sbitmap: Avoid leaving waitqueue in invalid state in __sbq_wake_up()
authorJan Kara <jack@suse.cz>
Thu, 8 Sep 2022 13:09:37 +0000 (15:09 +0200)
committerJens Axboe <axboe@kernel.dk>
Thu, 8 Sep 2022 14:39:04 +0000 (08:39 -0600)
commit48c033314f372478548203c583529f53080fd078
tree75c41982aac9bd59c1964c1fd1c9da8cd29afdc2
parent8807707df7ef09d679df2b21894d49a06fd8ba7e
sbitmap: Avoid leaving waitqueue in invalid state in __sbq_wake_up()

When __sbq_wake_up() decrements wait_cnt to 0 but races with someone
else waking the waiter on the waitqueue (so the waitqueue becomes
empty), it exits without reseting wait_cnt to wake_batch number. Once
wait_cnt is 0, nobody will ever reset the wait_cnt or wake the new
waiters resulting in possible deadlocks or busyloops. Fix the problem by
making sure we reset wait_cnt even if we didn't wake up anybody in the
end.

Fixes: 040b83fcecfb ("sbitmap: fix possible io hung due to lost wakeup")
Reported-by: Keith Busch <kbusch@kernel.org>
Signed-off-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20220908130937.2795-1-jack@suse.cz
Signed-off-by: Jens Axboe <axboe@kernel.dk>
lib/sbitmap.c