perf: Fix contexted inheritance

Linus reported that the RCU lockdep annotation bits triggered for this
rcu_dereference() because we're not holding rcu_read_lock().

Going over the code I cannot convince myself its correct:

 - holding a ref on the parent_ctx, doesn't avoid it being uncloned
   concurrently (as the comment says), so we can race with a free.

 - holding parent_ctx->mutex doesn't avoid the above free from taking
   place either, it would at best avoid parent_ctx from being freed.

I.e. the warning is correct. To fix the bug, serialize against the
unclone_ctx() call by extending the reach of the parent_ctx->lock.

Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
diff --git a/kernel/perf_event.c b/kernel/perf_event.c
index b782b7a..76be4c7 100644
--- a/kernel/perf_event.c
+++ b/kernel/perf_event.c
@@ -6494,7 +6494,6 @@
 
 	raw_spin_lock_irqsave(&parent_ctx->lock, flags);
 	parent_ctx->rotate_disable = 0;
-	raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
 
 	child_ctx = child->perf_event_ctxp[ctxn];
 
@@ -6502,12 +6501,11 @@
 		/*
 		 * Mark the child context as a clone of the parent
 		 * context, or of whatever the parent is a clone of.
-		 * Note that if the parent is a clone, it could get
-		 * uncloned at any point, but that doesn't matter
-		 * because the list of events and the generation
-		 * count can't have changed since we took the mutex.
+		 *
+		 * Note that if the parent is a clone, the holding of
+		 * parent_ctx->lock avoids it from being uncloned.
 		 */
-		cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
+		cloned_ctx = parent_ctx->parent_ctx;
 		if (cloned_ctx) {
 			child_ctx->parent_ctx = cloned_ctx;
 			child_ctx->parent_gen = parent_ctx->parent_gen;
@@ -6518,6 +6516,7 @@
 		get_ctx(child_ctx->parent_ctx);
 	}
 
+	raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
 	mutex_unlock(&parent_ctx->mutex);
 
 	perf_unpin_context(parent_ctx);