btrfs: avoid unnecessary btree search restarts when reading node

When reading a btree node, at read_block_for_search(), if we don't find
the node's (or leaf) extent buffer in the cache, we will read it from
disk. Since that requires waiting on IO, we release all upper level nodes
from our path before reading the target node/leaf, and then return -EAGAIN
to the caller, which will make the caller restart the while btree search.

However we are causing the restart of btree search even for cases where
it is not necessary:

1) We have a path with ->skip_locking set to true, typically when doing
   a search on a commit root, so we are never holding locks on any node;

2) We are doing a read search (the "ins_len" argument passed to
   btrfs_search_slot() is 0), or we are doing a search to modify an
   existing key (the "cow" argument passed to btrfs_search_slot() has
   a value of 1 and "ins_len" is 0), in which case we never hold locks
   for upper level nodes;

3) We are doing a search to insert or delete a key, in which case we may
   or may not have upper level nodes locked. That depends on the current
   minimum write lock levels at btrfs_search_slot(), if we had to split
   or merge parent nodes, if we had to COW upper level nodes and if
   we ever visited slot 0 of an upper level node. It's still common to
   not have upper level nodes locked, but our current node must be at
   least at level 1, for insertions, or at least at level 2 for deletions.
   In these cases when we have locks on upper level nodes, they are always
   write locks.

These cases where we are not holding locks on upper level nodes far
outweigh the cases where we are holding locks, so it's completely wasteful
to retry the whole search when we have no upper nodes locked.

So change the logic to not return -EAGAIN, and make the caller retry the
search, when we don't have the parent node locked - when it's not locked
it means no other upper level nodes are locked as well.
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

fs/btrfs/ctree.c

@@ -1447,19 +1447,22 @@ read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
 		return 0;
 	}
 
-	/*
-	 * reduce lock contention at high levels
-	 * of the btree by dropping locks before
-	 * we read.  Don't release the lock on the current
-	 * level because we need to walk this node to figure
-	 * out which blocks to read.
-	 */
-	btrfs_unlock_up_safe(p, level + 1);
+	if ((level + 1 < BTRFS_MAX_LEVEL) && p->locks[level + 1]) {
+		/*
+		 * Reduce lock contention at high levels of the btree by
+		 * dropping locks before we read.  Don't release the lock
+		 * on the current level because we need to walk this node
+		 * to figure out which blocks to read.
+		 */
+		btrfs_unlock_up_safe(p, level + 1);
+		ret = -EAGAIN;
+	} else {
+		ret = 0;
+	}
 
 	if (p->reada != READA_NONE)
 		reada_for_search(fs_info, p, level, slot, key->objectid);
 
-	ret = -EAGAIN;
 	tmp = read_tree_block(fs_info, blocknr, root->root_key.objectid,
 			      gen, parent_level - 1, &first_key);
 	if (IS_ERR(tmp)) {
@@ -1474,9 +1477,14 @@ read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
 	 */
 	if (!extent_buffer_uptodate(tmp))
 		ret = -EIO;
-	free_extent_buffer(tmp);
 
-	btrfs_release_path(p);
+	if (ret == 0) {
+		*eb_ret = tmp;
+	} else {
+		free_extent_buffer(tmp);
+		btrfs_release_path(p);
+	}
+
 	return ret;
 }