Messages in this thread | | | Date | Tue, 14 Nov 2023 17:10:49 +0000 | Subject | Re: [V14 3/8] drivers: perf: arm_pmuv3: Enable branch stack sampling framework | From | James Clark <> |
| |
On 14/11/2023 05:13, Anshuman Khandual wrote: [...] > > diff --git a/drivers/perf/arm_pmu.c b/drivers/perf/arm_pmu.c > index d712a19e47ac..76f1376ae594 100644 > --- a/drivers/perf/arm_pmu.c > +++ b/drivers/perf/arm_pmu.c > @@ -317,6 +317,15 @@ armpmu_del(struct perf_event *event, int flags) > struct hw_perf_event *hwc = &event->hw; > int idx = hwc->idx; > > + if (has_branch_stack(event)) { > + WARN_ON_ONCE(!hw_events->brbe_users); > + hw_events->brbe_users--; > + if (!hw_events->brbe_users) { > + hw_events->brbe_context = NULL; > + hw_events->brbe_sample_type = 0; > + } > + } > + > armpmu_stop(event, PERF_EF_UPDATE); > hw_events->events[idx] = NULL; > armpmu->clear_event_idx(hw_events, event); > @@ -333,6 +342,22 @@ armpmu_add(struct perf_event *event, int flags) > struct hw_perf_event *hwc = &event->hw; > int idx; > > + if (has_branch_stack(event)) { > + /* > + * Reset branch records buffer if a new task event gets > + * scheduled on a PMU which might have existing records. > + * Otherwise older branch records present in the buffer > + * might leak into the new task event. > + */ > + if (event->ctx->task && hw_events->brbe_context != event->ctx) { > + hw_events->brbe_context = event->ctx; > + if (armpmu->branch_reset) > + armpmu->branch_reset();
What about a per-thread event following a per-cpu event? Doesn't that also need to branch_reset()? If hw_events->brbe_context was already previously assigned, once the per-thread event is switched in it skips this reset following a per-cpu event on the same core.
I think it should be possible to add a test for this scenario by creating simulaneous per-cpu and per-thread events and checking for leakage.
> + } > + hw_events->brbe_users++; > + hw_events->brbe_sample_type = event->attr.branch_sample_type; > + } > + > /* An event following a process won't be stopped earlier */ > if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus)) > return -ENOENT; > @@ -512,13 +537,24 @@ static int armpmu_event_init(struct perf_event *event) > !cpumask_test_cpu(event->cpu, &armpmu->supported_cpus)) > return -ENOENT; > > - /* does not support taken branch sampling */ > - if (has_branch_stack(event)) > + /* > + * Branch stack sampling events are allowed > + * only on PMU which has required support. > + */ > + if (has_branch_stack(event) && !armpmu->has_branch_stack) > return -EOPNOTSUPP; > > return __hw_perf_event_init(event); > } >
[...] > +/* > + * This is a read only constant and safe during multi threaded access > + */ > +static struct perf_branch_stack zero_branch_stack = { .nr = 0, .hw_idx = -1ULL}; > + > +static void read_branch_records(struct pmu_hw_events *cpuc, > + struct perf_event *event, > + struct perf_sample_data *data, > + bool *branch_captured) > +{ > + /* > + * CPU specific branch records buffer must have been allocated already > + * for the hardware records to be captured and processed further. > + */ > + if (WARN_ON(!cpuc->branches)) > + return; > + > + /* > + * Overflowed event's branch_sample_type does not match the configured > + * branch filters in the BRBE HW. So the captured branch records here > + * cannot be co-related to the overflowed event. Report to the user as > + * if no branch records have been captured, and flush branch records. > + * The same scenario is applicable when the current task context does > + * not match with overflown event. > + */ > + if ((cpuc->brbe_sample_type != event->attr.branch_sample_type) || > + (event->ctx->task && cpuc->brbe_context != event->ctx)) { > + perf_sample_save_brstack(data, event, &zero_branch_stack); > + return; > + }
I think we should probably add a test for this scenario too. Like that the second event opened on the same thread as another event with different brbe settings always produces zero records.
I actually tried to reproduce this behaviour but couldn't. Not sure if I did something wrong though.
| |