Re: [PATCH] perf/core: Optimize event reschedule for a PMU

From: Peter Zijlstra
Date: Tue Jul 30 2024 - 16:41:19 EST


On Tue, Jul 30, 2024 at 12:19:25PM -0700, Namhyung Kim wrote:
> @@ -2728,13 +2727,62 @@ static void ctx_resched(struct perf_cpu_context *cpuctx,
> perf_ctx_enable(task_ctx, false);
> }
>
> +static void __perf_pmu_resched(struct pmu *pmu,
> + struct perf_event_context *task_ctx,
> + enum event_type_t event_type)
> +{
> + bool cpu_event = !!(event_type & EVENT_CPU);
> + struct perf_event_pmu_context *epc = NULL;
> + struct perf_cpu_pmu_context *cpc = this_cpu_ptr(pmu->cpu_pmu_context);
> +
> + /*
> + * If pinned groups are involved, flexible groups also need to be
> + * scheduled out.
> + */
> + if (event_type & EVENT_PINNED)
> + event_type |= EVENT_FLEXIBLE;
> +
> + event_type &= EVENT_ALL;
> +
> + perf_pmu_disable(pmu);
> + if (task_ctx) {
> + if (WARN_ON_ONCE(!cpc->task_epc || cpc->task_epc->ctx != task_ctx))
> + goto out;
> +
> + epc = cpc->task_epc;
> + __pmu_ctx_sched_out(epc, event_type);
> + }
> +
> + /*
> + * Decide which cpu ctx groups to schedule out based on the types
> + * of events that caused rescheduling:
> + * - EVENT_CPU: schedule out corresponding groups;
> + * - EVENT_PINNED task events: schedule out EVENT_FLEXIBLE groups;
> + * - otherwise, do nothing more.
> + */
> + if (cpu_event)
> + __pmu_ctx_sched_out(&cpc->epc, event_type);
> + else if (event_type & EVENT_PINNED)
> + __pmu_ctx_sched_out(&cpc->epc, EVENT_FLEXIBLE);
> +
> + __pmu_ctx_sched_in(&cpc->epc, EVENT_PINNED);
> + if (task_ctx)
> + __pmu_ctx_sched_in(epc, EVENT_PINNED);
> + __pmu_ctx_sched_in(&cpc->epc, EVENT_FLEXIBLE);
> + if (task_ctx)
> + __pmu_ctx_sched_in(epc, EVENT_FLEXIBLE);
> +
> +out:
> + perf_pmu_enable(pmu);
> +}

I so dislike duplication...

So lets see, ctx_resched() has pmu_ctx iterations in:

perf_ctx_{en,dis}able()
ctx_sched_{in,out}()

Can't we punch a 'struct pmu *pmu' argument through those callchains and
short-circuit the iteration when !NULL?

The alternative would be to lift the iteration I suppose.