Re: [PATCH net-next v18 12/25] ovpn: implement TCP transport

On 17/01/2025 18:14, Sabrina Dubroca wrote:

2025-01-13, 10:31:31 +0100, Antonio Quartulli wrote:

quoted

+static int ovpn_tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
+			    int flags, int *addr_len)
+{
+	int err = 0, off, copied = 0, ret;
+	struct ovpn_socket *sock;
+	struct ovpn_peer *peer;
+	struct sk_buff *skb;
+
+	rcu_read_lock();
+	sock = rcu_dereference_sk_user_data(sk);
+	if (!sock || !sock->peer) {
+		rcu_read_unlock();
+		return -EBADF;
+	}
+	/* we take a reference to the peer linked to this TCP socket, because
+	 * in turn the peer holds a reference to the socket itself.

Not anymore since v12? [*]

I think it's ok here because we're only using peer and sk (not
anything from ovpn_socket), but it is relevant in _sendmsg, which has
the same peer_hold pattern without this comment.

[*]
v11:
  - https://lore.kernel.org/netdev/20241029-b4-ovpn-v11-8-de4698c73a25@openvpn.net/ (local)
    ovpn_peer_release -> ovpn_socket_put

v12:
  - https://lore.kernel.org/netdev/20241202-b4-ovpn-v12-9-239ff733bf97@openvpn.net/ (local)
    ovpn_peer_release doesn't do ovpn_socket_put

  - https://lore.kernel.org/netdev/20241202-b4-ovpn-v12-7-239ff733bf97@openvpn.net/ (local)
    ovpn_socket_put is done directly at ovpn_peer_remove time, before the final peer_put

quoted

+	 * By doing so we also ensure that the peer stays alive along with
+	 * the socket while executing this function
+	 */
+	ovpn_peer_hold(sock->peer);
+	peer = sock->peer;
+	rcu_read_unlock();
+
+	skb = __skb_recv_datagram(sk, &peer->tcp.user_queue, flags, &off, &err);
+	if (!skb) {
+		if (err == -EAGAIN && sk->sk_shutdown & RCV_SHUTDOWN) {
+			ret = 0;
+			goto out;
+		}
+		ret = err;
+		goto out;
+	}
+
+	copied = len;
+	if (copied > skb->len)
+		copied = skb->len;
+	else if (copied < skb->len)
+		msg->msg_flags |= MSG_TRUNC;
+
+	err = skb_copy_datagram_msg(skb, 0, msg, copied);
+	if (unlikely(err)) {
+		kfree_skb(skb);
+		ret = err;
+		goto out;
+	}
+
+	if (flags & MSG_TRUNC)
+		copied = skb->len;
+	kfree_skb(skb);
+	ret = copied;
+out:
+	ovpn_peer_put(peer);
+	return ret;
+}

[...]

quoted

+static int ovpn_tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
+{
+	struct ovpn_socket *sock;
+	int ret, linear = PAGE_SIZE;
+	struct ovpn_peer *peer;
+	struct sk_buff *skb;
+
+	rcu_read_lock();
+	sock = rcu_dereference_sk_user_data(sk);
+	if (unlikely(!sock || !sock->peer || !ovpn_peer_hold(sock->peer))) {
+		rcu_read_unlock();
+		return -EIO;
+	}
+	peer = sock->peer;
+	rcu_read_unlock();
+
+	lock_sock(peer->sock->sock->sk);

Isn't that just sk?

Right - it's the same object at this point. I'll use sk.

quoted

+
+	if (msg->msg_flags & ~MSG_DONTWAIT) {
+		ret = -EOPNOTSUPP;
+		goto peer_free;
+	}
+
+	if (peer->tcp.out_msg.skb) {
+		ret = -EAGAIN;
+		goto peer_free;
+	}
+
+	if (size < linear)
+		linear = size;
+
+	skb = sock_alloc_send_pskb(sk, linear, size - linear,
+				   msg->msg_flags & MSG_DONTWAIT, &ret, 0);
+	if (!skb) {
+		net_err_ratelimited("%s: skb alloc failed: %d\n",
+				    netdev_name(sock->peer->ovpn->dev), ret);

Since we only have a ref on peer (but not on sock), I'd use
peer->... directly instead of sock->peer.

ACK

quoted

+		goto peer_free;
+	}
+
+	skb_put(skb, linear);
+	skb->len = size;
+	skb->data_len = size - linear;
+
+	ret = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
+	if (ret) {
+		kfree_skb(skb);
+		net_err_ratelimited("%s: skb copy from iter failed: %d\n",
+				    netdev_name(sock->peer->ovpn->dev), ret);

s/sock->//

ACK

quoted

+		goto peer_free;
+	}
+
+	ovpn_tcp_send_sock_skb(sock->peer, skb);

s/sock->//

ACK

quoted

+	ret = size;
+peer_free:
+	release_sock(peer->sock->sock->sk);
+	ovpn_peer_put(peer);
+	return ret;
+}

-- 
Antonio Quartulli
OpenVPN Inc.