Re: [PATCH v2 03/15] drivers/perf: apple_m1: Support per-implementation event tables

From: Myria Sarvay

Date: Wed Nov 18 2026 - 08:36:39 EST


On Wed, Nov 18, 2026 at 11:59:30AM +0000, Marc Zyngier wrote:
> On Sat, 07 Nov 2026 10:48:50 +0000,
> Myria Sarvay via B4 Relay <devnull+myrialsarvay.gmail.com@xxxxxxxxxx> wrote:
> >
> > From: Nick Chan <towinchenmi@xxxxxxxxx>
> >
> > Use per-implementation event tables to allow supporting implementations
> > with a different list of events and event affinities.
> >
> > Signed-off-by: Nick Chan <towinchenmi@xxxxxxxxx>
> > Co-developed-by: Myria Sarvay <myrialsarvay@xxxxxxxxx>
> > Signed-off-by: Myria Sarvay <myrialsarvay@xxxxxxxxx>
> > ---
> > drivers/perf/apple_m1_cpu_pmu.c | 113 ++++++++++++++++++++++++----------------
> > 1 file changed, 69 insertions(+), 44 deletions(-)
> >
> > diff --git a/drivers/perf/apple_m1_cpu_pmu.c b/drivers/perf/apple_m1_cpu_pmu.c
> > index de4026c63afd..2ff0c2e0b737 100644
> > --- a/drivers/perf/apple_m1_cpu_pmu.c
> > +++ b/drivers/perf/apple_m1_cpu_pmu.c
> > @@ -43,9 +43,6 @@
> > * moment, we don't really need to distinguish between the two because we
> > * know next to nothing about the events themselves, and we already have
> > * per cpu-type PMU abstractions.
> > - *
> > - * If we eventually find out that the events are different across
> > - * implementations, we'll have to introduce per cpu-type tables.
> > */
> > enum m1_pmu_events {
> > M1_PMU_PERFCTR_RETIRE_UOP = 0x1,
> > @@ -167,7 +164,7 @@ static const u16 m1_pmu_event_affinity[M1_PMU_PERFCTR_LAST + 1] = {
> > [M1_PMU_PERFCTR_UNKNOWN_fd] = ONLY_2_4_6,
> > };
> >
> > -static const unsigned m1_pmu_perf_map[PERF_COUNT_HW_MAX] = {
> > +static const unsigned int m1_pmu_perf_map[PERF_COUNT_HW_MAX] = {
> > PERF_MAP_ALL_UNSUPPORTED,
> > [PERF_COUNT_HW_CPU_CYCLES] = M1_PMU_PERFCTR_CORE_ACTIVE_CYCLE,
> > [PERF_COUNT_HW_INSTRUCTIONS] = M1_PMU_PERFCTR_INST_ALL,
> > @@ -486,11 +483,12 @@ static void m1_pmu_write_counter(struct perf_event *event, u64 value)
> > isb();
> > }
> >
> > -static int m1_pmu_get_event_idx(struct pmu_hw_events *cpuc,
> > - struct perf_event *event)
> > +static int apple_pmu_get_event_idx(struct pmu_hw_events *cpuc,
> > + struct perf_event *event,
> > + const u16 event_affinities[])
> > {
> > unsigned long evtype = event->hw.config_base & M1_PMU_CFG_EVENT;
> > - unsigned long affinity = m1_pmu_event_affinity[evtype];
> > + unsigned long affinity = event_affinities[evtype];
> > int idx;
> >
> > /*
> > @@ -509,6 +507,12 @@ static int m1_pmu_get_event_idx(struct pmu_hw_events *cpuc,
> > return -EAGAIN;
> > }
> >
> > +static int m1_pmu_get_event_idx(struct pmu_hw_events *cpuc,
> > + struct perf_event *event)
> > +{
> > + return apple_pmu_get_event_idx(cpuc, event, m1_pmu_event_affinity);
> > +}
> > +
> > static void m1_pmu_clear_event_idx(struct pmu_hw_events *cpuc,
> > struct perf_event *event)
> > {
> > @@ -536,44 +540,52 @@ static void m1_pmu_stop(struct arm_pmu *cpu_pmu)
> > __m1_pmu_set_mode(PMCR0_IMODE_OFF);
> > }
> >
> > +static int apple_pmu_map_event(struct perf_event *event,
> > + const unsigned int (*perf_map)[],
> > + unsigned int flags)
> > +{
> > + event->hw.flags |= flags;
> > + return armpmu_map_event(event, perf_map, NULL, M1_PMU_CFG_EVENT);
> > +}
> > +
> > +/*
> > + * Apple PMU counters have been observed to be either 48 or
> > + * 64 bits wide, depending on the SoC. However, the MSB is
> > + * what triggers an overflow interrupt, so we advertise the
> > + * counters as being 1-bit smaller than hardware so as to
> > + * mimick ARM PMU behaviour
> > + */
> > static int m1_pmu_map_event(struct perf_event *event)
> > {
> > - /*
> > - * Although the counters are 48bit wide, bit 47 is what
> > - * triggers the overflow interrupt. Advertise the counters
> > - * being 47bit wide to mimick the behaviour of the ARM PMU.
> > - */
> > - event->hw.flags |= ARMPMU_EVT_47BIT;
> > - return armpmu_map_event(event, &m1_pmu_perf_map, NULL, M1_PMU_CFG_EVENT);
> > + return apple_pmu_map_event(event, &m1_pmu_perf_map, ARMPMU_EVT_47BIT);
> > }
> >
> > static int m2_pmu_map_event(struct perf_event *event)
> > {
> > - /*
> > - * Same deal as the above, except that M2 has 64bit counters.
> > - * Which, as far as we're concerned, actually means 63 bits.
> > - * Yes, this is getting awkward.
> > - */
> > - event->hw.flags |= ARMPMU_EVT_63BIT;
> > - return armpmu_map_event(event, &m1_pmu_perf_map, NULL, M1_PMU_CFG_EVENT);
> > + return apple_pmu_map_event(event, &m1_pmu_perf_map, ARMPMU_EVT_63BIT);
> > }
> >
> > -static int m1_pmu_map_pmuv3_event(unsigned int eventsel)
> > +static int apple_pmu_map_pmuv3_event(unsigned int eventsel, const u16 pmceid_map[])
> > {
> > - u16 m1_event = HW_OP_UNSUPPORTED;
> > + u16 apple_event = HW_OP_UNSUPPORTED;
>
> nit: s/apple_event/hw_event/.

Thanks!
>
> >
> > if (eventsel < ARMV8_PMUV3_MAX_COMMON_EVENTS)
> > - m1_event = m1_pmu_pmceid_map[eventsel];
> > + apple_event = pmceid_map[eventsel];
> >
> > - return m1_event == HW_OP_UNSUPPORTED ? -EOPNOTSUPP : m1_event;
> > + return apple_event == HW_OP_UNSUPPORTED ? -EOPNOTSUPP : apple_event;
> > }
> >
> > -static void m1_pmu_init_pmceid(struct arm_pmu *pmu)
> > +static int m1_pmu_map_pmuv3_event(unsigned int eventsel)
> > +{
> > + return apple_pmu_map_pmuv3_event(eventsel, m1_pmu_pmceid_map);
> > +}
> > +
> > +static void apple_pmu_init_pmceid(struct arm_pmu *pmu)
> > {
> > unsigned int event;
> >
> > for (event = 0; event < ARMV8_PMUV3_MAX_COMMON_EVENTS; event++) {
> > - if (m1_pmu_map_pmuv3_event(event) >= 0)
> > + if (pmu->map_pmuv3_event(event) >= 0)
> > set_bit(event, pmu->pmceid_bitmap);
> > }
> > }
> > @@ -616,33 +628,22 @@ static int m1_pmu_set_event_filter(struct hw_perf_event *event,
> > return 0;
> > }
> >
> > -static int m1_pmu_init(struct arm_pmu *cpu_pmu, u32 flags)
> > +static int apple_pmu_init(struct arm_pmu *cpu_pmu)
> > {
> > cpu_pmu->handle_irq = m1_pmu_handle_irq;
> > cpu_pmu->enable = m1_pmu_enable_event;
> > cpu_pmu->disable = m1_pmu_disable_event;
> > cpu_pmu->read_counter = m1_pmu_read_counter;
> > cpu_pmu->write_counter = m1_pmu_write_counter;
> > - cpu_pmu->get_event_idx = m1_pmu_get_event_idx;
> > cpu_pmu->clear_event_idx = m1_pmu_clear_event_idx;
> > cpu_pmu->start = m1_pmu_start;
> > cpu_pmu->stop = m1_pmu_stop;
> > -
> > - if (flags & ARMPMU_EVT_47BIT)
> > - cpu_pmu->map_event = m1_pmu_map_event;
> > - else if (flags & ARMPMU_EVT_63BIT)
> > - cpu_pmu->map_event = m2_pmu_map_event;
> > - else
> > - return WARN_ON(-EINVAL);
> > -
> > cpu_pmu->reset = m1_pmu_reset;
> > cpu_pmu->set_event_filter = m1_pmu_set_event_filter;
> >
> > - cpu_pmu->map_pmuv3_event = m1_pmu_map_pmuv3_event;
> > - m1_pmu_init_pmceid(cpu_pmu);
> > + apple_pmu_init_pmceid(cpu_pmu);
> >
> > bitmap_set(cpu_pmu->cntr_mask, 0, M1_PMU_NR_COUNTERS);
> > - cpu_pmu->attr_groups[ARMPMU_ATTR_GROUP_EVENTS] = &m1_pmu_events_attr_group;
> > cpu_pmu->attr_groups[ARMPMU_ATTR_GROUP_FORMATS] = &m1_pmu_format_attr_group;
> > return 0;
> > }
> > @@ -651,25 +652,49 @@ static int m1_pmu_init(struct arm_pmu *cpu_pmu, u32 flags)
> > static int m1_pmu_ice_init(struct arm_pmu *cpu_pmu)
> > {
> > cpu_pmu->name = "apple_icestorm_pmu";
> > - return m1_pmu_init(cpu_pmu, ARMPMU_EVT_47BIT);
> > + cpu_pmu->get_event_idx = m1_pmu_get_event_idx;
> > + cpu_pmu->map_event = m1_pmu_map_event;
> > +
> > + cpu_pmu->map_pmuv3_event = m1_pmu_map_pmuv3_event;
> > +
> > + cpu_pmu->attr_groups[ARMPMU_ATTR_GROUP_EVENTS] = &m1_pmu_events_attr_group;
> > + return apple_pmu_init(cpu_pmu);
> > }
> >
>
> Looking at the following patches, I feel that this series is taking
> the long route to get to a relatively clear point: most of the
> invariant that we relied on for M1, slightly relaxed for M2 (counter
> width), are even more relaxed for M3+ (event width). We therefore need
> to indirect things. The way you do it is to mirror the current setup,
> with a whole set of callbacks per implementation and a bunch of
> additional helpers.
>
> But instead of having a set of helper functions, one per
> implementation, with an ever increasing set of parameters, it really
> feels like we could do with a more data-centric approach where we have
> a single data structure describing an implementation, containing all
> the relevant data (pointer to the event mapping table, flags, bit
> indices for the user/kernel/guest/host attributes, and so on. We could
> then have a single set of callbacks that work on everything.
>
> You could convert M1/M2 to that, then just add the relevant
> descriptions for M3+, which should just add data and the init
> functions, nothing else.
>
> Does this make sense to you?

Hi Marc,

It does make sense, and should reduce duplicated code, I'm just not
confident I have the right idea as to how to implement it. I don't want
to duplicate a lot of the generic arm_pmu code, so it looks like it'd
probably have to be modified to allow passing a hardware specific detail
struct to all the callbacks, likely stored as a new field (void *) in
the arm_pmu struct itself. Does that vaguely sound right to you?

Also, with this rewrite, if we're going to make the code a whole lot
more generic, and take hardware specific data from a static struct,
would it make sense to rename the driver/all the callbacks while we're
at it, now that we're adding a lot more than just M1/M2 pmu events?

Thanks,
Myria

>
> Thanks,
>
> M.
>
> --
> Without deviation from the norm, progress is not possible.