btrfs: fix race between finishing block group creation and its item update

Commit 675dfe12 ("btrfs: fix block group item corruption after
inserting new block group") fixed one race that resulted in not persisting
a block group's item when its "used" bytes field decreases to zero.
However there's another race that can happen in a much shorter time window
that results in the same problem. The following sequence of steps explains
how it can happen:

1) Task A creates a metadata block group X, its "used" and "commit_used"
   fields are initialized to 0;

2) Two extents are allocated from block group X, so its "used" field is
   updated to 32K, and its "commit_used" field remains as 0;

3) Transaction commit starts, by some task B, and it enters
   btrfs_start_dirty_block_groups(). There it tries to update the block
   group item for block group X, which currently has its "used" field with
   a value of 32K and its "commit_used" field with a value of 0. However
   that fails since the block group item was not yet inserted, so at
   update_block_group_item(), the btrfs_search_slot() call returns 1, and
   then we set 'ret' to -ENOENT. Before jumping to the label 'fail'...

4) The block group item is inserted by task A, when for example
   btrfs_create_pending_block_groups() is called when releasing its
   transaction handle. This results in insert_block_group_item() inserting
   the block group item in the extent tree (or block group tree), with a
   "used" field having a value of 32K and setting "commit_used", in struct
   btrfs_block_group, to the same value (32K);

5) Task B jumps to the 'fail' label and then resets the "commit_used"
   field to 0. At btrfs_start_dirty_block_groups(), because -ENOENT was
   returned from update_block_group_item(), we add the block group again
   to the list of dirty block groups, so that we will try again in the
   critical section of the transaction commit when calling
   btrfs_write_dirty_block_groups();

6) Later the two extents from block group X are freed, so its "used" field
   becomes 0;

7) If no more extents are allocated from block group X before we get into
   btrfs_write_dirty_block_groups(), then when we call
   update_block_group_item() again for block group X, we will not update
   the block group item to reflect that it has 0 bytes used, because the
   "used" and "commit_used" fields in struct btrfs_block_group have the
   same value, a value of 0.

   As a result after committing the transaction we have an empty block
   group with its block group item having a 32K value for its "used" field.
   This will trigger errors from fsck ("btrfs check" command) and after
   mounting again the fs, the cleaner kthread will not automatically delete
   the empty block group, since its "used" field is not 0. Possibly there
   are other issues due to this inconsistency.

   When this issue happens, the error reported by fsck is like this:

     [1/7] checking root items
     [2/7] checking extents
     block group [1104150528 1073741824] used 39796736 but extent items used 0
     ERROR: errors found in extent allocation tree or chunk allocation
     (...)

So fix this by not resetting the "commit_used" field of a block group when
we don't find the block group item at update_block_group_item().

Fixes: 7248e0ce ("btrfs: skip update of block group item if used bytes are the same")
CC: stable@vger.kernel.org # 6.2+
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

fs/btrfs/block-group.c

@@ -3028,8 +3028,16 @@ static int update_block_group_item(struct btrfs_trans_handle *trans,
 	btrfs_mark_buffer_dirty(leaf);
 fail:
 	btrfs_release_path(path);
-	/* We didn't update the block group item, need to revert @commit_used. */
-	if (ret < 0) {
+	/*
+	 * We didn't update the block group item, need to revert commit_used
+	 * unless the block group item didn't exist yet - this is to prevent a
+	 * race with a concurrent insertion of the block group item, with
+	 * insert_block_group_item(), that happened just after we attempted to
+	 * update. In that case we would reset commit_used to 0 just after the
+	 * insertion set it to a value greater than 0 - if the block group later
+	 * becomes with 0 used bytes, we would incorrectly skip its update.
+	 */
+	if (ret < 0 && ret != -ENOENT) {
 		spin_lock(&cache->lock);
 		cache->commit_used = old_commit_used;
 		spin_unlock(&cache->lock);