Messages in this thread | | | Subject | Re: [PATCH 3/6] arm_pmu: Add support for long event counters | From | Robin Murphy <> | Date | Fri, 18 May 2018 14:22:06 +0100 |
| |
On 18/05/18 11:22, Suzuki K Poulose wrote: > Each PMU has a set of fixed width event counters. But in some > special cases, the events could be counted using a counter which > effectively has twice the normal width of a coutner. > e.g, Arm V8 PMUv3 has a 64 bit cycle counter which can count > only the CPU cylces. Also, the PMU can chain the event counters > to effectively count as a 64bit counter.
Nit: a few typos in that paragraph.
> Add support for tracking the events that uses double the normal > counter size. This only affects the periods set for each counter. > > Cc: Mark Rutland <mark.rutland@arm.com> > Cc: Will Deacon <will.deacon@arm.com> > Signed-off-by: Suzuki K Poulose <suzuki.poulose@arm.com> > --- > drivers/perf/arm_pmu.c | 25 ++++++++++++++++++++++--- > include/linux/perf/arm_pmu.h | 6 ++++++ > 2 files changed, 28 insertions(+), 3 deletions(-) > > diff --git a/drivers/perf/arm_pmu.c b/drivers/perf/arm_pmu.c > index e23e1a1..1adabb5 100644 > --- a/drivers/perf/arm_pmu.c > +++ b/drivers/perf/arm_pmu.c > @@ -33,6 +33,21 @@ static inline u64 arm_pmu_max_period(struct arm_pmu *pmu) > return (((u64)1) << (pmu->counter_width)) - 1; > } > > +static inline u64 arm_pmu_get_event_max_period(struct arm_pmu *pmu,
The "get_" here seems a bit at odds with arm_pmu_max_period() - I'd be inlined to go for slightly more consistent naming (with a slight personal preference towards removing it here rather than adding it there)
> + struct perf_event *event) > +{ > + u64 period = arm_pmu_max_period(pmu); > + > + /* > + * To prevent shift-counter-overflow warning, create the > + * mask, by shift + OR sequence. > + */ > + if (event->hw.flags & ARMPMU_EVT_LONG) > + period = (period << pmu->counter_width) | period; > + > + return period; > +} > + > static int > armpmu_map_cache_event(const unsigned (*cache_map) > [PERF_COUNT_HW_CACHE_MAX] > @@ -122,7 +137,7 @@ int armpmu_event_set_period(struct perf_event *event) > u64 max_period; > int ret = 0; > > - max_period = arm_pmu_max_period(armpmu); > + max_period = arm_pmu_get_event_max_period(armpmu, event); > if (unlikely(left <= -period)) { > left = period; > local64_set(&hwc->period_left, left); > @@ -148,7 +163,7 @@ int armpmu_event_set_period(struct perf_event *event) > > local64_set(&hwc->prev_count, (u64)-left); > > - armpmu->write_counter(event, (u64)(-left) & 0xffffffff); > + armpmu->write_counter(event, (u64)(-left) & max_period); > > perf_event_update_userpage(event); > > @@ -160,7 +175,7 @@ u64 armpmu_event_update(struct perf_event *event) > struct arm_pmu *armpmu = to_arm_pmu(event->pmu); > struct hw_perf_event *hwc = &event->hw; > u64 delta, prev_raw_count, new_raw_count; > - u64 max_period = arm_pmu_max_period(armpmu); > + u64 max_period = arm_pmu_get_event_max_period(armpmu, event); > > again: > prev_raw_count = local64_read(&hwc->prev_count); > @@ -368,6 +383,7 @@ __hw_perf_event_init(struct perf_event *event) > struct hw_perf_event *hwc = &event->hw; > int mapping; > > + hwc->flags = 0; > mapping = armpmu->map_event(event); > > if (mapping < 0) { > @@ -670,6 +686,9 @@ static void cpu_pm_pmu_setup(struct arm_pmu *armpmu, unsigned long cmd) > continue; > > event = hw_events->events[idx]; > + /* Chained events could use multiple counters */ > + if (!event) > + continue;
This hunk looks a little out of place; does it perhaps belong to patch #6?
Robin.
> > switch (cmd) { > case CPU_PM_ENTER: > diff --git a/include/linux/perf/arm_pmu.h b/include/linux/perf/arm_pmu.h > index 705e8c3..ed7e3f7 100644 > --- a/include/linux/perf/arm_pmu.h > +++ b/include/linux/perf/arm_pmu.h > @@ -25,6 +25,12 @@ > */ > #define ARMPMU_MAX_HWEVENTS 32 > > +/* > + * ARM PMU hw_event flags > + */ > +/* Event uses a counter with double the normal width */ > +#define ARMPMU_EVT_LONG 1 > + > #define HW_OP_UNSUPPORTED 0xFFFF > #define C(_x) PERF_COUNT_HW_CACHE_##_x > #define CACHE_OP_UNSUPPORTED 0xFFFF >
| |