perf: Add a pmu capability for "exclusive" events

Usually, pmus that do, for example, instruction tracing, would only ever
be able to have one event per task per cpu (or per perf_event_context). For
such pmus it makes sense to disallow creating conflicting events early on,
so as to provide consistent behavior for the user.

This patch adds a pmu capability that indicates such constraint on event
creation.

Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Kaixu Xia <kaixu.xia@linaro.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Robert Richter <rric@kernel.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: acme@infradead.org
Cc: adrian.hunter@intel.com
Cc: kan.liang@intel.com
Cc: markus.t.metzger@intel.com
Cc: mathieu.poirier@linaro.org
Link: http://lkml.kernel.org/r/1422613866-113186-1-git-send-email-alexander.shishkin@linux.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>

include/linux/perf_event.h

@@ -176,6 +176,7 @@ struct perf_event;
 #define PERF_PMU_CAP_NO_NMI			0x02
 #define PERF_PMU_CAP_AUX_NO_SG			0x04
 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF		0x08
+#define PERF_PMU_CAP_EXCLUSIVE			0x10
 
 /**
  * struct pmu - generic performance monitoring unit
@@ -196,6 +197,7 @@ struct pmu {
 
 	int * __percpu			pmu_disable_count;
 	struct perf_cpu_context * __percpu pmu_cpu_context;
+	atomic_t			exclusive_cnt; /* < 0: cpu; > 0: tsk */
 	int				task_ctx_nr;
 	int				hrtimer_interval_ms;
 

kernel/events/core.c

@@ -3459,6 +3459,91 @@ static void unaccount_event(struct perf_event *event)
 	unaccount_event_cpu(event, event->cpu);
 }
 
+/*
+ * The following implement mutual exclusion of events on "exclusive" pmus
+ * (PERF_PMU_CAP_EXCLUSIVE). Such pmus can only have one event scheduled
+ * at a time, so we disallow creating events that might conflict, namely:
+ *
+ *  1) cpu-wide events in the presence of per-task events,
+ *  2) per-task events in the presence of cpu-wide events,
+ *  3) two matching events on the same context.
+ *
+ * The former two cases are handled in the allocation path (perf_event_alloc(),
+ * __free_event()), the latter -- before the first perf_install_in_context().
+ */
+static int exclusive_event_init(struct perf_event *event)
+{
+	struct pmu *pmu = event->pmu;
+
+	if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
+		return 0;
+
+	/*
+	 * Prevent co-existence of per-task and cpu-wide events on the
+	 * same exclusive pmu.
+	 *
+	 * Negative pmu::exclusive_cnt means there are cpu-wide
+	 * events on this "exclusive" pmu, positive means there are
+	 * per-task events.
+	 *
+	 * Since this is called in perf_event_alloc() path, event::ctx
+	 * doesn't exist yet; it is, however, safe to use PERF_ATTACH_TASK
+	 * to mean "per-task event", because unlike other attach states it
+	 * never gets cleared.
+	 */
+	if (event->attach_state & PERF_ATTACH_TASK) {
+		if (!atomic_inc_unless_negative(&pmu->exclusive_cnt))
+			return -EBUSY;
+	} else {
+		if (!atomic_dec_unless_positive(&pmu->exclusive_cnt))
+			return -EBUSY;
+	}
+
+	return 0;
+}
+
+static void exclusive_event_destroy(struct perf_event *event)
+{
+	struct pmu *pmu = event->pmu;
+
+	if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
+		return;
+
+	/* see comment in exclusive_event_init() */
+	if (event->attach_state & PERF_ATTACH_TASK)
+		atomic_dec(&pmu->exclusive_cnt);
+	else
+		atomic_inc(&pmu->exclusive_cnt);
+}
+
+static bool exclusive_event_match(struct perf_event *e1, struct perf_event *e2)
+{
+	if ((e1->pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) &&
+	    (e1->cpu == e2->cpu ||
+	     e1->cpu == -1 ||
+	     e2->cpu == -1))
+		return true;
+	return false;
+}
+
+/* Called under the same ctx::mutex as perf_install_in_context() */
+static bool exclusive_event_installable(struct perf_event *event,
+					struct perf_event_context *ctx)
+{
+	struct perf_event *iter_event;
+	struct pmu *pmu = event->pmu;
+
+	if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
+		return true;
+
+	list_for_each_entry(iter_event, &ctx->event_list, event_entry) {
+		if (exclusive_event_match(iter_event, event))
+			return false;
+	}
+
+	return true;
+}
+
 static void __free_event(struct perf_event *event)
 {
 	if (!event->parent) {
@@ -3472,8 +3557,10 @@ static void __free_event(struct perf_event *event)
 	if (event->ctx)
 		put_ctx(event->ctx);
 
-	if (event->pmu)
+	if (event->pmu) {
+		exclusive_event_destroy(event);
 		module_put(event->pmu->module);
+	}
 
 	call_rcu(&event->rcu_head, free_event_rcu);
 }
@@ -7150,6 +7237,7 @@ got_cpu_context:
 		pmu->event_idx = perf_event_idx_default;
 
 	list_add_rcu(&pmu->entry, &pmus);
+	atomic_set(&pmu->exclusive_cnt, 0);
 	ret = 0;
 unlock:
 	mutex_unlock(&pmus_lock);
@@ -7405,16 +7493,23 @@ perf_event_alloc(struct perf_event_attr *attr, int cpu,
 		goto err_ns;
 	}
 
+	err = exclusive_event_init(event);
+	if (err)
+		goto err_pmu;
+
 	if (!event->parent) {
 		if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
 			err = get_callchain_buffers();
 			if (err)
-				goto err_pmu;
+				goto err_per_task;
 		}
 	}
 
 	return event;
 
+err_per_task:
+	exclusive_event_destroy(event);
+
 err_pmu:
 	if (event->destroy)
 		event->destroy(event);
@@ -7819,6 +7914,11 @@ SYSCALL_DEFINE5(perf_event_open,
 		goto err_alloc;
 	}
 
+	if ((pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) && group_leader) {
+		err = -EBUSY;
+		goto err_context;
+	}
+
 	if (task) {
 		put_task_struct(task);
 		task = NULL;
@@ -7941,6 +8041,13 @@ SYSCALL_DEFINE5(perf_event_open,
 		get_ctx(ctx);
 	}
 
+	if (!exclusive_event_installable(event, ctx)) {
+		err = -EBUSY;
+		mutex_unlock(&ctx->mutex);
+		fput(event_file);
+		goto err_context;
+	}
+
 	perf_install_in_context(ctx, event, event->cpu);
 	perf_unpin_context(ctx);
 
@@ -8032,6 +8139,14 @@ perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
 
 	WARN_ON_ONCE(ctx->parent_ctx);
 	mutex_lock(&ctx->mutex);
+	if (!exclusive_event_installable(event, ctx)) {
+		mutex_unlock(&ctx->mutex);
+		perf_unpin_context(ctx);
+		put_ctx(ctx);
+		err = -EBUSY;
+		goto err_free;
+	}
+
 	perf_install_in_context(ctx, event, cpu);
 	perf_unpin_context(ctx);
 	mutex_unlock(&ctx->mutex);