Loading...
1/*
2 * Copyright (c) 2006 Oracle. 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
41int 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));
69bail:
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 */
82struct rds_tcp_connection *rds_tcp_accept_one_path(struct rds_connection *conn)
83{
84 int i;
85 bool peer_is_smaller = (conn->c_faddr < conn->c_laddr);
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 (!peer_is_smaller) {
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
114int rds_tcp_accept_one(struct socket *sock)
115{
116 struct socket *new_sock = NULL;
117 struct rds_connection *conn;
118 int ret;
119 struct inet_sock *inet;
120 struct rds_tcp_connection *rs_tcp = NULL;
121 int conn_state;
122 struct rds_conn_path *cp;
123
124 if (!sock) /* module unload or netns delete in progress */
125 return -ENETUNREACH;
126
127 ret = sock_create_kern(sock_net(sock->sk), sock->sk->sk_family,
128 sock->sk->sk_type, sock->sk->sk_protocol,
129 &new_sock);
130 if (ret)
131 goto out;
132
133 new_sock->type = sock->type;
134 new_sock->ops = sock->ops;
135 ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
136 if (ret < 0)
137 goto out;
138
139 ret = rds_tcp_keepalive(new_sock);
140 if (ret < 0)
141 goto out;
142
143 rds_tcp_tune(new_sock);
144
145 inet = inet_sk(new_sock->sk);
146
147 rdsdebug("accepted tcp %pI4:%u -> %pI4:%u\n",
148 &inet->inet_saddr, ntohs(inet->inet_sport),
149 &inet->inet_daddr, ntohs(inet->inet_dport));
150
151 conn = rds_conn_create(sock_net(sock->sk),
152 inet->inet_saddr, inet->inet_daddr,
153 &rds_tcp_transport, GFP_KERNEL);
154 if (IS_ERR(conn)) {
155 ret = PTR_ERR(conn);
156 goto out;
157 }
158 /* An incoming SYN request came in, and TCP just accepted it.
159 *
160 * If the client reboots, this conn will need to be cleaned up.
161 * rds_tcp_state_change() will do that cleanup
162 */
163 rs_tcp = rds_tcp_accept_one_path(conn);
164 if (!rs_tcp)
165 goto rst_nsk;
166 mutex_lock(&rs_tcp->t_conn_path_lock);
167 cp = rs_tcp->t_cpath;
168 conn_state = rds_conn_path_state(cp);
169 WARN_ON(conn_state == RDS_CONN_UP);
170 if (conn_state != RDS_CONN_CONNECTING && conn_state != RDS_CONN_ERROR)
171 goto rst_nsk;
172 if (rs_tcp->t_sock) {
173 /* Need to resolve a duelling SYN between peers.
174 * We have an outstanding SYN to this peer, which may
175 * potentially have transitioned to the RDS_CONN_UP state,
176 * so we must quiesce any send threads before resetting
177 * c_transport_data.
178 */
179 if (ntohl(inet->inet_saddr) < ntohl(inet->inet_daddr) ||
180 !cp->cp_outgoing) {
181 goto rst_nsk;
182 } else {
183 rds_tcp_reset_callbacks(new_sock, cp);
184 cp->cp_outgoing = 0;
185 /* rds_connect_path_complete() marks RDS_CONN_UP */
186 rds_connect_path_complete(cp, RDS_CONN_RESETTING);
187 }
188 } else {
189 rds_tcp_set_callbacks(new_sock, cp);
190 rds_connect_path_complete(cp, RDS_CONN_CONNECTING);
191 }
192 new_sock = NULL;
193 ret = 0;
194 goto out;
195rst_nsk:
196 /* reset the newly returned accept sock and bail */
197 kernel_sock_shutdown(new_sock, SHUT_RDWR);
198 ret = 0;
199out:
200 if (rs_tcp)
201 mutex_unlock(&rs_tcp->t_conn_path_lock);
202 if (new_sock)
203 sock_release(new_sock);
204 return ret;
205}
206
207void rds_tcp_listen_data_ready(struct sock *sk)
208{
209 void (*ready)(struct sock *sk);
210
211 rdsdebug("listen data ready sk %p\n", sk);
212
213 read_lock_bh(&sk->sk_callback_lock);
214 ready = sk->sk_user_data;
215 if (!ready) { /* check for teardown race */
216 ready = sk->sk_data_ready;
217 goto out;
218 }
219
220 /*
221 * ->sk_data_ready is also called for a newly established child socket
222 * before it has been accepted and the accepter has set up their
223 * data_ready.. we only want to queue listen work for our listening
224 * socket
225 */
226 if (sk->sk_state == TCP_LISTEN)
227 rds_tcp_accept_work(sk);
228 else
229 ready = rds_tcp_listen_sock_def_readable(sock_net(sk));
230
231out:
232 read_unlock_bh(&sk->sk_callback_lock);
233 ready(sk);
234}
235
236struct socket *rds_tcp_listen_init(struct net *net)
237{
238 struct sockaddr_in sin;
239 struct socket *sock = NULL;
240 int ret;
241
242 ret = sock_create_kern(net, PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
243 if (ret < 0)
244 goto out;
245
246 sock->sk->sk_reuse = SK_CAN_REUSE;
247 rds_tcp_nonagle(sock);
248
249 write_lock_bh(&sock->sk->sk_callback_lock);
250 sock->sk->sk_user_data = sock->sk->sk_data_ready;
251 sock->sk->sk_data_ready = rds_tcp_listen_data_ready;
252 write_unlock_bh(&sock->sk->sk_callback_lock);
253
254 sin.sin_family = PF_INET;
255 sin.sin_addr.s_addr = (__force u32)htonl(INADDR_ANY);
256 sin.sin_port = (__force u16)htons(RDS_TCP_PORT);
257
258 ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
259 if (ret < 0)
260 goto out;
261
262 ret = sock->ops->listen(sock, 64);
263 if (ret < 0)
264 goto out;
265
266 return sock;
267out:
268 if (sock)
269 sock_release(sock);
270 return NULL;
271}
272
273void rds_tcp_listen_stop(struct socket *sock)
274{
275 struct sock *sk;
276
277 if (!sock)
278 return;
279
280 sk = sock->sk;
281
282 /* serialize with and prevent further callbacks */
283 lock_sock(sk);
284 write_lock_bh(&sk->sk_callback_lock);
285 if (sk->sk_user_data) {
286 sk->sk_data_ready = sk->sk_user_data;
287 sk->sk_user_data = NULL;
288 }
289 write_unlock_bh(&sk->sk_callback_lock);
290 release_sock(sk);
291
292 /* wait for accepts to stop and close the socket */
293 flush_workqueue(rds_wq);
294 sock_release(sock);
295}
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
41void 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
47 sock_set_keepalive(sock->sk);
48 tcp_sock_set_keepcnt(sock->sk, keepcnt);
49 tcp_sock_set_keepidle(sock->sk, keepidle);
50 /* KEEPINTVL is the interval between successive probes. We follow
51 * the model in xs_tcp_finish_connecting() and re-use keepidle.
52 */
53 tcp_sock_set_keepintvl(sock->sk, keepidle);
54}
55
56/* rds_tcp_accept_one_path(): if accepting on cp_index > 0, make sure the
57 * client's ipaddr < server's ipaddr. Otherwise, close the accepted
58 * socket and force a reconneect from smaller -> larger ip addr. The reason
59 * we special case cp_index 0 is to allow the rds probe ping itself to itself
60 * get through efficiently.
61 * Since reconnects are only initiated from the node with the numerically
62 * smaller ip address, we recycle conns in RDS_CONN_ERROR on the passive side
63 * by moving them to CONNECTING in this function.
64 */
65static
66struct rds_tcp_connection *rds_tcp_accept_one_path(struct rds_connection *conn)
67{
68 int i;
69 int npaths = max_t(int, 1, conn->c_npaths);
70
71 /* for mprds, all paths MUST be initiated by the peer
72 * with the smaller address.
73 */
74 if (rds_addr_cmp(&conn->c_faddr, &conn->c_laddr) >= 0) {
75 /* Make sure we initiate at least one path if this
76 * has not already been done; rds_start_mprds() will
77 * take care of additional paths, if necessary.
78 */
79 if (npaths == 1)
80 rds_conn_path_connect_if_down(&conn->c_path[0]);
81 return NULL;
82 }
83
84 for (i = 0; i < npaths; i++) {
85 struct rds_conn_path *cp = &conn->c_path[i];
86
87 if (rds_conn_path_transition(cp, RDS_CONN_DOWN,
88 RDS_CONN_CONNECTING) ||
89 rds_conn_path_transition(cp, RDS_CONN_ERROR,
90 RDS_CONN_CONNECTING)) {
91 return cp->cp_transport_data;
92 }
93 }
94 return NULL;
95}
96
97int rds_tcp_accept_one(struct socket *sock)
98{
99 struct socket *new_sock = NULL;
100 struct rds_connection *conn;
101 int ret;
102 struct inet_sock *inet;
103 struct rds_tcp_connection *rs_tcp = NULL;
104 int conn_state;
105 struct rds_conn_path *cp;
106 struct in6_addr *my_addr, *peer_addr;
107#if !IS_ENABLED(CONFIG_IPV6)
108 struct in6_addr saddr, daddr;
109#endif
110 int dev_if = 0;
111
112 if (!sock) /* module unload or netns delete in progress */
113 return -ENETUNREACH;
114
115 ret = sock_create_lite(sock->sk->sk_family,
116 sock->sk->sk_type, sock->sk->sk_protocol,
117 &new_sock);
118 if (ret)
119 goto out;
120
121 ret = sock->ops->accept(sock, new_sock, O_NONBLOCK, true);
122 if (ret < 0)
123 goto out;
124
125 /* sock_create_lite() does not get a hold on the owner module so we
126 * need to do it here. Note that sock_release() uses sock->ops to
127 * determine if it needs to decrement the reference count. So set
128 * sock->ops after calling accept() in case that fails. And there's
129 * no need to do try_module_get() as the listener should have a hold
130 * already.
131 */
132 new_sock->ops = sock->ops;
133 __module_get(new_sock->ops->owner);
134
135 rds_tcp_keepalive(new_sock);
136 rds_tcp_tune(new_sock);
137
138 inet = inet_sk(new_sock->sk);
139
140#if IS_ENABLED(CONFIG_IPV6)
141 my_addr = &new_sock->sk->sk_v6_rcv_saddr;
142 peer_addr = &new_sock->sk->sk_v6_daddr;
143#else
144 ipv6_addr_set_v4mapped(inet->inet_saddr, &saddr);
145 ipv6_addr_set_v4mapped(inet->inet_daddr, &daddr);
146 my_addr = &saddr;
147 peer_addr = &daddr;
148#endif
149 rdsdebug("accepted family %d tcp %pI6c:%u -> %pI6c:%u\n",
150 sock->sk->sk_family,
151 my_addr, ntohs(inet->inet_sport),
152 peer_addr, ntohs(inet->inet_dport));
153
154#if IS_ENABLED(CONFIG_IPV6)
155 /* sk_bound_dev_if is not set if the peer address is not link local
156 * address. In this case, it happens that mcast_oif is set. So
157 * just use it.
158 */
159 if ((ipv6_addr_type(my_addr) & IPV6_ADDR_LINKLOCAL) &&
160 !(ipv6_addr_type(peer_addr) & IPV6_ADDR_LINKLOCAL)) {
161 struct ipv6_pinfo *inet6;
162
163 inet6 = inet6_sk(new_sock->sk);
164 dev_if = inet6->mcast_oif;
165 } else {
166 dev_if = new_sock->sk->sk_bound_dev_if;
167 }
168#endif
169
170 conn = rds_conn_create(sock_net(sock->sk),
171 my_addr, peer_addr,
172 &rds_tcp_transport, 0, GFP_KERNEL, dev_if);
173
174 if (IS_ERR(conn)) {
175 ret = PTR_ERR(conn);
176 goto out;
177 }
178 /* An incoming SYN request came in, and TCP just accepted it.
179 *
180 * If the client reboots, this conn will need to be cleaned up.
181 * rds_tcp_state_change() will do that cleanup
182 */
183 rs_tcp = rds_tcp_accept_one_path(conn);
184 if (!rs_tcp)
185 goto rst_nsk;
186 mutex_lock(&rs_tcp->t_conn_path_lock);
187 cp = rs_tcp->t_cpath;
188 conn_state = rds_conn_path_state(cp);
189 WARN_ON(conn_state == RDS_CONN_UP);
190 if (conn_state != RDS_CONN_CONNECTING && conn_state != RDS_CONN_ERROR)
191 goto rst_nsk;
192 if (rs_tcp->t_sock) {
193 /* Duelling SYN has been handled in rds_tcp_accept_one() */
194 rds_tcp_reset_callbacks(new_sock, cp);
195 /* rds_connect_path_complete() marks RDS_CONN_UP */
196 rds_connect_path_complete(cp, RDS_CONN_RESETTING);
197 } else {
198 rds_tcp_set_callbacks(new_sock, cp);
199 rds_connect_path_complete(cp, RDS_CONN_CONNECTING);
200 }
201 new_sock = NULL;
202 ret = 0;
203 if (conn->c_npaths == 0)
204 rds_send_ping(cp->cp_conn, cp->cp_index);
205 goto out;
206rst_nsk:
207 /* reset the newly returned accept sock and bail.
208 * It is safe to set linger on new_sock because the RDS connection
209 * has not been brought up on new_sock, so no RDS-level data could
210 * be pending on it. By setting linger, we achieve the side-effect
211 * of avoiding TIME_WAIT state on new_sock.
212 */
213 sock_no_linger(new_sock->sk);
214 kernel_sock_shutdown(new_sock, SHUT_RDWR);
215 ret = 0;
216out:
217 if (rs_tcp)
218 mutex_unlock(&rs_tcp->t_conn_path_lock);
219 if (new_sock)
220 sock_release(new_sock);
221 return ret;
222}
223
224void rds_tcp_listen_data_ready(struct sock *sk)
225{
226 void (*ready)(struct sock *sk);
227
228 rdsdebug("listen data ready sk %p\n", sk);
229
230 read_lock_bh(&sk->sk_callback_lock);
231 ready = sk->sk_user_data;
232 if (!ready) { /* check for teardown race */
233 ready = sk->sk_data_ready;
234 goto out;
235 }
236
237 /*
238 * ->sk_data_ready is also called for a newly established child socket
239 * before it has been accepted and the accepter has set up their
240 * data_ready.. we only want to queue listen work for our listening
241 * socket
242 *
243 * (*ready)() may be null if we are racing with netns delete, and
244 * the listen socket is being torn down.
245 */
246 if (sk->sk_state == TCP_LISTEN)
247 rds_tcp_accept_work(sk);
248 else
249 ready = rds_tcp_listen_sock_def_readable(sock_net(sk));
250
251out:
252 read_unlock_bh(&sk->sk_callback_lock);
253 if (ready)
254 ready(sk);
255}
256
257struct socket *rds_tcp_listen_init(struct net *net, bool isv6)
258{
259 struct socket *sock = NULL;
260 struct sockaddr_storage ss;
261 struct sockaddr_in6 *sin6;
262 struct sockaddr_in *sin;
263 int addr_len;
264 int ret;
265
266 ret = sock_create_kern(net, isv6 ? PF_INET6 : PF_INET, SOCK_STREAM,
267 IPPROTO_TCP, &sock);
268 if (ret < 0) {
269 rdsdebug("could not create %s listener socket: %d\n",
270 isv6 ? "IPv6" : "IPv4", ret);
271 goto out;
272 }
273
274 sock->sk->sk_reuse = SK_CAN_REUSE;
275 tcp_sock_set_nodelay(sock->sk);
276
277 write_lock_bh(&sock->sk->sk_callback_lock);
278 sock->sk->sk_user_data = sock->sk->sk_data_ready;
279 sock->sk->sk_data_ready = rds_tcp_listen_data_ready;
280 write_unlock_bh(&sock->sk->sk_callback_lock);
281
282 if (isv6) {
283 sin6 = (struct sockaddr_in6 *)&ss;
284 sin6->sin6_family = PF_INET6;
285 sin6->sin6_addr = in6addr_any;
286 sin6->sin6_port = (__force u16)htons(RDS_TCP_PORT);
287 sin6->sin6_scope_id = 0;
288 sin6->sin6_flowinfo = 0;
289 addr_len = sizeof(*sin6);
290 } else {
291 sin = (struct sockaddr_in *)&ss;
292 sin->sin_family = PF_INET;
293 sin->sin_addr.s_addr = INADDR_ANY;
294 sin->sin_port = (__force u16)htons(RDS_TCP_PORT);
295 addr_len = sizeof(*sin);
296 }
297
298 ret = sock->ops->bind(sock, (struct sockaddr *)&ss, addr_len);
299 if (ret < 0) {
300 rdsdebug("could not bind %s listener socket: %d\n",
301 isv6 ? "IPv6" : "IPv4", ret);
302 goto out;
303 }
304
305 ret = sock->ops->listen(sock, 64);
306 if (ret < 0)
307 goto out;
308
309 return sock;
310out:
311 if (sock)
312 sock_release(sock);
313 return NULL;
314}
315
316void rds_tcp_listen_stop(struct socket *sock, struct work_struct *acceptor)
317{
318 struct sock *sk;
319
320 if (!sock)
321 return;
322
323 sk = sock->sk;
324
325 /* serialize with and prevent further callbacks */
326 lock_sock(sk);
327 write_lock_bh(&sk->sk_callback_lock);
328 if (sk->sk_user_data) {
329 sk->sk_data_ready = sk->sk_user_data;
330 sk->sk_user_data = NULL;
331 }
332 write_unlock_bh(&sk->sk_callback_lock);
333 release_sock(sk);
334
335 /* wait for accepts to stop and close the socket */
336 flush_workqueue(rds_wq);
337 flush_work(acceptor);
338 sock_release(sock);
339}