Re: [RESEND v13 03/10] KVM: x86/pmu: Use IA32_PERF_CAPABILITIES to adjust features visibility

[Like Xu]

On 2021/1/26 17:42, Paolo Bonzini wrote:
> On 08/01/21 02:36, Like Xu wrote:
> > @@ -401,6 +398,9 @@ static void intel_pmu_init(struct kvm_vcpu *vcpu)
> >          pmu->fixed_counters[i].idx = i + INTEL_PMC_IDX_FIXED;
> >          pmu->fixed_counters[i].current_config = 0;
> >      }
> > +
> > +    vcpu->arch.perf_capabilities = guest_cpuid_has(vcpu, 
> > X86_FEATURE_PDCM) ?
> > +        vmx_get_perf_capabilities() : 0;
>
> There is one thing I don't understand with this patch: intel_pmu_init is 
> not called when CPUID is changed.  So I would have thought that anything 
> that uses guest_cpuid_has must stay in intel_pmu_refresh.  As I understand 
> it vcpu->arch.perf_capabilities is always set to 0 
> (vmx_get_perf_capabilities is never called), and kvm_set_msr_common would 
> fail to set any bit in the MSR.  What am I missing?
>
> In addition, the code of patch 4:
>
> +    if (!intel_pmu_lbr_is_enabled(vcpu)) {
> +        vcpu->arch.perf_capabilities &= ~PMU_CAP_LBR_FMT;
> +        lbr_desc->records.nr = 0;
> +    }
>
> is not okay after MSR changes.  The value written by the host must be 
> either rejected (with "return 1") or applied unchanged.
>
> Fortunately I think this code is dead if you move the check in kvm_set_msr 
> from patch 9 to patch 4.  However, in patch 9 vmx_get_perf_capabilities() 
> must only set the LBR format bits if intel_pmu_lbr_is_compatible(vcpu).

Thanks for the guidance. How about handling it in this way:

In the intel_pmu_init():

	vcpu->arch.perf_capabilities = 0;
	lbr_desc->records.nr = 0;

In the intel_pmu_refresh():

	if (guest_cpuid_has(vcpu, X86_FEATURE_PDCM)) {
		vcpu->arch.perf_capabilities = vmx_get_perf_capabilities();
		if (!lbr_desc->records.nr)
			vcpu->arch.perf_capabilities &= ~PMU_CAP_LBR_FMT;
	}

In the vmx_set_msr():

	case MSR_IA32_PERF_CAPABILITIES:
		// set up lbr_desc->records.nr
		if (!intel_pmu_lbr_is_compatible(vcpu))
			return 1;
		ret = kvm_set_msr_common(vcpu, msr_info);

In the kvm_set_msr_common():

	case MSR_IA32_PERF_CAPABILITIES:
		vcpu->arch.perf_capabilities = data;
		kvm_pmu_refresh(vcpu);

> The patches look good apart from these issues and the other nits I pointed 
> out.  However, you need testcases here, for both kvm-unit-tests and 
> tools/testing/selftests/kvm.
>
> For KVM, it would be at least a basic check that looks for the MSR LBR 
> (using the MSR indices for the various processors), does a branch, and 
> checks that the FROM_IP/TO_IP are good.  You can write the kvm-unit-tests 
> using the QEMU option "-cpu host,migratable=no": if you do this, QEMU will 
> pick the KVM_GET_SUPPORTED_CPUID bits and move them more or less directly 
> into the guest CPUID.
>
> For tools/testing/selftests/kvm, your test need to check the effect of 
> various CPUID settings on the PERF_CAPABILITIES MSR, check that whatever 
> you write with KVM_SET_MSR is _not_ modified and can be retrieved with 
> KVM_GET_MSR, and check that invalid LBR formats are rejected.

Thanks, I will add the above tests in the next version.

> I'm really, really sorry for leaving these patches on my todo list for 
> months, but you guys need to understand the main reason for this: they come 
> with no testcases.  A large patch series adding userspace APIs and 
> complicated CPUID/MSR processing *automatically* goes to the bottom of my 
> queue, because:
>
> - I need to go with a fine comb over all the userspace API changes, I 
> cannot just look at test code and see if it works.
>
> - I will have no way to test its correctness after it's committed.
>
> For you, the work ends when your patch is accepted.  For me, that's when 
> the work begins, and I need to make sure that the patch will be 
> maintainable in the future.
>
> Thanks, and sorry again for the delay.
>
> Paolo