On 24/02/2020 21:33, Jens Axboe wrote:
> On 2/24/20 11:26 AM, Pavel Begunkov wrote:
>> On 24/02/2020 21:23, Jens Axboe wrote:
>>> On 2/24/20 10:55 AM, Pavel Begunkov wrote:
>>>> +static int copy_single(struct io_uring *ring,
>>>> +			int fd_in, loff_t off_in,
>>>> +			int fd_out, loff_t off_out,
>>>> +			int pipe_fds[2],
>>>> +			unsigned int len,
>>>> +			unsigned flags1, unsigned flags2)
>>>> +{
>>>> +	struct io_uring_cqe *cqe;
>>>> +	struct io_uring_sqe *sqe;
>>>> +	int i, ret = -1;
>>>> +
>>>> +	sqe = io_uring_get_sqe(ring);
>>>> +	if (!sqe) {
>>>> +		fprintf(stderr, "get sqe failed\n");
>>>> +		return -1;
>>>> +	}
>>>> +	io_uring_prep_splice(sqe, fd_in, off_in, pipe_fds[1], -1,
>>>> +			     len, flags1);
>>>> +	sqe->flags = IOSQE_IO_LINK;
>>>> +
>>>> +	sqe = io_uring_get_sqe(ring);
>>>> +	if (!sqe) {
>>>> +		fprintf(stderr, "get sqe failed\n");
>>>> +		return -1;
>>>> +	}
>>>> +	io_uring_prep_splice(sqe, pipe_fds[0], -1, fd_out, off_out,
>>>> +			     len, flags2);
>>>> +
>>>> +	ret = io_uring_submit(ring);
>>>> +	if (ret <= 1) {
>>>> +		fprintf(stderr, "sqe submit failed: %d\n", ret);
>>>> +		return -1;
>>>> +	}
>>>
>>> This seems wrong, you prep one and submit, the right return value would
>>> be 1. This check should be < 1, not <= 1. I'll make the change, rest
>>> looks good to me. Thanks!
>>>
>>
>> There are 2 sqes, "fd_in -> pipe" and "pipe -> fd_out".
> 
> I must be blind... I guess the failure case is that it doesn't work so well
> on the kernels not supporting splice, as we'll return 1 there as the first

I've wanted for long to kill this weird behaviour, it should consume the whole
link. Can't imagine any userspace app handling all edge-case errors right...

> submit fails. I'll clean up that bit.

...I should have tested better. Thanks!

-- 
Pavel Begunkov