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 /* Send out corked data remaining in sndbuf */ 61 smc_tx_pending(&smc->conn); 62 63 smc->wait_close_tx_prepared = 1; 64 add_wait_queue(sk_sleep(sk), &wait); 65 while (!signal_pending(current) && timeout) { 66 int rc; 67 68 rc = sk_wait_event(sk, &timeout, 69 !smc_tx_prepared_sends(&smc->conn) || 70 sk->sk_err == ECONNABORTED || 71 sk->sk_err == ECONNRESET || 72 smc->conn.killed, 73 &wait); 74 if (rc) 75 break; 76 } 77 remove_wait_queue(sk_sleep(sk), &wait); 78 smc->wait_close_tx_prepared = 0; 79 } 80 81 void smc_close_wake_tx_prepared(struct smc_sock *smc) 82 { 83 if (smc->wait_close_tx_prepared) 84 /* wake up socket closing */ 85 smc->sk.sk_state_change(&smc->sk); 86 } 87 88 static int smc_close_wr(struct smc_connection *conn) 89 { 90 conn->local_tx_ctrl.conn_state_flags.peer_done_writing = 1; 91 92 return smc_cdc_get_slot_and_msg_send(conn); 93 } 94 95 static int smc_close_final(struct smc_connection *conn) 96 { 97 if (atomic_read(&conn->bytes_to_rcv)) 98 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 99 else 100 conn->local_tx_ctrl.conn_state_flags.peer_conn_closed = 1; 101 if (conn->killed) 102 return -EPIPE; 103 104 return smc_cdc_get_slot_and_msg_send(conn); 105 } 106 107 int smc_close_abort(struct smc_connection *conn) 108 { 109 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1; 110 111 return smc_cdc_get_slot_and_msg_send(conn); 112 } 113 114 static void smc_close_cancel_work(struct smc_sock *smc) 115 { 116 struct sock *sk = &smc->sk; 117 118 release_sock(sk); 119 cancel_work_sync(&smc->conn.close_work); 120 cancel_delayed_work_sync(&smc->conn.tx_work); 121 lock_sock(sk); 122 } 123 124 /* terminate smc socket abnormally - active abort 125 * link group is terminated, i.e. RDMA communication no longer possible 126 */ 127 void smc_close_active_abort(struct smc_sock *smc) 128 { 129 struct sock *sk = &smc->sk; 130 bool release_clcsock = false; 131 132 if (sk->sk_state != SMC_INIT && smc->clcsock && smc->clcsock->sk) { 133 sk->sk_err = ECONNABORTED; 134 if (smc->clcsock && smc->clcsock->sk) 135 tcp_abort(smc->clcsock->sk, ECONNABORTED); 136 } 137 switch (sk->sk_state) { 138 case SMC_ACTIVE: 139 case SMC_APPCLOSEWAIT1: 140 case SMC_APPCLOSEWAIT2: 141 sk->sk_state = SMC_PEERABORTWAIT; 142 smc_close_cancel_work(smc); 143 if (sk->sk_state != SMC_PEERABORTWAIT) 144 break; 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 if (sk->sk_state != SMC_PEERABORTWAIT) 154 break; 155 sk->sk_state = SMC_CLOSED; 156 smc_conn_free(&smc->conn); 157 release_clcsock = true; 158 sock_put(sk); /* passive closing */ 159 break; 160 case SMC_PROCESSABORT: 161 case SMC_APPFINCLOSEWAIT: 162 sk->sk_state = SMC_PEERABORTWAIT; 163 smc_close_cancel_work(smc); 164 if (sk->sk_state != SMC_PEERABORTWAIT) 165 break; 166 sk->sk_state = SMC_CLOSED; 167 smc_conn_free(&smc->conn); 168 release_clcsock = true; 169 break; 170 case SMC_INIT: 171 case SMC_PEERABORTWAIT: 172 case SMC_CLOSED: 173 break; 174 } 175 176 sock_set_flag(sk, SOCK_DEAD); 177 sk->sk_state_change(sk); 178 179 if (release_clcsock) { 180 release_sock(sk); 181 smc_clcsock_release(smc); 182 lock_sock(sk); 183 } 184 } 185 186 static inline bool smc_close_sent_any_close(struct smc_connection *conn) 187 { 188 return conn->local_tx_ctrl.conn_state_flags.peer_conn_abort || 189 conn->local_tx_ctrl.conn_state_flags.peer_conn_closed; 190 } 191 192 int smc_close_active(struct smc_sock *smc) 193 { 194 struct smc_cdc_conn_state_flags *txflags = 195 &smc->conn.local_tx_ctrl.conn_state_flags; 196 struct smc_connection *conn = &smc->conn; 197 struct sock *sk = &smc->sk; 198 int old_state; 199 long timeout; 200 int rc = 0; 201 int rc1 = 0; 202 203 timeout = current->flags & PF_EXITING ? 204 0 : sock_flag(sk, SOCK_LINGER) ? 205 sk->sk_lingertime : SMC_MAX_STREAM_WAIT_TIMEOUT; 206 207 old_state = sk->sk_state; 208 again: 209 switch (sk->sk_state) { 210 case SMC_INIT: 211 sk->sk_state = SMC_CLOSED; 212 break; 213 case SMC_LISTEN: 214 sk->sk_state = SMC_CLOSED; 215 sk->sk_state_change(sk); /* wake up accept */ 216 if (smc->clcsock && smc->clcsock->sk) { 217 smc->clcsock->sk->sk_data_ready = smc->clcsk_data_ready; 218 smc->clcsock->sk->sk_user_data = NULL; 219 rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR); 220 } 221 smc_close_cleanup_listen(sk); 222 release_sock(sk); 223 flush_work(&smc->tcp_listen_work); 224 lock_sock(sk); 225 break; 226 case SMC_ACTIVE: 227 smc_close_stream_wait(smc, timeout); 228 release_sock(sk); 229 cancel_delayed_work_sync(&conn->tx_work); 230 lock_sock(sk); 231 if (sk->sk_state == SMC_ACTIVE) { 232 /* send close request */ 233 rc = smc_close_final(conn); 234 sk->sk_state = SMC_PEERCLOSEWAIT1; 235 236 /* actively shutdown clcsock before peer close it, 237 * prevent peer from entering TIME_WAIT state. 238 */ 239 if (smc->clcsock && smc->clcsock->sk) { 240 rc1 = kernel_sock_shutdown(smc->clcsock, 241 SHUT_RDWR); 242 rc = rc ? rc : rc1; 243 } 244 } else { 245 /* peer event has changed the state */ 246 goto again; 247 } 248 break; 249 case SMC_APPFINCLOSEWAIT: 250 /* socket already shutdown wr or both (active close) */ 251 if (txflags->peer_done_writing && 252 !smc_close_sent_any_close(conn)) { 253 /* just shutdown wr done, send close request */ 254 rc = smc_close_final(conn); 255 } 256 sk->sk_state = SMC_CLOSED; 257 break; 258 case SMC_APPCLOSEWAIT1: 259 case SMC_APPCLOSEWAIT2: 260 if (!smc_cdc_rxed_any_close(conn)) 261 smc_close_stream_wait(smc, timeout); 262 release_sock(sk); 263 cancel_delayed_work_sync(&conn->tx_work); 264 lock_sock(sk); 265 if (sk->sk_state != SMC_APPCLOSEWAIT1 && 266 sk->sk_state != SMC_APPCLOSEWAIT2) 267 goto again; 268 /* confirm close from peer */ 269 rc = smc_close_final(conn); 270 if (smc_cdc_rxed_any_close(conn)) { 271 /* peer has closed the socket already */ 272 sk->sk_state = SMC_CLOSED; 273 sock_put(sk); /* postponed passive closing */ 274 } else { 275 /* peer has just issued a shutdown write */ 276 sk->sk_state = SMC_PEERFINCLOSEWAIT; 277 } 278 break; 279 case SMC_PEERCLOSEWAIT1: 280 case SMC_PEERCLOSEWAIT2: 281 if (txflags->peer_done_writing && 282 !smc_close_sent_any_close(conn)) { 283 /* just shutdown wr done, send close request */ 284 rc = smc_close_final(conn); 285 } 286 /* peer sending PeerConnectionClosed will cause transition */ 287 break; 288 case SMC_PEERFINCLOSEWAIT: 289 /* peer sending PeerConnectionClosed will cause transition */ 290 break; 291 case SMC_PROCESSABORT: 292 rc = smc_close_abort(conn); 293 sk->sk_state = SMC_CLOSED; 294 break; 295 case SMC_PEERABORTWAIT: 296 sk->sk_state = SMC_CLOSED; 297 break; 298 case SMC_CLOSED: 299 /* nothing to do, add tracing in future patch */ 300 break; 301 } 302 303 if (old_state != sk->sk_state) 304 sk->sk_state_change(sk); 305 return rc; 306 } 307 308 static void smc_close_passive_abort_received(struct smc_sock *smc) 309 { 310 struct smc_cdc_conn_state_flags *txflags = 311 &smc->conn.local_tx_ctrl.conn_state_flags; 312 struct sock *sk = &smc->sk; 313 314 switch (sk->sk_state) { 315 case SMC_INIT: 316 case SMC_ACTIVE: 317 case SMC_APPCLOSEWAIT1: 318 sk->sk_state = SMC_PROCESSABORT; 319 sock_put(sk); /* passive closing */ 320 break; 321 case SMC_APPFINCLOSEWAIT: 322 sk->sk_state = SMC_PROCESSABORT; 323 break; 324 case SMC_PEERCLOSEWAIT1: 325 case SMC_PEERCLOSEWAIT2: 326 if (txflags->peer_done_writing && 327 !smc_close_sent_any_close(&smc->conn)) 328 /* just shutdown, but not yet closed locally */ 329 sk->sk_state = SMC_PROCESSABORT; 330 else 331 sk->sk_state = SMC_CLOSED; 332 sock_put(sk); /* passive closing */ 333 break; 334 case SMC_APPCLOSEWAIT2: 335 case SMC_PEERFINCLOSEWAIT: 336 sk->sk_state = SMC_CLOSED; 337 sock_put(sk); /* passive closing */ 338 break; 339 case SMC_PEERABORTWAIT: 340 sk->sk_state = SMC_CLOSED; 341 break; 342 case SMC_PROCESSABORT: 343 /* nothing to do, add tracing in future patch */ 344 break; 345 } 346 } 347 348 /* Either some kind of closing has been received: peer_conn_closed, 349 * peer_conn_abort, or peer_done_writing 350 * or the link group of the connection terminates abnormally. 351 */ 352 static void smc_close_passive_work(struct work_struct *work) 353 { 354 struct smc_connection *conn = container_of(work, 355 struct smc_connection, 356 close_work); 357 struct smc_sock *smc = container_of(conn, struct smc_sock, conn); 358 struct smc_cdc_conn_state_flags *rxflags; 359 bool release_clcsock = false; 360 struct sock *sk = &smc->sk; 361 int old_state; 362 363 lock_sock(sk); 364 old_state = sk->sk_state; 365 366 rxflags = &conn->local_rx_ctrl.conn_state_flags; 367 if (rxflags->peer_conn_abort) { 368 /* peer has not received all data */ 369 smc_close_passive_abort_received(smc); 370 release_sock(sk); 371 cancel_delayed_work_sync(&conn->tx_work); 372 lock_sock(sk); 373 goto wakeup; 374 } 375 376 switch (sk->sk_state) { 377 case SMC_INIT: 378 sk->sk_state = SMC_APPCLOSEWAIT1; 379 break; 380 case SMC_ACTIVE: 381 sk->sk_state = SMC_APPCLOSEWAIT1; 382 /* postpone sock_put() for passive closing to cover 383 * received SEND_SHUTDOWN as well 384 */ 385 break; 386 case SMC_PEERCLOSEWAIT1: 387 if (rxflags->peer_done_writing) 388 sk->sk_state = SMC_PEERCLOSEWAIT2; 389 fallthrough; 390 /* to check for closing */ 391 case SMC_PEERCLOSEWAIT2: 392 if (!smc_cdc_rxed_any_close(conn)) 393 break; 394 if (sock_flag(sk, SOCK_DEAD) && 395 smc_close_sent_any_close(conn)) { 396 /* smc_release has already been called locally */ 397 sk->sk_state = SMC_CLOSED; 398 } else { 399 /* just shutdown, but not yet closed locally */ 400 sk->sk_state = SMC_APPFINCLOSEWAIT; 401 } 402 sock_put(sk); /* passive closing */ 403 break; 404 case SMC_PEERFINCLOSEWAIT: 405 if (smc_cdc_rxed_any_close(conn)) { 406 sk->sk_state = SMC_CLOSED; 407 sock_put(sk); /* passive closing */ 408 } 409 break; 410 case SMC_APPCLOSEWAIT1: 411 case SMC_APPCLOSEWAIT2: 412 /* postpone sock_put() for passive closing to cover 413 * received SEND_SHUTDOWN as well 414 */ 415 break; 416 case SMC_APPFINCLOSEWAIT: 417 case SMC_PEERABORTWAIT: 418 case SMC_PROCESSABORT: 419 case SMC_CLOSED: 420 /* nothing to do, add tracing in future patch */ 421 break; 422 } 423 424 wakeup: 425 sk->sk_data_ready(sk); /* wakeup blocked rcvbuf consumers */ 426 sk->sk_write_space(sk); /* wakeup blocked sndbuf producers */ 427 428 if (old_state != sk->sk_state) { 429 sk->sk_state_change(sk); 430 if ((sk->sk_state == SMC_CLOSED) && 431 (sock_flag(sk, SOCK_DEAD) || !sk->sk_socket)) { 432 smc_conn_free(conn); 433 if (smc->clcsock) 434 release_clcsock = true; 435 } 436 } 437 release_sock(sk); 438 if (release_clcsock) 439 smc_clcsock_release(smc); 440 sock_put(sk); /* sock_hold done by schedulers of close_work */ 441 } 442 443 int smc_close_shutdown_write(struct smc_sock *smc) 444 { 445 struct smc_connection *conn = &smc->conn; 446 struct sock *sk = &smc->sk; 447 int old_state; 448 long timeout; 449 int rc = 0; 450 451 timeout = current->flags & PF_EXITING ? 452 0 : sock_flag(sk, SOCK_LINGER) ? 453 sk->sk_lingertime : SMC_MAX_STREAM_WAIT_TIMEOUT; 454 455 old_state = sk->sk_state; 456 again: 457 switch (sk->sk_state) { 458 case SMC_ACTIVE: 459 smc_close_stream_wait(smc, timeout); 460 release_sock(sk); 461 cancel_delayed_work_sync(&conn->tx_work); 462 lock_sock(sk); 463 if (sk->sk_state != SMC_ACTIVE) 464 goto again; 465 /* send close wr request */ 466 rc = smc_close_wr(conn); 467 sk->sk_state = SMC_PEERCLOSEWAIT1; 468 break; 469 case SMC_APPCLOSEWAIT1: 470 /* passive close */ 471 if (!smc_cdc_rxed_any_close(conn)) 472 smc_close_stream_wait(smc, timeout); 473 release_sock(sk); 474 cancel_delayed_work_sync(&conn->tx_work); 475 lock_sock(sk); 476 if (sk->sk_state != SMC_APPCLOSEWAIT1) 477 goto again; 478 /* confirm close from peer */ 479 rc = smc_close_wr(conn); 480 sk->sk_state = SMC_APPCLOSEWAIT2; 481 break; 482 case SMC_APPCLOSEWAIT2: 483 case SMC_PEERFINCLOSEWAIT: 484 case SMC_PEERCLOSEWAIT1: 485 case SMC_PEERCLOSEWAIT2: 486 case SMC_APPFINCLOSEWAIT: 487 case SMC_PROCESSABORT: 488 case SMC_PEERABORTWAIT: 489 /* nothing to do, add tracing in future patch */ 490 break; 491 } 492 493 if (old_state != sk->sk_state) 494 sk->sk_state_change(sk); 495 return rc; 496 } 497 498 /* Initialize close properties on connection establishment. */ 499 void smc_close_init(struct smc_sock *smc) 500 { 501 INIT_WORK(&smc->conn.close_work, smc_close_passive_work); 502 } 503