[RFC,4/4] btrfs: skip subtree scan if it's too high to avoid low stall in btrfs_commit_transaction()

Commit Message

Qu Wenruo Aug. 22, 2021, 7:02 a.m. UTC

Btrfs qgroup has a long history of bringing performance penalty in
btrfs_commit_transaction().

Although we tried our best to migrate such impact, there is still a
unsolved call site, btrfs_drop_snapshot().

This function will find the highest shared tree block and modify its
extent ownership to do a subvolume/snapshot dropping.

Such change will affect the whole subtree, and cause tons of qgroup
dirty extents and stall btrfs_commit_transaction().

To avoid such problem, we can simply skip such subtree accounting if
it's too high.
Of course, the cost is to mark qgroup inconsistent.

Signed-off-by: Qu Wenruo <[e-mail]>
---
 fs/btrfs/qgroup.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

Patch

diff --git a/fs/btrfs/qgroup.c b/fs/btrfs/qgroup.c
index 291c404e8718..c650258f5cec 100644
--- a/fs/btrfs/qgroup.c
+++ b/fs/btrfs/qgroup.c
@@ -2248,6 +2248,19 @@  int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
 	if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
 		return 0;
 
+	/*
+	 * This function only get called for snapshot drop, if we hit a high
+	 * node here, it means we are going to change ownership for quite a lot
+	 * of extents, which will greatly slow down btrfs_commit_transaction().
+	 *
+	 * So here if we find a high tree here, we just skip the accounting and
+	 * mark qgroup inconsistent.
+	 */
+	if (root_level >= 4) {
+		qgroup_mark_inconsistent(fs_info);
+		return 0;
+	}
+
 	if (!extent_buffer_uptodate(root_eb)) {
 		ret = btrfs_read_buffer(root_eb, root_gen, root_level, NULL);
 		if (ret)