Re: [PATCH RFC] io_uring: remove retries from io_wq_submit_work()

[Jens Axboe <axboe@kernel.dk>]

On 2/24/20 11:33 AM, Pavel Begunkov wrote:
> On 24/02/2020 21:16, Jens Axboe wrote:
>> On 2/24/20 8:40 AM, Pavel Begunkov wrote:
>>> On 24/02/2020 18:27, Jens Axboe wrote:
>>>> On 2/24/20 2:15 AM, Pavel Begunkov wrote:
>>>>> It seems no opcode may return -EAGAIN for non-blocking case and expect
>>>>> to be reissued. Remove retry code from io_wq_submit_work().
>>>>
>>>> There's actually a comment right there on how that's possible :-)
>>>
>>> Yeah, I saw it and understand the motive, and how it may happen, but can't
>>> find a line, which can actually return -EAGAIN. Could you please point to an
>>> example?
>>
>> Just give it a whirl, should be easy to reproduce if you just do:
>>
>> # echo 2 > /sys/block/nvme0n1/queue/nr_requests
>> # fio/t/io_uring /dev/nvme0n1
>>
>> or something like that. It's propagated from the kiocb endio handler,
>> through, req->result at the bottom of io_issue_sqe()
> 
> I see now, thanks! What a jungle

The problem is, as you alluded to, that errors are returned through the
callback. Not sure if you recall, but I did make an attempt at doing
inline -EAGAIN returns, but had to abort and never got around to fixing
it up. See the revert here:

commit 7b6620d7db566a46f49b4b9deab9fa061fd4b59b
Author: Jens Axboe <axboe@kernel.dk>
Date:   Thu Aug 15 11:09:16 2019 -0600

    block: remove REQ_NOWAIT_INLINE

Without that, we can't do better than what we do now. With the inline
NOWAIT, we'd get the same treatment on polled and non-polled, and we
could kill that check. Which would be lovely...

-- 
Jens Axboe