btrfs: qgroup: Fix qgroup incorrectness caused by log replay
authorQu Wenruo <quwenruo@cn.fujitsu.com>
Mon, 15 Aug 2016 02:36:52 +0000 (10:36 +0800)
committerChris Mason <clm@fb.com>
Thu, 25 Aug 2016 10:58:23 +0000 (03:58 -0700)
When doing log replay at mount time(after power loss), qgroup will leak
numbers of replayed data extents.

The cause is almost the same of balance.
So fix it by manually informing qgroup for owner changed extents.

The bug can be detected by btrfs/119 test case.

Cc: Mark Fasheh <mfasheh@suse.de>
Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Reviewed-and-Tested-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/tree-log.c

index fff3f3efa43602e0c04f9e5e019bf0e85a239d6f..ffe92da81b8ace42dfd6f15f9d261fb62c113d2c 100644 (file)
@@ -27,6 +27,7 @@
 #include "backref.h"
 #include "hash.h"
 #include "compression.h"
+#include "qgroup.h"
 
 /* magic values for the inode_only field in btrfs_log_inode:
  *
@@ -680,6 +681,21 @@ static noinline int replay_one_extent(struct btrfs_trans_handle *trans,
                ins.type = BTRFS_EXTENT_ITEM_KEY;
                offset = key->offset - btrfs_file_extent_offset(eb, item);
 
+               /*
+                * Manually record dirty extent, as here we did a shallow
+                * file extent item copy and skip normal backref update,
+                * but modifying extent tree all by ourselves.
+                * So need to manually record dirty extent for qgroup,
+                * as the owner of the file extent changed from log tree
+                * (doesn't affect qgroup) to fs/file tree(affects qgroup)
+                */
+               ret = btrfs_qgroup_insert_dirty_extent(trans, root->fs_info,
+                               btrfs_file_extent_disk_bytenr(eb, item),
+                               btrfs_file_extent_disk_num_bytes(eb, item),
+                               GFP_NOFS);
+               if (ret < 0)
+                       goto out;
+
                if (ins.objectid > 0) {
                        u64 csum_start;
                        u64 csum_end;