Batch enqueue/dequeue for bqueue

The Blocking Queue (bqueue) code is used by zfs send/receive to send
messages between the various threads.  It uses a shared linked list,
which is locked whenever we enqueue or dequeue.  For workloads which
process many blocks per second, the locking on the shared list can be
quite expensive.

This commit changes the bqueue logic to have 3 linked lists:
1. An enquing list, which is used only by the (single) enquing thread,
   and thus needs no locks.
2. A shared list, with an associated lock.
3. A dequing list, which is used only by the (single) dequing thread,
   and thus needs no locks.

The entire enquing list can be moved to the shared list in constant
time, and the entire shared list can be moved to the dequing list in
constant time.  These operations only happen when the `fill_fraction` is
reached, or on an explicit flush request.  Therefore, the lock only
needs to be acquired infrequently.

The API already allows for dequing to block until an explicit flush, so
callers don't need to be changed.

Signed-off-by: Matthew Ahrens <[email protected]>

Author: ahrens

include/sys/bqueue.h

@@ -27,10 +27,14 @@ extern "C" {
 
 typedef struct bqueue {
 	list_t bq_list;
+	size_t bq_size;
+	list_t bq_dequeuing_list;
+	size_t bq_dequeuing_size;
+	list_t bq_enqueuing_list;
+	size_t bq_enqueuing_size;
 	kmutex_t bq_lock;
 	kcondvar_t bq_add_cv;
 	kcondvar_t bq_pop_cv;
-	size_t bq_size;
 	size_t bq_maxsize;
 	uint_t bq_fill_fraction;
 	size_t bq_node_offset;
@@ -47,7 +51,6 @@ void bqueue_destroy(bqueue_t *);
 void bqueue_enqueue(bqueue_t *, void *, size_t);
 void bqueue_enqueue_flush(bqueue_t *, void *, size_t);
 void *bqueue_dequeue(bqueue_t *);
-boolean_t bqueue_empty(bqueue_t *);
 
 #ifdef	__cplusplus
 }

module/zfs/bqueue.c

@@ -27,19 +27,26 @@ obj2node(bqueue_t *q, void *data)
 
 /*
  * Initialize a blocking queue  The maximum capacity of the queue is set to
- * size.  Types that are stored in a bqueue must contain a bqueue_node_t,
- * and node_offset must be its offset from the start of the struct.
- * fill_fraction is a performance tuning value; when the queue is full, any
- * threads attempting to enqueue records will block.  They will block until
- * they're signaled, which will occur when the queue is at least 1/fill_fraction
+ * size.  Types that are stored in a bqueue must contain a bqueue_node_t, and
+ * node_offset must be its offset from the start of the struct. fill_fraction
+ * is a performance tuning value; when the queue is full, any threads
+ * attempting to enqueue records will block.  They will block until they're
+ * signaled, which will occur when the queue is at least 1/fill_fraction
  * empty.  Similar behavior occurs on dequeue; if the queue is empty, threads
- * block.  They will be signalled when the queue has 1/fill_fraction full, or
- * when bqueue_flush is called.  As a result, you must call bqueue_flush when
- * you enqueue your final record on a thread, in case the dequeueing threads are
- * currently blocked and that enqueue does not cause them to be awoken.
- * Alternatively, this behavior can be disabled (causing signaling to happen
- * immediately) by setting fill_fraction to any value larger than size.
- * Return 0 on success, or -1 on failure.
+ * block.  They will be signalled when the queue has 1/fill_fraction full.
+ * As a result, you must call bqueue_enqueue_flush() when you enqueue your
+ * final record on a thread, in case the dequeuing threads are currently
+ * blocked and that enqueue does not cause them to be woken. Alternatively,
+ * this behavior can be disabled (causing signaling to happen immediately) by
+ * setting fill_fraction to any value larger than size. Return 0 on success,
+ * or -1 on failure.
+ *
+ * Note: The caller must ensure that for a given bqueue_t, there's only a
+ * single call to bqueue_enqueue() running at a time (e.g. by calling only
+ * from a single thread, or with locking around the call). Similarly, the
+ * caller must ensure that there's only a single call to bqueue_dequeue()
+ * running at a time. However, the one call to bqueue_enqueue() may be
+ * invoked concurrently with the one call to bqueue_dequeue().
  */
 int
 bqueue_init(bqueue_t *q, uint_t fill_fraction, size_t size, size_t node_offset)
@@ -49,11 +56,17 @@ bqueue_init(bqueue_t *q, uint_t fill_fraction, size_t size, size_t node_offset)
 	}
 	list_create(&q->bq_list, node_offset + sizeof (bqueue_node_t),
 	    node_offset + offsetof(bqueue_node_t, bqn_node));
+	list_create(&q->bq_dequeuing_list, node_offset + sizeof (bqueue_node_t),
+	    node_offset + offsetof(bqueue_node_t, bqn_node));
+	list_create(&q->bq_enqueuing_list, node_offset + sizeof (bqueue_node_t),
+	    node_offset + offsetof(bqueue_node_t, bqn_node));
 	cv_init(&q->bq_add_cv, NULL, CV_DEFAULT, NULL);
 	cv_init(&q->bq_pop_cv, NULL, CV_DEFAULT, NULL);
 	mutex_init(&q->bq_lock, NULL, MUTEX_DEFAULT, NULL);
 	q->bq_node_offset = node_offset;
 	q->bq_size = 0;
+	q->bq_dequeuing_size = 0;
+	q->bq_enqueuing_size = 0;
 	q->bq_maxsize = size;
 	q->bq_fill_fraction = fill_fraction;
 	return (0);
@@ -69,9 +82,13 @@ bqueue_destroy(bqueue_t *q)
 {
 	mutex_enter(&q->bq_lock);
 	ASSERT0(q->bq_size);
+	ASSERT0(q->bq_dequeuing_size);
+	ASSERT0(q->bq_enqueuing_size);
 	cv_destroy(&q->bq_add_cv);
 	cv_destroy(&q->bq_pop_cv);
 	list_destroy(&q->bq_list);
+	list_destroy(&q->bq_dequeuing_list);
+	list_destroy(&q->bq_enqueuing_list);
 	mutex_exit(&q->bq_lock);
 	mutex_destroy(&q->bq_lock);
 }
@@ -81,23 +98,24 @@ bqueue_enqueue_impl(bqueue_t *q, void *data, size_t item_size, boolean_t flush)
 {
 	ASSERT3U(item_size, >, 0);
 	ASSERT3U(item_size, <=, q->bq_maxsize);
-	mutex_enter(&q->bq_lock);
+
 	obj2node(q, data)->bqn_size = item_size;
-	while (q->bq_size && q->bq_size + item_size > q->bq_maxsize) {
-		/*
-		 * Wake up bqueue_dequeue() thread if already sleeping in order
-		 * to prevent the deadlock condition
-		 */
-		cv_signal(&q->bq_pop_cv);
-		cv_wait_sig(&q->bq_add_cv, &q->bq_lock);
-	}
-	q->bq_size += item_size;
-	list_insert_tail(&q->bq_list, data);
-	if (flush)
+	q->bq_enqueuing_size += item_size;
+	list_insert_tail(&q->bq_enqueuing_list, data);
+
+	if (flush ||
+	    q->bq_enqueuing_size >= q->bq_maxsize / q->bq_fill_fraction) {
+		/* Append the enquing list to the shared list. */
+		mutex_enter(&q->bq_lock);
+		while (q->bq_size > q->bq_maxsize) {
+			cv_wait_sig(&q->bq_add_cv, &q->bq_lock);
+		}
+		q->bq_size += q->bq_enqueuing_size;
+		list_move_tail(&q->bq_list, &q->bq_enqueuing_list);
+		q->bq_enqueuing_size = 0;
 		cv_broadcast(&q->bq_pop_cv);
-	else if (q->bq_size >= q->bq_maxsize / q->bq_fill_fraction)
-		cv_signal(&q->bq_pop_cv);
-	mutex_exit(&q->bq_lock);
+		mutex_exit(&q->bq_lock);
+	}
 }
 
 /*
@@ -115,8 +133,8 @@ bqueue_enqueue(bqueue_t *q, void *data, size_t item_size)
  * Enqueue an entry, and then flush the queue.  This forces the popping threads
  * to wake up, even if we're below the fill fraction.  We have this in a single
  * function, rather than having a separate call, because it prevents race
- * conditions between the enqueuing thread and the dequeueing thread, where the
- * enqueueing thread will wake up the dequeueing thread, that thread will
+ * conditions between the enqueuing thread and the dequeuing thread, where the
+ * enqueueing thread will wake up the dequeuing thread, that thread will
  * destroy the condvar before the enqueuing thread is done.
  */
 void
@@ -132,27 +150,26 @@ bqueue_enqueue_flush(bqueue_t *q, void *data, size_t item_size)
 void *
 bqueue_dequeue(bqueue_t *q)
 {
-	void *ret = NULL;
-	size_t item_size;
-	mutex_enter(&q->bq_lock);
-	while (q->bq_size == 0) {
-		cv_wait_sig(&q->bq_pop_cv, &q->bq_lock);
+	void *ret = list_remove_head(&q->bq_dequeuing_list);
+	if (ret == NULL) {
+		/*
+		 * Dequeuing list is empty.  Wait for there to be something on
+		 * the shared list, then move the entire shared list to the
+		 * dequeuing list.
+		 */
+		mutex_enter(&q->bq_lock);
+		while (q->bq_size == 0) {
+			cv_wait_sig(&q->bq_pop_cv, &q->bq_lock);
+		}
+		ASSERT0(q->bq_dequeuing_size);
+		ASSERT(list_is_empty(&q->bq_dequeuing_list));
+		list_move_tail(&q->bq_dequeuing_list, &q->bq_list);
+		q->bq_dequeuing_size = q->bq_size;
+		q->bq_size = 0;
+		cv_broadcast(&q->bq_add_cv);
+		mutex_exit(&q->bq_lock);
+		ret = list_remove_head(&q->bq_dequeuing_list);
 	}
-	ret = list_remove_head(&q->bq_list);
-	ASSERT3P(ret, !=, NULL);
-	item_size = obj2node(q, ret)->bqn_size;
-	q->bq_size -= item_size;
-	if (q->bq_size <= q->bq_maxsize - (q->bq_maxsize / q->bq_fill_fraction))
-		cv_signal(&q->bq_add_cv);
-	mutex_exit(&q->bq_lock);
+	q->bq_dequeuing_size -= obj2node(q, ret)->bqn_size;
 	return (ret);
 }
-
-/*
- * Returns true if the space used is 0.
- */
-boolean_t
-bqueue_empty(bqueue_t *q)
-{
-	return (q->bq_size == 0);
-}