Re: [PATCH] io_uring: uring_cmd: add multishot support without poll

[Jens Axboe <axboe@kernel.dk>]

On 8/19/25 8:40 AM, Ming Lei wrote:
> On Tue, Aug 19, 2025 at 08:31:32AM -0600, Jens Axboe wrote:
>> On 8/19/25 8:28 AM, Ming Lei wrote:
>>> On Tue, Aug 19, 2025 at 08:01:18AM -0600, Jens Axboe wrote:
>>>> On 8/19/25 5:19 AM, Ming Lei wrote:
>>>>>>> @@ -251,6 +264,11 @@ int io_uring_cmd(struct io_kiocb *req, unsigned int issue_flags)
>>>>>>>  	}
>>>>>>>  
>>>>>>>  	ret = file->f_op->uring_cmd(ioucmd, issue_flags);
>>>>>>> +	if (ioucmd->flags & IORING_URING_CMD_MULTISHOT) {
>>>>>>> +		if (ret >= 0)
>>>>>>> +			return IOU_ISSUE_SKIP_COMPLETE;
>>>>>>> +		io_kbuf_recycle(req, issue_flags);
>>>>>>> +	}
>>>>>>>  	if (ret == -EAGAIN) {
>>>>>>>  		ioucmd->flags |= IORING_URING_CMD_REISSUE;
>>>>>>>  		return ret;
>>>>>>
>>>>>> Missing recycle for -EAGAIN?
>>>>>
>>>>> io_kbuf_recycle() is done above if `ret < 0`
>>>>
>>>> Inside the multishot case. I don't see anywhere where it's forbidden to
>>>> use IOSQE_BUFFER_SELECT without having multishot set? Either that needs
>>>
>>> REQ_F_BUFFER_SELECT is supposed to be allowed for IORING_URING_CMD_MULTISHOT
>>> only, and it is checked in io_uring_cmd_prep().
>>>
>>>> to be explicit for now, or the recycling should happen generically.
>>>> Probably the former I would suspect.
>>>
>>> Yes, the former is exactly what the patch is doing.
>>
>> Is it? Because looking at v2, you check if IORING_URING_CMD_FIXED is
>> set, and you fail for that case if REQ_F_BUFFER_SELECT is set. Then you
>> have a IORING_URING_CMD_MULTISHOT where the opposite is true, which
>> obviously makes sense.
>>
>> But no checks if neither is set?
> 
> Indeed, thanks for the catch, and the REQ_F_BUFFER_SELECT check in IORING_URING_CMD_FIXED
> branch can be moved to the branch of !IORING_URING_CMD_MULTISHOT.
> 
>>
>> You could add that in io_uring_cmd_select_buffer(), eg fail if
>> IORING_URING_CMD_MULTISHOT isn't set. Which if done, then the prep side
>> checking could probably just go away.
> 
> Looks this way is good too.

Want to spin a v3 for this?

-- 
Jens Axboe