xfs_bunmapi can legitimately return before all work is done, to
avoid deadlocks across AGs.
Sadly nobody told xfs_repair, so it fires an assert if this happens:
phase6.c:1410: longform_dir2_rebuild: Assertion `done' failed.
Fix this by calling back in until all work is done, as we do
in the kernel.
Fixes: 5a8bcc ("xfs: fix multi-AG deadlock in xfs_bunmapi")
Bugzilla: https://bugzilla.redhat.com/show_bug.cgi?id=1641116
Reported-by: Tomasz Torcz <tomek@xxxxxxxxxxxxxx>
Signed-off-by: Eric Sandeen <sandeen@xxxxxxxxxx>
---
V2: libxfs-ify the code, remove now-pointless assert, thanks Darrick!
diff --git a/repair/phase6.c b/repair/phase6.c
index e017326..bdbbbaa 100644
--- a/repair/phase6.c
+++ b/repair/phase6.c
@@ -1317,7 +1317,7 @@ longform_dir2_rebuild(
xfs_fileoff_t lastblock;
xfs_inode_t pip;
dir_hash_ent_t *p;
- int done;
+ int done = 0;
/*
* trash directory completely and rebuild from scratch using the
@@ -1352,14 +1352,25 @@ longform_dir2_rebuild(
error);
/* free all data, leaf, node and freespace blocks */
- error = -libxfs_bunmapi(tp, ip, 0, lastblock, XFS_BMAPI_METADATA, 0,
- &done);
- if (error) {
- do_warn(_("xfs_bunmapi failed -- error - %d\n"), error);
- goto out_bmap_cancel;
- }
-
- ASSERT(done);
+ while (!done) {
+ error = -libxfs_bunmapi(tp, ip, 0, lastblock, XFS_BMAPI_METADATA,
+ 0, &done);
+ if (error) {
+ do_warn(_("xfs_bunmapi failed -- error - %d\n"), error);
+ goto out_bmap_cancel;
+ }
+ error = -libxfs_defer_finish(&tp);
+ if (error) {
+ do_warn(("defer_finish failed -- error - %d\n"), error);
+ goto out_bmap_cancel;
+ }
+ /*
+ * Close out trans and start the next one in the chain.
+ */
+ error = -libxfs_trans_roll_inode(&tp, ip);
+ if (error)
+ goto out_bmap_cancel;
+ }
error = -libxfs_dir_init(tp, ip, &pip);
if (error) {
