Re: [PATCH 6/6] io_uring: enable multishot mode for accept

[Jens Axboe <axboe@kernel.dk>]

On 9/6/21 9:32 AM, Pavel Begunkov wrote:
> On 9/6/21 1:35 PM, Hao Xu wrote:
>> 在 2021/9/6 上午3:44, Jens Axboe 写道:
>>> On 9/4/21 4:46 PM, Pavel Begunkov wrote:
>>>> On 9/4/21 7:40 PM, Jens Axboe wrote:
>>>>> On 9/4/21 9:34 AM, Hao Xu wrote:
>>>>>> 在 2021/9/4 上午12:29, Jens Axboe 写道:
>>>>>>> On 9/3/21 5:00 AM, Hao Xu wrote:
>>>>>>>> Update io_accept_prep() to enable multishot mode for accept operation.
>>>>>>>>
>>>>>>>> Signed-off-by: Hao Xu <haoxu@linux.alibaba.com>
>>>>>>>> ---
>>>>>>>>    fs/io_uring.c | 14 ++++++++++++--
>>>>>>>>    1 file changed, 12 insertions(+), 2 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>>>>>>>> index eb81d37dce78..34612646ae3c 100644
>>>>>>>> --- a/fs/io_uring.c
>>>>>>>> +++ b/fs/io_uring.c
>>>>>>>> @@ -4861,6 +4861,7 @@ static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
>>>>>>>>    static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
>>>>>>>>    {
>>>>>>>>        struct io_accept *accept = &req->accept;
>>>>>>>> +    bool is_multishot;
>>>>>>>>           if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
>>>>>>>>            return -EINVAL;
>>>>>>>> @@ -4872,14 +4873,23 @@ static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
>>>>>>>>        accept->flags = READ_ONCE(sqe->accept_flags);
>>>>>>>>        accept->nofile = rlimit(RLIMIT_NOFILE);
>>>>>>>>    +    is_multishot = accept->flags & IORING_ACCEPT_MULTISHOT;
>>>>>>>> +    if (is_multishot && (req->flags & REQ_F_FORCE_ASYNC))
>>>>>>>> +        return -EINVAL;
>>>>>>>
>>>>>>> I like the idea itself as I think it makes a lot of sense to just have
>>>>>>> an accept sitting there and generating multiple CQEs, but I'm a bit
>>>>>>> puzzled by how you pass it in. accept->flags is the accept4(2) flags,
>>>>>>> which can currently be:
>>>>>>>
>>>>>>> SOCK_NONBLOCK
>>>>>>> SOCK_CLOEXEC
>>>>>>>
>>>>>>> While there's not any overlap here, that is mostly by chance I think. A
>>>>>>> cleaner separation is needed here, what happens if some other accept4(2)
>>>>>>> flag is enabled and it just happens to be the same as
>>>>>>> IORING_ACCEPT_MULTISHOT?
>>>>>> Make sense, how about a new IOSQE flag, I saw not many
>>>>>> entries left there.
>>>>>
>>>>> Not quite sure what the best approach would be... The mshot flag only
>>>>> makes sense for a few request types, so a bit of a shame to have to
>>>>> waste an IOSQE flag on it. Especially when the flags otherwise passed in
>>>>> are so sparse, there's plenty of bits there.
>>>>>
>>>>> Hence while it may not be the prettiest, perhaps using accept->flags is
>>>>> ok and we just need some careful code to ensure that we never have any
>>>>> overlap.
>>>>
>>>> Or we can alias with some of the almost-never-used fields like
>>>> ->ioprio or ->buf_index.
>>>
>>> It's not a bad idea, as long as we can safely use flags from eg ioprio
>>> for cases where ioprio would never be used. In that sense it's probably
>>> safer than using buf_index.
>>>
>>> The alternative is, as has been brougt up before, adding a flags2 and
>>> reserving the last flag in ->flags to say "there are flags in flags2".
>>> Not exactly super pretty either, but we'll need to extend them at some
>>> point.
>> I'm going to do it in this way, there is another thing we have to do:
>> extend req->flags too, since flags we already used > 32 if we add
>> sqe->ext_flags
> 
> We still have 2 bits left, and IIRC you wanted to take only 1 of them.
> We don't need extending it at the moment, it sounded to me like a plan
> for the future. No extra trouble for now

Right, and it should be a separate thing anyway. But as you say, there's
still 2 bits left, this is more about longer term planning than this
particular patchset.

> Anyway, I can't think of many requests working in this mode, and I think
> sqe_flags should be taken only for features applicable to all (~most) of
> requests. Maybe we'd better to fit it individually into accept in the
> end? Sounds more plausible tbh

That's why I suggested making it op private instead, I don't
particularly like having io_uring wide flags that are only applicable to
a (very) small subset of requests. And there's also precedence here
already in terms of poll supporting mshot with a private flag.

-- 
Jens Axboe