1 /*
2 * Copyright (c) 2006, 2018 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/gfp.h>
35 #include <linux/in.h>
36 #include <net/tcp.h>
37
38 #include "rds.h"
39 #include "tcp.h"
40
rds_tcp_keepalive(struct socket * sock)41 int rds_tcp_keepalive(struct socket *sock)
42 {
43 /* values below based on xs_udp_default_timeout */
44 int keepidle = 5; /* send a probe 'keepidle' secs after last data */
45 int keepcnt = 5; /* number of unack'ed probes before declaring dead */
46 int keepalive = 1;
47 int ret = 0;
48
49 ret = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
50 (char *)&keepalive, sizeof(keepalive));
51 if (ret < 0)
52 goto bail;
53
54 ret = kernel_setsockopt(sock, IPPROTO_TCP, TCP_KEEPCNT,
55 (char *)&keepcnt, sizeof(keepcnt));
56 if (ret < 0)
57 goto bail;
58
59 ret = kernel_setsockopt(sock, IPPROTO_TCP, TCP_KEEPIDLE,
60 (char *)&keepidle, sizeof(keepidle));
61 if (ret < 0)
62 goto bail;
63
64 /* KEEPINTVL is the interval between successive probes. We follow
65 * the model in xs_tcp_finish_connecting() and re-use keepidle.
66 */
67 ret = kernel_setsockopt(sock, IPPROTO_TCP, TCP_KEEPINTVL,
68 (char *)&keepidle, sizeof(keepidle));
69 bail:
70 return ret;
71 }
72
73 /* rds_tcp_accept_one_path(): if accepting on cp_index > 0, make sure the
74 * client's ipaddr < server's ipaddr. Otherwise, close the accepted
75 * socket and force a reconneect from smaller -> larger ip addr. The reason
76 * we special case cp_index 0 is to allow the rds probe ping itself to itself
77 * get through efficiently.
78 * Since reconnects are only initiated from the node with the numerically
79 * smaller ip address, we recycle conns in RDS_CONN_ERROR on the passive side
80 * by moving them to CONNECTING in this function.
81 */
82 static
rds_tcp_accept_one_path(struct rds_connection * conn)83 struct rds_tcp_connection *rds_tcp_accept_one_path(struct rds_connection *conn)
84 {
85 int i;
86 int npaths = max_t(int, 1, conn->c_npaths);
87
88 /* for mprds, all paths MUST be initiated by the peer
89 * with the smaller address.
90 */
91 if (rds_addr_cmp(&conn->c_faddr, &conn->c_laddr) >= 0) {
92 /* Make sure we initiate at least one path if this
93 * has not already been done; rds_start_mprds() will
94 * take care of additional paths, if necessary.
95 */
96 if (npaths == 1)
97 rds_conn_path_connect_if_down(&conn->c_path[0]);
98 return NULL;
99 }
100
101 for (i = 0; i < npaths; i++) {
102 struct rds_conn_path *cp = &conn->c_path[i];
103
104 if (rds_conn_path_transition(cp, RDS_CONN_DOWN,
105 RDS_CONN_CONNECTING) ||
106 rds_conn_path_transition(cp, RDS_CONN_ERROR,
107 RDS_CONN_CONNECTING)) {
108 return cp->cp_transport_data;
109 }
110 }
111 return NULL;
112 }
113
rds_tcp_set_linger(struct socket * sock)114 void rds_tcp_set_linger(struct socket *sock)
115 {
116 struct linger no_linger = {
117 .l_onoff = 1,
118 .l_linger = 0,
119 };
120
121 kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER,
122 (char *)&no_linger, sizeof(no_linger));
123 }
124
rds_tcp_accept_one(struct socket * sock)125 int rds_tcp_accept_one(struct socket *sock)
126 {
127 struct socket *new_sock = NULL;
128 struct rds_connection *conn;
129 int ret;
130 struct inet_sock *inet;
131 struct rds_tcp_connection *rs_tcp = NULL;
132 int conn_state;
133 struct rds_conn_path *cp;
134 struct in6_addr *my_addr, *peer_addr;
135 #if !IS_ENABLED(CONFIG_IPV6)
136 struct in6_addr saddr, daddr;
137 #endif
138 int dev_if = 0;
139
140 if (!sock) /* module unload or netns delete in progress */
141 return -ENETUNREACH;
142
143 ret = sock_create_lite(sock->sk->sk_family,
144 sock->sk->sk_type, sock->sk->sk_protocol,
145 &new_sock);
146 if (ret)
147 goto out;
148
149 ret = sock->ops->accept(sock, new_sock, O_NONBLOCK, true);
150 if (ret < 0)
151 goto out;
152
153 /* sock_create_lite() does not get a hold on the owner module so we
154 * need to do it here. Note that sock_release() uses sock->ops to
155 * determine if it needs to decrement the reference count. So set
156 * sock->ops after calling accept() in case that fails. And there's
157 * no need to do try_module_get() as the listener should have a hold
158 * already.
159 */
160 new_sock->ops = sock->ops;
161 __module_get(new_sock->ops->owner);
162
163 ret = rds_tcp_keepalive(new_sock);
164 if (ret < 0)
165 goto out;
166
167 rds_tcp_tune(new_sock);
168
169 inet = inet_sk(new_sock->sk);
170
171 #if IS_ENABLED(CONFIG_IPV6)
172 my_addr = &new_sock->sk->sk_v6_rcv_saddr;
173 peer_addr = &new_sock->sk->sk_v6_daddr;
174 #else
175 ipv6_addr_set_v4mapped(inet->inet_saddr, &saddr);
176 ipv6_addr_set_v4mapped(inet->inet_daddr, &daddr);
177 my_addr = &saddr;
178 peer_addr = &daddr;
179 #endif
180 rdsdebug("accepted family %d tcp %pI6c:%u -> %pI6c:%u\n",
181 sock->sk->sk_family,
182 my_addr, ntohs(inet->inet_sport),
183 peer_addr, ntohs(inet->inet_dport));
184
185 #if IS_ENABLED(CONFIG_IPV6)
186 /* sk_bound_dev_if is not set if the peer address is not link local
187 * address. In this case, it happens that mcast_oif is set. So
188 * just use it.
189 */
190 if ((ipv6_addr_type(my_addr) & IPV6_ADDR_LINKLOCAL) &&
191 !(ipv6_addr_type(peer_addr) & IPV6_ADDR_LINKLOCAL)) {
192 struct ipv6_pinfo *inet6;
193
194 inet6 = inet6_sk(new_sock->sk);
195 dev_if = inet6->mcast_oif;
196 } else {
197 dev_if = new_sock->sk->sk_bound_dev_if;
198 }
199 #endif
200
201 if (!rds_tcp_laddr_check(sock_net(sock->sk), peer_addr, dev_if)) {
202 /* local address connection is only allowed via loopback */
203 ret = -EOPNOTSUPP;
204 goto out;
205 }
206
207 conn = rds_conn_create(sock_net(sock->sk),
208 my_addr, peer_addr,
209 &rds_tcp_transport, 0, GFP_KERNEL, dev_if);
210
211 if (IS_ERR(conn)) {
212 ret = PTR_ERR(conn);
213 goto out;
214 }
215 /* An incoming SYN request came in, and TCP just accepted it.
216 *
217 * If the client reboots, this conn will need to be cleaned up.
218 * rds_tcp_state_change() will do that cleanup
219 */
220 rs_tcp = rds_tcp_accept_one_path(conn);
221 if (!rs_tcp)
222 goto rst_nsk;
223 mutex_lock(&rs_tcp->t_conn_path_lock);
224 cp = rs_tcp->t_cpath;
225 conn_state = rds_conn_path_state(cp);
226 WARN_ON(conn_state == RDS_CONN_UP);
227 if (conn_state != RDS_CONN_CONNECTING && conn_state != RDS_CONN_ERROR)
228 goto rst_nsk;
229 if (rs_tcp->t_sock) {
230 /* Duelling SYN has been handled in rds_tcp_accept_one() */
231 rds_tcp_reset_callbacks(new_sock, cp);
232 /* rds_connect_path_complete() marks RDS_CONN_UP */
233 rds_connect_path_complete(cp, RDS_CONN_RESETTING);
234 } else {
235 rds_tcp_set_callbacks(new_sock, cp);
236 rds_connect_path_complete(cp, RDS_CONN_CONNECTING);
237 }
238 new_sock = NULL;
239 ret = 0;
240 if (conn->c_npaths == 0)
241 rds_send_ping(cp->cp_conn, cp->cp_index);
242 goto out;
243 rst_nsk:
244 /* reset the newly returned accept sock and bail.
245 * It is safe to set linger on new_sock because the RDS connection
246 * has not been brought up on new_sock, so no RDS-level data could
247 * be pending on it. By setting linger, we achieve the side-effect
248 * of avoiding TIME_WAIT state on new_sock.
249 */
250 rds_tcp_set_linger(new_sock);
251 kernel_sock_shutdown(new_sock, SHUT_RDWR);
252 ret = 0;
253 out:
254 if (rs_tcp)
255 mutex_unlock(&rs_tcp->t_conn_path_lock);
256 if (new_sock)
257 sock_release(new_sock);
258 return ret;
259 }
260
rds_tcp_listen_data_ready(struct sock * sk)261 void rds_tcp_listen_data_ready(struct sock *sk)
262 {
263 void (*ready)(struct sock *sk);
264
265 rdsdebug("listen data ready sk %p\n", sk);
266
267 read_lock_bh(&sk->sk_callback_lock);
268 ready = sk->sk_user_data;
269 if (!ready) { /* check for teardown race */
270 ready = sk->sk_data_ready;
271 goto out;
272 }
273
274 /*
275 * ->sk_data_ready is also called for a newly established child socket
276 * before it has been accepted and the accepter has set up their
277 * data_ready.. we only want to queue listen work for our listening
278 * socket
279 *
280 * (*ready)() may be null if we are racing with netns delete, and
281 * the listen socket is being torn down.
282 */
283 if (sk->sk_state == TCP_LISTEN)
284 rds_tcp_accept_work(sk);
285 else
286 ready = rds_tcp_listen_sock_def_readable(sock_net(sk));
287
288 out:
289 read_unlock_bh(&sk->sk_callback_lock);
290 if (ready)
291 ready(sk);
292 }
293
rds_tcp_listen_init(struct net * net,bool isv6)294 struct socket *rds_tcp_listen_init(struct net *net, bool isv6)
295 {
296 struct socket *sock = NULL;
297 struct sockaddr_storage ss;
298 struct sockaddr_in6 *sin6;
299 struct sockaddr_in *sin;
300 int addr_len;
301 int ret;
302
303 ret = sock_create_kern(net, isv6 ? PF_INET6 : PF_INET, SOCK_STREAM,
304 IPPROTO_TCP, &sock);
305 if (ret < 0) {
306 rdsdebug("could not create %s listener socket: %d\n",
307 isv6 ? "IPv6" : "IPv4", ret);
308 goto out;
309 }
310
311 sock->sk->sk_reuse = SK_CAN_REUSE;
312 rds_tcp_nonagle(sock);
313
314 write_lock_bh(&sock->sk->sk_callback_lock);
315 sock->sk->sk_user_data = sock->sk->sk_data_ready;
316 sock->sk->sk_data_ready = rds_tcp_listen_data_ready;
317 write_unlock_bh(&sock->sk->sk_callback_lock);
318
319 if (isv6) {
320 sin6 = (struct sockaddr_in6 *)&ss;
321 sin6->sin6_family = PF_INET6;
322 sin6->sin6_addr = in6addr_any;
323 sin6->sin6_port = (__force u16)htons(RDS_TCP_PORT);
324 sin6->sin6_scope_id = 0;
325 sin6->sin6_flowinfo = 0;
326 addr_len = sizeof(*sin6);
327 } else {
328 sin = (struct sockaddr_in *)&ss;
329 sin->sin_family = PF_INET;
330 sin->sin_addr.s_addr = INADDR_ANY;
331 sin->sin_port = (__force u16)htons(RDS_TCP_PORT);
332 addr_len = sizeof(*sin);
333 }
334
335 ret = kernel_bind(sock, (struct sockaddr *)&ss, addr_len);
336 if (ret < 0) {
337 rdsdebug("could not bind %s listener socket: %d\n",
338 isv6 ? "IPv6" : "IPv4", ret);
339 goto out;
340 }
341
342 ret = sock->ops->listen(sock, 64);
343 if (ret < 0)
344 goto out;
345
346 return sock;
347 out:
348 if (sock)
349 sock_release(sock);
350 return NULL;
351 }
352
rds_tcp_listen_stop(struct socket * sock,struct work_struct * acceptor)353 void rds_tcp_listen_stop(struct socket *sock, struct work_struct *acceptor)
354 {
355 struct sock *sk;
356
357 if (!sock)
358 return;
359
360 sk = sock->sk;
361
362 /* serialize with and prevent further callbacks */
363 lock_sock(sk);
364 write_lock_bh(&sk->sk_callback_lock);
365 if (sk->sk_user_data) {
366 sk->sk_data_ready = sk->sk_user_data;
367 sk->sk_user_data = NULL;
368 }
369 write_unlock_bh(&sk->sk_callback_lock);
370 release_sock(sk);
371
372 /* wait for accepts to stop and close the socket */
373 flush_workqueue(rds_wq);
374 flush_work(acceptor);
375 sock_release(sock);
376 }
377