Message-ID: <897dc597-fc0a-34ec-84b8-7e1c4901e0fc@leemhuis.info>
Date:   Sun, 15 May 2022 10:20:43 +0200
From:   Thorsten Leemhuis <regressions@...mhuis.info>
To:     Daniel Harding <dharding@...ing180.net>,
        Pavel Begunkov <asml.silence@...il.com>,
        Jens Axboe <axboe@...nel.dk>
Cc:     regressions@...ts.linux.dev, io-uring@...r.kernel.org,
        linux-kernel@...r.kernel.org
Subject: Re: [REGRESSION] lxc-stop hang on 5.17.x kernels

On 04.05.22 08:54, Daniel Harding wrote:
> On 5/3/22 17:14, Pavel Begunkov wrote:
>> On 5/3/22 08:37, Daniel Harding wrote:
>>> [Resend with a smaller trace]
>>> On 5/3/22 02:14, Pavel Begunkov wrote:
>>>> On 5/2/22 19:49, Daniel Harding wrote:
>>>>> On 5/2/22 20:40, Pavel Begunkov wrote:
>>>>>> On 5/2/22 18:00, Jens Axboe wrote:
>>>>>>> On 5/2/22 7:59 AM, Jens Axboe wrote:
>>>>>>>> On 5/2/22 7:36 AM, Daniel Harding wrote:
>>>>>>>>> On 5/2/22 16:26, Jens Axboe wrote:
>>>>>>>>>> On 5/2/22 7:17 AM, Daniel Harding wrote:
>>>>>>>>>>> I use lxc-4.0.12 on Gentoo, built with io-uring support
>>>>>>>>>>> (--enable-liburing), targeting liburing-2.1.  My kernel
>>>>>>>>>>> config is a
>>>>>>>>>>> very lightly modified version of Fedora's generic kernel
>>>>>>>>>>> config. After
>>>>>>>>>>> moving from the 5.16.x series to the 5.17.x kernel series, I
>>>>>>>>>>> started
>>>>>>>>>>> noticed frequent hangs in lxc-stop.  It doesn't happen 100%
>>>>>>>>>>> of the
>>>>>>>>>>> time, but definitely more than 50% of the time. Bisecting
>>>>>>>>>>> narrowed
>>>>>>>>>>> down the issue to commit
>>>>>>>>>>> aa43477b040251f451db0d844073ac00a8ab66ee:
>>>>>>>>>>> io_uring: poll rework. Testing indicates the problem is still
>>>>>>>>>>> present
>>>>>>>>>>> in 5.18-rc5. Unfortunately I do not have the expertise with the
>>>>>>>>>>> codebases of either lxc or io-uring to try to debug the problem
>>>>>>>>>>> further on my own, but I can easily apply patches to any of the
>>>>>>>>>>> involved components (lxc, liburing, kernel) and rebuild for
>>>>>>>>>>> testing or
>>>>>>>>>>> validation.  I am also happy to provide any further
>>>>>>>>>>> information that
>>>>>>>>>>> would be helpful with reproducing or debugging the problem.
>>>>>>>>>> Do you have a recipe to reproduce the hang? That would make it
>>>>>>>>>> significantly easier to figure out.
>>>>>>>>>
>>>>>>>>> I can reproduce it with just the following:
>>>>>>>>>
>>>>>>>>>      sudo lxc-create --n lxc-test --template download --bdev
>>>>>>>>> dir --dir /var/lib/lxc/lxc-test/rootfs -- -d ubuntu -r bionic
>>>>>>>>> -a amd64
>>>>>>>>>      sudo lxc-start -n lxc-test
>>>>>>>>>      sudo lxc-stop -n lxc-test
>>>>>>>>>
>>>>>>>>> The lxc-stop command never exits and the container continues
>>>>>>>>> running.
>>>>>>>>> If that isn't sufficient to reproduce, please let me know.
>>>>>>>>
>>>>>>>> Thanks, that's useful! I'm at a conference this week and hence have
>>>>>>>> limited amount of time to debug, hopefully Pavel has time to
>>>>>>>> take a look
>>>>>>>> at this.
>>>>>>>
>>>>>>> Didn't manage to reproduce. Can you try, on both the good and bad
>>>>>>> kernel, to do:
>>>>>>
>>>>>> Same here, it doesn't reproduce for me
>>>>> OK, sorry it wasn't something simple.
>>>>>> # echo 1 > /sys/kernel/debug/tracing/events/io_uring/enable
>>>>>>>
>>>>>>> run lxc-stop
>>>>>>>
>>>>>>> # cp /sys/kernel/debug/tracing/trace ~/iou-trace
>>>>>>>
>>>>>>> so we can see what's going on? Looking at the source, lxc is just
>>>>>>> using
>>>>>>> plain POLL_ADD, so I'm guessing it's not getting a notification
>>>>>>> when it
>>>>>>> expects to, or it's POLL_REMOVE not doing its job. If we have a
>>>>>>> trace
>>>>>>> from both a working and broken kernel, that might shed some light
>>>>>>> on it.
>>>>> It's late in my timezone, but I'll try to work on getting those
>>>>> traces tomorrow.
>>>>
>>>> I think I got it, I've attached a trace.
>>>>
>>>> What's interesting is that it issues a multi shot poll but I don't
>>>> see any kind of cancellation, neither cancel requests nor task/ring
>>>> exit. Perhaps have to go look at lxc to see how it's supposed
>>>> to work
>>>
>>> Yes, that looks exactly like my bad trace.  I've attached good trace
>>> (captured with linux-5.16.19) and a bad trace (captured with
>>> linux-5.17.5).  These are the differences I noticed with just a
>>> visual scan:
>>>
>>> * Both traces have three io_uring_submit_sqe calls at the very
>>> beginning, but in the good trace, there are further
>>> io_uring_submit_sqe calls throughout the trace, while in the bad
>>> trace, there are none.
>>> * The good trace uses a mask of c3 for io_uring_task_add much more
>>> often than the bad trace:  the bad trace uses a mask of c3 only for
>>> the very last call to io_uring_task_add, but a mask of 41 for the
>>> other calls.
>>> * In the good trace, many of the io_uring_complete calls have a
>>> result of 195, while in the bad trace, they all have a result of 1.
>>>
>>> I don't know whether any of those things are significant or not, but
>>> that's what jumped out at me.
>>>
>>> I have also attached a copy of the script I used to generate the
>>> traces.  If there is anything further I can to do help debug, please
>>> let me know.
>>
>> Good observations! thanks for traces.
>>
>> It sounds like multi-shot poll requests were getting downgraded
>> to one-shot, which is a valid behaviour and was so because we
>> didn't fully support some cases. If that's the reason, than
>> the userspace/lxc is misusing the ABI. At least, that's the
>> working hypothesis for now, need to check lxc.
> 
> So, I looked at the lxc source code, and it appears to at least try to
> handle the case of multi-shot being downgraded to one-shot.  I don't
> know enough to know if the code is actually correct however:
> 
> https://github.com/lxc/lxc/blob/7e37cc96bb94175a8e351025d26cc35dc2d10543/src/lxc/mainloop.c#L165-L189
> https://github.com/lxc/lxc/blob/7e37cc96bb94175a8e351025d26cc35dc2d10543/src/lxc/mainloop.c#L254
> https://github.com/lxc/lxc/blob/7e37cc96bb94175a8e351025d26cc35dc2d10543/src/lxc/mainloop.c#L288-L290

Hi, this is your Linux kernel regression tracker. Nothing happened here
for round about ten days now afaics; or did the discussion continue
somewhere else.

>From what I gathered from this discussion is seems the root cause might
be in LXC, but it was exposed by kernel change. That makes it sill a
kernel regression that should be fixed; or is there a strong reason why
we should let this one slip?

Ciao, Thorsten (wearing his 'the Linux kernel's regression tracker' hat)

P.S.: As the Linux kernel's regression tracker I deal with a lot of
reports and sometimes miss something important when writing mails like
this. If that's the case here, don't hesitate to tell me in a public
reply, it's in everyone's interest to set the public record straight.

#regzbot poke

Powered by blists - more mailing lists