > +	if (static_call(kvm_x86_has_emulated_msr)(NULL, MSR_VIRTUAL_ENUMERATION))
> > > +		data |= ARCH_CAP_VIRTUAL_ENUM;
>
> the feature is exposed on Intel CPUs only.
>
> Do you mean AMD L1 created on Intel L0? and Intel L0 even emulates
> nested (SVM) support for the L1? This sounds a very contrived case.

you are right. I was thinking of an rare case but ignored the fact that 
VMX doesn't nested svm.

Sorry for it.

> > > +
> > >    	/*
> > >    	 * If we're doing cache flushes (either "always" or "cond")
> > >    	 * we will do one whenever the guest does a vmlaunch/vmresume.