Btrfs: Fix race against disk_i_size updates

The code to update the on disk i_size happens before the
ordered_extent record is removed.  So, it is possible for multiple
ordered_extent completion routines to run at the same time, and to
find each other in the ordered tree.

The end result is they both decide not to update disk_i_size, leaving
it too small.  This temporary fix just puts the updates inside
the extent_mutex.  A real solution would be stronger ordering of
disk_i_size updates against removing the ordered extent from the tree.

Signed-off-by: Chris Mason <chris.mason@oracle.com>

fs/btrfs/inode.c

@@ -608,9 +608,11 @@ static int btrfs_finish_ordered_io(struct inode *inode, u64 start, u64 end)
 	add_pending_csums(trans, inode, ordered_extent->file_offset,
 			  &ordered_extent->list);
 
+	mutex_lock(&BTRFS_I(inode)->extent_mutex);
 	btrfs_ordered_update_i_size(inode, ordered_extent);
 	btrfs_update_inode(trans, root, inode);
 	btrfs_remove_ordered_extent(inode, ordered_extent);
+	mutex_unlock(&BTRFS_I(inode)->extent_mutex);
 
 	/* once for us */
 	btrfs_put_ordered_extent(ordered_extent);