[PATCH v3 04/10] perf: Add per perf_cpu_context min_heap storage

[Ian Rogers <irogers@google.com>]

Based-on-work-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ian Rogers <irogers@google.com>
---
 include/linux/perf_event.h |  7 ++++++
 kernel/events/core.c       | 49 ++++++++++++++++++++++++++++----------
 2 files changed, 44 insertions(+), 12 deletions(-)

diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index 011dcbdbccc2..b3580afbf358 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -835,6 +835,13 @@ struct perf_cpu_context {
 	int				sched_cb_usage;
 
 	int				online;
+	/*
+	 * Per-CPU storage for iterators used in visit_groups_merge. The default
+	 * storage is of size 2 to hold the CPU and any CPU event iterators.
+	 */
+	int				itr_storage_cap;
+	struct perf_event		**itr_storage;
+	struct perf_event		*itr_default[2];
 };
 
 struct perf_output_handle {
diff --git a/kernel/events/core.c b/kernel/events/core.c
index a5a3d349a8f1..0dab60bf5935 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -3403,30 +3403,46 @@ static void __heap_add(struct min_max_heap *heap, struct perf_event *event)
 	}
 }
 
-static noinline int visit_groups_merge(struct perf_event_groups *groups, int cpu,
-			      int (*func)(struct perf_event *, void *), void *data)
+static noinline int visit_groups_merge(struct perf_cpu_context *cpuctx,
+				struct perf_event_groups *groups, int cpu,
+				int (*func)(struct perf_event *, void *),
+				void *data)
 {
 	/* Space for per CPU and/or any CPU event iterators. */
 	struct perf_event *itrs[2];
-	struct min_max_heap event_heap = {
-		.data = itrs,
-		.size = 0,
-		.cap = ARRAY_SIZE(itrs),
-	};
+	struct min_max_heap event_heap;
+	struct perf_event **evt;
 	struct perf_event *next;
 	int ret;
 
-	__heap_add(&event_heap, perf_event_groups_first(groups, -1));
+	if (cpuctx) {
+		event_heap = (struct min_max_heap){
+			.data = cpuctx->itr_storage,
+			.size = 0,
+			.cap = cpuctx->itr_storage_cap,
+		};
+	} else {
+		event_heap = (struct min_max_heap){
+			.data = itrs,
+			.size = 0,
+			.cap = ARRAY_SIZE(itrs),
+		};
+		/* Events not within a CPU context may be on any CPU. */
+		__heap_add(&event_heap, perf_event_groups_first(groups, -1));
+
+	}
+	evt = event_heap.data;
+
 	__heap_add(&event_heap, perf_event_groups_first(groups, cpu));
 
 	heapify_all(&event_heap, &perf_min_heap);
 
 	while (event_heap.size) {
-		ret = func(itrs[0], data);
+		ret = func(*evt, data);
 		if (ret)
 			return ret;
 
-		next = perf_event_groups_next(itrs[0]);
+		next = perf_event_groups_next(*evt);
 		if (next)
 			heap_pop_push(&event_heap, &next, &perf_min_heap);
 		else
@@ -3500,7 +3516,10 @@ ctx_pinned_sched_in(struct perf_event_context *ctx,
 		.can_add_hw = 1,
 	};
 
-	visit_groups_merge(&ctx->pinned_groups,
+	if (ctx != &cpuctx->ctx)
+		cpuctx = NULL;
+
+	visit_groups_merge(cpuctx, &ctx->pinned_groups,
 			   smp_processor_id(),
 			   pinned_sched_in, &sid);
 }
@@ -3515,7 +3534,10 @@ ctx_flexible_sched_in(struct perf_event_context *ctx,
 		.can_add_hw = 1,
 	};
 
-	visit_groups_merge(&ctx->flexible_groups,
+	if (ctx != &cpuctx->ctx)
+		cpuctx = NULL;
+
+	visit_groups_merge(cpuctx, &ctx->flexible_groups,
 			   smp_processor_id(),
 			   flexible_sched_in, &sid);
 }
@@ -10182,6 +10204,9 @@ int perf_pmu_register(struct pmu *pmu, const char *name, int type)
 		cpuctx->online = cpumask_test_cpu(cpu, perf_online_mask);
 
 		__perf_mux_hrtimer_init(cpuctx, cpu);
+
+		cpuctx->itr_storage_cap = ARRAY_SIZE(cpuctx->itr_default);
+		cpuctx->itr_storage = cpuctx->itr_default;
 	}
 
 got_cpu_context:
-- 
2.24.0.432.g9d3f5f5b63-goog