xfs: mark i_mode=0 and shut down on attr fork inactivation failure

xfs_attr_inactive() can fail on an I/O error reading the attribute
fork's btree blocks, for example from fault injection such as the
dm-flakey errors used by xfstests generic/475. On failure,
xfs_ifork_zap_attr() zeros af_nextents without freeing the attribute
blocks or decrementing i_nblocks, leaving the inode in an
inconsistent in-memory state.

The inode is marked XFS_IRECLAIMABLE and remains in the cache. If a
concurrent operation (e.g., XFS_IOC_FSBULKSTAT) calls xfs_iget() on
this inode, xfs_iget_check_free_state() will allow recycling if
i_mode != 0. The recycled inode reuses the stale in-memory state
(af_nextents=0, i_nblocks=1).

When the recycled inode is queued for inactivation again,
xfs_attr_inactive() sees af_nextents=0 and skips truncation, leaving
i_nblocks non-zero. This triggers the assertion in xfs_ifree():

  XFS: Assertion failed: ip->i_nblocks == 0

Clear i_mode so the stale inode cannot be recycled, and force a
shutdown so log recovery frees the still-unlinked inode on the next
mount.

Suggested-by: Dave Chinner <dgc@kernel.org>
Signed-off-by: Ravi Singh <ravising@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Carlos Maiolino <cem@kernel.org>
This commit is contained in:
Ravi Singh 2026-10-21 08:33:29 +02:00 committed by Carlos Maiolino
parent 0344a9492e
commit bde0010631

View File

@ -1481,8 +1481,17 @@ xfs_inactive(
*/
if (xfs_inode_has_attr_fork(ip)) {
error = xfs_attr_inactive(ip);
if (error)
if (error) {
/*
* Attr teardown failed, leaking blocks with no in-core
* record. Clear i_mode to block xfs_iget() recycling,
* then shut down so recovery cleans up the inode.
*/
VFS_I(ip)->i_mode = 0;
if (!xfs_is_shutdown(mp))
xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
goto out;
}
}
ASSERT(ip->i_forkoff == 0);