[PATCH 3.2 112/126] perf: Fix event->ctx locking

[Ben Hutchings]

3.2.85-rc1 review patch.  If anyone has any objections, please let me know.

------------------

From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>

commit f63a8daa5812afef4f06c962351687e1ff9ccb2b upstream.

There have been a few reported issues wrt. the lack of locking around
changing event->ctx. This patch tries to address those.

It avoids the whole rwsem thing; and while it appears to work, please
give it some thought in review.

What I did fail at is sensible runtime checks on the use of
event->ctx, the RCU use makes it very hard.

Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
Cc: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>
Cc: Jiri Olsa <jolsa@xxxxxxxxxx>
Cc: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
Cc: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
Link: http://lkml.kernel.org/r/20150123125834.209535886@xxxxxxxxxxxxx
Signed-off-by: Ingo Molnar <mingo@xxxxxxxxxx>
[bwh: Backported to 3.2:
 - We don't have perf_pmu_migrate_context()
 - Adjust context]
Signed-off-by: Ben Hutchings <ben@xxxxxxxxxxxxxxx>
---
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -665,6 +665,76 @@ static void put_ctx(struct perf_event_co
 	}
 }
 
+/*
+ * Because of perf_event::ctx migration in sys_perf_event_open::move_group we
+ * need some magic.
+ *
+ * Those places that change perf_event::ctx will hold both
+ * perf_event_ctx::mutex of the 'old' and 'new' ctx value.
+ *
+ * Lock ordering is by mutex address. There is one other site where
+ * perf_event_context::mutex nests and that is put_event(). But remember that
+ * that is a parent<->child context relation, and migration does not affect
+ * children, therefore these two orderings should not interact.
+ *
+ * The change in perf_event::ctx does not affect children (as claimed above)
+ * because the sys_perf_event_open() case will install a new event and break
+ * the ctx parent<->child relation.
+ *
+ * The places that change perf_event::ctx will issue:
+ *
+ *   perf_remove_from_context();
+ *   synchronize_rcu();
+ *   perf_install_in_context();
+ *
+ * to affect the change. The remove_from_context() + synchronize_rcu() should
+ * quiesce the event, after which we can install it in the new location. This
+ * means that only external vectors (perf_fops, prctl) can perturb the event
+ * while in transit. Therefore all such accessors should also acquire
+ * perf_event_context::mutex to serialize against this.
+ *
+ * However; because event->ctx can change while we're waiting to acquire
+ * ctx->mutex we must be careful and use the below perf_event_ctx_lock()
+ * function.
+ *
+ * Lock order:
+ *	task_struct::perf_event_mutex
+ *	  perf_event_context::mutex
+ *	    perf_event_context::lock
+ *	    perf_event::child_mutex;
+ *	    perf_event::mmap_mutex
+ *	    mmap_sem
+ */
+static struct perf_event_context *perf_event_ctx_lock(struct perf_event *event)
+{
+	struct perf_event_context *ctx;
+
+again:
+	rcu_read_lock();
+	ctx = ACCESS_ONCE(event->ctx);
+	if (!atomic_inc_not_zero(&ctx->refcount)) {
+		rcu_read_unlock();
+		goto again;
+	}
+	rcu_read_unlock();
+
+	mutex_lock(&ctx->mutex);
+	if (event->ctx != ctx) {
+		mutex_unlock(&ctx->mutex);
+		put_ctx(ctx);
+		goto again;
+	}
+
+	return ctx;
+}
+
+static void perf_event_ctx_unlock(struct perf_event *event,
+				  struct perf_event_context *ctx)
+{
+	mutex_unlock(&ctx->mutex);
+	put_ctx(ctx);
+}
+
 static void unclone_ctx(struct perf_event_context *ctx)
 {
 	if (ctx->parent_ctx) {
@@ -1325,7 +1395,7 @@ static int __perf_event_disable(void *in
  * is the current context on this CPU and preemption is disabled,
  * hence we can't get into perf_event_task_sched_out for this context.
  */
-void perf_event_disable(struct perf_event *event)
+static void _perf_event_disable(struct perf_event *event)
 {
 	struct perf_event_context *ctx = event->ctx;
 	struct task_struct *task = ctx->task;
@@ -1367,6 +1437,19 @@ retry:
 	raw_spin_unlock_irq(&ctx->lock);
 }
 
+/*
+ * Strictly speaking kernel users cannot create groups and therefore this
+ * interface does not need the perf_event_ctx_lock() magic.
+ */
+void perf_event_disable(struct perf_event *event)
+{
+	struct perf_event_context *ctx;
+
+	ctx = perf_event_ctx_lock(event);
+	_perf_event_disable(event);
+	perf_event_ctx_unlock(event, ctx);
+}
+
 static void perf_set_shadow_time(struct perf_event *event,
 				 struct perf_event_context *ctx,
 				 u64 tstamp)
@@ -1813,7 +1896,7 @@ unlock:
  * perf_event_for_each_child or perf_event_for_each as described
  * for perf_event_disable.
  */
-void perf_event_enable(struct perf_event *event)
+static void _perf_event_enable(struct perf_event *event)
 {
 	struct perf_event_context *ctx = event->ctx;
 	struct task_struct *task = ctx->task;
@@ -1870,7 +1953,19 @@ out:
 	raw_spin_unlock_irq(&ctx->lock);
 }
 
-int perf_event_refresh(struct perf_event *event, int refresh)
+/*
+ * See perf_event_disable();
+ */
+void perf_event_enable(struct perf_event *event)
+{
+	struct perf_event_context *ctx;
+
+	ctx = perf_event_ctx_lock(event);
+	_perf_event_enable(event);
+	perf_event_ctx_unlock(event, ctx);
+}
+
+static int _perf_event_refresh(struct perf_event *event, int refresh)
 {
 	/*
 	 * not supported on inherited events
@@ -1879,10 +1974,25 @@ int perf_event_refresh(struct perf_event
 		return -EINVAL;
 
 	atomic_add(refresh, &event->event_limit);
-	perf_event_enable(event);
+	_perf_event_enable(event);
 
 	return 0;
 }
+
+/*
+ * See perf_event_disable()
+ */
+int perf_event_refresh(struct perf_event *event, int refresh)
+{
+	struct perf_event_context *ctx;
+	int ret;
+
+	ctx = perf_event_ctx_lock(event);
+	ret = _perf_event_refresh(event, refresh);
+	perf_event_ctx_unlock(event, ctx);
+
+	return ret;
+}
 EXPORT_SYMBOL_GPL(perf_event_refresh);
 
 static void ctx_sched_out(struct perf_event_context *ctx,
@@ -3110,7 +3220,16 @@ static void put_event(struct perf_event
 	rcu_read_unlock();
 
 	if (owner) {
-		mutex_lock(&owner->perf_event_mutex);
+		/*
+		 * If we're here through perf_event_exit_task() we're already
+		 * holding ctx->mutex which would be an inversion wrt. the
+		 * normal lock order.
+		 *
+		 * However we can safely take this lock because its the child
+		 * ctx->mutex.
+		 */
+		mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING);
+
 		/*
 		 * We have to re-check the event->owner field, if it is cleared
 		 * we raced with perf_event_exit_task(), acquiring the mutex
@@ -3162,12 +3281,13 @@ static int perf_event_read_group(struct
 				   u64 read_format, char __user *buf)
 {
 	struct perf_event *leader = event->group_leader, *sub;
-	int n = 0, size = 0, ret = -EFAULT;
 	struct perf_event_context *ctx = leader->ctx;
-	u64 values[5];
+	int n = 0, size = 0, ret;
 	u64 count, enabled, running;
+	u64 values[5];
+
+	lockdep_assert_held(&ctx->mutex);
 
-	mutex_lock(&ctx->mutex);
 	count = perf_event_read_value(leader, &enabled, &running);
 
 	values[n++] = 1 + leader->nr_siblings;
@@ -3182,7 +3302,7 @@ static int perf_event_read_group(struct
 	size = n * sizeof(u64);
 
 	if (copy_to_user(buf, values, size))
-		goto unlock;
+		return -EFAULT;
 
 	ret = size;
 
@@ -3196,14 +3316,11 @@ static int perf_event_read_group(struct
 		size = n * sizeof(u64);
 
 		if (copy_to_user(buf + ret, values, size)) {
-			ret = -EFAULT;
-			goto unlock;
+			return -EFAULT;
 		}
 
 		ret += size;
 	}
-unlock:
-	mutex_unlock(&ctx->mutex);
 
 	return ret;
 }
@@ -3262,8 +3379,14 @@ static ssize_t
 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 {
 	struct perf_event *event = file->private_data;
+	struct perf_event_context *ctx;
+	int ret;
 
-	return perf_read_hw(event, buf, count);
+	ctx = perf_event_ctx_lock(event);
+	ret = perf_read_hw(event, buf, count);
+	perf_event_ctx_unlock(event, ctx);
+
+	return ret;
 }
 
 static unsigned int perf_poll(struct file *file, poll_table *wait)
@@ -3287,7 +3410,7 @@ static unsigned int perf_poll(struct fil
 	return events;
 }
 
-static void perf_event_reset(struct perf_event *event)
+static void _perf_event_reset(struct perf_event *event)
 {
 	(void)perf_event_read(event);
 	local64_set(&event->count, 0);
@@ -3306,6 +3429,7 @@ static void perf_event_for_each_child(st
 	struct perf_event *child;
 
 	WARN_ON_ONCE(event->ctx->parent_ctx);
+
 	mutex_lock(&event->child_mutex);
 	func(event);
 	list_for_each_entry(child, &event->child_list, child_list)
@@ -3319,15 +3443,14 @@ static void perf_event_for_each(struct p
 	struct perf_event_context *ctx = event->ctx;
 	struct perf_event *sibling;
 
-	WARN_ON_ONCE(ctx->parent_ctx);
-	mutex_lock(&ctx->mutex);
+	lockdep_assert_held(&ctx->mutex);
+
 	event = event->group_leader;
 
 	perf_event_for_each_child(event, func);
 	func(event);
 	list_for_each_entry(sibling, &event->sibling_list, group_entry)
 		perf_event_for_each_child(sibling, func);
-	mutex_unlock(&ctx->mutex);
 }
 
 static int perf_event_period(struct perf_event *event, u64 __user *arg)
@@ -3386,25 +3509,24 @@ static int perf_event_set_output(struct
 				 struct perf_event *output_event);
 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
 
-static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
+static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg)
 {
-	struct perf_event *event = file->private_data;
 	void (*func)(struct perf_event *);
 	u32 flags = arg;
 
 	switch (cmd) {
 	case PERF_EVENT_IOC_ENABLE:
-		func = perf_event_enable;
+		func = _perf_event_enable;
 		break;
 	case PERF_EVENT_IOC_DISABLE:
-		func = perf_event_disable;
+		func = _perf_event_disable;
 		break;
 	case PERF_EVENT_IOC_RESET:
-		func = perf_event_reset;
+		func = _perf_event_reset;
 		break;
 
 	case PERF_EVENT_IOC_REFRESH:
-		return perf_event_refresh(event, arg);
+		return _perf_event_refresh(event, arg);
 
 	case PERF_EVENT_IOC_PERIOD:
 		return perf_event_period(event, (u64 __user *)arg);
@@ -3445,6 +3567,19 @@ static long perf_ioctl(struct file *file
 	return 0;
 }
 
+static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
+{
+	struct perf_event *event = file->private_data;
+	struct perf_event_context *ctx;
+	long ret;
+
+	ctx = perf_event_ctx_lock(event);
+	ret = _perf_ioctl(event, cmd, arg);
+	perf_event_ctx_unlock(event, ctx);
+
+	return ret;
+}
+
 #ifdef CONFIG_COMPAT
 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
 				unsigned long arg)
@@ -3466,11 +3601,15 @@ static long perf_compat_ioctl(struct fil
 
 int perf_event_task_enable(void)
 {
+	struct perf_event_context *ctx;
 	struct perf_event *event;
 
 	mutex_lock(&current->perf_event_mutex);
-	list_for_each_entry(event, &current->perf_event_list, owner_entry)
-		perf_event_for_each_child(event, perf_event_enable);
+	list_for_each_entry(event, &current->perf_event_list, owner_entry) {
+		ctx = perf_event_ctx_lock(event);
+		perf_event_for_each_child(event, _perf_event_enable);
+		perf_event_ctx_unlock(event, ctx);
+	}
 	mutex_unlock(&current->perf_event_mutex);
 
 	return 0;
@@ -3478,11 +3617,15 @@ int perf_event_task_enable(void)
 
 int perf_event_task_disable(void)
 {
+	struct perf_event_context *ctx;
 	struct perf_event *event;
 
 	mutex_lock(&current->perf_event_mutex);
-	list_for_each_entry(event, &current->perf_event_list, owner_entry)
-		perf_event_for_each_child(event, perf_event_disable);
+	list_for_each_entry(event, &current->perf_event_list, owner_entry) {
+		ctx = perf_event_ctx_lock(event);
+		perf_event_for_each_child(event, _perf_event_disable);
+		perf_event_ctx_unlock(event, ctx);
+	}
 	mutex_unlock(&current->perf_event_mutex);
 
 	return 0;
@@ -6322,6 +6465,15 @@ out:
 	return ret;
 }
 
+static void mutex_lock_double(struct mutex *a, struct mutex *b)
+{
+	if (b < a)
+		swap(a, b);
+
+	mutex_lock(a);
+	mutex_lock_nested(b, SINGLE_DEPTH_NESTING);
+}
+
 /**
  * sys_perf_event_open - open a performance event, associate it to a task/cpu
  *
@@ -6337,7 +6489,7 @@ SYSCALL_DEFINE5(perf_event_open,
 	struct perf_event *group_leader = NULL, *output_event = NULL;
 	struct perf_event *event, *sibling;
 	struct perf_event_attr attr;
-	struct perf_event_context *ctx;
+	struct perf_event_context *ctx, *uninitialized_var(gctx);
 	struct file *event_file = NULL;
 	struct file *group_file = NULL;
 	struct task_struct *task = NULL;
@@ -6509,9 +6661,14 @@ SYSCALL_DEFINE5(perf_event_open,
 	}
 
 	if (move_group) {
-		struct perf_event_context *gctx = group_leader->ctx;
+		gctx = group_leader->ctx;
+
+		/*
+		 * See perf_event_ctx_lock() for comments on the details
+		 * of swizzling perf_event::ctx.
+		 */
+		mutex_lock_double(&gctx->mutex, &ctx->mutex);
 
-		mutex_lock(&gctx->mutex);
 		perf_remove_from_context(group_leader, false);
 
 		/*
@@ -6526,14 +6683,19 @@ SYSCALL_DEFINE5(perf_event_open,
 			perf_event__state_init(sibling);
 			put_ctx(gctx);
 		}
-		mutex_unlock(&gctx->mutex);
-		put_ctx(gctx);
+	} else {
+		mutex_lock(&ctx->mutex);
 	}
 
 	WARN_ON_ONCE(ctx->parent_ctx);
-	mutex_lock(&ctx->mutex);
 
 	if (move_group) {
+		/*
+		 * Wait for everybody to stop referencing the events through
+		 * the old lists, before installing it on new lists.
+		 */
+		synchronize_rcu();
+
 		perf_install_in_context(ctx, group_leader, cpu);
 		get_ctx(ctx);
 		list_for_each_entry(sibling, &group_leader->sibling_list,
@@ -6546,6 +6708,11 @@ SYSCALL_DEFINE5(perf_event_open,
 	perf_install_in_context(ctx, event, cpu);
 	++ctx->generation;
 	perf_unpin_context(ctx);
+
+	if (move_group) {
+		mutex_unlock(&gctx->mutex);
+		put_ctx(gctx);
+	}
 	mutex_unlock(&ctx->mutex);
 
 	event->owner = current;