Re: [PATCH 5.13 v2] io_uring: maintain drain requests' logic

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

On 09/04/2021 08:05, Hao Xu wrote:
> 在 2021/4/9 下午2:15, Hao Xu 写道:
>> 在 2021/4/9 上午12:18, Jens Axboe 写道:
>>> On 4/8/21 6:22 AM, Pavel Begunkov wrote:
>>>> On 08/04/2021 12:43, Hao Xu wrote:
>>>>> 在 2021/4/8 下午6:16, Hao Xu 写道:
>>>>>> 在 2021/4/7 下午11:49, Jens Axboe 写道:
>>>>>>> On 4/7/21 5:23 AM, Hao Xu wrote:
>>>>>>>> more tests comming, send this out first for comments.
>>>>>>>>
>>>>>>>> Hao Xu (3):
>>>>>>>>     io_uring: add IOSQE_MULTI_CQES/REQ_F_MULTI_CQES for multishot requests
>>>>>>>>     io_uring: maintain drain logic for multishot requests
>>>>>>>>     io_uring: use REQ_F_MULTI_CQES for multipoll IORING_OP_ADD
>>>>>>>>
>>>>>>>>    fs/io_uring.c                 | 34 +++++++++++++++++++++++++++++-----
>>>>>>>>    include/uapi/linux/io_uring.h |  8 +++-----
>>>>>>>>    2 files changed, 32 insertions(+), 10 deletions(-)
>>>>>>>
>>>>>>> Let's do the simple cq_extra first. I don't see a huge need to add an
>>>>>>> IOSQE flag for this, probably best to just keep this on a per opcode
>>>>>>> basis for now, which also then limits the code path to just touching
>>>>>>> poll for now, as nothing else supports multishot CQEs at this point.
>>>>>>>
>>>>>> gotcha.
>>>>>> a small issue here:
>>>>>>    sqe-->sqe(link)-->sqe(link)-->sqe(link, multishot)-->sqe(drain)
>>>>>>
>>>>>> in the above case, assume the first 3 single-shot reqs have completed.
>>>>>> then I think the drian request won't be issued now unless the multishot request in the linkchain has been issued. The trick is: a multishot req
>>>>>> in a linkchain consumes cached_sq_head when io_get_sqe(), which means it
>>>>>> is counted in seq, but we will deduct the sqe when it is issued if we
>>>>>> want to do the job per opcode not in the main code path.
>>>>>> before the multishot req issued:
>>>>>>        all_sqes(4) - multishot_sqes(0) == all_cqes(3) - multishot_cqes(0)
>>>>>> after the multishot req issued:
>>>>>>        all_sqes(4) - multishot_sqes(1) == all_cqes(3) - multishot_cqes(0)
>>>>>
>>>>> Sorry, my statement is wrong. It's not "won't be issued now unless the
>>>>> multishot request in the linkchain has been issued". Actually I now
>>>>> think the drain req won't be issued unless the multishot request in the
>>>>> linkchain has completed. Because we may first check req_need_defer()
>>>>> then issue(req->link), so:
>>>>>     sqe0-->sqe1(link)-->sqe2(link)-->sqe3(link, multishot)-->sqe4(drain)
>>>>>
>>>>>    sqe2 is completed:
>>>>>      call req_need_defer:
>>>>>      all_sqes(4) - multishot_sqes(0) == all_cqes(3) - multishot_cqes(0)
>>>>>    sqe3 is issued:
>>>>>      all_sqes(4) - multishot_sqes(1) == all_cqes(3) - multishot_cqes(0)
>>>>>    sqe3 is completed:
>>>>>      call req_need_defer:
>>>>>      all_sqes(4) - multishot_sqes(1) == all_cqes(3) - multishot_cqes(0)
>>>>>
>>>>> sqe4 shouldn't wait sqe3.
>>>>
>>>> Do you mean it wouldn't if the patch is applied? Because any drain
>>>> request must wait for all requests submitted before to complete. And
>>>> so before issuing sqe4 it must wait for sqe3 __request__ to die, and
>>>> so for all sqe3's CQEs.
>>>>
>>>> previously
>>>
>>> I think we need to agree on what multishot means for dependencies. Does
>>> it mean it just needs to trigger once? Or does it mean that it needs to
>>> be totally finished. The latter may obviously never happen, depending on
>>> the use case. Or it may be an expected condition because the caller will
>>> cancel it at some point.
>>>
>>> The most logical view imho is that multishot changes nothing wrt drain.
>>> If you ask for drain before something executes and you are using
>>> multishot, then you need to understand that the multishot request needs
>>> to fully complete before that condition is true and your dependency can
>>> execute.
>> This makes sense, and the implementation would be quite simpler. but we
>> really need to document it somewhere so that users easily get to know
>> that they cannot put a drain req after some multishot reqs if they don't
>> want it to wait for them. Otherwise I worry about wrong use of it since
>> the meaning of 'put a drain req after some multishot reqs' isn't so
>> obvious:
>>     - does it waits for those multishot reqs to complete once
>>     - or does it waits for those ones to fully complete
>>     - or does it ignore those ones at all
>>
> I realised that if a drain req has to wait for multishot reqs' fully
>  completion, then users have to explicitly cancel all the previous
> multishot reqs, otherwise it won't execute forever:
>     sqe0(multishot)-->sqe1(drain)-->sqe2(cancel multishot)    stuck

And it's not a new behaviour, e.g. read(pipe); drain(); where nobody
writes to the pipe will stuck as well.

I like that it currently provides a full barrier between requests, are
there other patterns used by someone?

-- 
Pavel Begunkov