Re: [PATCH net-next v15 11/22] ovpn: implement TCP transport

On 16/12/2024 14:59, Sabrina Dubroca wrote:

2024-12-11, 22:15:15 +0100, Antonio Quartulli wrote:

quoted

@@ -42,6 +56,31 @@ struct ovpn_peer {
  		struct in6_addr ipv6;
  	} vpn_addrs;
  	struct ovpn_socket *sock;
+
+	/* state of the TCP reading. Needed to keep track of how much of a
+	 * single packet has already been read from the stream and how much is
+	 * missing
+	 */

nit: not so accurate since the switch to strp, can probably be dropped
since @tcp has a kdoc entry

right - dropping it.

quoted

+	struct {
+		struct strparser strp;
+		struct work_struct tx_work;
+		struct sk_buff_head user_queue;
+		struct sk_buff_head out_queue;
+		bool tx_in_progress;
+
+		struct {
+			struct sk_buff *skb;
+			int offset;
+			int len;
+		} out_msg;
+
+		struct {
+			void (*sk_data_ready)(struct sock *sk);
+			void (*sk_write_space)(struct sock *sk);
+			struct proto *prot;
+			const struct proto_ops *ops;
+		} sk_cb;
+	} tcp;

[...]

quoted

+static void ovpn_tcp_send_sock_skb(struct ovpn_peer *peer, struct sk_buff *skb)
+{
+	if (peer->tcp.out_msg.skb)
+		ovpn_tcp_send_sock(peer);
+
+	if (peer->tcp.out_msg.skb) {
+		dev_core_stats_rx_dropped_inc(peer->ovpn->dev);

tx_dropped?

ACK

quoted

+		kfree_skb(skb);
+		return;
+	}
+
+	peer->tcp.out_msg.skb = skb;
+	peer->tcp.out_msg.len = skb->len;
+	peer->tcp.out_msg.offset = 0;
+	ovpn_tcp_send_sock(peer);
+}
+
+void ovpn_tcp_send_skb(struct ovpn_peer *peer, struct sk_buff *skb)
+{
+	u16 len = skb->len;
+
+	*(__be16 *)__skb_push(skb, sizeof(u16)) = htons(len);
+
+	bh_lock_sock(peer->sock->sock->sk);
+	if (sock_owned_by_user(peer->sock->sock->sk)) {
+		if (skb_queue_len(&peer->tcp.out_queue) >=
+		    READ_ONCE(net_hotdata.max_backlog)) {
+			dev_core_stats_rx_dropped_inc(peer->ovpn->dev);

tx_dropped?

ACK

quoted

+			kfree_skb(skb);
+			goto unlock;
+		}
+		__skb_queue_tail(&peer->tcp.out_queue, skb);
+	} else {
+		ovpn_tcp_send_sock_skb(peer, skb);
+	}
+unlock:
+	bh_unlock_sock(peer->sock->sock->sk);
+}

[...]

quoted

+static void ovpn_tcp_close(struct sock *sk, long timeout)
+{
+	struct ovpn_socket *sock;
+
+	rcu_read_lock();

[can't sleep until unlock]

quoted

+	sock = rcu_dereference_sk_user_data(sk);
+
+	strp_stop(&sock->peer->tcp.strp);
+
+	tcp_close(sk, timeout);

     void tcp_close(struct sock *sk, long timeout)
     {
     	lock_sock(sk);

but this can sleep.

Ouch.. I wonder why I have never seen the might_sleep() trigger this, but probably that's due to the fact that we hardly hit this cb in the classic use case.

Is there anything that prevents delaying tcp_close until after
ovpn_peer_del and rcu_read_unlock?

not really.

quoted

+	ovpn_peer_del(sock->peer, OVPN_DEL_PEER_REASON_TRANSPORT_ERROR);
+	rcu_read_unlock();

I will move the tcp_close() here.

quoted

+}

[...]

quoted

+void __init ovpn_tcp_init(void)
+{
+	ovpn_tcp_build_protos(&ovpn_tcp_prot, &ovpn_tcp_ops, &tcp_prot,
+			      &inet_stream_ops);
+
+#if IS_ENABLED(CONFIG_IPV6)
+	ovpn_tcp_build_protos(&ovpn_tcp6_prot, &ovpn_tcp6_ops, &tcpv6_prot,
+			      &inet6_stream_ops);

I don't think that works for CONFIG_OVPN=y and CONFIG_IPV6=m. You can
either go back to the ugly thing espintcp and tls do, or use the
traditional Kconfig hack:

	depends on IPV6 || !IPV6

(you can find it sprinkled in various places of drivers/net/Kconfig
and net/)

I'll go for the Kconfig hack. Hopefully one day IPV6 will become bool..

Thanks!

Regards,


-- 
Antonio Quartulli
OpenVPN Inc.