xfs: fix xfs_inodegc_stop racing with mod_delayed_work
authorDarrick J. Wong <djwong@kernel.org>
Mon, 1 May 2023 23:16:14 +0000 (09:16 +1000)
committerDave Chinner <dchinner@redhat.com>
Mon, 1 May 2023 23:16:14 +0000 (09:16 +1000)
commit2254a7396a0ca6309854948ee1c0a33fa4268cec
treee006067c21fc3c92b270b41ac2b2aa5fe9b0598e
parent2d5f38a31980d7090f5bf91021488dc61a0ba8ee
xfs: fix xfs_inodegc_stop racing with mod_delayed_work

syzbot reported this warning from the faux inodegc shrinker that tries
to kick off inodegc work:

------------[ cut here ]------------
WARNING: CPU: 1 PID: 102 at kernel/workqueue.c:1445 __queue_work+0xd44/0x1120 kernel/workqueue.c:1444
RIP: 0010:__queue_work+0xd44/0x1120 kernel/workqueue.c:1444
Call Trace:
 __queue_delayed_work+0x1c8/0x270 kernel/workqueue.c:1672
 mod_delayed_work_on+0xe1/0x220 kernel/workqueue.c:1746
 xfs_inodegc_shrinker_scan fs/xfs/xfs_icache.c:2212 [inline]
 xfs_inodegc_shrinker_scan+0x250/0x4f0 fs/xfs/xfs_icache.c:2191
 do_shrink_slab+0x428/0xaa0 mm/vmscan.c:853
 shrink_slab+0x175/0x660 mm/vmscan.c:1013
 shrink_one+0x502/0x810 mm/vmscan.c:5343
 shrink_many mm/vmscan.c:5394 [inline]
 lru_gen_shrink_node mm/vmscan.c:5511 [inline]
 shrink_node+0x2064/0x35f0 mm/vmscan.c:6459
 kswapd_shrink_node mm/vmscan.c:7262 [inline]
 balance_pgdat+0xa02/0x1ac0 mm/vmscan.c:7452
 kswapd+0x677/0xd60 mm/vmscan.c:7712
 kthread+0x2e8/0x3a0 kernel/kthread.c:376
 ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:308

This warning corresponds to this code in __queue_work:

/*
 * For a draining wq, only works from the same workqueue are
 * allowed. The __WQ_DESTROYING helps to spot the issue that
 * queues a new work item to a wq after destroy_workqueue(wq).
 */
if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
     WARN_ON_ONCE(!is_chained_work(wq))))
return;

For this to trip, we must have a thread draining the inodedgc workqueue
and a second thread trying to queue inodegc work to that workqueue.
This can happen if freezing or a ro remount race with reclaim poking our
faux inodegc shrinker and another thread dropping an unlinked O_RDONLY
file:

Thread 0 Thread 1 Thread 2

xfs_inodegc_stop

xfs_inodegc_shrinker_scan
xfs_is_inodegc_enabled
<yes, will continue>

xfs_clear_inodegc_enabled
xfs_inodegc_queue_all
<list empty, do not queue inodegc worker>

xfs_inodegc_queue
<add to list>
xfs_is_inodegc_enabled
<no, returns>

drain_workqueue
<set WQ_DRAINING>

llist_empty
<no, will queue list>
mod_delayed_work_on(..., 0)
__queue_work
<sees WQ_DRAINING, kaboom>

In other words, everything between the access to inodegc_enabled state
and the decision to poke the inodegc workqueue requires some kind of
coordination to avoid the WQ_DRAINING state.  We could perhaps introduce
a lock here, but we could also try to eliminate WQ_DRAINING from the
picture.

We could replace the drain_workqueue call with a loop that flushes the
workqueue and queues workers as long as there is at least one inode
present in the per-cpu inodegc llists.  We've disabled inodegc at this
point, so we know that the number of queued inodes will eventually hit
zero as long as xfs_inodegc_start cannot reactivate the workers.

There are four callers of xfs_inodegc_start.  Three of them come from the
VFS with s_umount held: filesystem thawing, failed filesystem freezing,
and the rw remount transition.  The fourth caller is mounting rw (no
remount or freezing possible).

There are three callers ofs xfs_inodegc_stop.  One is unmounting (no
remount or thaw possible).  Two of them come from the VFS with s_umount
held: fs freezing and ro remount transition.

Hence, it is correct to replace the drain_workqueue call with a loop
that drains the inodegc llists.

Fixes: 6191cf3ad59f ("xfs: flush inodegc workqueue tasks before cancel")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
fs/xfs/xfs_icache.c