[PATCH v6 04/44] perf: Add APIs to create/release mediated guest vPMUs

[Sean Christopherson]

On Mon, Dec 08, 2025, Peter Zijlstra wrote:
> On Fri, Dec 05, 2025 at 04:16:40PM -0800, Sean Christopherson wrote:
> 
> > +static atomic_t nr_include_guest_events __read_mostly;
> > +
> > +static atomic_t nr_mediated_pmu_vms __read_mostly;
> > +static DEFINE_MUTEX(perf_mediated_pmu_mutex);
> 
> > +static int mediated_pmu_account_event(struct perf_event *event)
> > +{
> > +	if (!is_include_guest_event(event))
> > +		return 0;
> > +
> > +	guard(mutex)(&perf_mediated_pmu_mutex);
> > +
> > +	if (atomic_read(&nr_mediated_pmu_vms))
> > +		return -EOPNOTSUPP;
> > +
> > +	atomic_inc(&nr_include_guest_events);
> > +	return 0;
> > +}
> > +
> > +static void mediated_pmu_unaccount_event(struct perf_event *event)
> > +{
> > +	if (!is_include_guest_event(event))
> > +		return;
> > +
> > +	atomic_dec(&nr_include_guest_events);
> > +}
> 
> > +int perf_create_mediated_pmu(void)
> > +{
> > +	guard(mutex)(&perf_mediated_pmu_mutex);
> > +	if (atomic_inc_not_zero(&nr_mediated_pmu_vms))
> > +		return 0;
> > +
> > +	if (atomic_read(&nr_include_guest_events))
> > +		return -EBUSY;
> > +
> > +	atomic_inc(&nr_mediated_pmu_vms);
> > +	return 0;
> > +}
> > +EXPORT_SYMBOL_GPL(perf_create_mediated_pmu);
> > +
> > +void perf_release_mediated_pmu(void)
> > +{
> > +	if (WARN_ON_ONCE(!atomic_read(&nr_mediated_pmu_vms)))
> > +		return;
> > +
> > +	atomic_dec(&nr_mediated_pmu_vms);
> > +}
> > +EXPORT_SYMBOL_GPL(perf_release_mediated_pmu);
> 
> These two things are supposed to be symmetric, but are implemented
> differently; what gives?
> 
> That is, should not both have the general shape:
> 
> 	if (atomic_inc_not_zero(&A))
> 		return 0;
> 
> 	guard(mutex)(&lock);
> 
> 	if (atomic_read(&B))
> 		return -EBUSY;
> 
> 	atomic_inc(&A);
> 	return 0;
> 
> Similarly, I would imagine both release variants to have the underflow
> warn on like:
> 
> 	if (WARN_ON_ONCE(!atomic_read(&A)))
> 		return;
> 
> 	atomic_dec(&A);
> 
> Hmm?

IIUC, you're suggesting someting like this?  If so, that makes perfect sense to me.

diff --git a/kernel/events/core.c b/kernel/events/core.c
index c6368c64b866..fa2e7b722283 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -6356,7 +6356,8 @@ static int mediated_pmu_account_event(struct perf_event *event)
 
 static void mediated_pmu_unaccount_event(struct perf_event *event)
 {
-       if (!is_include_guest_event(event))
+       if (!is_include_guest_event(event) ||
+           WARN_ON_ONCE(!atomic_read(&nr_include_guest_events)))
                return;
 
        atomic_dec(&nr_include_guest_events);

> Also, EXPORT_SYMBOL_FOR_KVM() ?

Ya, for sure.  I posted this against a branch without EXPORT_SYMBOL_FOR_KVM(),
because there are also hard dependencies on the for-6.19 KVM pull requests, and
I didn't want to wait to post until 6.19-rc1 because of the impending winter
break.  Though I also simply forgot about these exports :-(

These could also use EXPORT_SYMBOL_FOR_KVM():

  EXPORT_SYMBOL_FOR_MODULES(perf_load_guest_lvtpc, "kvm");
  EXPORT_SYMBOL_FOR_MODULES(perf_put_guest_lvtpc, "kvm");


> I can make these edits when applying, if/when we get to applying. Let me
> continue reading.
>