Re: [PATCH net-next 1/2] vsock: report pending receive data to io_uring

[Stefano Garzarella <sgarzare@redhat.com>]

On Sun, Sep 20, 2026 at 08:47:09PM +0100, David Carlier wrote:
>AF_VSOCK stream receives never fill msghdr.msg_inq, so io_uring cannot set
>IORING_CQE_F_SOCK_NONEMPTY and retries a multishot receive even after the
>queue has been drained.
>
>Fill the hint at the common receive exit using the transport callback that
>SIOCINQ already uses, and report 1 once the connection is finished so the
>caller performs the receive which observes EOF, as TCP does after a FIN.
>
>A vsock loopback ping-pong with io_uring multishot receive drops entries
>into __vsock_connectible_recvmsg from 1.97 to 1.00 per delivered message,
>and receiver CPU time by about 3% (25 runs of 50000 messages, p=0.006).
>
>Signed-off-by: David Carlier <devnexen@gmail.com>
>---
> net/vmw_vsock/af_vsock.c | 39 +++++++++++++++++++++++++++++++++++++++
> 1 file changed, 39 insertions(+)
>
>diff --git a/net/vmw_vsock/af_vsock.c b/net/vmw_vsock/af_vsock.c
>index f840498b58af..63309708c916 100644
>--- a/net/vmw_vsock/af_vsock.c
>+++ b/net/vmw_vsock/af_vsock.c
>@@ -2543,6 +2543,38 @@ static int __vsock_seqpacket_recvmsg(struct sock *sk, struct msghdr *msg,
> 	return err;
> }
>
>+/* Bytes a following receive can consume, 1 if it would only see EOF, or a
>+ * negative value to leave the caller's hint alone.
>+ *
>+ * Called under the socket lock after a nonnegative stream receive, so a NULL
>+ * transport implies SOCK_DONE.
>+ */
>+static int vsock_inq_hint(struct sock *sk)

can we add "stream" in the name so it's clear that is stream only?

(e.g. vsock_stream_inq_hint)

>+{
>+	struct vsock_sock *vsk = vsock_sk(sk);
>+	s64 data;
>+
>+	if ((sk->sk_shutdown & RCV_SHUTDOWN) ||
>+	    (sock_flag(sk, SOCK_DONE) && sk->sk_state != TCP_ESTABLISHED))
>+		return 1;
>+
>+	if (!vsk->transport)
>+		return 1;

Can we squash this in the previous block?

>+
>+	data = vsock_stream_has_data(vsk);
>+	if (data < 0)
>+		return -1;
>+	if (data > 0)
>+		return (int)min_t(s64, data, INT_MAX);

nit: IMO we can remove the int cast.

>+
>+	/* Empty but finished: keep the caller reading so it sees EOF. */
>+	if (sock_flag(sk, SOCK_DONE) ||
>+	    (READ_ONCE(vsk->peer_shutdown) & SEND_SHUTDOWN))
>+		return 1;
>+
>+	return 0;
>+}
>+
> int
> __vsock_connectible_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
> 			    int flags)
>@@ -2606,6 +2638,13 @@ __vsock_connectible_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
> 		err = __vsock_seqpacket_recvmsg(sk, msg, len, flags);
>
> out:
>+	if (msg->msg_get_inq && err >= 0 && sk->sk_type == SOCK_STREAM) {

Why only STREAM?
A comment here and in the commit would be nice.



>+		int inq = vsock_inq_hint(sk);
>+
>+		if (inq >= 0)

Why assigning only >= 0 values?

Thanks,
Stefano


>+			msg->msg_inq = inq;
>+	}
>+
> 	release_sock(sk);
> 	return err;
> }
>-- 
>2.55.0
>