xref: /linux/net/rds/tcp_connect.c (revision 1c4b5ecb7ea190fa3e9f9d6891e6c90b60e04f24)
1 /*
2  * Copyright (c) 2006, 2017 Oracle and/or its affiliates. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  *
32  */
33 #include <linux/kernel.h>
34 #include <linux/in.h>
35 #include <net/tcp.h>
36 
37 #include "rds.h"
38 #include "tcp.h"
39 
40 void rds_tcp_state_change(struct sock *sk)
41 {
42 	void (*state_change)(struct sock *sk);
43 	struct rds_conn_path *cp;
44 	struct rds_tcp_connection *tc;
45 
46 	read_lock_bh(&sk->sk_callback_lock);
47 	cp = sk->sk_user_data;
48 	if (!cp) {
49 		state_change = sk->sk_state_change;
50 		goto out;
51 	}
52 	tc = cp->cp_transport_data;
53 	state_change = tc->t_orig_state_change;
54 
55 	rdsdebug("sock %p state_change to %d\n", tc->t_sock, sk->sk_state);
56 
57 	switch (sk->sk_state) {
58 	/* ignore connecting sockets as they make progress */
59 	case TCP_SYN_SENT:
60 	case TCP_SYN_RECV:
61 		break;
62 	case TCP_ESTABLISHED:
63 		/* Force the peer to reconnect so that we have the
64 		 * TCP ports going from <smaller-ip>.<transient> to
65 		 * <larger-ip>.<RDS_TCP_PORT>. We avoid marking the
66 		 * RDS connection as RDS_CONN_UP until the reconnect,
67 		 * to avoid RDS datagram loss.
68 		 */
69 		if (rds_addr_cmp(&cp->cp_conn->c_laddr,
70 				 &cp->cp_conn->c_faddr) >= 0 &&
71 		    rds_conn_path_transition(cp, RDS_CONN_CONNECTING,
72 					     RDS_CONN_ERROR)) {
73 			rds_conn_path_drop(cp, false);
74 		} else {
75 			rds_connect_path_complete(cp, RDS_CONN_CONNECTING);
76 		}
77 		break;
78 	case TCP_CLOSE_WAIT:
79 	case TCP_CLOSE:
80 		rds_conn_path_drop(cp, false);
81 		break;
82 	default:
83 		break;
84 	}
85 out:
86 	read_unlock_bh(&sk->sk_callback_lock);
87 	state_change(sk);
88 }
89 
90 int rds_tcp_conn_path_connect(struct rds_conn_path *cp)
91 {
92 	struct socket *sock = NULL;
93 	struct sockaddr_in6 sin6;
94 	struct sockaddr_in sin;
95 	struct sockaddr *addr;
96 	int addrlen;
97 	bool isv6;
98 	int ret;
99 	struct rds_connection *conn = cp->cp_conn;
100 	struct rds_tcp_connection *tc = cp->cp_transport_data;
101 
102 	/* for multipath rds,we only trigger the connection after
103 	 * the handshake probe has determined the number of paths.
104 	 */
105 	if (cp->cp_index > 0 && cp->cp_conn->c_npaths < 2)
106 		return -EAGAIN;
107 
108 	mutex_lock(&tc->t_conn_path_lock);
109 
110 	if (rds_conn_path_up(cp)) {
111 		mutex_unlock(&tc->t_conn_path_lock);
112 		return 0;
113 	}
114 	if (ipv6_addr_v4mapped(&conn->c_laddr)) {
115 		ret = sock_create_kern(rds_conn_net(conn), PF_INET,
116 				       SOCK_STREAM, IPPROTO_TCP, &sock);
117 		isv6 = false;
118 	} else {
119 		ret = sock_create_kern(rds_conn_net(conn), PF_INET6,
120 				       SOCK_STREAM, IPPROTO_TCP, &sock);
121 		isv6 = true;
122 	}
123 
124 	if (ret < 0)
125 		goto out;
126 
127 	rds_tcp_tune(sock);
128 
129 	if (isv6) {
130 		sin6.sin6_family = AF_INET6;
131 		sin6.sin6_addr = conn->c_laddr;
132 		sin6.sin6_port = 0;
133 		sin6.sin6_flowinfo = 0;
134 		sin6.sin6_scope_id = conn->c_dev_if;
135 		addr = (struct sockaddr *)&sin6;
136 		addrlen = sizeof(sin6);
137 	} else {
138 		sin.sin_family = AF_INET;
139 		sin.sin_addr.s_addr = conn->c_laddr.s6_addr32[3];
140 		sin.sin_port = 0;
141 		addr = (struct sockaddr *)&sin;
142 		addrlen = sizeof(sin);
143 	}
144 
145 	ret = sock->ops->bind(sock, addr, addrlen);
146 	if (ret) {
147 		rdsdebug("bind failed with %d at address %pI6c\n",
148 			 ret, &conn->c_laddr);
149 		goto out;
150 	}
151 
152 	if (isv6) {
153 		sin6.sin6_family = AF_INET6;
154 		sin6.sin6_addr = conn->c_faddr;
155 		sin6.sin6_port = htons(RDS_TCP_PORT);
156 		sin6.sin6_flowinfo = 0;
157 		sin6.sin6_scope_id = conn->c_dev_if;
158 		addr = (struct sockaddr *)&sin6;
159 		addrlen = sizeof(sin6);
160 	} else {
161 		sin.sin_family = AF_INET;
162 		sin.sin_addr.s_addr = conn->c_faddr.s6_addr32[3];
163 		sin.sin_port = htons(RDS_TCP_PORT);
164 		addr = (struct sockaddr *)&sin;
165 		addrlen = sizeof(sin);
166 	}
167 
168 	/*
169 	 * once we call connect() we can start getting callbacks and they
170 	 * own the socket
171 	 */
172 	rds_tcp_set_callbacks(sock, cp);
173 	ret = sock->ops->connect(sock, addr, addrlen, O_NONBLOCK);
174 
175 	rdsdebug("connect to address %pI6c returned %d\n", &conn->c_faddr, ret);
176 	if (ret == -EINPROGRESS)
177 		ret = 0;
178 	if (ret == 0) {
179 		rds_tcp_keepalive(sock);
180 		sock = NULL;
181 	} else {
182 		rds_tcp_restore_callbacks(sock, cp->cp_transport_data);
183 	}
184 
185 out:
186 	mutex_unlock(&tc->t_conn_path_lock);
187 	if (sock)
188 		sock_release(sock);
189 	return ret;
190 }
191 
192 /*
193  * Before killing the tcp socket this needs to serialize with callbacks.  The
194  * caller has already grabbed the sending sem so we're serialized with other
195  * senders.
196  *
197  * TCP calls the callbacks with the sock lock so we hold it while we reset the
198  * callbacks to those set by TCP.  Our callbacks won't execute again once we
199  * hold the sock lock.
200  */
201 void rds_tcp_conn_path_shutdown(struct rds_conn_path *cp)
202 {
203 	struct rds_tcp_connection *tc = cp->cp_transport_data;
204 	struct socket *sock = tc->t_sock;
205 
206 	rdsdebug("shutting down conn %p tc %p sock %p\n",
207 		 cp->cp_conn, tc, sock);
208 
209 	if (sock) {
210 		if (rds_destroy_pending(cp->cp_conn))
211 			sock_no_linger(sock->sk);
212 		sock->ops->shutdown(sock, RCV_SHUTDOWN | SEND_SHUTDOWN);
213 		lock_sock(sock->sk);
214 		rds_tcp_restore_callbacks(sock, tc); /* tc->tc_sock = NULL */
215 
216 		release_sock(sock->sk);
217 		sock_release(sock);
218 	}
219 
220 	if (tc->t_tinc) {
221 		rds_inc_put(&tc->t_tinc->ti_inc);
222 		tc->t_tinc = NULL;
223 	}
224 	tc->t_tinc_hdr_rem = sizeof(struct rds_header);
225 	tc->t_tinc_data_rem = 0;
226 }
227