diff options
| author | Joseph Qi <joseph.qi@linux.alibaba.com> | 2026-09-22 17:34:37 +0800 |
|---|---|---|
| committer | Andrew Morton <akpm@linux-foundation.org> | 2026-10-13 21:25:47 -0700 |
| commit | 169bc4d16a8efa7852b101fdaa4c028505cd6750 (patch) | |
| tree | 8eabffd0c13a83041cc5dfcdc8196915639373ae /rust/kernel/alloc | |
| parent | 60f50cdd5ac54bcbc21a4b01b1c2f6409d4563a7 (diff) | |
ocfs2: fix possible deadlock between nfs_sync_rwlock and fs_reclaim
syzbot detected a circular locking dependency on &osb->nfs_sync_rwlock:
CPU0 CPU1
---- ----
lock(fs_reclaim);
lock(&ocfs2_sysfile_lock_key[INODE_ALLOC_SYSTEM_INODE]);
lock(fs_reclaim);
rlock(&osb->nfs_sync_rwlock);
*** DEADLOCK ***
Chain exists of:
&osb->nfs_sync_rwlock --> &ocfs2_sysfile_lock_key[INODE_ALLOC_SYSTEM_INODE] --> fs_reclaim
CPU0 is kswapd. The dentry shrinker runs under fs_reclaim and reaches
ocfs2_delete_inode() through ->evict_inode():
kswapd
balance_pgdat
shrink_node
shrink_slab
super_cache_scan
prune_dcache_sb
shrink_dentry_list
dentry_kill
ocfs2_dentry_iput
evict
ocfs2_evict_inode
ocfs2_delete_inode
ocfs2_nfs_sync_lock
down_read(&osb->nfs_sync_rwlock) //C0: grabbing
CPU1 is a task deleting an inode. It takes nfs_sync_rwlock first, then
the orphan dir and inode alloc system inode i_rwsems, and allocates the
metadata reservation with GFP_KERNEL while holding them:
evict
ocfs2_evict_inode
ocfs2_delete_inode
+ ocfs2_nfs_sync_lock
| down_read(&osb->nfs_sync_rwlock) //C1: hold
+ ocfs2_wipe_inode
+ inode_lock(orphan_dir_inode) //C1: hold
+ ocfs2_truncate_for_delete
| ocfs2_commit_truncate
| ocfs2_remove_btree_range
| ocfs2_reserve_blocks_for_rec_trunc
| ocfs2_reserve_new_metadata_blocks
| kzalloc_obj() //C1: grabbing
| // GFP_KERNEL -> might_alloc -> fs_reclaim_acquire
+ ocfs2_remove_inode
inode_lock(inode_alloc_inode)
jbd2 pins allocations to GFP_NOFS for the lifetime of a transaction
handle, but these reservations are made before ocfs2_start_trans(), so
fs_reclaim is acquired with ocfs2 locks still held and the cycle closes.
The same cycle is reachable from ocfs2_get_dentry() and
ocfs2_get_parent(), which hold nfs_sync_rwlock for write across
ocfs2_test_inode_bit() and ocfs2_iget().
This is more than a lockdep artifact. A GFP_KERNEL allocation under the
orphan dir i_rwsem enters direct reclaim, which runs the shrinkers, which
can evict another ocfs2 inode and re-enter ocfs2_wipe_inode() on the same
task; the second inode_lock() on the singleton orphan dir inode then
self-deadlocks, since i_rwsem is not recursive.
Establish a GFP_NOFS allocation context for the whole nfs_sync_rwlock
critical section so that nothing holding it can recurse into filesystem
reclaim. Do it in ocfs2_nfs_sync_lock()/ocfs2_nfs_sync_unlock() rather
than at the call sites, so that future callers cannot forget it.
Link: https://lore.kernel.org/20260922093437.954127-1-joseph.qi@linux.alibaba.com
Fixes: 6ca497a83e59 ("ocfs2: fix rare stale inode errors when exporting via nfs")
Signed-off-by: Joseph Qi <joseph.qi@linux.alibaba.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reported-by: syzbot+a68ce48df87b8e36e915@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=a68ce48df87b8e36e915
Cc: Heming Zhao <heming.zhao@suse.com>
Cc: Mark Fasheh <mark@fasheh.com>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Junxiao Bi <junxiao.bi@oracle.com>
Diffstat (limited to 'rust/kernel/alloc')
0 files changed, 0 insertions, 0 deletions
