Btrfs: attach delayed ref updates to delayed ref heads

Currently we have two rb-trees, one for delayed ref heads and one for all of the
delayed refs, including the delayed ref heads.  When we process the delayed refs
we have to hold onto the delayed ref lock for all of the selecting and merging
and such, which results in quite a bit of lock contention.  This was solved by
having a waitqueue and only one flusher at a time, however this hurts if we get
a lot of delayed refs queued up.

So instead just have an rb tree for the delayed ref heads, and then attach the
delayed ref updates to an rb tree that is per delayed ref head.  Then we only
need to take the delayed ref lock when adding new delayed refs and when
selecting a delayed ref head to process, all the rest of the time we deal with a
per delayed ref head lock which will be much less contentious.

The locking rules for this get a little more complicated since we have to lock
up to 3 things to properly process delayed refs, but I will address that problem
later.  For now this passes all of xfstests and my overnight stress tests.
Thanks,

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>

6 files changed, 267 insertions, 405 deletions

diff --git a/fs/btrfs/backref.c b/fs/btrfs/backref.c
index 835b6c9a26a8..34a8952de8dd 100644
--- a/fs/btrfs/backref.c
+++ b/fs/btrfs/backref.c
@@ -538,14 +538,13 @@ static int __add_delayed_refs(struct btrfs_delayed_ref_head *head, u64 seq,
 	if (extent_op && extent_op->update_key)
 		btrfs_disk_key_to_cpu(&op_key, &extent_op->key);
 
-	while ((n = rb_prev(n))) {
+	spin_lock(&head->lock);
+	n = rb_first(&head->ref_root);
+	while (n) {
 		struct btrfs_delayed_ref_node *node;
 		node = rb_entry(n, struct btrfs_delayed_ref_node,
 				rb_node);
-		if (node->bytenr != head->node.bytenr)
-			break;
-		WARN_ON(node->is_head);
-
+		n = rb_next(n);
 		if (node->seq > seq)
 			continue;
 
@@ -612,10 +611,10 @@ static int __add_delayed_refs(struct btrfs_delayed_ref_head *head, u64 seq,
 			WARN_ON(1);
 		}
 		if (ret)
-			return ret;
+			break;
 	}
-
-	return 0;
+	spin_unlock(&head->lock);
+	return ret;
 }
 
 /*
@@ -882,15 +881,15 @@ again:
 				btrfs_put_delayed_ref(&head->node);
 				goto again;
 			}
+			spin_unlock(&delayed_refs->lock);
 			ret = __add_delayed_refs(head, time_seq,
 						 &prefs_delayed);
 			mutex_unlock(&head->mutex);
-			if (ret) {
-				spin_unlock(&delayed_refs->lock);
+			if (ret)
 				goto out;
-			}
+		} else {
+			spin_unlock(&delayed_refs->lock);
 		}
-		spin_unlock(&delayed_refs->lock);
 	}
 
 	if (path->slots[0]) {
diff --git a/fs/btrfs/delayed-ref.c b/fs/btrfs/delayed-ref.c
index fab60c1ba065..f3bff89eecf0 100644
--- a/fs/btrfs/delayed-ref.c
+++ b/fs/btrfs/delayed-ref.c
@@ -269,39 +269,38 @@ int btrfs_delayed_ref_lock(struct btrfs_trans_handle *trans,
 
 static inline void drop_delayed_ref(struct btrfs_trans_handle *trans,
 				    struct btrfs_delayed_ref_root *delayed_refs,
+				    struct btrfs_delayed_ref_head *head,
 				    struct btrfs_delayed_ref_node *ref)
 {
-	rb_erase(&ref->rb_node, &delayed_refs->root);
 	if (btrfs_delayed_ref_is_head(ref)) {
-		struct btrfs_delayed_ref_head *head;
-
 		head = btrfs_delayed_node_to_head(ref);
 		rb_erase(&head->href_node, &delayed_refs->href_root);
+	} else {
+		assert_spin_locked(&head->lock);
+		rb_erase(&ref->rb_node, &head->ref_root);
 	}
 	ref->in_tree = 0;
 	btrfs_put_delayed_ref(ref);
-	delayed_refs->num_entries--;
+	atomic_dec(&delayed_refs->num_entries);
 	if (trans->delayed_ref_updates)
 		trans->delayed_ref_updates--;
 }
 
 static int merge_ref(struct btrfs_trans_handle *trans,
 		     struct btrfs_delayed_ref_root *delayed_refs,
+		     struct btrfs_delayed_ref_head *head,
 		     struct btrfs_delayed_ref_node *ref, u64 seq)
 {
 	struct rb_node *node;
-	int merged = 0;
 	int mod = 0;
 	int done = 0;
 
-	node = rb_prev(&ref->rb_node);
-	while (node) {
+	node = rb_next(&ref->rb_node);
+	while (!done && node) {
 		struct btrfs_delayed_ref_node *next;
 
 		next = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
-		node = rb_prev(node);
-		if (next->bytenr != ref->bytenr)
-			break;
+		node = rb_next(node);
 		if (seq && next->seq >= seq)
 			break;
 		if (comp_entry(ref, next, 0))
@@ -321,12 +320,11 @@ static int merge_ref(struct btrfs_trans_handle *trans,
 			mod = -next->ref_mod;
 		}
 
-		merged++;
-		drop_delayed_ref(trans, delayed_refs, next);
+		drop_delayed_ref(trans, delayed_refs, head, next);
 		ref->ref_mod += mod;
 		if (ref->ref_mod == 0) {
-			drop_delayed_ref(trans, delayed_refs, ref);
-			break;
+			drop_delayed_ref(trans, delayed_refs, head, ref);
+			done = 1;
 		} else {
 			/*
 			 * You can't have multiples of the same ref on a tree
@@ -335,13 +333,8 @@ static int merge_ref(struct btrfs_trans_handle *trans,
 			WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
 				ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
 		}
-
-		if (done)
-			break;
-		node = rb_prev(&ref->rb_node);
 	}
-
-	return merged;
+	return done;
 }
 
 void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
@@ -352,6 +345,7 @@ void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
 	struct rb_node *node;
 	u64 seq = 0;
 
+	assert_spin_locked(&head->lock);
 	/*
 	 * We don't have too much refs to merge in the case of delayed data
 	 * refs.
@@ -369,22 +363,19 @@ void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
 	}
 	spin_unlock(&fs_info->tree_mod_seq_lock);
 
-	node = rb_prev(&head->node.rb_node);
+	node = rb_first(&head->ref_root);
 	while (node) {
 		struct btrfs_delayed_ref_node *ref;
 
 		ref = rb_entry(node, struct btrfs_delayed_ref_node,
 			       rb_node);
-		if (ref->bytenr != head->node.bytenr)
-			break;
-
 		/* We can't merge refs that are outside of our seq count */
 		if (seq && ref->seq >= seq)
 			break;
-		if (merge_ref(trans, delayed_refs, ref, seq))
-			node = rb_prev(&head->node.rb_node);
+		if (merge_ref(trans, delayed_refs, head, ref, seq))
+			node = rb_first(&head->ref_root);
 		else
-			node = rb_prev(node);
+			node = rb_next(&ref->rb_node);
 	}
 }
 
@@ -412,64 +403,52 @@ int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info,
 	return ret;
 }
 
-int btrfs_find_ref_cluster(struct btrfs_trans_handle *trans,
-			   struct list_head *cluster, u64 start)
+struct btrfs_delayed_ref_head *
+btrfs_select_ref_head(struct btrfs_trans_handle *trans)
 {
-	int count = 0;
 	struct btrfs_delayed_ref_root *delayed_refs;
-	struct rb_node *node;
-	struct btrfs_delayed_ref_head *head = NULL;
+	struct btrfs_delayed_ref_head *head;
+	u64 start;
+	bool loop = false;
 
 	delayed_refs = &trans->transaction->delayed_refs;
-	node = rb_first(&delayed_refs->href_root);
 
-	if (start) {
-		find_ref_head(&delayed_refs->href_root, start + 1, &head, 1);
-		if (head)
-			node = &head->href_node;
-	}
 again:
-	while (node && count < 32) {
-		head = rb_entry(node, struct btrfs_delayed_ref_head, href_node);
-		if (list_empty(&head->cluster)) {
-			list_add_tail(&head->cluster, cluster);
-			delayed_refs->run_delayed_start =
-				head->node.bytenr;
-			count++;
-
-			WARN_ON(delayed_refs->num_heads_ready == 0);
-			delayed_refs->num_heads_ready--;
-		} else if (count) {
-			/* the goal of the clustering is to find extents
-			 * that are likely to end up in the same extent
-			 * leaf on disk.  So, we don't want them spread
-			 * all over the tree.  Stop now if we've hit
-			 * a head that was already in use
-			 */
-			break;
-		}
-		node = rb_next(node);
-	}
-	if (count) {
-		return 0;
-	} else if (start) {
-		/*
-		 * we've gone to the end of the rbtree without finding any
-		 * clusters.  start from the beginning and try again
-		 */
+	start = delayed_refs->run_delayed_start;
+	head = find_ref_head(&delayed_refs->href_root, start, NULL, 1);
+	if (!head && !loop) {
+		delayed_refs->run_delayed_start = 0;
 		start = 0;
-		node = rb_first(&delayed_refs->href_root);
-		goto again;
+		loop = true;
+		head = find_ref_head(&delayed_refs->href_root, start, NULL, 1);
+		if (!head)
+			return NULL;
+	} else if (!head && loop) {
+		return NULL;
 	}
-	return 1;
-}
 
-void btrfs_release_ref_cluster(struct list_head *cluster)
-{
-	struct list_head *pos, *q;
+	while (head->processing) {
+		struct rb_node *node;
+
+		node = rb_next(&head->href_node);
+		if (!node) {
+			if (loop)
+				return NULL;
+			delayed_refs->run_delayed_start = 0;
+			start = 0;
+			loop = true;
+			goto again;
+		}
+		head = rb_entry(node, struct btrfs_delayed_ref_head,
+				href_node);
+	}
 
-	list_for_each_safe(pos, q, cluster)
-		list_del_init(pos);
+	head->processing = 1;
+	WARN_ON(delayed_refs->num_heads_ready == 0);
+	delayed_refs->num_heads_ready--;
+	delayed_refs->run_delayed_start = head->node.bytenr +
+		head->node.num_bytes;
+	return head;
 }
 
 /*
@@ -483,6 +462,7 @@ void btrfs_release_ref_cluster(struct list_head *cluster)
 static noinline void
 update_existing_ref(struct btrfs_trans_handle *trans,
 		    struct btrfs_delayed_ref_root *delayed_refs,
+		    struct btrfs_delayed_ref_head *head,
 		    struct btrfs_delayed_ref_node *existing,
 		    struct btrfs_delayed_ref_node *update)
 {
@@ -495,7 +475,7 @@ update_existing_ref(struct btrfs_trans_handle *trans,
 		 */
 		existing->ref_mod--;
 		if (existing->ref_mod == 0)
-			drop_delayed_ref(trans, delayed_refs, existing);
+			drop_delayed_ref(trans, delayed_refs, head, existing);
 		else
 			WARN_ON(existing->type == BTRFS_TREE_BLOCK_REF_KEY ||
 				existing->type == BTRFS_SHARED_BLOCK_REF_KEY);
@@ -565,9 +545,13 @@ update_existing_head_ref(struct btrfs_delayed_ref_node *existing,
 		}
 	}
 	/*
-	 * update the reference mod on the head to reflect this new operation
+	 * update the reference mod on the head to reflect this new operation,
+	 * only need the lock for this case cause we could be processing it
+	 * currently, for refs we just added we know we're a-ok.
 	 */
+	spin_lock(&existing_ref->lock);
 	existing->ref_mod += update->ref_mod;
+	spin_unlock(&existing_ref->lock);
 }
 
 /*
@@ -575,13 +559,13 @@ update_existing_head_ref(struct btrfs_delayed_ref_node *existing,
  * this does all the dirty work in terms of maintaining the correct
  * overall modification count.
  */
-static noinline void add_delayed_ref_head(struct btrfs_fs_info *fs_info,
-					struct btrfs_trans_handle *trans,
-					struct btrfs_delayed_ref_node *ref,
-					u64 bytenr, u64 num_bytes,
-					int action, int is_data)
+static noinline struct btrfs_delayed_ref_head *
+add_delayed_ref_head(struct btrfs_fs_info *fs_info,
+		     struct btrfs_trans_handle *trans,
+		     struct btrfs_delayed_ref_node *ref, u64 bytenr,
+		     u64 num_bytes, int action, int is_data)
 {
-	struct btrfs_delayed_ref_node *existing;
+	struct btrfs_delayed_ref_head *existing;
 	struct btrfs_delayed_ref_head *head_ref = NULL;
 	struct btrfs_delayed_ref_root *delayed_refs;
 	int count_mod = 1;
@@ -628,39 +612,43 @@ static noinline void add_delayed_ref_head(struct btrfs_fs_info *fs_info,
 	head_ref = btrfs_delayed_node_to_head(ref);
 	head_ref->must_insert_reserved = must_insert_reserved;
 	head_ref->is_data = is_data;
+	head_ref->ref_root = RB_ROOT;
+	head_ref->processing = 0;
 
-	INIT_LIST_HEAD(&head_ref->cluster);
+	spin_lock_init(&head_ref->lock);
 	mutex_init(&head_ref->mutex);
 
 	trace_add_delayed_ref_head(ref, head_ref, action);
 
-	existing = tree_insert(&delayed_refs->root, &ref->rb_node);
-
+	existing = htree_insert(&delayed_refs->href_root,
+				&head_ref->href_node);
 	if (existing) {
-		update_existing_head_ref(existing, ref);
+		update_existing_head_ref(&existing->node, ref);
 		/*
 		 * we've updated the existing ref, free the newly
 		 * allocated ref
 		 */
 		kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
+		head_ref = existing;
 	} else {
-		htree_insert(&delayed_refs->href_root, &head_ref->href_node);
 		delayed_refs->num_heads++;
 		delayed_refs->num_heads_ready++;
-		delayed_refs->num_entries++;
+		atomic_inc(&delayed_refs->num_entries);
 		trans->delayed_ref_updates++;
 	}
+	return head_ref;
 }
 
 /*
  * helper to insert a delayed tree ref into the rbtree.
  */
-static noinline void add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
-					 struct btrfs_trans_handle *trans,
-					 struct btrfs_delayed_ref_node *ref,
-					 u64 bytenr, u64 num_bytes, u64 parent,
-					 u64 ref_root, int level, int action,
-					 int for_cow)
+static noinline void
+add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
+		     struct btrfs_trans_handle *trans,
+		     struct btrfs_delayed_ref_head *head_ref,
+		     struct btrfs_delayed_ref_node *ref, u64 bytenr,
+		     u64 num_bytes, u64 parent, u64 ref_root, int level,
+		     int action, int for_cow)
 {
 	struct btrfs_delayed_ref_node *existing;
 	struct btrfs_delayed_tree_ref *full_ref;
@@ -696,30 +684,33 @@ static noinline void add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
 
 	trace_add_delayed_tree_ref(ref, full_ref, action);
 
-	existing = tree_insert(&delayed_refs->root, &ref->rb_node);
-
+	spin_lock(&head_ref->lock);
+	existing = tree_insert(&head_ref->ref_root, &ref->rb_node);
 	if (existing) {
-		update_existing_ref(trans, delayed_refs, existing, ref);
+		update_existing_ref(trans, delayed_refs, head_ref, existing,
+				    ref);
 		/*
 		 * we've updated the existing ref, free the newly
 		 * allocated ref
 		 */
 		kmem_cache_free(btrfs_delayed_tree_ref_cachep, full_ref);
 	} else {
-		delayed_refs->num_entries++;
+		atomic_inc(&delayed_refs->num_entries);
 		trans->delayed_ref_updates++;
 	}
+	spin_unlock(&head_ref->lock);
 }
 
 /*
  * helper to insert a delayed data ref into the rbtree.
  */
-static noinline void add_delayed_data_ref(struct btrfs_fs_info *fs_info,
-					 struct btrfs_trans_handle *trans,
-					 struct btrfs_delayed_ref_node *ref,
-					 u64 bytenr, u64 num_bytes, u64 parent,
-					 u64 ref_root, u64 owner, u64 offset,
-					 int action, int for_cow)
+static noinline void
+add_delayed_data_ref(struct btrfs_fs_info *fs_info,
+		     struct btrfs_trans_handle *trans,
+		     struct btrfs_delayed_ref_head *head_ref,
+		     struct btrfs_delayed_ref_node *ref, u64 bytenr,
+		     u64 num_bytes, u64 parent, u64 ref_root, u64 owner,
+		     u64 offset, int action, int for_cow)
 {
 	struct btrfs_delayed_ref_node *existing;
 	struct btrfs_delayed_data_ref *full_ref;
@@ -757,19 +748,21 @@ static noinline void add_delayed_data_ref(struct btrfs_fs_info *fs_info,
 
 	trace_add_delayed_data_ref(ref, full_ref, action);
 
-	existing = tree_insert(&delayed_refs->root, &ref->rb_node);
-
+	spin_lock(&head_ref->lock);
+	existing = tree_insert(&head_ref->ref_root, &ref->rb_node);
 	if (existing) {
-		update_existing_ref(trans, delayed_refs, existing, ref);
+		update_existing_ref(trans, delayed_refs, head_ref, existing,
+				    ref);
 		/*
 		 * we've updated the existing ref, free the newly
 		 * allocated ref
 		 */
 		kmem_cache_free(btrfs_delayed_data_ref_cachep, full_ref);
 	} else {
-		delayed_refs->num_entries++;
+		atomic_inc(&delayed_refs->num_entries);
 		trans->delayed_ref_updates++;
 	}
+	spin_unlock(&head_ref->lock);
 }
 
 /*
@@ -808,10 +801,10 @@ int btrfs_add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
 	 * insert both the head node and the new ref without dropping
 	 * the spin lock
 	 */
-	add_delayed_ref_head(fs_info, trans, &head_ref->node, bytenr,
-				   num_bytes, action, 0);
+	head_ref = add_delayed_ref_head(fs_info, trans, &head_ref->node,
+					bytenr, num_bytes, action, 0);
 
-	add_delayed_tree_ref(fs_info, trans, &ref->node, bytenr,
+	add_delayed_tree_ref(fs_info, trans, head_ref, &ref->node, bytenr,
 				   num_bytes, parent, ref_root, level, action,
 				   for_cow);
 	spin_unlock(&delayed_refs->lock);
@@ -856,10 +849,10 @@ int btrfs_add_delayed_data_ref(struct btrfs_fs_info *fs_info,
 	 * insert both the head node and the new ref without dropping
 	 * the spin lock
 	 */
-	add_delayed_ref_head(fs_info, trans, &head_ref->node, bytenr,
-				   num_bytes, action, 1);
+	head_ref = add_delayed_ref_head(fs_info, trans, &head_ref->node,
+					bytenr, num_bytes, action, 1);
 
-	add_delayed_data_ref(fs_info, trans, &ref->node, bytenr,
+	add_delayed_data_ref(fs_info, trans, head_ref, &ref->node, bytenr,
 				   num_bytes, parent, ref_root, owner, offset,
 				   action, for_cow);
 	spin_unlock(&delayed_refs->lock);
diff --git a/fs/btrfs/delayed-ref.h b/fs/btrfs/delayed-ref.h
index a54c9d47918f..4ba9b93022ff 100644
--- a/fs/btrfs/delayed-ref.h
+++ b/fs/btrfs/delayed-ref.h
@@ -81,7 +81,8 @@ struct btrfs_delayed_ref_head {
 	 */
 	struct mutex mutex;
 
-	struct list_head cluster;
+	spinlock_t lock;
+	struct rb_root ref_root;
 
 	struct rb_node href_node;
 
@@ -100,6 +101,7 @@ struct btrfs_delayed_ref_head {
 	 */
 	unsigned int must_insert_reserved:1;
 	unsigned int is_data:1;
+	unsigned int processing:1;
 };
 
 struct btrfs_delayed_tree_ref {
@@ -118,8 +120,6 @@ struct btrfs_delayed_data_ref {
 };
 
 struct btrfs_delayed_ref_root {
-	struct rb_root root;
-
 	/* head ref rbtree */
 	struct rb_root href_root;
 
@@ -129,7 +129,7 @@ struct btrfs_delayed_ref_root {
 	/* how many delayed ref updates we've queued, used by the
 	 * throttling code
 	 */
-	unsigned long num_entries;
+	atomic_t num_entries;
 
 	/* total number of head nodes in tree */
 	unsigned long num_heads;
@@ -138,15 +138,6 @@ struct btrfs_delayed_ref_root {
 	unsigned long num_heads_ready;
 
 	/*
-	 * bumped when someone is making progress on the delayed
-	 * refs, so that other procs know they are just adding to
-	 * contention intead of helping
-	 */
-	atomic_t procs_running_refs;
-	atomic_t ref_seq;
-	wait_queue_head_t wait;
-
-	/*
 	 * set when the tree is flushing before a transaction commit,
 	 * used by the throttling code to decide if new updates need
 	 * to be run right away
@@ -231,9 +222,9 @@ static inline void btrfs_delayed_ref_unlock(struct btrfs_delayed_ref_head *head)
 	mutex_unlock(&head->mutex);
 }
 
-int btrfs_find_ref_cluster(struct btrfs_trans_handle *trans,
-			   struct list_head *cluster, u64 search_start);
-void btrfs_release_ref_cluster(struct list_head *cluster);
+
+struct btrfs_delayed_ref_head *
+btrfs_select_ref_head(struct btrfs_trans_handle *trans);
 
 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info,
 			    struct btrfs_delayed_ref_root *delayed_refs,
diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
index 9850a51a3f19..ed23127a4b02 100644
--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c
@@ -3801,58 +3801,55 @@ static int btrfs_destroy_delayed_refs(struct btrfs_transaction *trans,
 	delayed_refs = &trans->delayed_refs;
 
 	spin_lock(&delayed_refs->lock);
-	if (delayed_refs->num_entries == 0) {
+	if (atomic_read(&delayed_refs->num_entries) == 0) {
 		spin_unlock(&delayed_refs->lock);
 		btrfs_info(root->fs_info, "delayed_refs has NO entry");
 		return ret;
 	}
 
-	while ((node = rb_first(&delayed_refs->root)) != NULL) {
-		struct btrfs_delayed_ref_head *head = NULL;
+	while ((node = rb_first(&delayed_refs->href_root)) != NULL) {
+		struct btrfs_delayed_ref_head *head;
 		bool pin_bytes = false;
 
-		ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
-		atomic_set(&ref->refs, 1);
-		if (btrfs_delayed_ref_is_head(ref)) {
+		head = rb_entry(node, struct btrfs_delayed_ref_head,
+				href_node);
+		if (!mutex_trylock(&head->mutex)) {
+			atomic_inc(&head->node.refs);
+			spin_unlock(&delayed_refs->lock);
 
-			head = btrfs_delayed_node_to_head(ref);
-			if (!mutex_trylock(&head->mutex)) {
-				atomic_inc(&ref->refs);
-				spin_unlock(&delayed_refs->lock);
-
-				/* Need to wait for the delayed ref to run */
-				mutex_lock(&head->mutex);
-				mutex_unlock(&head->mutex);
-				btrfs_put_delayed_ref(ref);
-
-				spin_lock(&delayed_refs->lock);
-				continue;
-			}
-
-			if (head->must_insert_reserved)
-				pin_bytes = true;
-			btrfs_free_delayed_extent_op(head->extent_op);
-			delayed_refs->num_heads--;
-			if (list_empty(&head->cluster))
-				delayed_refs->num_heads_ready--;
-			list_del_init(&head->cluster);
-		}
-
-		ref->in_tree = 0;
-		rb_erase(&ref->rb_node, &delayed_refs->root);
-		if (head)
-			rb_erase(&head->href_node, &delayed_refs->href_root);
-
-		delayed_refs->num_entries--;
-		spin_unlock(&delayed_refs->lock);
-		if (head) {
-			if (pin_bytes)
-				btrfs_pin_extent(root, ref->bytenr,
-						 ref->num_bytes, 1);
+			mutex_lock(&head->mutex);
 			mutex_unlock(&head->mutex);
+			btrfs_put_delayed_ref(&head->node);
+			spin_lock(&delayed_refs->lock);
+			continue;
+		}
+		spin_lock(&head->lock);
+		while ((node = rb_first(&head->ref_root)) != NULL) {
+			ref = rb_entry(node, struct btrfs_delayed_ref_node,
+				       rb_node);
+			ref->in_tree = 0;
+			rb_erase(&ref->rb_node, &head->ref_root);
+			atomic_dec(&delayed_refs->num_entries);
+			btrfs_put_delayed_ref(ref);
+			cond_resched_lock(&head->lock);
 		}
-		btrfs_put_delayed_ref(ref);
+		if (head->must_insert_reserved)
+			pin_bytes = true;
+		btrfs_free_delayed_extent_op(head->extent_op);
+		delayed_refs->num_heads--;
+		if (head->processing == 0)
+			delayed_refs->num_heads_ready--;
+		atomic_dec(&delayed_refs->num_entries);
+		head->node.in_tree = 0;
+		rb_erase(&head->href_node, &delayed_refs->href_root);
+		spin_unlock(&head->lock);
+		spin_unlock(&delayed_refs->lock);
+		mutex_unlock(&head->mutex);
 
+		if (pin_bytes)
+			btrfs_pin_extent(root, head->node.bytenr,
+					 head->node.num_bytes, 1);
+		btrfs_put_delayed_ref(&head->node);
 		cond_resched();
 		spin_lock(&delayed_refs->lock);
 	}
diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c
index 77acc0873895..c77156c77de7 100644
--- a/fs/btrfs/extent-tree.c
+++ b/fs/btrfs/extent-tree.c
@@ -857,12 +857,14 @@ again:
 			btrfs_put_delayed_ref(&head->node);
 			goto search_again;
 		}
+		spin_lock(&head->lock);
 		if (head->extent_op && head->extent_op->update_flags)
 			extent_flags |= head->extent_op->flags_to_set;
 		else
 			BUG_ON(num_refs == 0);
 
 		num_refs += head->node.ref_mod;
+		spin_unlock(&head->lock);
 		mutex_unlock(&head->mutex);
 	}
 	spin_unlock(&delayed_refs->lock);
@@ -2287,40 +2289,33 @@ static noinline struct btrfs_delayed_ref_node *
 select_delayed_ref(struct btrfs_delayed_ref_head *head)
 {
 	struct rb_node *node;
-	struct btrfs_delayed_ref_node *ref;
-	int action = BTRFS_ADD_DELAYED_REF;
-again:
+	struct btrfs_delayed_ref_node *ref, *last = NULL;;
+
 	/*
 	 * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
 	 * this prevents ref count from going down to zero when
 	 * there still are pending delayed ref.
 	 */
-	node = rb_prev(&head->node.rb_node);
-	while (1) {
-		if (!node)
-			break;
+	node = rb_first(&head->ref_root);
+	while (node) {
 		ref = rb_entry(node, struct btrfs_delayed_ref_node,
 				rb_node);
-		if (ref->bytenr != head->node.bytenr)
-			break;
-		if (ref->action == action)
+		if (ref->action == BTRFS_ADD_DELAYED_REF)
 			return ref;
-		node = rb_prev(node);
-	}
-	if (action == BTRFS_ADD_DELAYED_REF) {
-		action = BTRFS_DROP_DELAYED_REF;
-		goto again;
+		else if (last == NULL)
+			last = ref;
+		node = rb_next(node);
 	}
-	return NULL;
+	return last;
 }
 
 /*
  * Returns 0 on success or if called with an already aborted transaction.
  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
  */
-static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
-				       struct btrfs_root *root,
-				       struct list_head *cluster)
+static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
+					     struct btrfs_root *root,
+					     unsigned long nr)
 {
 	struct btrfs_delayed_ref_root *delayed_refs;
 	struct btrfs_delayed_ref_node *ref;
@@ -2328,23 +2323,26 @@ static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
 	struct btrfs_delayed_extent_op *extent_op;
 	struct btrfs_fs_info *fs_info = root->fs_info;
 	int ret;
-	int count = 0;
+	unsigned long count = 0;
 	int must_insert_reserved = 0;
 
 	delayed_refs = &trans->transaction->delayed_refs;
 	while (1) {
 		if (!locked_ref) {
-			/* pick a new head ref from the cluster list */
-			if (list_empty(cluster))
+			if (count >= nr)
 				break;
 
-			locked_ref = list_entry(cluster->next,
-				     struct btrfs_delayed_ref_head, cluster);
+			spin_lock(&delayed_refs->lock);
+			locked_ref = btrfs_select_ref_head(trans);
+			if (!locked_ref) {
+				spin_unlock(&delayed_refs->lock);
+				break;
+			}
 
 			/* grab the lock that says we are going to process
 			 * all the refs for this head */
 			ret = btrfs_delayed_ref_lock(trans, locked_ref);
-
+			spin_unlock(&delayed_refs->lock);
 			/*
 			 * we may have dropped the spin lock to get the head
 			 * mutex lock, and that might have given someone else
@@ -2365,6 +2363,7 @@ static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
 		 * finish.  If we merged anything we need to re-loop so we can
 		 * get a good ref.
 		 */
+		spin_lock(&locked_ref->lock);
 		btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
 					 locked_ref);
 
@@ -2376,17 +2375,14 @@ static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
 
 		if (ref && ref->seq &&
 		    btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
-			/*
-			 * there are still refs with lower seq numbers in the
-			 * process of being added. Don't run this ref yet.
-			 */
-			list_del_init(&locked_ref->cluster);
+			spin_unlock(&locked_ref->lock);
 			btrfs_delayed_ref_unlock(locked_ref);
-			locked_ref = NULL;
+			spin_lock(&delayed_refs->lock);
+			locked_ref->processing = 0;
 			delayed_refs->num_heads_ready++;
 			spin_unlock(&delayed_refs->lock);
+			locked_ref = NULL;
 			cond_resched();
-			spin_lock(&delayed_refs->lock);
 			continue;
 		}
 
@@ -2401,6 +2397,8 @@ static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
 		locked_ref->extent_op = NULL;
 
 		if (!ref) {
+
+
 			/* All delayed refs have been processed, Go ahead
 			 * and send the head node to run_one_delayed_ref,
 			 * so that any accounting fixes can happen
@@ -2413,8 +2411,7 @@ static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
 			}
 
 			if (extent_op) {
-				spin_unlock(&delayed_refs->lock);
-
+				spin_unlock(&locked_ref->lock);
 				ret = run_delayed_extent_op(trans, root,
 							    ref, extent_op);
 				btrfs_free_delayed_extent_op(extent_op);
@@ -2428,23 +2425,38 @@ static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
 					 */
 					if (must_insert_reserved)
 						locked_ref->must_insert_reserved = 1;
+					locked_ref->processing = 0;
 					btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
-					spin_lock(&delayed_refs->lock);
 					btrfs_delayed_ref_unlock(locked_ref);
 					return ret;
 				}
-
-				goto next;
+				continue;
 			}
-		}
 
-		ref->in_tree = 0;
-		rb_erase(&ref->rb_node, &delayed_refs->root);
-		if (btrfs_delayed_ref_is_head(ref)) {
+			/*
+			 * Need to drop our head ref lock and re-aqcuire the
+			 * delayed ref lock and then re-check to make sure
+			 * nobody got added.
+			 */
+			spin_unlock(&locked_ref->lock);
+			spin_lock(&delayed_refs->lock);
+			spin_lock(&locked_ref->lock);
+			if (rb_first(&locked_ref->ref_root)) {
+				spin_unlock(&locked_ref->lock);
+				spin_unlock(&delayed_refs->lock);
+				continue;
+			}
+			ref->in_tree = 0;
+			delayed_refs->num_heads--;
 			rb_erase(&locked_ref->href_node,
 				 &delayed_refs->href_root);
+			spin_unlock(&delayed_refs->lock);
+		} else {
+			ref->in_tree = 0;
+			rb_erase(&ref->rb_node, &locked_ref->ref_root);
 		}
-		delayed_refs->num_entries--;
+		atomic_dec(&delayed_refs->num_entries);
+
 		if (!btrfs_delayed_ref_is_head(ref)) {
 			/*
 			 * when we play the delayed ref, also correct the
@@ -2461,20 +2473,18 @@ static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
 			default:
 				WARN_ON(1);
 			}
-		} else {
-			list_del_init(&locked_ref->cluster);
 		}
-		spin_unlock(&delayed_refs->lock);
+		spin_unlock(&locked_ref->lock);
 
 		ret = run_one_delayed_ref(trans, root, ref, extent_op,
 					  must_insert_reserved);
 
 		btrfs_free_delayed_extent_op(extent_op);
 		if (ret) {
+			locked_ref->processing = 0;
 			btrfs_delayed_ref_unlock(locked_ref);
 			btrfs_put_delayed_ref(ref);
 			btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
-			spin_lock(&delayed_refs->lock);
 			return ret;
 		}
 
@@ -2490,11 +2500,9 @@ static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
 		}
 		btrfs_put_delayed_ref(ref);
 		count++;
-next:
 		cond_resched();
-		spin_lock(&delayed_refs->lock);
 	}
-	return count;
+	return 0;
 }
 
 #ifdef SCRAMBLE_DELAYED_REFS
@@ -2576,16 +2584,6 @@ int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
 	return ret;
 }
 
-static int refs_newer(struct btrfs_delayed_ref_root *delayed_refs, int seq,
-		      int count)
-{
-	int val = atomic_read(&delayed_refs->ref_seq);
-
-	if (val < seq || val >= seq + count)
-		return 1;
-	return 0;
-}
-
 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
 {
 	u64 num_bytes;
@@ -2647,12 +2645,9 @@ int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
 	struct rb_node *node;
 	struct btrfs_delayed_ref_root *delayed_refs;
 	struct btrfs_delayed_ref_head *head;
-	struct list_head cluster;
 	int ret;
-	u64 delayed_start;
 	int run_all = count == (unsigned long)-1;
 	int run_most = 0;
-	int loops;
 
 	/* We'll clean this up in btrfs_cleanup_transaction */
 	if (trans->aborted)
@@ -2664,121 +2659,31 @@ int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
 	btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
 
 	delayed_refs = &trans->transaction->delayed_refs;
-	INIT_LIST_HEAD(&cluster);
 	if (count == 0) {
-		count = delayed_refs->num_entries * 2;
+		count = atomic_read(&delayed_refs->num_entries) * 2;
 		run_most = 1;
 	}
 
-	if (!run_all && !run_most) {
-		int old;
-		int seq = atomic_read(&delayed_refs->ref_seq);
-
-progress:
-		old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
-		if (old) {
-			DEFINE_WAIT(__wait);
-			if (delayed_refs->flushing ||
-			    !btrfs_should_throttle_delayed_refs(trans, root))
-				return 0;
-
-			prepare_to_wait(&delayed_refs->wait, &__wait,
-					TASK_UNINTERRUPTIBLE);
-
-			old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
-			if (old) {
-				schedule();
-				finish_wait(&delayed_refs->wait, &__wait);
-
-				if (!refs_newer(delayed_refs, seq, 256))
-					goto progress;
-				else
-					return 0;
-			} else {
-				finish_wait(&delayed_refs->wait, &__wait);
-				goto again;
-			}
-		}
-
-	} else {
-		atomic_inc(&delayed_refs->procs_running_refs);
-	}
-
 again:
-	loops = 0;
-	spin_lock(&delayed_refs->lock);
-
 #ifdef SCRAMBLE_DELAYED_REFS
 	delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
 #endif
-
-	while (1) {
-		if (!(run_all || run_most) &&
-		    !btrfs_should_throttle_delayed_refs(trans, root))
-			break;
-
-		/*
-		 * go find something we can process in the rbtree.  We start at
-		 * the beginning of the tree, and then build a cluster
-		 * of refs to process starting at the first one we are able to
-		 * lock
-		 */
-		delayed_start = delayed_refs->run_delayed_start;
-		ret = btrfs_find_ref_cluster(trans, &cluster,
-					     delayed_refs->run_delayed_start);
-		if (ret)
-			break;
-
-		ret = run_clustered_refs(trans, root, &cluster);
-		if (ret < 0) {
-			btrfs_release_ref_cluster(&cluster);
-			spin_unlock(&delayed_refs->lock);
-			btrfs_abort_transaction(trans, root, ret);
-			atomic_dec(&delayed_refs->procs_running_refs);
-			wake_up(&delayed_refs->wait);
-			return ret;
-		}
-
-		atomic_add(ret, &delayed_refs->ref_seq);
-
-		count -= min_t(unsigned long, ret, count);
-
-		if (count == 0)
-			break;
-
-		if (delayed_start >= delayed_refs->run_delayed_start) {
-			if (loops == 0) {
-				/*
-				 * btrfs_find_ref_cluster looped. let's do one
-				 * more cycle. if we don't run any delayed ref
-				 * during that cycle (because we can't because
-				 * all of them are blocked), bail out.
-				 */
-				loops = 1;
-			} else {
-				/*
-				 * no runnable refs left, stop trying
-				 */
-				BUG_ON(run_all);
-				break;
-			}
-		}
-		if (ret) {
-			/* refs were run, let's reset staleness detection */
-			loops = 0;
-		}
+	ret = __btrfs_run_delayed_refs(trans, root, count);
+	if (ret < 0) {
+		btrfs_abort_transaction(trans, root, ret);
+		return ret;
 	}
 
 	if (run_all) {
-		if (!list_empty(&trans->new_bgs)) {
-			spin_unlock(&delayed_refs->lock);
+		if (!list_empty(&trans->new_bgs))
 			btrfs_create_pending_block_groups(trans, root);
-			spin_lock(&delayed_refs->lock);
-		}
 
+		spin_lock(&delayed_refs->lock);
 		node = rb_first(&delayed_refs->href_root);
-		if (!node)
+		if (!node) {
+			spin_unlock(&delayed_refs->lock);
 			goto out;
+		}
 		count = (unsigned long)-1;
 
 		while (node) {
@@ -2807,16 +2712,10 @@ again:
 			node = rb_next(node);
 		}
 		spin_unlock(&delayed_refs->lock);
-		schedule_timeout(1);
+		cond_resched();
 		goto again;
 	}
 out:
-	atomic_dec(&delayed_refs->procs_running_refs);
-	smp_mb();
-	if (waitqueue_active(&delayed_refs->wait))
-		wake_up(&delayed_refs->wait);
-
-	spin_unlock(&delayed_refs->lock);
 	assert_qgroups_uptodate(trans);
 	return 0;
 }
@@ -2858,12 +2757,13 @@ static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
 	struct rb_node *node;
 	int ret = 0;
 
-	ret = -ENOENT;
 	delayed_refs = &trans->transaction->delayed_refs;
 	spin_lock(&delayed_refs->lock);
 	head = btrfs_find_delayed_ref_head(trans, bytenr);
-	if (!head)
-		goto out;
+	if (!head) {
+		spin_unlock(&delayed_refs->lock);
+		return 0;
+	}
 
 	if (!mutex_trylock(&head->mutex)) {
 		atomic_inc(&head->node.refs);
@@ -2880,40 +2780,35 @@ static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
 		btrfs_put_delayed_ref(&head->node);
 		return -EAGAIN;
 	}
+	spin_unlock(&delayed_refs->lock);
 
-	node = rb_prev(&head->node.rb_node);
-	if (!node)
-		goto out_unlock;
-
-	ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
-
-	if (ref->bytenr != bytenr)
-		goto out_unlock;
+	spin_lock(&head->lock);
+	node = rb_first(&head->ref_root);
+	while (node) {
+		ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
+		node = rb_next(node);
 
-	ret = 1;
-	if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
-		goto out_unlock;
+		/* If it's a shared ref we know a cross reference exists */
+		if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
+			ret = 1;
+			break;
+		}
 
-	data_ref = btrfs_delayed_node_to_data_ref(ref);
+		data_ref = btrfs_delayed_node_to_data_ref(ref);
 
-	node = rb_prev(node);
-	if (node) {
-		int seq = ref->seq;
-
-		ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
-		if (ref->bytenr == bytenr && ref->seq == seq)
-			goto out_unlock;
+		/*
+		 * If our ref doesn't match the one we're currently looking at
+		 * then we have a cross reference.
+		 */
+		if (data_ref->root != root->root_key.objectid ||
+		    data_ref->objectid != objectid ||
+		    data_ref->offset != offset) {
+			ret = 1;
+			break;
+		}
 	}
-
-	if (data_ref->root != root->root_key.objectid ||
-	    data_ref->objectid != objectid || data_ref->offset != offset)
-		goto out_unlock;
-
-	ret = 0;
-out_unlock:
+	spin_unlock(&head->lock);
 	mutex_unlock(&head->mutex);
-out:
-	spin_unlock(&delayed_refs->lock);
 	return ret;
 }
 
@@ -5953,8 +5848,6 @@ static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
 {
 	struct btrfs_delayed_ref_head *head;
 	struct btrfs_delayed_ref_root *delayed_refs;
-	struct btrfs_delayed_ref_node *ref;
-	struct rb_node *node;
 	int ret = 0;
 
 	delayed_refs = &trans->transaction->delayed_refs;
@@ -5963,14 +5856,8 @@ static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
 	if (!head)
 		goto out;
 
-	node = rb_prev(&head->node.rb_node);
-	if (!node)
-		goto out;
-
-	ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
-
-	/* there are still entries for this ref, we can't drop it */
-	if (ref->bytenr == bytenr)
+	spin_lock(&head->lock);
+	if (rb_first(&head->ref_root))
 		goto out;
 
 	if (head->extent_op) {
@@ -5992,20 +5879,19 @@ static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
 	 * ahead and process it.
 	 */
 	head->node.in_tree = 0;
-	rb_erase(&head->node.rb_node, &delayed_refs->root);
 	rb_erase(&head->href_node, &delayed_refs->href_root);
 
-	delayed_refs->num_entries--;
+	atomic_dec(&delayed_refs->num_entries);
 
 	/*
 	 * we don't take a ref on the node because we're removing it from the
 	 * tree, so we just steal the ref the tree was holding.
 	 */
 	delayed_refs->num_heads--;
-	if (list_empty(&head->cluster))
+	if (head->processing == 0)
 		delayed_refs->num_heads_ready--;
-
-	list_del_init(&head->cluster);
+	head->processing = 0;
+	spin_unlock(&head->lock);
 	spin_unlock(&delayed_refs->lock);
 
 	BUG_ON(head->extent_op);
@@ -6016,6 +5902,7 @@ static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
 	btrfs_put_delayed_ref(&head->node);
 	return ret;
 out:
+	spin_unlock(&head->lock);
 	spin_unlock(&delayed_refs->lock);
 	return 0;
 }
diff --git a/fs/btrfs/transaction.c b/fs/btrfs/transaction.c
index b16352ce0f73..fd1446496fe8 100644
--- a/fs/btrfs/transaction.c
+++ b/fs/btrfs/transaction.c
@@ -62,7 +62,6 @@ void btrfs_put_transaction(struct btrfs_transaction *transaction)
 	WARN_ON(atomic_read(&transaction->use_count) == 0);
 	if (atomic_dec_and_test(&transaction->use_count)) {
 		BUG_ON(!list_empty(&transaction->list));
-		WARN_ON(!RB_EMPTY_ROOT(&transaction->delayed_refs.root));
 		WARN_ON(!RB_EMPTY_ROOT(&transaction->delayed_refs.href_root));
 		while (!list_empty(&transaction->pending_chunks)) {
 			struct extent_map *em;
@@ -184,9 +183,8 @@ loop:
 	atomic_set(&cur_trans->use_count, 2);
 	cur_trans->start_time = get_seconds();
 
-	cur_trans->delayed_refs.root = RB_ROOT;
 	cur_trans->delayed_refs.href_root = RB_ROOT;
-	cur_trans->delayed_refs.num_entries = 0;
+	atomic_set(&cur_trans->delayed_refs.num_entries, 0);
 	cur_trans->delayed_refs.num_heads_ready = 0;
 	cur_trans->delayed_refs.num_heads = 0;
 	cur_trans->delayed_refs.flushing = 0;
@@ -206,9 +204,6 @@ loop:
 	atomic64_set(&fs_info->tree_mod_seq, 0);
 
 	spin_lock_init(&cur_trans->delayed_refs.lock);
-	atomic_set(&cur_trans->delayed_refs.procs_running_refs, 0);
-	atomic_set(&cur_trans->delayed_refs.ref_seq, 0);
-	init_waitqueue_head(&cur_trans->delayed_refs.wait);
 
 	INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 	INIT_LIST_HEAD(&cur_trans->ordered_operations);