[PATCH] Btrfs: fix ENOSPC in qgroups (metadata)

[Jan Schmidt <list.btrfs@jan-o-sch.net>]

When start_transaction() returns ENOSPC after btrfs_qgroup_reserve(), we
must call btrfs_qgroup_free() to avoid the qgroup counters increasing when
there's actually no data being written.

Signed-off-by: Jan Schmidt <list.btrfs@jan-o-sch.net>
---

Josef: It looks like these places (except the first) are candidates for
btrfs_block_rsv_release, please check.

---
 fs/btrfs/transaction.c |    9 +++++++--
 1 files changed, 7 insertions(+), 2 deletions(-)

diff --git a/fs/btrfs/transaction.c b/fs/btrfs/transaction.c
index 87fac9a..f2430fd 100644
--- a/fs/btrfs/transaction.c
+++ b/fs/btrfs/transaction.c
@@ -332,13 +332,17 @@ start_transaction(struct btrfs_root *root, u64 num_items, int type,
 		ret = btrfs_block_rsv_add(root,
 					  &root->fs_info->trans_block_rsv,
 					  num_bytes, flush);
-		if (ret)
+		if (ret) {
+			btrfs_qgroup_free(root, qgroup_reserved);
 			return ERR_PTR(ret);
+		}
 	}
 again:
 	h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
-	if (!h)
+	if (!h) {
+		btrfs_qgroup_free(root, qgroup_reserved);
 		return ERR_PTR(-ENOMEM);
+	}
 
 	/*
 	 * If we are JOIN_NOLOCK we're already committing a transaction and
@@ -369,6 +373,7 @@ again:
 		if (type < TRANS_JOIN_NOLOCK)
 			sb_end_intwrite(root->fs_info->sb);
 		kmem_cache_free(btrfs_trans_handle_cachep, h);
+		btrfs_qgroup_free(root, qgroup_reserved);
 		return ERR_PTR(ret);
 	}
 
-- 
1.7.1