perf: Rework perf_event_exit_event()

Make perf_event_exit_event() more robust, such that we can use it from
other contexts. Specifically the up and coming remove_on_exec.

For this to work we need to address a few issues. Remove_on_exec will
not destroy the entire context, so we cannot rely on TASK_TOMBSTONE to
disable event_function_call() and we thus have to use
perf_remove_from_context().

When using perf_remove_from_context(), there's two races to consider.
The first is against close(), where we can have concurrent tear-down
of the event. The second is against child_list iteration, which should
not find a half baked event.

To address this, teach perf_remove_from_context() to special case
!ctx->is_active and about DETACH_CHILD.

[ elver@google.com: fix racing parent/child exit in sync_child_event(). ]
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20210408103605.1676875-2-elver@google.com

include/linux/perf_event.h

@@ -607,6 +607,7 @@ struct swevent_hlist {
 #define PERF_ATTACH_TASK_DATA	0x08
 #define PERF_ATTACH_ITRACE	0x10
 #define PERF_ATTACH_SCHED_CB	0x20
+#define PERF_ATTACH_CHILD	0x40
 
 struct perf_cgroup;
 struct perf_buffer;

kernel/events/core.c

@@ -2205,6 +2205,26 @@ static void perf_group_detach(struct perf_event *event)
 	perf_event__header_size(leader);
 }
 
+static void sync_child_event(struct perf_event *child_event);
+
+static void perf_child_detach(struct perf_event *event)
+{
+	struct perf_event *parent_event = event->parent;
+
+	if (!(event->attach_state & PERF_ATTACH_CHILD))
+		return;
+
+	event->attach_state &= ~PERF_ATTACH_CHILD;
+
+	if (WARN_ON_ONCE(!parent_event))
+		return;
+
+	lockdep_assert_held(&parent_event->child_mutex);
+
+	sync_child_event(event);
+	list_del_init(&event->child_list);
+}
+
 static bool is_orphaned_event(struct perf_event *event)
 {
 	return event->state == PERF_EVENT_STATE_DEAD;
@@ -2312,6 +2332,7 @@ group_sched_out(struct perf_event *group_event,
 }
 
 #define DETACH_GROUP	0x01UL
+#define DETACH_CHILD	0x02UL
 
 /*
  * Cross CPU call to remove a performance event
@@ -2335,6 +2356,8 @@ __perf_remove_from_context(struct perf_event *event,
 	event_sched_out(event, cpuctx, ctx);
 	if (flags & DETACH_GROUP)
 		perf_group_detach(event);
+	if (flags & DETACH_CHILD)
+		perf_child_detach(event);
 	list_del_event(event, ctx);
 
 	if (!ctx->nr_events && ctx->is_active) {
@@ -2363,25 +2386,21 @@ static void perf_remove_from_context(struct perf_event *event, unsigned long fla
 
 	lockdep_assert_held(&ctx->mutex);
 
-	event_function_call(event, __perf_remove_from_context, (void *)flags);
-
 	/*
-	 * The above event_function_call() can NO-OP when it hits
+	 * Because of perf_event_exit_task(), perf_remove_from_context() ought
-	 * TASK_TOMBSTONE. In that case we must already have been detached
+	 * to work in the face of TASK_TOMBSTONE, unlike every other
-	 * from the context (by perf_event_exit_event()) but the grouping
+	 * event_function_call() user.
-	 * might still be in-tact.
 	 */
-	WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
+	raw_spin_lock_irq(&ctx->lock);
-	if ((flags & DETACH_GROUP) &&
+	if (!ctx->is_active) {
-	    (event->attach_state & PERF_ATTACH_GROUP)) {
+		__perf_remove_from_context(event, __get_cpu_context(ctx),
-		/*
+					   ctx, (void *)flags);
-		 * Since in that case we cannot possibly be scheduled, simply
-		 * detach now.
-		 */
-		raw_spin_lock_irq(&ctx->lock);
-		perf_group_detach(event);
 		raw_spin_unlock_irq(&ctx->lock);
+		return;
 	}
+	raw_spin_unlock_irq(&ctx->lock);
+
+	event_function_call(event, __perf_remove_from_context, (void *)flags);
 }
 
 /*
@@ -12377,14 +12396,17 @@ void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
 }
 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
 
-static void sync_child_event(struct perf_event *child_event,
+static void sync_child_event(struct perf_event *child_event)
-			       struct task_struct *child)
 {
 	struct perf_event *parent_event = child_event->parent;
 	u64 child_val;
 
-	if (child_event->attr.inherit_stat)
+	if (child_event->attr.inherit_stat) {
-		perf_event_read_event(child_event, child);
+		struct task_struct *task = child_event->ctx->task;
+
+		if (task && task != TASK_TOMBSTONE)
+			perf_event_read_event(child_event, task);
+	}
 
 	child_val = perf_event_count(child_event);
 
@@ -12399,60 +12421,53 @@ static void sync_child_event(struct perf_event *child_event,
 }
 
 static void
-perf_event_exit_event(struct perf_event *child_event,
+perf_event_exit_event(struct perf_event *event, struct perf_event_context *ctx)
-		      struct perf_event_context *child_ctx,
-		      struct task_struct *child)
 {
-	struct perf_event *parent_event = child_event->parent;
+	struct perf_event *parent_event = event->parent;
+	unsigned long detach_flags = 0;
+
+	if (parent_event) {
+		/*
+		 * Do not destroy the 'original' grouping; because of the
+		 * context switch optimization the original events could've
+		 * ended up in a random child task.
+		 *
+		 * If we were to destroy the original group, all group related
+		 * operations would cease to function properly after this
+		 * random child dies.
+		 *
+		 * Do destroy all inherited groups, we don't care about those
+		 * and being thorough is better.
+		 */
+		detach_flags = DETACH_GROUP | DETACH_CHILD;
+		mutex_lock(&parent_event->child_mutex);
+	}
+
+	perf_remove_from_context(event, detach_flags);
+
+	raw_spin_lock_irq(&ctx->lock);
+	if (event->state > PERF_EVENT_STATE_EXIT)
+		perf_event_set_state(event, PERF_EVENT_STATE_EXIT);
+	raw_spin_unlock_irq(&ctx->lock);
 
 	/*
-	 * Do not destroy the 'original' grouping; because of the context
+	 * Child events can be freed.
-	 * switch optimization the original events could've ended up in a
-	 * random child task.
-	 *
-	 * If we were to destroy the original group, all group related
-	 * operations would cease to function properly after this random
-	 * child dies.
-	 *
-	 * Do destroy all inherited groups, we don't care about those
-	 * and being thorough is better.
 	 */
-	raw_spin_lock_irq(&child_ctx->lock);
+	if (parent_event) {
-	WARN_ON_ONCE(child_ctx->is_active);
+		mutex_unlock(&parent_event->child_mutex);
-
+		/*
-	if (parent_event)
+		 * Kick perf_poll() for is_event_hup();
-		perf_group_detach(child_event);
+		 */
-	list_del_event(child_event, child_ctx);
+		perf_event_wakeup(parent_event);
-	perf_event_set_state(child_event, PERF_EVENT_STATE_EXIT); /* is_event_hup() */
+		free_event(event);
-	raw_spin_unlock_irq(&child_ctx->lock);
+		put_event(parent_event);
+		return;
+	}
 
 	/*
 	 * Parent events are governed by their filedesc, retain them.
 	 */
-	if (!parent_event) {
+	perf_event_wakeup(event);
-		perf_event_wakeup(child_event);
-		return;
-	}
-	/*
-	 * Child events can be cleaned up.
-	 */
-
-	sync_child_event(child_event, child);
-
-	/*
-	 * Remove this event from the parent's list
-	 */
-	WARN_ON_ONCE(parent_event->ctx->parent_ctx);
-	mutex_lock(&parent_event->child_mutex);
-	list_del_init(&child_event->child_list);
-	mutex_unlock(&parent_event->child_mutex);
-
-	/*
-	 * Kick perf_poll() for is_event_hup().
-	 */
-	perf_event_wakeup(parent_event);
-	free_event(child_event);
-	put_event(parent_event);
 }
 
 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
@@ -12509,7 +12524,7 @@ static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
 	perf_event_task(child, child_ctx, 0);
 
 	list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
-		perf_event_exit_event(child_event, child_ctx, child);
+		perf_event_exit_event(child_event, child_ctx);
 
 	mutex_unlock(&child_ctx->mutex);
 
@@ -12769,6 +12784,7 @@ inherit_event(struct perf_event *parent_event,
 	 */
 	raw_spin_lock_irqsave(&child_ctx->lock, flags);
 	add_event_to_ctx(child_event, child_ctx);
+	child_event->attach_state |= PERF_ATTACH_CHILD;
 	raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
 
 	/*