From: Darrick J. Wong <djwong@xxxxxxxxxx>
Dave Chinner complained about xfs_scrub failures coming from xfs/158.
That test induces xfs_repair to fail while upgrading a filesystem to
have the inobtcount feature, and then restarts xfs_repair to finish the
upgrade. When the second xfs_repair run starts, it will find that the
primary super has NEEDSREPAIR set, along with whatever new feature that
we were trying to add to the filesystem.
>From there, repair completes the upgrade in much the same manner as the
first repair run would have, with one big exception -- it forgets to set
features_changed to trigger rewriting of the secondary supers at the end
of repair. This results in discrepancies between the supers:
# XFS_REPAIR_FAIL_AFTER_PHASE=2 xfs_repair -c inobtcount=1 /dev/sdf
Phase 1 - find and verify superblock...
Phase 2 - using internal log
- zero log...
- scan filesystem freespace and inode maps...
- found root inode chunk
Adding inode btree counts to filesystem.
Killed
# xfs_repair /dev/sdf
Phase 1 - find and verify superblock...
Phase 2 - using internal log
- zero log...
- scan filesystem freespace and inode maps...
clearing needsrepair flag and regenerating metadata
bad inobt block count 0, saw 1
bad finobt block count 0, saw 1
bad inobt block count 0, saw 1
bad finobt block count 0, saw 1
bad inobt block count 0, saw 1
bad finobt block count 0, saw 1
bad inobt block count 0, saw 1
bad finobt block count 0, saw 1
- found root inode chunk
Phase 3 - for each AG...
- scan and clear agi unlinked lists...
- process known inodes and perform inode discovery...
- agno = 0
- agno = 1
- agno = 2
- agno = 3
- process newly discovered inodes...
Phase 4 - check for duplicate blocks...
- setting up duplicate extent list...
- check for inodes claiming duplicate blocks...
- agno = 1
- agno = 2
- agno = 0
- agno = 3
Phase 5 - rebuild AG headers and trees...
- reset superblock...
Phase 6 - check inode connectivity...
- resetting contents of realtime bitmap and summary inodes
- traversing filesystem ...
- traversal finished ...
- moving disconnected inodes to lost+found ...
Phase 7 - verify and correct link counts...
done
# xfs_db -c 'sb 0' -c 'print' -c 'sb 1' -c 'print' /dev/sdf | \
egrep '(features_ro_compat|features_incompat)'
features_ro_compat = 0xd
features_incompat = 0xb
features_ro_compat = 0x5
features_incompat = 0xb
Curiously, re-running xfs_repair will not trigger any warnings about the
featureset mismatch between the primary and secondary supers. xfs_scrub
immediately notices, which is what causes xfs/158 to fail.
This discrepancy doesn't happen when the upgrade completes successfully
in a single repair run, so we need to teach repair to rewrite the
secondaries at the end of repair any time needsrepair was set.
Reported-by: Dave Chinner <david@xxxxxxxxxxxxx>
Signed-off-by: Darrick J. Wong <djwong@xxxxxxxxxx>
---
repair/agheader.c | 8 ++++++++
1 file changed, 8 insertions(+)
diff --git a/repair/agheader.c b/repair/agheader.c
index d8f912f2..7da5641b 100644
--- a/repair/agheader.c
+++ b/repair/agheader.c
@@ -460,6 +460,14 @@ secondary_sb_whack(
else
do_warn(
_("would clear needsrepair flag and regenerate metadata\n"));
+ /*
+ * If needsrepair is set on the primary super, there's
+ * a possibility that repair crashed during an upgrade.
+ * Set features_changed to ensure that the secondary
+ * supers are rewritten with the new feature bits once
+ * we've finished the upgrade.
+ */
+ features_changed = true;
} else {
/*
* Quietly clear needsrepair on the secondary supers as
