Re: INFO: task hung in io_uring_cancel_sqpoll

[Pavel Begunkov <asml.silence@gmail.com>]

On 5/10/21 5:47 AM, Palash Oswal wrote:
> On Mon, May 3, 2021 at 4:15 PM Pavel Begunkov <asml.silence@gmail.com>
> wrote:
> 
>> On 5/3/21 5:41 AM, Palash Oswal wrote:
>>> On Mon, May 3, 2021 at 3:42 AM Pavel Begunkov <asml.silence@gmail.com>
>> wrote:
>>>> The test might be very useful. Would you send a patch to
>>>> liburing? Or mind the repro being taken as a test?
>>>
>>> Pavel,
>>>
>>> I'm working on a PR for liburing with this test. Do you know how I can
>>
>> Perfect, thanks
>>
>>> address this behavior?
>>
>> As the hang is sqpoll cancellations, it's most probably
>> fixed in 5.13 + again awaits to be taken for stable.
>>
>> Don't know about segfaults, but it was so for long, and
>> syz reproducers are ofthen faults for me, and exit with 0
>> in the end. So, I wouldn't worry about it.
>>
>>
> Hey Pavel,
> The bug actually fails to reproduce on 5.12 when the fork() call is made by
> the runtests.sh script. This causes the program to end correctly, and the
> hang does not occur. I verified this on 5.12 where the bug isn't patched.
> Just running the `sqpoll-cancel-hang` triggers the bug; whereas running it
> after being forked from runtests.sh does not trigger the bug.

I see. fyi, it's always good to wait for 5 minutes, because some useful
logs are not generated immediately but do timeout based hang detection.

I'd think that may be due CLONE_IO and how to whom it binds workers,
but can you try first:

1) timeout -s INT -k $TIMEOUT $TIMEOUT ./sqpoll-cancel-hang

2) remove timeout from <liburing>/tests/Makefile and run
"./runtests.sh sqpoll-cancel-hang" again looking for faults?

diff --git a/test/runtests.sh b/test/runtests.sh
index e8f4ae5..2b51dca 100755
--- a/test/runtests.sh
+++ b/test/runtests.sh
@@ -91,7 +91,8 @@ run_test()
 	fi
 
 	# Run the test
-	timeout -s INT -k $TIMEOUT $TIMEOUT ./$test_name $dev
+	# timeout -s INT -k $TIMEOUT $TIMEOUT ./$test_name $dev
+	./$test_name $dev
 	local status=$?
 
 	# Check test status


> 
> The segfaults are benign, but notice the "All tests passed" in the previous
> mail. It should not have passed, as the run was on 5.12. Therefore I wanted
> to ask your input on how to address this odd behaviour, where the
> involvement of runtests.sh actually mitigated the bug.
> 
> 
> 
>>> root@syzkaller:~/liburing/test# ./runtests.sh sqpoll-cancel-hang
>>> Running test sqp[   15.310997] Running test sqpoll-cancel-hang:
>>> oll-cancel-hang:
>>> [   15.333348] sqpoll-cancel-h[305]: segfault at 0 ip 000055ad00e265e3
>> sp]
>>> [   15.334940] Code: 89 d8 8d 34 90 8b 45 04 ba 03 00 00 00 c1 e0 04 03
>> 46
>>> All tests passed
>>>
>>> root@syzkaller:~/liburing/test# ./sqpoll-cancel-hang
>>> [   13.572639] sqpoll-cancel-h[298]: segfault at 0 ip 00005634c4a455e3
>> sp]
>>> [   13.576506] Code: 89 d8 8d 34 90 8b 45 04 ba 03 00 00 00 c1 e0 04 03
>> 46
>>> [   23.350459] random: crng init done
>>> [   23.352837] random: 7 urandom warning(s) missed due to ratelimiting
>>> [  243.090865] INFO: task iou-sqp-298:299 blocked for more than 120
>> secon.
>>> [  243.095187]       Not tainted 5.12.0 #142
>>> [  243.099800] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs"
>> disable.
>>> [  243.105928] task:iou-sqp-298     state:D stack:    0 pid:  299 ppid:
>> 4
>>> [  243.111044] Call Trace:
>>> [  243.112855]  __schedule+0xb1d/0x1130
>>> [  243.115549]  ? __sched_text_start+0x8/0x8
>>> [  243.118328]  ? io_wq_worker_sleeping+0x145/0x500
>>> [  243.121790]  schedule+0x131/0x1c0
>>> [  243.123698]  io_uring_cancel_sqpoll+0x288/0x350
>>> [  243.125977]  ? io_sq_thread_unpark+0xd0/0xd0
>>> [  243.128966]  ? mutex_lock+0xbb/0x130
>>> [  243.132572]  ? init_wait_entry+0xe0/0xe0
>>> [  243.135429]  ? wait_for_completion_killable_timeout+0x20/0x20
>>> [  243.138303]  io_sq_thread+0x174c/0x18c0
>>> [  243.140162]  ? io_rsrc_put_work+0x380/0x380
>>> [  243.141613]  ? init_wait_entry+0xe0/0xe0
>>> [  243.143686]  ? _raw_spin_lock_irq+0xa5/0x180
>>> [  243.145619]  ? _raw_spin_lock_irqsave+0x190/0x190
>>> [  243.147671]  ? calculate_sigpending+0x6b/0xa0
>>> [  243.149036]  ? io_rsrc_put_work+0x380/0x380
>>> [  243.150694]  ret_from_fork+0x22/0x30
>>>
>>> Palash
>>>
>>
>> --
>> Pavel Begunkov
>>
> 

-- 
Pavel Begunkov