Btrfs: don't delay inode ref updates during log replay
authorChris Mason <clm@fb.com>
Wed, 31 Dec 2014 17:18:29 +0000 (12:18 -0500)
committerChris Mason <clm@fb.com>
Fri, 2 Jan 2015 19:47:56 +0000 (14:47 -0500)
Commit 1d52c78afbb (Btrfs: try not to ENOSPC on log replay) added a
check to skip delayed inode updates during log replay because it
confuses the enospc code.  But the delayed processing will end up
ignoring delayed refs from log replay because the inode itself wasn't
put through the delayed code.

This can end up triggering a warning at commit time:

WARNING: CPU: 2 PID: 778 at fs/btrfs/delayed-inode.c:1410 btrfs_assert_delayed_root_empty+0x32/0x34()

Which is repeated for each commit because we never process the delayed
inode ref update.

The fix used here is to change btrfs_delayed_delete_inode_ref to return
an error if we're currently in log replay.  The caller will do the ref
deletion immediately and everything will work properly.

Signed-off-by: Chris Mason <clm@fb.com>
cc: stable@vger.kernel.org # v3.18 and any stable series that picked 1d52c78afbbf80b58299e076a159617d6b42fe3c

fs/btrfs/delayed-inode.c

index 054577bddaf27869d9a524a73d4df5a76072e4e1..de4e70fb3cbbd4a5c28d13f1fe3aec16733ed49f 100644 (file)
@@ -1857,6 +1857,14 @@ int btrfs_delayed_delete_inode_ref(struct inode *inode)
 {
        struct btrfs_delayed_node *delayed_node;
 
+       /*
+        * we don't do delayed inode updates during log recovery because it
+        * leads to enospc problems.  This means we also can't do
+        * delayed inode refs
+        */
+       if (BTRFS_I(inode)->root->fs_info->log_root_recovering)
+               return -EAGAIN;
+
        delayed_node = btrfs_get_or_create_delayed_node(inode);
        if (IS_ERR(delayed_node))
                return PTR_ERR(delayed_node);