Re: [PATCH] perf/x86/amd: cpu_hw_events::perf_ctr_virt_mask should only be used on host

[Like Xu]

On 12/4/2022 9:25 am, Dongli Si wrote:
> On Mon 11 Apr 2022 22:29:18 +0800, Like Xu wrote:
> > Or you can work it out to make nested vPMU functional on AMD.
> Unless in the future kvm wants to emulate on L1 HV the behavior of not
> count when HO bit is set and SVM is disabled, otherwise it doesn't make

In fact, I would prefer that we make a little effort to enable nested vPMU.

> sense to use perf_ctr_virt_mask in the guest to mask HO bit. At least for
> now, this patch helps clarify what perf_ctr_virt_mask actually does.

This has been clarified by the commit 1018faa6cf23.

> > This is not a typical revert commit.
> Thanks for pointing out the problem with my patch,
> I will write another patch specifically to revert this commit.

The indispensable commit df51fe7ea1c1 is concerned with the symmetric use of
'disable_mask' in both __x86_pmu_enable_event() and x86_pmu_disable_event().

I had this false assumption at that time, and we'd better support HOST, GUEST}ONLY
bits in the L1 for L2 guest, and if not, it's bug. Please help. :D

> > Please check the chronological order of the related commits and the motivations.
> I know that commit df51fe7ea1c1c fixed the problem of use vPMU on old KVM,
> but I think it's a speculative way and make things a little obscure,
> because this #GP is actually a KVM problem rather than a guest problem,

And we have 9b026073db2f to fix KVM for older guest kernel.

> I think it is the user's responsibility to update their host kernel.
>
> > > +	/*
> > > +	 * When SVM is disabled, set the Host-Only bit will cause the
> > > +	 * performance counter to not work.
> > It's ridiculous. Based on the AMD APM Table 13-3. Host/Guest Only Bits,
> > the performance counter would count "Host events" rather than "not work".
> You are wrong, you can test it on the host, and the description of the
> commit 1018faa6cf23 also pointed out this problem, this is the result of an
> experiment, AMD APM has not documented this problem.

I have to say it's true on a ZEN3 host after a quick experiment.

> I forgot to say this is the behavior on the host, I will improve this
> comment to specify 'why' more clearly, like this:
> 	/*
> 	 * It turns out that when SVM is disabled on the host (L0), set the

Again, we need the semantics to hold true on L1.

> 	 * Host-Only bit will cause the performance counter to not count.
> 	 */
>
> > Note, your proposal change should work on the L0, L1 and L2.
> Yes, I tested it on L0, L1, L2 with 5.18-rc1 and it works as expected.

Specifically for L1, we need rely on the EFER[SVME] check instead of the
meaningless "boot_cpu_has(X86_FEATURE_HYPERVISOR)".

> There is a related discussion here:
> https://lore.kernel.org/all/20220320002106.1800166-1-sidongli1997@xxxxxxxxx/
>
> Regards,
> Dongli