btrfs: qgroup: Finish rescan when hit the last leaf of extent tree
authorQu Wenruo <wqu@suse.com>
Mon, 14 May 2018 01:38:13 +0000 (09:38 +0800)
committerDavid Sterba <dsterba@suse.com>
Mon, 28 May 2018 16:23:03 +0000 (18:23 +0200)
Under the following case, qgroup rescan can double account cowed tree
blocks:

In this case, extent tree only has one tree block.

-
| transid=5 last committed=4
| btrfs_qgroup_rescan_worker()
| |- btrfs_start_transaction()
| |  transid = 5
| |- qgroup_rescan_leaf()
|    |- btrfs_search_slot_for_read() on extent tree
|       Get the only extent tree block from commit root (transid = 4).
|       Scan it, set qgroup_rescan_progress to the last
|       EXTENT/META_ITEM + 1
|       now qgroup_rescan_progress = A + 1.
|
| fs tree get CoWed, new tree block is at A + 16K
| transid 5 get committed
-
| transid=6 last committed=5
| btrfs_qgroup_rescan_worker()
| btrfs_qgroup_rescan_worker()
| |- btrfs_start_transaction()
| |  transid = 5
| |- qgroup_rescan_leaf()
|    |- btrfs_search_slot_for_read() on extent tree
|       Get the only extent tree block from commit root (transid = 5).
|       scan it using qgroup_rescan_progress (A + 1).
|       found new tree block beyong A, and it's fs tree block,
|       account it to increase qgroup numbers.
-

In above case, tree block A, and tree block A + 16K get accounted twice,
while qgroup rescan should stop when it already reach the last leaf,
other than continue using its qgroup_rescan_progress.

Such case could happen by just looping btrfs/017 and with some
possibility it can hit such double qgroup accounting problem.

Fix it by checking the path to determine if we should finish qgroup
rescan, other than relying on next loop to exit.

Reported-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
fs/btrfs/qgroup.c

index 641ef2679bb1832891990fdf75e17afc92a28c38..457219d6a16b6b7ced8c4a1993123c7846377bc5 100644 (file)
@@ -2579,6 +2579,21 @@ out:
        spin_unlock(&fs_info->qgroup_lock);
 }
 
+/*
+ * Check if the leaf is the last leaf. Which means all node pointers
+ * are at their last position.
+ */
+static bool is_last_leaf(struct btrfs_path *path)
+{
+       int i;
+
+       for (i = 1; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
+               if (path->slots[i] != btrfs_header_nritems(path->nodes[i]) - 1)
+                       return false;
+       }
+       return true;
+}
+
 /*
  * returns < 0 on error, 0 when more leafs are to be scanned.
  * returns 1 when done.
@@ -2591,6 +2606,7 @@ qgroup_rescan_leaf(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
        struct extent_buffer *scratch_leaf = NULL;
        struct ulist *roots = NULL;
        u64 num_bytes;
+       bool done;
        int slot;
        int ret;
 
@@ -2619,6 +2635,7 @@ qgroup_rescan_leaf(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
                mutex_unlock(&fs_info->qgroup_rescan_lock);
                return ret;
        }
+       done = is_last_leaf(path);
 
        btrfs_item_key_to_cpu(path->nodes[0], &found,
                              btrfs_header_nritems(path->nodes[0]) - 1);
@@ -2663,6 +2680,8 @@ out:
                free_extent_buffer(scratch_leaf);
        }
 
+       if (done && !ret)
+               ret = 1;
        return ret;
 }