xfs: fix race between healthmon unmount and read_iter
PATCH v2 1/1 xfs: fix race between healthmon unmount and read_iter
From: Darrick J. Wong <[email protected]>
xfs/1879 on one of my test VMs got stuck due to the xfs_io healthmon subcommand sleeping in wait_event_interruptible at:
xfs_healthmon_read_iter+0x558/0x5f8 [xfs] vfs_read+0x248/0x320 ksys_read+0x78/0x120
Looking at xfs_healthmon_read_iter, in !O_NONBLOCK mode it will sleep until the mount cookie == DETACHED_MOUNT_COOKIE, there are events waiting to be formatted, or there are formatted events in the read buffer that could be copied to userspace.
Poking into the running kernel, I see that there are zero events in the list, the read buffer is empty, and the mount cookie is indeed in DETACHED state. IOWs, xfs_healthmon_has_eventdata should have returned true, but instead we're asleep waiting for a wakeup.
I think what happened here is that xfs_healthmon_read_iter and xfs_healthmon_unmount were racing with each other, and _read_iter lost the race. _unmount queued an unmount event, which woke up _read_iter. It found, formatted, and copied the event out to userspace. That cleared out the pending event list and emptied the read buffer. xfs_io then called read() again, so _has_eventdata decided that we should sleep on the empty event queue.
Next, _unmount called xfs_healthmon_detach, which set the mount cookie to DETACHED. Unfortunately, it didn't call wake_up_all on the hm, so the wait_event_interruptible in the _read_iter thread remains asleep. That's why the test stalled.
Fix this by moving the wake_up_all call to xfs_healthmon_detach.
---
v2: add review tags
--- fs/xfs/xfs_healthmon.c | 17 ++++++++++------- 1 file changed, 10 insertions(+), 7 deletions(-)
fs/xfs/xfs_healthmon.c +10 -7
@@ -141,6 +141,16 @@ xfs_healthmon_detach( hm->mount_cookie =DETACHED_MOUNT_COOKIE ;spin_unlock (&xfs_healthmon_lock); +/* +* Wake up any readers that might remain. This can happen if unmount +* races with the healthmon fd owner entering ->read_iter, having +* already emptied the event queue. +* +* In the ->release case there shouldn't be any readers because the +* only users of the waiter are read and poll. +*/ +wake_up_all (&hm->wait); +trace_xfs_healthmon_detach (hm);xfs_healthmon_put (hm); }
@@ -1027,13 +1037,6 @@ xfs_healthmon_release(* process can create another health monitor file. */ xfs_healthmon_detach (hm); - -/* -* Wake up any readers that might be left. There shouldn't be any -* because the only users of the waiter are read and poll. -*/ -wake_up_all (&hm->wait); -xfs_healthmon_put (hm);return 0 ; }
Quoted 8 lines · Darrick J. Wong
xfs/1879 on one of my test VMs got stuck due to the
xfs_iohealthmon subcommand sleeping inwait_event_interruptibleat:xfs_healthmon_read_iter+0x558/0x5f8 [xfs] vfs_read+0x248/0x320 ksys_read+0x78/0x120[...]
Applied to for-next, thanks!
[1/1] xfs: fix race between healthmon unmount and read_iter
commit: 0ca1a8331c0fa5e57844e003a5d667a15b1e002c
Best regards,
Signature · 1 line
Carlos Maiolino <[email protected]>