Re: [PATCH 12/19] perf: Ignore event state for group validation

[Mark Rutland <mark.rutland@arm.com>]

On Tue, Aug 26, 2025 at 11:48:48AM -0700, Ian Rogers wrote:
> On Tue, Aug 26, 2025 at 8:32 AM Robin Murphy <robin.murphy@arm.com> wrote:
> >
> > On 2025-08-26 2:03 pm, Peter Zijlstra wrote:
> > > On Wed, Aug 13, 2025 at 06:01:04PM +0100, Robin Murphy wrote:
> > >> It may have been different long ago, but today it seems wrong for these
> > >> drivers to skip counting disabled sibling events in group validation,
> > >> given that perf_event_enable() could make them schedulable again, and
> > >> thus increase the effective size of the group later. Conversely, if a
> > >> sibling event is truly dead then it stands to reason that the whole
> > >> group is dead, so it's not worth going to any special effort to try to
> > >> squeeze in a new event that's never going to run anyway. Thus, we can
> > >> simply remove all these checks.
> > >
> > > So currently you can do sort of a manual event rotation inside an
> > > over-sized group and have it work.
> > >
> > > I'm not sure if anybody actually does this, but its possible.
> > >
> > > Eg. on a PMU that supports only 4 counters, create a group of 5 and
> > > periodically cycle which of the 5 events is off.
> 
> I'm not sure this is true, I thought this would fail in the
> perf_event_open when adding the 5th event and there being insufficient
> counters for the group.

We're talking specifically about cases where the logic in a pmu's
pmu::event_init() callback doesn't count events in specific states, and
hence the 5th even doesn't get rejected when it is initialised.

For example, in arch/x86/events/core.c, validate_group() uses
collect_events(), which has:

	for_each_sibling_event(event, leader) {
		if (!is_x86_event(event) || event->state <= PERF_EVENT_STATE_OFF)
			continue;

		if (collect_event(cpuc, event, max_count, n))
			return -EINVAL;

		n++;
	}

... and so where an event's state is <= PERF_EVENT_STATE_OFF at init
time, that event is not counted to see if it fits into HW counters.

Mark.