1 /* 2 * Copyright (c) 2006, 2018 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 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 83 struct rds_tcp_connection *rds_tcp_accept_one_path(struct rds_connection *conn) 84 { 85 int i; 86 bool peer_is_smaller = IS_CANONICAL(conn->c_faddr, conn->c_laddr); 87 int npaths = max_t(int, 1, conn->c_npaths); 88 89 /* for mprds, all paths MUST be initiated by the peer 90 * with the smaller address. 91 */ 92 if (!peer_is_smaller) { 93 /* Make sure we initiate at least one path if this 94 * has not already been done; rds_start_mprds() will 95 * take care of additional paths, if necessary. 96 */ 97 if (npaths == 1) 98 rds_conn_path_connect_if_down(&conn->c_path[0]); 99 return NULL; 100 } 101 102 for (i = 0; i < npaths; i++) { 103 struct rds_conn_path *cp = &conn->c_path[i]; 104 105 if (rds_conn_path_transition(cp, RDS_CONN_DOWN, 106 RDS_CONN_CONNECTING) || 107 rds_conn_path_transition(cp, RDS_CONN_ERROR, 108 RDS_CONN_CONNECTING)) { 109 return cp->cp_transport_data; 110 } 111 } 112 return NULL; 113 } 114 115 void rds_tcp_set_linger(struct socket *sock) 116 { 117 struct linger no_linger = { 118 .l_onoff = 1, 119 .l_linger = 0, 120 }; 121 122 kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER, 123 (char *)&no_linger, sizeof(no_linger)); 124 } 125 126 int rds_tcp_accept_one(struct socket *sock) 127 { 128 struct socket *new_sock = NULL; 129 struct rds_connection *conn; 130 int ret; 131 struct inet_sock *inet; 132 struct rds_tcp_connection *rs_tcp = NULL; 133 int conn_state; 134 struct rds_conn_path *cp; 135 136 if (!sock) /* module unload or netns delete in progress */ 137 return -ENETUNREACH; 138 139 ret = sock_create_lite(sock->sk->sk_family, 140 sock->sk->sk_type, sock->sk->sk_protocol, 141 &new_sock); 142 if (ret) 143 goto out; 144 145 ret = sock->ops->accept(sock, new_sock, O_NONBLOCK, true); 146 if (ret < 0) 147 goto out; 148 149 /* sock_create_lite() does not get a hold on the owner module so we 150 * need to do it here. Note that sock_release() uses sock->ops to 151 * determine if it needs to decrement the reference count. So set 152 * sock->ops after calling accept() in case that fails. And there's 153 * no need to do try_module_get() as the listener should have a hold 154 * already. 155 */ 156 new_sock->ops = sock->ops; 157 __module_get(new_sock->ops->owner); 158 159 ret = rds_tcp_keepalive(new_sock); 160 if (ret < 0) 161 goto out; 162 163 rds_tcp_tune(new_sock); 164 165 inet = inet_sk(new_sock->sk); 166 167 rdsdebug("accepted tcp %pI4:%u -> %pI4:%u\n", 168 &inet->inet_saddr, ntohs(inet->inet_sport), 169 &inet->inet_daddr, ntohs(inet->inet_dport)); 170 171 conn = rds_conn_create(sock_net(sock->sk), 172 inet->inet_saddr, inet->inet_daddr, 173 &rds_tcp_transport, GFP_KERNEL); 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; 206 rst_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 rds_tcp_set_linger(new_sock); 214 kernel_sock_shutdown(new_sock, SHUT_RDWR); 215 ret = 0; 216 out: 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 } 223 224 void 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)); 250 251 out: 252 read_unlock_bh(&sk->sk_callback_lock); 253 if (ready) 254 ready(sk); 255 } 256 257 struct socket *rds_tcp_listen_init(struct net *net) 258 { 259 struct sockaddr_in sin; 260 struct socket *sock = NULL; 261 int ret; 262 263 ret = sock_create_kern(net, PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock); 264 if (ret < 0) 265 goto out; 266 267 sock->sk->sk_reuse = SK_CAN_REUSE; 268 rds_tcp_nonagle(sock); 269 270 write_lock_bh(&sock->sk->sk_callback_lock); 271 sock->sk->sk_user_data = sock->sk->sk_data_ready; 272 sock->sk->sk_data_ready = rds_tcp_listen_data_ready; 273 write_unlock_bh(&sock->sk->sk_callback_lock); 274 275 sin.sin_family = PF_INET; 276 sin.sin_addr.s_addr = (__force u32)htonl(INADDR_ANY); 277 sin.sin_port = (__force u16)htons(RDS_TCP_PORT); 278 279 ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin)); 280 if (ret < 0) 281 goto out; 282 283 ret = sock->ops->listen(sock, 64); 284 if (ret < 0) 285 goto out; 286 287 return sock; 288 out: 289 if (sock) 290 sock_release(sock); 291 return NULL; 292 } 293 294 void rds_tcp_listen_stop(struct socket *sock, struct work_struct *acceptor) 295 { 296 struct sock *sk; 297 298 if (!sock) 299 return; 300 301 sk = sock->sk; 302 303 /* serialize with and prevent further callbacks */ 304 lock_sock(sk); 305 write_lock_bh(&sk->sk_callback_lock); 306 if (sk->sk_user_data) { 307 sk->sk_data_ready = sk->sk_user_data; 308 sk->sk_user_data = NULL; 309 } 310 write_unlock_bh(&sk->sk_callback_lock); 311 release_sock(sk); 312 313 /* wait for accepts to stop and close the socket */ 314 flush_workqueue(rds_wq); 315 flush_work(acceptor); 316 sock_release(sock); 317 } 318