perf: Fix AUX buffer refcounting

Its currently possible to drop the last refcount to the aux buffer
from NMI context, which results in the expected fireworks.

The refcounting needs a bigger overhaul, but to cure the immediate
problem, delay the freeing by using an irq_work.

Reviewed-and-tested-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Reported-by: Vince Weaver <vincent.weaver@maine.edu>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20150618103249.GK19282@twins.programming.kicks-ass.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>

kernel/events/core.c

@@ -4358,14 +4358,6 @@ static void ring_buffer_wakeup(struct perf_event *event)
 	rcu_read_unlock();
 }
 
-static void rb_free_rcu(struct rcu_head *rcu_head)
-{
-	struct ring_buffer *rb;
-
-	rb = container_of(rcu_head, struct ring_buffer, rcu_head);
-	rb_free(rb);
-}
-
 struct ring_buffer *ring_buffer_get(struct perf_event *event)
 {
 	struct ring_buffer *rb;

kernel/events/internal.h

@@ -11,6 +11,7 @@
 struct ring_buffer {
 	atomic_t			refcount;
 	struct rcu_head			rcu_head;
+	struct irq_work			irq_work;
 #ifdef CONFIG_PERF_USE_VMALLOC
 	struct work_struct		work;
 	int				page_order;	/* allocation order  */
@@ -55,6 +56,15 @@ struct ring_buffer {
 };
 
 extern void rb_free(struct ring_buffer *rb);
+
+static inline void rb_free_rcu(struct rcu_head *rcu_head)
+{
+	struct ring_buffer *rb;
+
+	rb = container_of(rcu_head, struct ring_buffer, rcu_head);
+	rb_free(rb);
+}
+
 extern struct ring_buffer *
 rb_alloc(int nr_pages, long watermark, int cpu, int flags);
 extern void perf_event_wakeup(struct perf_event *event);

kernel/events/ring_buffer.c

@@ -221,6 +221,8 @@ void perf_output_end(struct perf_output_handle *handle)
 	rcu_read_unlock();
 }
 
+static void rb_irq_work(struct irq_work *work);
+
 static void
 ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
 {
@@ -241,6 +243,16 @@ ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
 
 	INIT_LIST_HEAD(&rb->event_list);
 	spin_lock_init(&rb->event_lock);
+	init_irq_work(&rb->irq_work, rb_irq_work);
+}
+
+static void ring_buffer_put_async(struct ring_buffer *rb)
+{
+	if (!atomic_dec_and_test(&rb->refcount))
+		return;
+
+	rb->rcu_head.next = (void *)rb;
+	irq_work_queue(&rb->irq_work);
 }
 
 /*
@@ -319,7 +331,7 @@ err_put:
 	rb_free_aux(rb);
 
 err:
-	ring_buffer_put(rb);
+	ring_buffer_put_async(rb);
 	handle->event = NULL;
 
 	return NULL;
@@ -370,7 +382,7 @@ void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
 
 	local_set(&rb->aux_nest, 0);
 	rb_free_aux(rb);
-	ring_buffer_put(rb);
+	ring_buffer_put_async(rb);
 }
 
 /*
@@ -557,7 +569,18 @@ static void __rb_free_aux(struct ring_buffer *rb)
 void rb_free_aux(struct ring_buffer *rb)
 {
 	if (atomic_dec_and_test(&rb->aux_refcount))
+		irq_work_queue(&rb->irq_work);
+}
+
+static void rb_irq_work(struct irq_work *work)
+{
+	struct ring_buffer *rb = container_of(work, struct ring_buffer, irq_work);
+
+	if (!atomic_read(&rb->aux_refcount))
 		__rb_free_aux(rb);
+
+	if (rb->rcu_head.next == (void *)rb)
+		call_rcu(&rb->rcu_head, rb_free_rcu);
 }
 
 #ifndef CONFIG_PERF_USE_VMALLOC