Re: [PATCH v3] io_uring/net: Allow to do vectorized send

[Jens Axboe <axboe@kernel.dk>]

On 7/29/25 12:59 AM, norman.maurer@googlemail.com wrote:
> From: Norman Maurer <norman_maurer@apple.com>
> 
> At the moment you have to use sendmsg for vectorized send.
> While this works it's suboptimal as it also means you need to
> allocate a struct msghdr that needs to be kept alive until a
> submission happens. We can remove this limitation by just
> allowing to use send directly.
> 
> Signed-off-by: Norman Maurer <norman_maurer@apple.com>
> ---
> Changes since v1: Simplify flag check and fix line length of commit message
> Changes since v2: Adjust SENDMSG_FLAGS  
> 
> ---
>  include/uapi/linux/io_uring.h | 4 ++++
>  io_uring/net.c                | 9 ++++++++-
>  2 files changed, 12 insertions(+), 1 deletion(-)
> 
> diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
> index b8a0e70ee2fd..6957dc539d83 100644
> --- a/include/uapi/linux/io_uring.h
> +++ b/include/uapi/linux/io_uring.h
> @@ -392,12 +392,16 @@ enum io_uring_op {
>   *				the starting buffer ID in cqe->flags as per
>   *				usual for provided buffer usage. The buffers
>   *				will be	contiguous from the starting buffer ID.
> + *
> + * IORING_SEND_VECTORIZED	If set, SEND[_ZC] will take a pointer to a io_vec
> + * 				to allow vectorized send operations.
>   */
>  #define IORING_RECVSEND_POLL_FIRST	(1U << 0)
>  #define IORING_RECV_MULTISHOT		(1U << 1)
>  #define IORING_RECVSEND_FIXED_BUF	(1U << 2)
>  #define IORING_SEND_ZC_REPORT_USAGE	(1U << 3)
>  #define IORING_RECVSEND_BUNDLE		(1U << 4)
> +#define IORING_SEND_VECTORIZED		(1U << 5)
>  
>  /*
>   * cqe.res for IORING_CQE_F_NOTIF if
> diff --git a/io_uring/net.c b/io_uring/net.c
> index ba2d0abea349..3ce5478438f0 100644
> --- a/io_uring/net.c
> +++ b/io_uring/net.c
> @@ -382,6 +382,10 @@ static int io_send_setup(struct io_kiocb *req, const struct io_uring_sqe *sqe)
>  	}
>  	if (req->flags & REQ_F_BUFFER_SELECT)
>  		return 0;
> +
> +	if (sr->flags & IORING_SEND_VECTORIZED)
> +               return io_net_import_vec(req, kmsg, sr->buf, sr->len, ITER_SOURCE);
> +
>  	return import_ubuf(ITER_SOURCE, sr->buf, sr->len, &kmsg->msg.msg_iter);
>  }
>  
> @@ -409,7 +413,7 @@ static int io_sendmsg_setup(struct io_kiocb *req, const struct io_uring_sqe *sqe
>  	return io_net_import_vec(req, kmsg, msg.msg_iov, msg.msg_iovlen, ITER_SOURCE);
>  }
>  
> -#define SENDMSG_FLAGS (IORING_RECVSEND_POLL_FIRST | IORING_RECVSEND_BUNDLE)
> +#define SENDMSG_FLAGS (IORING_RECVSEND_POLL_FIRST | IORING_RECVSEND_BUNDLE | IORING_SEND_VECTORIZED)
>  
>  int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
>  {
> @@ -420,6 +424,9 @@ int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
>  	sr->flags = READ_ONCE(sqe->ioprio);
>  	if (sr->flags & ~SENDMSG_FLAGS)
>  		return -EINVAL;
> +	if (req->opcode == IORING_OP_SENDMSG && sr->flags & IORING_SEND_VECTORIZED)
> +		return -EINVAL;
> +
>  	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
>  	if (sr->msg_flags & MSG_DONTWAIT)
>  		req->flags |= REQ_F_NOWAIT;

I think this looks simple enough, but after pondering this a bit, I also
think we can just skip the check right above here. OP_SENDMSG is, by
definition, working on IORING_SEND_VECTORIZED data. Hence returning
-EINVAL from that seems a bit redundant. Maybe just delete this hunk?
What do you think?

No need for v3 for that, I can just edit out that hunk with a note. I'll
run some testing today.

-- 
Jens Axboe