Re: [PATCH 5.10 1/5] io_uring: always let io_iopoll_complete() complete polled io.

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

On 08/12/2020 21:10, Jens Axboe wrote:
> On 12/8/20 12:24 PM, Pavel Begunkov wrote:
>> On 08/12/2020 19:17, Jens Axboe wrote:
>>> On 12/8/20 12:12 PM, Pavel Begunkov wrote:
>>>> On 07/12/2020 16:28, Jens Axboe wrote:
>>>>> On Sun, Dec 6, 2020 at 3:26 PM Pavel Begunkov <asml.silence@gmail.com> wrote:
>>>>>> From: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
>>>>>>
>>>>>> The reason is that once we got a non EAGAIN error in io_wq_submit_work(),
>>>>>> we'll complete req by calling io_req_complete(), which will hold completion_lock
>>>>>> to call io_commit_cqring(), but for polled io, io_iopoll_complete() won't
>>>>>> hold completion_lock to call io_commit_cqring(), then there maybe concurrent
>>>>>> access to ctx->defer_list, double free may happen.
>>>>>>
>>>>>> To fix this bug, we always let io_iopoll_complete() complete polled io.
>>>>>
>>>>> This patch is causing hangs with iopoll testing, if you end up getting
>>>>> -EAGAIN on request submission. I've dropped it.
>>>>
>>>> I fail to understand without debugging how does it happen, especially since
>>>> it shouldn't even get out of the while in io_wq_submit_work(). Is that
>>>> something obvious I've missed?
>>>
>>> I didn't have time to look into it, and haven't yet, just reporting that
>>> it very reliably fails (and under what conditions).
>>
>> Yeah, I get it, asked just in case.
>> I'll see what's going on if Xiaoguang wouldn't handle it before.
> 
> Should be trivial to reproduce on eg nvme by doing:
> 
> echo mq-deadline > /sys/block/nvme0n1/queue/scheduler
> echo 2 > /sys/block/nvme0n1/queue/nr_requests
> 
> and then run test/iopoll on that device. I'll try and take a look
> tomorrow unless someone beats me to it.

Tried out with iopoll-enabled null_blk. test/iopoll fails with
"test_io_uring_submit_enters failed", but if remove iteration limit
from the test, it completes... eventually.

premise: io_complete_rw_iopoll() gets -EAGAIN but returns 0 to
io_wq_submit_work().
The old version happily completes IO with that 0, but the patch delays
it to do_iopoll(), which retries and so all that repeats. And, I believe,
that the behaviour that io_wq_submit_work()'s -EGAIN check was trying to
achieve...

The question left is why no one progresses. May even be something in block.
Need to trace further.

-- 
Pavel Begunkov