1 files changed, 37 insertions, 6 deletions

diff --git a/fs/btrfs/send.c b/fs/btrfs/send.c
index 8f323859156b..55741adf9071 100644
--- a/fs/btrfs/send.c
+++ b/fs/btrfs/send.c
@@ -6650,6 +6650,7 @@ static int full_send_tree(struct send_ctx *sctx)
 	path = alloc_path_for_send();
 	if (!path)
 		return -ENOMEM;
+	path->reada = READA_FORWARD_ALWAYS;
 
 	key.objectid = BTRFS_FIRST_FREE_OBJECTID;
 	key.type = BTRFS_INODE_ITEM_KEY;
@@ -6688,15 +6689,35 @@ out:
 	return ret;
 }
 
-static int tree_move_down(struct btrfs_path *path, int *level)
+static int tree_move_down(struct btrfs_path *path, int *level, u64 reada_min_gen)
 {
 	struct extent_buffer *eb;
+	struct extent_buffer *parent = path->nodes[*level];
+	int slot = path->slots[*level];
+	const int nritems = btrfs_header_nritems(parent);
+	u64 reada_max;
+	u64 reada_done = 0;
 
 	BUG_ON(*level == 0);
-	eb = btrfs_read_node_slot(path->nodes[*level], path->slots[*level]);
+	eb = btrfs_read_node_slot(parent, slot);
 	if (IS_ERR(eb))
 		return PTR_ERR(eb);
 
+	/*
+	 * Trigger readahead for the next leaves we will process, so that it is
+	 * very likely that when we need them they are already in memory and we
+	 * will not block on disk IO. For nodes we only do readahead for one,
+	 * since the time window between processing nodes is typically larger.
+	 */
+	reada_max = (*level == 1 ? SZ_128K : eb->fs_info->nodesize);
+
+	for (slot++; slot < nritems && reada_done < reada_max; slot++) {
+		if (btrfs_node_ptr_generation(parent, slot) > reada_min_gen) {
+			btrfs_readahead_node_child(parent, slot);
+			reada_done += eb->fs_info->nodesize;
+		}
+	}
+
 	path->nodes[*level - 1] = eb;
 	path->slots[*level - 1] = 0;
 	(*level)--;
@@ -6736,14 +6757,15 @@ static int tree_move_next_or_upnext(struct btrfs_path *path,
 static int tree_advance(struct btrfs_path *path,
 			int *level, int root_level,
 			int allow_down,
-			struct btrfs_key *key)
+			struct btrfs_key *key,
+			u64 reada_min_gen)
 {
 	int ret;
 
 	if (*level == 0 || !allow_down) {
 		ret = tree_move_next_or_upnext(path, level, root_level);
 	} else {
-		ret = tree_move_down(path, level);
+		ret = tree_move_down(path, level, reada_min_gen);
 	}
 	if (ret >= 0) {
 		if (*level == 0)
@@ -6817,6 +6839,7 @@ static int btrfs_compare_trees(struct btrfs_root *left_root,
 	u64 right_blockptr;
 	u64 left_gen;
 	u64 right_gen;
+	u64 reada_min_gen;
 
 	left_path = btrfs_alloc_path();
 	if (!left_path) {
@@ -6896,6 +6919,14 @@ static int btrfs_compare_trees(struct btrfs_root *left_root,
 		ret = -ENOMEM;
 		goto out;
 	}
+	/*
+	 * Our right root is the parent root, while the left root is the "send"
+	 * root. We know that all new nodes/leaves in the left root must have
+	 * a generation greater than the right root's generation, so we trigger
+	 * readahead for those nodes and leaves of the left root, as we know we
+	 * will need to read them at some point.
+	 */
+	reada_min_gen = btrfs_header_generation(right_root->commit_root);
 	up_read(&fs_info->commit_root_sem);
 
 	if (left_level == 0)
@@ -6920,7 +6951,7 @@ static int btrfs_compare_trees(struct btrfs_root *left_root,
 			ret = tree_advance(left_path, &left_level,
 					left_root_level,
 					advance_left != ADVANCE_ONLY_NEXT,
-					&left_key);
+					&left_key, reada_min_gen);
 			if (ret == -1)
 				left_end_reached = ADVANCE;
 			else if (ret < 0)
@@ -6931,7 +6962,7 @@ static int btrfs_compare_trees(struct btrfs_root *left_root,
 			ret = tree_advance(right_path, &right_level,
 					right_root_level,
 					advance_right != ADVANCE_ONLY_NEXT,
-					&right_key);
+					&right_key, reada_min_gen);
 			if (ret == -1)
 				right_end_reached = ADVANCE;
 			else if (ret < 0)