1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Shared Memory Communications over RDMA (SMC-R) and RoCE 4 * 5 * Socket Closing - normal and abnormal 6 * 7 * Copyright IBM Corp. 2016 8 * 9 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com> 10 */ 11 12 #include <linux/workqueue.h> 13 #include <linux/sched/signal.h> 14 15 #include <net/sock.h> 16 #include <net/tcp.h> 17 18 #include "smc.h" 19 #include "smc_tx.h" 20 #include "smc_cdc.h" 21 #include "smc_close.h" 22 23 /* release the clcsock that is assigned to the smc_sock */ 24 void smc_clcsock_release(struct smc_sock *smc) 25 { 26 struct socket *tcp; 27 28 if (smc->listen_smc && current_work() != &smc->smc_listen_work) 29 cancel_work_sync(&smc->smc_listen_work); 30 mutex_lock(&smc->clcsock_release_lock); 31 if (smc->clcsock) { 32 tcp = smc->clcsock; 33 smc->clcsock = NULL; 34 sock_release(tcp); 35 } 36 mutex_unlock(&smc->clcsock_release_lock); 37 } 38 39 static void smc_close_cleanup_listen(struct sock *parent) 40 { 41 struct sock *sk; 42 43 /* Close non-accepted connections */ 44 while ((sk = smc_accept_dequeue(parent, NULL))) 45 smc_close_non_accepted(sk); 46 } 47 48 /* wait for sndbuf data being transmitted */ 49 static void smc_close_stream_wait(struct smc_sock *smc, long timeout) 50 { 51 DEFINE_WAIT_FUNC(wait, woken_wake_function); 52 struct sock *sk = &smc->sk; 53 54 if (!timeout) 55 return; 56 57 if (!smc_tx_prepared_sends(&smc->conn)) 58 return; 59 60 smc->wait_close_tx_prepared = 1; 61 add_wait_queue(sk_sleep(sk), &wait); 62 while (!signal_pending(current) && timeout) { 63 int rc; 64 65 rc = sk_wait_event(sk, &timeout, 66 !smc_tx_prepared_sends(&smc->conn) || 67 sk->sk_err == ECONNABORTED || 68 sk->sk_err == ECONNRESET || 69 smc->conn.killed, 70 &wait); 71 if (rc) 72 break; 73 } 74 remove_wait_queue(sk_sleep(sk), &wait); 75 smc->wait_close_tx_prepared = 0; 76 } 77 78 void smc_close_wake_tx_prepared(struct smc_sock *smc) 79 { 80 if (smc->wait_close_tx_prepared) 81 /* wake up socket closing */ 82 smc->sk.sk_state_change(&smc->sk); 83 } 84 85 static int smc_close_wr(struct smc_connection *conn) 86 { 87 conn->local_tx_ctrl.conn_state_flags.peer_done_writing = 1; 88 89 return smc_cdc_get_slot_and_msg_send(conn); 90 } 91 92 static int smc_close_final(struct smc_connection *conn) 93 { 94 if (atomic_read(&conn->bytes_to_rcv)) 95 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 96 else 97 conn->local_tx_ctrl.conn_state_flags.peer_conn_closed = 1; 98 if (conn->killed) 99 return -EPIPE; 100 101 return smc_cdc_get_slot_and_msg_send(conn); 102 } 103 104 int smc_close_abort(struct smc_connection *conn) 105 { 106 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 107 108 return smc_cdc_get_slot_and_msg_send(conn); 109 } 110 111 static void smc_close_cancel_work(struct smc_sock *smc) 112 { 113 struct sock *sk = &smc->sk; 114 115 release_sock(sk); 116 cancel_work_sync(&smc->conn.close_work); 117 cancel_delayed_work_sync(&smc->conn.tx_work); 118 lock_sock(sk); 119 sk->sk_state = SMC_CLOSED; 120 } 121 122 /* terminate smc socket abnormally - active abort 123 * link group is terminated, i.e. RDMA communication no longer possible 124 */ 125 void smc_close_active_abort(struct smc_sock *smc) 126 { 127 struct sock *sk = &smc->sk; 128 bool release_clcsock = false; 129 130 if (sk->sk_state != SMC_INIT && smc->clcsock && smc->clcsock->sk) { 131 sk->sk_err = ECONNABORTED; 132 if (smc->clcsock && smc->clcsock->sk) 133 tcp_abort(smc->clcsock->sk, ECONNABORTED); 134 } 135 switch (sk->sk_state) { 136 case SMC_ACTIVE: 137 sk->sk_state = SMC_PEERABORTWAIT; 138 smc_close_cancel_work(smc); 139 sk->sk_state = SMC_CLOSED; 140 sock_put(sk); /* passive closing */ 141 break; 142 case SMC_APPCLOSEWAIT1: 143 case SMC_APPCLOSEWAIT2: 144 smc_close_cancel_work(smc); 145 sk->sk_state = SMC_CLOSED; 146 sock_put(sk); /* postponed passive closing */ 147 break; 148 case SMC_PEERCLOSEWAIT1: 149 case SMC_PEERCLOSEWAIT2: 150 case SMC_PEERFINCLOSEWAIT: 151 sk->sk_state = SMC_PEERABORTWAIT; 152 smc_close_cancel_work(smc); 153 sk->sk_state = SMC_CLOSED; 154 smc_conn_free(&smc->conn); 155 release_clcsock = true; 156 sock_put(sk); /* passive closing */ 157 break; 158 case SMC_PROCESSABORT: 159 case SMC_APPFINCLOSEWAIT: 160 sk->sk_state = SMC_PEERABORTWAIT; 161 smc_close_cancel_work(smc); 162 sk->sk_state = SMC_CLOSED; 163 smc_conn_free(&smc->conn); 164 release_clcsock = true; 165 break; 166 case SMC_INIT: 167 case SMC_PEERABORTWAIT: 168 case SMC_CLOSED: 169 break; 170 } 171 172 sock_set_flag(sk, SOCK_DEAD); 173 sk->sk_state_change(sk); 174 175 if (release_clcsock) { 176 release_sock(sk); 177 smc_clcsock_release(smc); 178 lock_sock(sk); 179 } 180 } 181 182 static inline bool smc_close_sent_any_close(struct smc_connection *conn) 183 { 184 return conn->local_tx_ctrl.conn_state_flags.peer_conn_abort || 185 conn->local_tx_ctrl.conn_state_flags.peer_conn_closed; 186 } 187 188 int smc_close_active(struct smc_sock *smc) 189 { 190 struct smc_cdc_conn_state_flags *txflags = 191 &smc->conn.local_tx_ctrl.conn_state_flags; 192 struct smc_connection *conn = &smc->conn; 193 struct sock *sk = &smc->sk; 194 int old_state; 195 long timeout; 196 int rc = 0; 197 198 timeout = current->flags & PF_EXITING ? 199 0 : sock_flag(sk, SOCK_LINGER) ? 200 sk->sk_lingertime : SMC_MAX_STREAM_WAIT_TIMEOUT; 201 202 old_state = sk->sk_state; 203 again: 204 switch (sk->sk_state) { 205 case SMC_INIT: 206 sk->sk_state = SMC_CLOSED; 207 break; 208 case SMC_LISTEN: 209 sk->sk_state = SMC_CLOSED; 210 sk->sk_state_change(sk); /* wake up accept */ 211 if (smc->clcsock && smc->clcsock->sk) { 212 rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR); 213 /* wake up kernel_accept of smc_tcp_listen_worker */ 214 smc->clcsock->sk->sk_data_ready(smc->clcsock->sk); 215 } 216 smc_close_cleanup_listen(sk); 217 release_sock(sk); 218 flush_work(&smc->tcp_listen_work); 219 lock_sock(sk); 220 break; 221 case SMC_ACTIVE: 222 smc_close_stream_wait(smc, timeout); 223 release_sock(sk); 224 cancel_delayed_work_sync(&conn->tx_work); 225 lock_sock(sk); 226 if (sk->sk_state == SMC_ACTIVE) { 227 /* send close request */ 228 rc = smc_close_final(conn); 229 sk->sk_state = SMC_PEERCLOSEWAIT1; 230 } else { 231 /* peer event has changed the state */ 232 goto again; 233 } 234 break; 235 case SMC_APPFINCLOSEWAIT: 236 /* socket already shutdown wr or both (active close) */ 237 if (txflags->peer_done_writing && 238 !smc_close_sent_any_close(conn)) { 239 /* just shutdown wr done, send close request */ 240 rc = smc_close_final(conn); 241 } 242 sk->sk_state = SMC_CLOSED; 243 break; 244 case SMC_APPCLOSEWAIT1: 245 case SMC_APPCLOSEWAIT2: 246 if (!smc_cdc_rxed_any_close(conn)) 247 smc_close_stream_wait(smc, timeout); 248 release_sock(sk); 249 cancel_delayed_work_sync(&conn->tx_work); 250 lock_sock(sk); 251 if (sk->sk_state != SMC_APPCLOSEWAIT1 && 252 sk->sk_state != SMC_APPCLOSEWAIT2) 253 goto again; 254 /* confirm close from peer */ 255 rc = smc_close_final(conn); 256 if (smc_cdc_rxed_any_close(conn)) { 257 /* peer has closed the socket already */ 258 sk->sk_state = SMC_CLOSED; 259 sock_put(sk); /* postponed passive closing */ 260 } else { 261 /* peer has just issued a shutdown write */ 262 sk->sk_state = SMC_PEERFINCLOSEWAIT; 263 } 264 break; 265 case SMC_PEERCLOSEWAIT1: 266 case SMC_PEERCLOSEWAIT2: 267 if (txflags->peer_done_writing && 268 !smc_close_sent_any_close(conn)) { 269 /* just shutdown wr done, send close request */ 270 rc = smc_close_final(conn); 271 } 272 /* peer sending PeerConnectionClosed will cause transition */ 273 break; 274 case SMC_PEERFINCLOSEWAIT: 275 /* peer sending PeerConnectionClosed will cause transition */ 276 break; 277 case SMC_PROCESSABORT: 278 rc = smc_close_abort(conn); 279 sk->sk_state = SMC_CLOSED; 280 break; 281 case SMC_PEERABORTWAIT: 282 sk->sk_state = SMC_CLOSED; 283 break; 284 case SMC_CLOSED: 285 /* nothing to do, add tracing in future patch */ 286 break; 287 } 288 289 if (old_state != sk->sk_state) 290 sk->sk_state_change(sk); 291 return rc; 292 } 293 294 static void smc_close_passive_abort_received(struct smc_sock *smc) 295 { 296 struct smc_cdc_conn_state_flags *txflags = 297 &smc->conn.local_tx_ctrl.conn_state_flags; 298 struct sock *sk = &smc->sk; 299 300 switch (sk->sk_state) { 301 case SMC_INIT: 302 case SMC_ACTIVE: 303 case SMC_APPCLOSEWAIT1: 304 sk->sk_state = SMC_PROCESSABORT; 305 sock_put(sk); /* passive closing */ 306 break; 307 case SMC_APPFINCLOSEWAIT: 308 sk->sk_state = SMC_PROCESSABORT; 309 break; 310 case SMC_PEERCLOSEWAIT1: 311 case SMC_PEERCLOSEWAIT2: 312 if (txflags->peer_done_writing && 313 !smc_close_sent_any_close(&smc->conn)) 314 /* just shutdown, but not yet closed locally */ 315 sk->sk_state = SMC_PROCESSABORT; 316 else 317 sk->sk_state = SMC_CLOSED; 318 sock_put(sk); /* passive closing */ 319 break; 320 case SMC_APPCLOSEWAIT2: 321 case SMC_PEERFINCLOSEWAIT: 322 sk->sk_state = SMC_CLOSED; 323 sock_put(sk); /* passive closing */ 324 break; 325 case SMC_PEERABORTWAIT: 326 sk->sk_state = SMC_CLOSED; 327 break; 328 case SMC_PROCESSABORT: 329 /* nothing to do, add tracing in future patch */ 330 break; 331 } 332 } 333 334 /* Either some kind of closing has been received: peer_conn_closed, 335 * peer_conn_abort, or peer_done_writing 336 * or the link group of the connection terminates abnormally. 337 */ 338 static void smc_close_passive_work(struct work_struct *work) 339 { 340 struct smc_connection *conn = container_of(work, 341 struct smc_connection, 342 close_work); 343 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 344 struct smc_cdc_conn_state_flags *rxflags; 345 bool release_clcsock = false; 346 struct sock *sk = &smc->sk; 347 int old_state; 348 349 lock_sock(sk); 350 old_state = sk->sk_state; 351 352 rxflags = &conn->local_rx_ctrl.conn_state_flags; 353 if (rxflags->peer_conn_abort) { 354 /* peer has not received all data */ 355 smc_close_passive_abort_received(smc); 356 release_sock(&smc->sk); 357 cancel_delayed_work_sync(&conn->tx_work); 358 lock_sock(&smc->sk); 359 goto wakeup; 360 } 361 362 switch (sk->sk_state) { 363 case SMC_INIT: 364 sk->sk_state = SMC_APPCLOSEWAIT1; 365 break; 366 case SMC_ACTIVE: 367 sk->sk_state = SMC_APPCLOSEWAIT1; 368 /* postpone sock_put() for passive closing to cover 369 * received SEND_SHUTDOWN as well 370 */ 371 break; 372 case SMC_PEERCLOSEWAIT1: 373 if (rxflags->peer_done_writing) 374 sk->sk_state = SMC_PEERCLOSEWAIT2; 375 fallthrough; 376 /* to check for closing */ 377 case SMC_PEERCLOSEWAIT2: 378 if (!smc_cdc_rxed_any_close(conn)) 379 break; 380 if (sock_flag(sk, SOCK_DEAD) && 381 smc_close_sent_any_close(conn)) { 382 /* smc_release has already been called locally */ 383 sk->sk_state = SMC_CLOSED; 384 } else { 385 /* just shutdown, but not yet closed locally */ 386 sk->sk_state = SMC_APPFINCLOSEWAIT; 387 } 388 sock_put(sk); /* passive closing */ 389 break; 390 case SMC_PEERFINCLOSEWAIT: 391 if (smc_cdc_rxed_any_close(conn)) { 392 sk->sk_state = SMC_CLOSED; 393 sock_put(sk); /* passive closing */ 394 } 395 break; 396 case SMC_APPCLOSEWAIT1: 397 case SMC_APPCLOSEWAIT2: 398 /* postpone sock_put() for passive closing to cover 399 * received SEND_SHUTDOWN as well 400 */ 401 break; 402 case SMC_APPFINCLOSEWAIT: 403 case SMC_PEERABORTWAIT: 404 case SMC_PROCESSABORT: 405 case SMC_CLOSED: 406 /* nothing to do, add tracing in future patch */ 407 break; 408 } 409 410 wakeup: 411 sk->sk_data_ready(sk); /* wakeup blocked rcvbuf consumers */ 412 sk->sk_write_space(sk); /* wakeup blocked sndbuf producers */ 413 414 if (old_state != sk->sk_state) { 415 sk->sk_state_change(sk); 416 if ((sk->sk_state == SMC_CLOSED) && 417 (sock_flag(sk, SOCK_DEAD) || !sk->sk_socket)) { 418 smc_conn_free(conn); 419 if (smc->clcsock) 420 release_clcsock = true; 421 } 422 } 423 release_sock(sk); 424 if (release_clcsock) 425 smc_clcsock_release(smc); 426 sock_put(sk); /* sock_hold done by schedulers of close_work */ 427 } 428 429 int smc_close_shutdown_write(struct smc_sock *smc) 430 { 431 struct smc_connection *conn = &smc->conn; 432 struct sock *sk = &smc->sk; 433 int old_state; 434 long timeout; 435 int rc = 0; 436 437 timeout = current->flags & PF_EXITING ? 438 0 : sock_flag(sk, SOCK_LINGER) ? 439 sk->sk_lingertime : SMC_MAX_STREAM_WAIT_TIMEOUT; 440 441 old_state = sk->sk_state; 442 again: 443 switch (sk->sk_state) { 444 case SMC_ACTIVE: 445 smc_close_stream_wait(smc, timeout); 446 release_sock(sk); 447 cancel_delayed_work_sync(&conn->tx_work); 448 lock_sock(sk); 449 if (sk->sk_state != SMC_ACTIVE) 450 goto again; 451 /* send close wr request */ 452 rc = smc_close_wr(conn); 453 sk->sk_state = SMC_PEERCLOSEWAIT1; 454 break; 455 case SMC_APPCLOSEWAIT1: 456 /* passive close */ 457 if (!smc_cdc_rxed_any_close(conn)) 458 smc_close_stream_wait(smc, timeout); 459 release_sock(sk); 460 cancel_delayed_work_sync(&conn->tx_work); 461 lock_sock(sk); 462 if (sk->sk_state != SMC_APPCLOSEWAIT1) 463 goto again; 464 /* confirm close from peer */ 465 rc = smc_close_wr(conn); 466 sk->sk_state = SMC_APPCLOSEWAIT2; 467 break; 468 case SMC_APPCLOSEWAIT2: 469 case SMC_PEERFINCLOSEWAIT: 470 case SMC_PEERCLOSEWAIT1: 471 case SMC_PEERCLOSEWAIT2: 472 case SMC_APPFINCLOSEWAIT: 473 case SMC_PROCESSABORT: 474 case SMC_PEERABORTWAIT: 475 /* nothing to do, add tracing in future patch */ 476 break; 477 } 478 479 if (old_state != sk->sk_state) 480 sk->sk_state_change(sk); 481 return rc; 482 } 483 484 /* Initialize close properties on connection establishment. */ 485 void smc_close_init(struct smc_sock *smc) 486 { 487 INIT_WORK(&smc->conn.close_work, smc_close_passive_work); 488 } 489