perf: Fix the perf context switch optimization

Currently we only optimize the context switch between two
contexts that have the same parent; this forgoes the
optimization between parent and child context, even though these
contexts could be equivalent too.

Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Shishkin, Alexander <alexander.shishkin@intel.com>
Link: http://lkml.kernel.org/r/20131007164257.GH3081@twins.programming.kicks-ass.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>

kernel/events/core.c

@@ -899,6 +899,7 @@ static void unclone_ctx(struct perf_event_context *ctx)
 		put_ctx(ctx->parent_ctx);
 		ctx->parent_ctx = NULL;
 	}
+	ctx->generation++;
 }
 
 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
@@ -1136,6 +1137,8 @@ list_add_event(struct perf_event *event, struct perf_event_context *ctx)
 	ctx->nr_events++;
 	if (event->attr.inherit_stat)
 		ctx->nr_stat++;
+
+	ctx->generation++;
 }
 
 /*
@@ -1313,6 +1316,8 @@ list_del_event(struct perf_event *event, struct perf_event_context *ctx)
 	 */
 	if (event->state > PERF_EVENT_STATE_OFF)
 		event->state = PERF_EVENT_STATE_OFF;
+
+	ctx->generation++;
 }
 
 static void perf_group_detach(struct perf_event *event)
@@ -2149,22 +2154,38 @@ static void ctx_sched_out(struct perf_event_context *ctx,
 }
 
 /*
- * Test whether two contexts are equivalent, i.e. whether they
- * have both been cloned from the same version of the same context
- * and they both have the same number of enabled events.
- * If the number of enabled events is the same, then the set
- * of enabled events should be the same, because these are both
- * inherited contexts, therefore we can't access individual events
- * in them directly with an fd; we can only enable/disable all
- * events via prctl, or enable/disable all events in a family
- * via ioctl, which will have the same effect on both contexts.
+ * Test whether two contexts are equivalent, i.e. whether they have both been
+ * cloned from the same version of the same context.
+ *
+ * Equivalence is measured using a generation number in the context that is
+ * incremented on each modification to it; see unclone_ctx(), list_add_event()
+ * and list_del_event().
  */
 static int context_equiv(struct perf_event_context *ctx1,
 			 struct perf_event_context *ctx2)
 {
-	return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
-		&& ctx1->parent_gen == ctx2->parent_gen
-		&& !ctx1->pin_count && !ctx2->pin_count;
+	/* Pinning disables the swap optimization */
+	if (ctx1->pin_count || ctx2->pin_count)
+		return 0;
+
+	/* If ctx1 is the parent of ctx2 */
+	if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
+		return 1;
+
+	/* If ctx2 is the parent of ctx1 */
+	if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
+		return 1;
+
+	/*
+	 * If ctx1 and ctx2 have the same parent; we flatten the parent
+	 * hierarchy, see perf_event_init_context().
+	 */
+	if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
+			ctx1->parent_gen == ctx2->parent_gen)
+		return 1;
+
+	/* Unmatched */
+	return 0;
 }
 
 static void __perf_event_sync_stat(struct perf_event *event,
@@ -2247,7 +2268,7 @@ static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
 {
 	struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
 	struct perf_event_context *next_ctx;
-	struct perf_event_context *parent;
+	struct perf_event_context *parent, *next_parent;
 	struct perf_cpu_context *cpuctx;
 	int do_switch = 1;
 
@@ -2259,10 +2280,18 @@ static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
 		return;
 
 	rcu_read_lock();
-	parent = rcu_dereference(ctx->parent_ctx);
 	next_ctx = next->perf_event_ctxp[ctxn];
-	if (parent && next_ctx &&
-	    rcu_dereference(next_ctx->parent_ctx) == parent) {
+	if (!next_ctx)
+		goto unlock;
+
+	parent = rcu_dereference(ctx->parent_ctx);
+	next_parent = rcu_dereference(next_ctx->parent_ctx);
+
+	/* If neither context have a parent context; they cannot be clones. */
+	if (!parent && !next_parent)
+		goto unlock;
+
+	if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
 		/*
 		 * Looks like the two contexts are clones, so we might be
 		 * able to optimize the context switch.  We lock both
@@ -2290,6 +2319,7 @@ static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
 		raw_spin_unlock(&next_ctx->lock);
 		raw_spin_unlock(&ctx->lock);
 	}
+unlock:
 	rcu_read_unlock();
 
 	if (do_switch) {
@@ -7136,7 +7166,6 @@ SYSCALL_DEFINE5(perf_event_open,
 	}
 
 	perf_install_in_context(ctx, event, event->cpu);
-	++ctx->generation;
 	perf_unpin_context(ctx);
 	mutex_unlock(&ctx->mutex);
 
@@ -7219,7 +7248,6 @@ perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
 	WARN_ON_ONCE(ctx->parent_ctx);
 	mutex_lock(&ctx->mutex);
 	perf_install_in_context(ctx, event, cpu);
-	++ctx->generation;
 	perf_unpin_context(ctx);
 	mutex_unlock(&ctx->mutex);