On Tue, Jun 17, 2025 at 12:36:18PM +0200, Peter Zijlstra wrote:
> On Tue, Jun 17, 2025 at 07:27:50PM +0900, Koichi Okuno wrote:
> > + pcipmu->pmu = (struct pmu) {
> > + .parent = dev,
> > + .task_ctx_nr = perf_invalid_context,
> > +
> > + .pmu_enable = fujitsu_pci__pmu_enable,
> > + .pmu_disable = fujitsu_pci__pmu_disable,
> > + .event_init = fujitsu_pci__event_init,
> > + .add = fujitsu_pci__event_add,
> > + .del = fujitsu_pci__event_del,
> > + .start = fujitsu_pci__event_start,
> > + .stop = fujitsu_pci__event_stop,
> > + .read = fujitsu_pci__event_read,
> > +
> > + .attr_groups = fujitsu_pci_pmu_attr_grps,
> > + .capabilities = PERF_PMU_CAP_NO_EXCLUDE,
>
> Should these drivers not also have PERF_PMU_CAP_NO_INTERRUPT ? Per them
> being uncore they cannot generate samples.
These PMUs actually have an interrupt, so that might be a bit confusing.
That said, the counters seem to be 64-bit, so the interrupt doesn't seem
practically necessary today.
Either way, the fujitsu_pci__event_init() function rejects sampling
events, so it correctly rejects sampling events.
IMO, it'd be a bit nicer to do the inverse, and have the perf core
reject sampling events by default for non-CPU PMUs.
Mark.
