Re: io_uring process termination/killing is not working

[Jens Axboe <axboe@kernel.dk>]

On 8/16/20 6:45 AM, Jens Axboe wrote:
> On 8/15/20 9:48 AM, Pavel Begunkov wrote:
>> On 15/08/2020 18:12, Jens Axboe wrote:
>>> On 8/15/20 12:45 AM, Pavel Begunkov wrote:
>>>> On 13/08/2020 02:32, Jens Axboe wrote:
>>>>> On 8/12/20 12:28 PM, Pavel Begunkov wrote:
>>>>>> On 12/08/2020 21:22, Pavel Begunkov wrote:
>>>>>>> On 12/08/2020 21:20, Pavel Begunkov wrote:
>>>>>>>> On 12/08/2020 21:05, Jens Axboe wrote:
>>>>>>>>> On 8/12/20 11:58 AM, Josef wrote:
>>>>>>>>>> Hi,
>>>>>>>>>>
>>>>>>>>>> I have a weird issue on kernel 5.8.0/5.8.1, SIGINT even SIGKILL
>>>>>>>>>> doesn't work to kill this process(always state D or D+), literally I
>>>>>>>>>> have to terminate my VM because even the kernel can't kill the process
>>>>>>>>>> and no issue on 5.7.12-201, however if IOSQE_IO_LINK is not set, it
>>>>>>>>>> works
>>>>>>>>>>
>>>>>>>>>> I've attached a file to reproduce it
>>>>>>>>>> or here
>>>>>>>>>> https://gist.github.com/1Jo1/15cb3c63439d0c08e3589cfa98418b2c
>>>>>>>>>
>>>>>>>>> Thanks, I'll take a look at this. It's stuck in uninterruptible
>>>>>>>>> state, which is why you can't kill it.
>>>>>>>>
>>>>>>>> It looks like one of the hangs I've been talking about a few days ago,
>>>>>>>> an accept is inflight but can't be found by cancel_files() because it's
>>>>>>>> in a link.
>>>>>>>
>>>>>>> BTW, I described it a month ago, there were more details.
>>>>>>
>>>>>> https://lore.kernel.org/io-uring/34eb5e5a-8d37-0cae-be6c-c6ac4d85b5d4@gmail.com
>>>>>
>>>>> Yeah I think you're right. How about something like the below? That'll
>>>>> potentially cancel more than just the one we're looking for, but seems
>>>>> kind of silly to only cancel from the file table holding request and to
>>>>> the end.
>>>>
>>>> The bug is not poll/t-out related, IIRC my test reproduces it with
>>>> read(pipe)->open(). See the previously sent link.
>>>
>>> Right, but in this context for poll, I just mean any request that has a
>>> poll handler armed. Not necessarily only a pure poll. The patch should
>>> fix your case, too.
>>
>> Ok. I was thinking about sleeping in io_read(), etc. from io-wq context.
>> That should have the same effect.
> 
> We already cancel any blocking work for the exiting task - but we do
> that _after_ trying to cancel files, so we should probably just swap
> those around in io_uring_flush(). That'll remove any need to find and
> cancel those explicitly in io_uring_cancel_files().

I guess there's still the case of the task just closing the fd, not
necessarily exiting. So I do agree with you that the io-wq case is still
unhandled. I'll take a look...

-- 
Jens Axboe