[PATCH RFC] KVM: arm64: PMU: Use multiple host PMUs

[Oliver Upton]

On Thu, Mar 20, 2025 at 09:19:02AM +0000, Marc Zyngier wrote:
> On Wed, 19 Mar 2025 18:51:28 +0000, Oliver Upton <oliver.upton at linux.dev> wrote:
> > I'm at least willing to plug my nose and do the following:
> > 
> >  1) When the VMM does not specify a vPMU type:
> > 
> >    - We continue to present the 'default' PMU (including event counters)
> >      to the VM
> > 
> >    - KVM ensures that the fixed CPU cycle counter works on any PMUv3
> >      implementation in the system, even if it is different from the
> >      default
> > 
> >    - Otherwise, event counters will only count on the default
> >      implementation and will not count on different PMUs
> 
> I think this is confusing. The CC is counting, but nothing else, and
> people using the cycle counters in conjunction with other events (a
> very common use case) will not be able to correlate things correctly.
> The current behaviour is, for all its sins, at least consistent.

You of course have a good point. What Windows is doing is definitely an
outlier.

> > 
> >  2) Implement your suggestion of a UAPI where the VMM can select a PMU
> >     that only has the CPU cycle counter and works on any PMUv3
> >     implementation.
> > 
> > Either way KVM will need to have some special case handling of the fixed
> > CPU cycle counter. That'd allow users to actually run Windows *now* and
> > provide a clear mechanism for userspace to present a less-broken vPMU if
> > it cares.
> 
> Honestly, I don't care about one guest or another. My point is that if
> we are changing the behaviour of the PMU to deal with this sort of
> things, then it has to be a userspace buy-in.

I'm fine with just the user buy-in then. But I still do care about the
guest compatibility issue, especially since the end user of all this
crap is unlikely to know/care about the fine details of the
implementation.

So, Akihiko, I would *greatly* appreciate it if you propose a complete
solution to the problem, including the KVM and VMM patches to make it
all work.

Thanks,
Oliver