1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause 2 3 /* Authors: Bernard Metzler <bmt@zurich.ibm.com> */ 4 /* Fredy Neeser */ 5 /* Greg Joyce <greg@opengridcomputing.com> */ 6 /* Copyright (c) 2008-2019, IBM Corporation */ 7 /* Copyright (c) 2017, Open Grid Computing, Inc. */ 8 9 #include <linux/errno.h> 10 #include <linux/types.h> 11 #include <linux/net.h> 12 #include <linux/inetdevice.h> 13 #include <net/addrconf.h> 14 #include <linux/workqueue.h> 15 #include <net/sock.h> 16 #include <net/tcp.h> 17 #include <linux/inet.h> 18 #include <linux/tcp.h> 19 #include <trace/events/sock.h> 20 21 #include <rdma/iw_cm.h> 22 #include <rdma/ib_verbs.h> 23 #include <rdma/ib_user_verbs.h> 24 25 #include "siw.h" 26 #include "siw_cm.h" 27 28 /* 29 * Set to any combination of 30 * MPA_V2_RDMA_NO_RTR, MPA_V2_RDMA_READ_RTR, MPA_V2_RDMA_WRITE_RTR 31 */ 32 static __be16 rtr_type = MPA_V2_RDMA_READ_RTR | MPA_V2_RDMA_WRITE_RTR; 33 static const bool relaxed_ird_negotiation = true; 34 35 static void siw_cm_llp_state_change(struct sock *s); 36 static void siw_cm_llp_data_ready(struct sock *s); 37 static void siw_cm_llp_write_space(struct sock *s); 38 static void siw_cm_llp_error_report(struct sock *s); 39 static int siw_cm_upcall(struct siw_cep *cep, enum iw_cm_event_type reason, 40 int status); 41 42 static void siw_sk_assign_cm_upcalls(struct sock *sk) 43 { 44 write_lock_bh(&sk->sk_callback_lock); 45 sk->sk_state_change = siw_cm_llp_state_change; 46 sk->sk_data_ready = siw_cm_llp_data_ready; 47 sk->sk_write_space = siw_cm_llp_write_space; 48 sk->sk_error_report = siw_cm_llp_error_report; 49 write_unlock_bh(&sk->sk_callback_lock); 50 } 51 52 static void siw_sk_save_upcalls(struct sock *sk) 53 { 54 struct siw_cep *cep = sk_to_cep(sk); 55 56 write_lock_bh(&sk->sk_callback_lock); 57 cep->sk_state_change = sk->sk_state_change; 58 cep->sk_data_ready = sk->sk_data_ready; 59 cep->sk_write_space = sk->sk_write_space; 60 cep->sk_error_report = sk->sk_error_report; 61 write_unlock_bh(&sk->sk_callback_lock); 62 } 63 64 static void siw_sk_restore_upcalls(struct sock *sk, struct siw_cep *cep) 65 { 66 sk->sk_state_change = cep->sk_state_change; 67 sk->sk_data_ready = cep->sk_data_ready; 68 sk->sk_write_space = cep->sk_write_space; 69 sk->sk_error_report = cep->sk_error_report; 70 sk->sk_user_data = NULL; 71 } 72 73 static void siw_qp_socket_assoc(struct siw_cep *cep, struct siw_qp *qp) 74 { 75 struct socket *s = cep->sock; 76 struct sock *sk = s->sk; 77 78 write_lock_bh(&sk->sk_callback_lock); 79 80 qp->attrs.sk = s; 81 sk->sk_data_ready = siw_qp_llp_data_ready; 82 sk->sk_write_space = siw_qp_llp_write_space; 83 84 write_unlock_bh(&sk->sk_callback_lock); 85 } 86 87 static void siw_socket_disassoc(struct socket *s) 88 { 89 struct sock *sk = s->sk; 90 struct siw_cep *cep; 91 92 if (sk) { 93 write_lock_bh(&sk->sk_callback_lock); 94 cep = sk_to_cep(sk); 95 if (cep) { 96 siw_sk_restore_upcalls(sk, cep); 97 siw_cep_put(cep); 98 } else { 99 pr_warn("siw: cannot restore sk callbacks: no ep\n"); 100 } 101 write_unlock_bh(&sk->sk_callback_lock); 102 } else { 103 pr_warn("siw: cannot restore sk callbacks: no sk\n"); 104 } 105 } 106 107 static void siw_rtr_data_ready(struct sock *sk) 108 { 109 struct siw_cep *cep; 110 struct siw_qp *qp = NULL; 111 read_descriptor_t rd_desc; 112 113 trace_sk_data_ready(sk); 114 115 read_lock(&sk->sk_callback_lock); 116 117 cep = sk_to_cep(sk); 118 if (!cep) { 119 WARN(1, "No connection endpoint\n"); 120 goto out; 121 } 122 qp = sk_to_qp(sk); 123 124 memset(&rd_desc, 0, sizeof(rd_desc)); 125 rd_desc.arg.data = qp; 126 rd_desc.count = 1; 127 128 tcp_read_sock(sk, &rd_desc, siw_tcp_rx_data); 129 /* 130 * Check if first frame was successfully processed. 131 * Signal connection full establishment if yes. 132 * Failed data processing would have already scheduled 133 * connection drop. 134 */ 135 if (!qp->rx_stream.rx_suspend) 136 siw_cm_upcall(cep, IW_CM_EVENT_ESTABLISHED, 0); 137 out: 138 read_unlock(&sk->sk_callback_lock); 139 if (qp) 140 siw_qp_socket_assoc(cep, qp); 141 } 142 143 static void siw_sk_assign_rtr_upcalls(struct siw_cep *cep) 144 { 145 struct sock *sk = cep->sock->sk; 146 147 write_lock_bh(&sk->sk_callback_lock); 148 sk->sk_data_ready = siw_rtr_data_ready; 149 sk->sk_write_space = siw_qp_llp_write_space; 150 write_unlock_bh(&sk->sk_callback_lock); 151 } 152 153 static void siw_cep_socket_assoc(struct siw_cep *cep, struct socket *s) 154 { 155 cep->sock = s; 156 siw_cep_get(cep); 157 s->sk->sk_user_data = cep; 158 159 siw_sk_save_upcalls(s->sk); 160 siw_sk_assign_cm_upcalls(s->sk); 161 } 162 163 static struct siw_cep *siw_cep_alloc(struct siw_device *sdev) 164 { 165 struct siw_cep *cep = kzalloc(sizeof(*cep), GFP_KERNEL); 166 unsigned long flags; 167 168 if (!cep) 169 return NULL; 170 171 INIT_LIST_HEAD(&cep->listenq); 172 INIT_LIST_HEAD(&cep->devq); 173 INIT_LIST_HEAD(&cep->work_freelist); 174 175 kref_init(&cep->ref); 176 cep->state = SIW_EPSTATE_IDLE; 177 init_waitqueue_head(&cep->waitq); 178 spin_lock_init(&cep->lock); 179 cep->sdev = sdev; 180 cep->enhanced_rdma_conn_est = false; 181 182 spin_lock_irqsave(&sdev->lock, flags); 183 list_add_tail(&cep->devq, &sdev->cep_list); 184 spin_unlock_irqrestore(&sdev->lock, flags); 185 186 siw_dbg_cep(cep, "new endpoint\n"); 187 return cep; 188 } 189 190 static void siw_cm_free_work(struct siw_cep *cep) 191 { 192 struct list_head *w, *tmp; 193 struct siw_cm_work *work; 194 195 list_for_each_safe(w, tmp, &cep->work_freelist) { 196 work = list_entry(w, struct siw_cm_work, list); 197 list_del(&work->list); 198 kfree(work); 199 } 200 } 201 202 static void siw_cancel_mpatimer(struct siw_cep *cep) 203 { 204 spin_lock_bh(&cep->lock); 205 if (cep->mpa_timer) { 206 if (cancel_delayed_work(&cep->mpa_timer->work)) { 207 siw_cep_put(cep); 208 kfree(cep->mpa_timer); /* not needed again */ 209 } 210 cep->mpa_timer = NULL; 211 } 212 spin_unlock_bh(&cep->lock); 213 } 214 215 static void siw_put_work(struct siw_cm_work *work) 216 { 217 INIT_LIST_HEAD(&work->list); 218 spin_lock_bh(&work->cep->lock); 219 list_add(&work->list, &work->cep->work_freelist); 220 spin_unlock_bh(&work->cep->lock); 221 } 222 223 static void siw_cep_set_inuse(struct siw_cep *cep) 224 { 225 unsigned long flags; 226 retry: 227 spin_lock_irqsave(&cep->lock, flags); 228 229 if (cep->in_use) { 230 spin_unlock_irqrestore(&cep->lock, flags); 231 wait_event_interruptible(cep->waitq, !cep->in_use); 232 if (signal_pending(current)) 233 flush_signals(current); 234 goto retry; 235 } else { 236 cep->in_use = 1; 237 spin_unlock_irqrestore(&cep->lock, flags); 238 } 239 } 240 241 static void siw_cep_set_free(struct siw_cep *cep) 242 { 243 unsigned long flags; 244 245 spin_lock_irqsave(&cep->lock, flags); 246 cep->in_use = 0; 247 spin_unlock_irqrestore(&cep->lock, flags); 248 249 wake_up(&cep->waitq); 250 } 251 252 static void __siw_cep_dealloc(struct kref *ref) 253 { 254 struct siw_cep *cep = container_of(ref, struct siw_cep, ref); 255 struct siw_device *sdev = cep->sdev; 256 unsigned long flags; 257 258 WARN_ON(cep->listen_cep); 259 260 /* kfree(NULL) is safe */ 261 kfree(cep->mpa.pdata); 262 spin_lock_bh(&cep->lock); 263 if (!list_empty(&cep->work_freelist)) 264 siw_cm_free_work(cep); 265 spin_unlock_bh(&cep->lock); 266 267 spin_lock_irqsave(&sdev->lock, flags); 268 list_del(&cep->devq); 269 spin_unlock_irqrestore(&sdev->lock, flags); 270 271 siw_dbg_cep(cep, "free endpoint\n"); 272 kfree(cep); 273 } 274 275 static struct siw_cm_work *siw_get_work(struct siw_cep *cep) 276 { 277 struct siw_cm_work *work = NULL; 278 279 spin_lock_bh(&cep->lock); 280 if (!list_empty(&cep->work_freelist)) { 281 work = list_entry(cep->work_freelist.next, struct siw_cm_work, 282 list); 283 list_del_init(&work->list); 284 } 285 spin_unlock_bh(&cep->lock); 286 return work; 287 } 288 289 static int siw_cm_alloc_work(struct siw_cep *cep, int num) 290 { 291 struct siw_cm_work *work; 292 293 while (num--) { 294 work = kmalloc(sizeof(*work), GFP_KERNEL); 295 if (!work) { 296 if (!(list_empty(&cep->work_freelist))) 297 siw_cm_free_work(cep); 298 return -ENOMEM; 299 } 300 work->cep = cep; 301 INIT_LIST_HEAD(&work->list); 302 list_add(&work->list, &cep->work_freelist); 303 } 304 return 0; 305 } 306 307 /* 308 * siw_cm_upcall() 309 * 310 * Upcall to IWCM to inform about async connection events 311 */ 312 static int siw_cm_upcall(struct siw_cep *cep, enum iw_cm_event_type reason, 313 int status) 314 { 315 struct iw_cm_event event; 316 struct iw_cm_id *id; 317 318 memset(&event, 0, sizeof(event)); 319 event.status = status; 320 event.event = reason; 321 322 if (reason == IW_CM_EVENT_CONNECT_REQUEST) { 323 event.provider_data = cep; 324 id = cep->listen_cep->cm_id; 325 } else { 326 id = cep->cm_id; 327 } 328 /* Signal IRD and ORD */ 329 if (reason == IW_CM_EVENT_ESTABLISHED || 330 reason == IW_CM_EVENT_CONNECT_REPLY) { 331 /* Signal negotiated IRD/ORD values we will use */ 332 event.ird = cep->ird; 333 event.ord = cep->ord; 334 } else if (reason == IW_CM_EVENT_CONNECT_REQUEST) { 335 event.ird = cep->ord; 336 event.ord = cep->ird; 337 } 338 /* Signal private data and address information */ 339 if (reason == IW_CM_EVENT_CONNECT_REQUEST || 340 reason == IW_CM_EVENT_CONNECT_REPLY) { 341 u16 pd_len = be16_to_cpu(cep->mpa.hdr.params.pd_len); 342 343 if (pd_len) { 344 /* 345 * hand over MPA private data 346 */ 347 event.private_data_len = pd_len; 348 event.private_data = cep->mpa.pdata; 349 350 /* Hide MPA V2 IRD/ORD control */ 351 if (cep->enhanced_rdma_conn_est) { 352 event.private_data_len -= 353 sizeof(struct mpa_v2_data); 354 event.private_data += 355 sizeof(struct mpa_v2_data); 356 } 357 } 358 getname_local(cep->sock, &event.local_addr); 359 getname_peer(cep->sock, &event.remote_addr); 360 } 361 siw_dbg_cep(cep, "[QP %u]: reason=%d, status=%d\n", 362 cep->qp ? qp_id(cep->qp) : UINT_MAX, reason, status); 363 364 return id->event_handler(id, &event); 365 } 366 367 /* 368 * siw_qp_cm_drop() 369 * 370 * Drops established LLP connection if present and not already 371 * scheduled for dropping. Called from user context, SQ workqueue 372 * or receive IRQ. Caller signals if socket can be immediately 373 * closed (basically, if not in IRQ). 374 */ 375 void siw_qp_cm_drop(struct siw_qp *qp, int schedule) 376 { 377 struct siw_cep *cep = qp->cep; 378 379 qp->rx_stream.rx_suspend = 1; 380 qp->tx_ctx.tx_suspend = 1; 381 382 if (!qp->cep) 383 return; 384 385 if (schedule) { 386 siw_cm_queue_work(cep, SIW_CM_WORK_CLOSE_LLP); 387 } else { 388 siw_cep_set_inuse(cep); 389 390 if (cep->state == SIW_EPSTATE_CLOSED) { 391 siw_dbg_cep(cep, "already closed\n"); 392 goto out; 393 } 394 siw_dbg_cep(cep, "immediate close, state %d\n", cep->state); 395 396 if (qp->term_info.valid) 397 siw_send_terminate(qp); 398 399 if (cep->cm_id) { 400 switch (cep->state) { 401 case SIW_EPSTATE_AWAIT_MPAREP: 402 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, 403 -EINVAL); 404 break; 405 406 case SIW_EPSTATE_RDMA_MODE: 407 siw_cm_upcall(cep, IW_CM_EVENT_CLOSE, 0); 408 break; 409 410 case SIW_EPSTATE_IDLE: 411 case SIW_EPSTATE_LISTENING: 412 case SIW_EPSTATE_CONNECTING: 413 case SIW_EPSTATE_AWAIT_MPAREQ: 414 case SIW_EPSTATE_RECVD_MPAREQ: 415 case SIW_EPSTATE_CLOSED: 416 default: 417 break; 418 } 419 cep->cm_id->rem_ref(cep->cm_id); 420 cep->cm_id = NULL; 421 siw_cep_put(cep); 422 } 423 cep->state = SIW_EPSTATE_CLOSED; 424 425 if (cep->sock) { 426 siw_socket_disassoc(cep->sock); 427 /* 428 * Immediately close socket 429 */ 430 sock_release(cep->sock); 431 cep->sock = NULL; 432 } 433 if (cep->qp) { 434 cep->qp = NULL; 435 siw_qp_put(qp); 436 } 437 out: 438 siw_cep_set_free(cep); 439 } 440 } 441 442 void siw_cep_put(struct siw_cep *cep) 443 { 444 WARN_ON(kref_read(&cep->ref) < 1); 445 kref_put(&cep->ref, __siw_cep_dealloc); 446 } 447 448 void siw_cep_get(struct siw_cep *cep) 449 { 450 kref_get(&cep->ref); 451 } 452 453 /* 454 * Expects params->pd_len in host byte order 455 */ 456 static int siw_send_mpareqrep(struct siw_cep *cep, const void *pdata, u8 pd_len) 457 { 458 struct socket *s = cep->sock; 459 struct mpa_rr *rr = &cep->mpa.hdr; 460 struct kvec iov[3]; 461 struct msghdr msg; 462 int rv; 463 int iovec_num = 0; 464 int mpa_len; 465 466 memset(&msg, 0, sizeof(msg)); 467 468 iov[iovec_num].iov_base = rr; 469 iov[iovec_num].iov_len = sizeof(*rr); 470 mpa_len = sizeof(*rr); 471 472 if (cep->enhanced_rdma_conn_est) { 473 iovec_num++; 474 iov[iovec_num].iov_base = &cep->mpa.v2_ctrl; 475 iov[iovec_num].iov_len = sizeof(cep->mpa.v2_ctrl); 476 mpa_len += sizeof(cep->mpa.v2_ctrl); 477 } 478 if (pd_len) { 479 iovec_num++; 480 iov[iovec_num].iov_base = (char *)pdata; 481 iov[iovec_num].iov_len = pd_len; 482 mpa_len += pd_len; 483 } 484 if (cep->enhanced_rdma_conn_est) 485 pd_len += sizeof(cep->mpa.v2_ctrl); 486 487 rr->params.pd_len = cpu_to_be16(pd_len); 488 489 rv = kernel_sendmsg(s, &msg, iov, iovec_num + 1, mpa_len); 490 491 return rv < 0 ? rv : 0; 492 } 493 494 /* 495 * Receive MPA Request/Reply header. 496 * 497 * Returns 0 if complete MPA Request/Reply header including 498 * eventual private data was received. Returns -EAGAIN if 499 * header was partially received or negative error code otherwise. 500 * 501 * Context: May be called in process context only 502 */ 503 static int siw_recv_mpa_rr(struct siw_cep *cep) 504 { 505 struct mpa_rr *hdr = &cep->mpa.hdr; 506 struct socket *s = cep->sock; 507 u16 pd_len; 508 int rcvd, to_rcv; 509 510 if (cep->mpa.bytes_rcvd < sizeof(struct mpa_rr)) { 511 rcvd = ksock_recv(s, (char *)hdr + cep->mpa.bytes_rcvd, 512 sizeof(struct mpa_rr) - cep->mpa.bytes_rcvd, 513 0); 514 if (rcvd <= 0) 515 return -ECONNABORTED; 516 517 cep->mpa.bytes_rcvd += rcvd; 518 519 if (cep->mpa.bytes_rcvd < sizeof(struct mpa_rr)) 520 return -EAGAIN; 521 522 if (be16_to_cpu(hdr->params.pd_len) > MPA_MAX_PRIVDATA) 523 return -EPROTO; 524 } 525 pd_len = be16_to_cpu(hdr->params.pd_len); 526 527 /* 528 * At least the MPA Request/Reply header (frame not including 529 * private data) has been received. 530 * Receive (or continue receiving) any private data. 531 */ 532 to_rcv = pd_len - (cep->mpa.bytes_rcvd - sizeof(struct mpa_rr)); 533 534 if (!to_rcv) { 535 /* 536 * We must have hdr->params.pd_len == 0 and thus received a 537 * complete MPA Request/Reply frame. 538 * Check against peer protocol violation. 539 */ 540 u32 word; 541 542 rcvd = ksock_recv(s, (char *)&word, sizeof(word), MSG_DONTWAIT); 543 if (rcvd == -EAGAIN) 544 return 0; 545 546 if (rcvd == 0) { 547 siw_dbg_cep(cep, "peer EOF\n"); 548 return -EPIPE; 549 } 550 if (rcvd < 0) { 551 siw_dbg_cep(cep, "error: %d\n", rcvd); 552 return rcvd; 553 } 554 siw_dbg_cep(cep, "peer sent extra data: %d\n", rcvd); 555 556 return -EPROTO; 557 } 558 559 /* 560 * At this point, we must have hdr->params.pd_len != 0. 561 * A private data buffer gets allocated if hdr->params.pd_len != 0. 562 */ 563 if (!cep->mpa.pdata) { 564 cep->mpa.pdata = kmalloc(pd_len + 4, GFP_KERNEL); 565 if (!cep->mpa.pdata) 566 return -ENOMEM; 567 } 568 rcvd = ksock_recv( 569 s, cep->mpa.pdata + cep->mpa.bytes_rcvd - sizeof(struct mpa_rr), 570 to_rcv + 4, MSG_DONTWAIT); 571 572 if (rcvd < 0) 573 return rcvd; 574 575 if (rcvd > to_rcv) 576 return -EPROTO; 577 578 cep->mpa.bytes_rcvd += rcvd; 579 580 if (to_rcv == rcvd) { 581 siw_dbg_cep(cep, "%d bytes private data received\n", pd_len); 582 return 0; 583 } 584 return -EAGAIN; 585 } 586 587 /* 588 * siw_proc_mpareq() 589 * 590 * Read MPA Request from socket and signal new connection to IWCM 591 * if success. Caller must hold lock on corresponding listening CEP. 592 */ 593 static int siw_proc_mpareq(struct siw_cep *cep) 594 { 595 struct mpa_rr *req; 596 int version, rv; 597 u16 pd_len; 598 599 rv = siw_recv_mpa_rr(cep); 600 if (rv) 601 return rv; 602 603 req = &cep->mpa.hdr; 604 605 version = __mpa_rr_revision(req->params.bits); 606 pd_len = be16_to_cpu(req->params.pd_len); 607 608 if (version > MPA_REVISION_2) 609 /* allow for 0, 1, and 2 only */ 610 return -EPROTO; 611 612 if (memcmp(req->key, MPA_KEY_REQ, 16)) 613 return -EPROTO; 614 615 /* Prepare for sending MPA reply */ 616 memcpy(req->key, MPA_KEY_REP, 16); 617 618 if (version == MPA_REVISION_2 && 619 (req->params.bits & MPA_RR_FLAG_ENHANCED)) { 620 /* 621 * MPA version 2 must signal IRD/ORD values and P2P mode 622 * in private data if header flag MPA_RR_FLAG_ENHANCED 623 * is set. 624 */ 625 if (pd_len < sizeof(struct mpa_v2_data)) 626 goto reject_conn; 627 628 cep->enhanced_rdma_conn_est = true; 629 } 630 631 /* MPA Markers: currently not supported. Marker TX to be added. */ 632 if (req->params.bits & MPA_RR_FLAG_MARKERS) 633 goto reject_conn; 634 635 if (req->params.bits & MPA_RR_FLAG_CRC) { 636 /* 637 * RFC 5044, page 27: CRC MUST be used if peer requests it. 638 * siw specific: 'mpa_crc_strict' parameter to reject 639 * connection with CRC if local CRC off enforced by 640 * 'mpa_crc_strict' module parameter. 641 */ 642 if (!mpa_crc_required && mpa_crc_strict) 643 goto reject_conn; 644 645 /* Enable CRC if requested by module parameter */ 646 if (mpa_crc_required) 647 req->params.bits |= MPA_RR_FLAG_CRC; 648 } 649 if (cep->enhanced_rdma_conn_est) { 650 struct mpa_v2_data *v2 = (struct mpa_v2_data *)cep->mpa.pdata; 651 652 /* 653 * Peer requested ORD becomes requested local IRD, 654 * peer requested IRD becomes requested local ORD. 655 * IRD and ORD get limited by global maximum values. 656 */ 657 cep->ord = ntohs(v2->ird) & MPA_IRD_ORD_MASK; 658 cep->ord = min(cep->ord, SIW_MAX_ORD_QP); 659 cep->ird = ntohs(v2->ord) & MPA_IRD_ORD_MASK; 660 cep->ird = min(cep->ird, SIW_MAX_IRD_QP); 661 662 /* May get overwritten by locally negotiated values */ 663 cep->mpa.v2_ctrl.ird = htons(cep->ird); 664 cep->mpa.v2_ctrl.ord = htons(cep->ord); 665 666 /* 667 * Support for peer sent zero length Write or Read to 668 * let local side enter RTS. Writes are preferred. 669 * Sends would require pre-posting a Receive and are 670 * not supported. 671 * Propose zero length Write if none of Read and Write 672 * is indicated. 673 */ 674 if (v2->ird & MPA_V2_PEER_TO_PEER) { 675 cep->mpa.v2_ctrl.ird |= MPA_V2_PEER_TO_PEER; 676 677 if (v2->ord & MPA_V2_RDMA_WRITE_RTR) 678 cep->mpa.v2_ctrl.ord |= MPA_V2_RDMA_WRITE_RTR; 679 else if (v2->ord & MPA_V2_RDMA_READ_RTR) 680 cep->mpa.v2_ctrl.ord |= MPA_V2_RDMA_READ_RTR; 681 else 682 cep->mpa.v2_ctrl.ord |= MPA_V2_RDMA_WRITE_RTR; 683 } 684 } 685 686 cep->state = SIW_EPSTATE_RECVD_MPAREQ; 687 688 /* Keep reference until IWCM accepts/rejects */ 689 siw_cep_get(cep); 690 rv = siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REQUEST, 0); 691 if (rv) 692 siw_cep_put(cep); 693 694 return rv; 695 696 reject_conn: 697 siw_dbg_cep(cep, "reject: crc %d:%d:%d, m %d:%d\n", 698 req->params.bits & MPA_RR_FLAG_CRC ? 1 : 0, 699 mpa_crc_required, mpa_crc_strict, 700 req->params.bits & MPA_RR_FLAG_MARKERS ? 1 : 0, 0); 701 702 req->params.bits &= ~MPA_RR_FLAG_MARKERS; 703 req->params.bits |= MPA_RR_FLAG_REJECT; 704 705 if (!mpa_crc_required && mpa_crc_strict) 706 req->params.bits &= ~MPA_RR_FLAG_CRC; 707 708 if (pd_len) 709 kfree(cep->mpa.pdata); 710 711 cep->mpa.pdata = NULL; 712 713 siw_send_mpareqrep(cep, NULL, 0); 714 715 return -EOPNOTSUPP; 716 } 717 718 static int siw_proc_mpareply(struct siw_cep *cep) 719 { 720 struct siw_qp_attrs qp_attrs; 721 enum siw_qp_attr_mask qp_attr_mask; 722 struct siw_qp *qp = cep->qp; 723 struct mpa_rr *rep; 724 int rv; 725 u16 rep_ord; 726 u16 rep_ird; 727 bool ird_insufficient = false; 728 enum mpa_v2_ctrl mpa_p2p_mode = MPA_V2_RDMA_NO_RTR; 729 730 rv = siw_recv_mpa_rr(cep); 731 if (rv) 732 goto out_err; 733 734 siw_cancel_mpatimer(cep); 735 736 rep = &cep->mpa.hdr; 737 738 if (__mpa_rr_revision(rep->params.bits) > MPA_REVISION_2) { 739 /* allow for 0, 1, and 2 only */ 740 rv = -EPROTO; 741 goto out_err; 742 } 743 if (memcmp(rep->key, MPA_KEY_REP, 16)) { 744 siw_init_terminate(qp, TERM_ERROR_LAYER_LLP, LLP_ETYPE_MPA, 745 LLP_ECODE_INVALID_REQ_RESP, 0); 746 siw_send_terminate(qp); 747 rv = -EPROTO; 748 goto out_err; 749 } 750 if (rep->params.bits & MPA_RR_FLAG_REJECT) { 751 siw_dbg_cep(cep, "got mpa reject\n"); 752 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, -ECONNRESET); 753 754 return -ECONNRESET; 755 } 756 if (try_gso && rep->params.bits & MPA_RR_FLAG_GSO_EXP) { 757 siw_dbg_cep(cep, "peer allows GSO on TX\n"); 758 qp->tx_ctx.gso_seg_limit = 0; 759 } 760 if ((rep->params.bits & MPA_RR_FLAG_MARKERS) || 761 (mpa_crc_required && !(rep->params.bits & MPA_RR_FLAG_CRC)) || 762 (mpa_crc_strict && !mpa_crc_required && 763 (rep->params.bits & MPA_RR_FLAG_CRC))) { 764 siw_dbg_cep(cep, "reply unsupp: crc %d:%d:%d, m %d:%d\n", 765 rep->params.bits & MPA_RR_FLAG_CRC ? 1 : 0, 766 mpa_crc_required, mpa_crc_strict, 767 rep->params.bits & MPA_RR_FLAG_MARKERS ? 1 : 0, 0); 768 769 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, -ECONNREFUSED); 770 771 return -EINVAL; 772 } 773 if (cep->enhanced_rdma_conn_est) { 774 struct mpa_v2_data *v2; 775 776 if (__mpa_rr_revision(rep->params.bits) < MPA_REVISION_2 || 777 !(rep->params.bits & MPA_RR_FLAG_ENHANCED)) { 778 /* 779 * Protocol failure: The responder MUST reply with 780 * MPA version 2 and MUST set MPA_RR_FLAG_ENHANCED. 781 */ 782 siw_dbg_cep(cep, "mpa reply error: vers %d, enhcd %d\n", 783 __mpa_rr_revision(rep->params.bits), 784 rep->params.bits & MPA_RR_FLAG_ENHANCED ? 785 1 : 786 0); 787 788 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, 789 -ECONNRESET); 790 return -EINVAL; 791 } 792 v2 = (struct mpa_v2_data *)cep->mpa.pdata; 793 rep_ird = ntohs(v2->ird) & MPA_IRD_ORD_MASK; 794 rep_ord = ntohs(v2->ord) & MPA_IRD_ORD_MASK; 795 796 if (cep->ird < rep_ord && 797 (relaxed_ird_negotiation == false || 798 rep_ord > cep->sdev->attrs.max_ird)) { 799 siw_dbg_cep(cep, "ird %d, rep_ord %d, max_ord %d\n", 800 cep->ird, rep_ord, 801 cep->sdev->attrs.max_ord); 802 ird_insufficient = true; 803 } 804 if (cep->ord > rep_ird && relaxed_ird_negotiation == false) { 805 siw_dbg_cep(cep, "ord %d, rep_ird %d\n", cep->ord, 806 rep_ird); 807 ird_insufficient = true; 808 } 809 /* 810 * Always report negotiated peer values to user, 811 * even if IRD/ORD negotiation failed 812 */ 813 cep->ird = rep_ord; 814 cep->ord = rep_ird; 815 816 if (ird_insufficient) { 817 /* 818 * If the initiator IRD is insuffient for the 819 * responder ORD, send a TERM. 820 */ 821 siw_init_terminate(qp, TERM_ERROR_LAYER_LLP, 822 LLP_ETYPE_MPA, 823 LLP_ECODE_INSUFFICIENT_IRD, 0); 824 siw_send_terminate(qp); 825 rv = -ENOMEM; 826 goto out_err; 827 } 828 if (cep->mpa.v2_ctrl_req.ird & MPA_V2_PEER_TO_PEER) 829 mpa_p2p_mode = 830 cep->mpa.v2_ctrl_req.ord & 831 (MPA_V2_RDMA_WRITE_RTR | MPA_V2_RDMA_READ_RTR); 832 833 /* 834 * Check if we requested P2P mode, and if peer agrees 835 */ 836 if (mpa_p2p_mode != MPA_V2_RDMA_NO_RTR) { 837 if ((mpa_p2p_mode & v2->ord) == 0) { 838 /* 839 * We requested RTR mode(s), but the peer 840 * did not pick any mode we support. 841 */ 842 siw_dbg_cep(cep, 843 "rtr mode: req %2x, got %2x\n", 844 mpa_p2p_mode, 845 v2->ord & (MPA_V2_RDMA_WRITE_RTR | 846 MPA_V2_RDMA_READ_RTR)); 847 848 siw_init_terminate(qp, TERM_ERROR_LAYER_LLP, 849 LLP_ETYPE_MPA, 850 LLP_ECODE_NO_MATCHING_RTR, 851 0); 852 siw_send_terminate(qp); 853 rv = -EPROTO; 854 goto out_err; 855 } 856 mpa_p2p_mode = v2->ord & (MPA_V2_RDMA_WRITE_RTR | 857 MPA_V2_RDMA_READ_RTR); 858 } 859 } 860 memset(&qp_attrs, 0, sizeof(qp_attrs)); 861 862 if (rep->params.bits & MPA_RR_FLAG_CRC) 863 qp_attrs.flags = SIW_MPA_CRC; 864 865 qp_attrs.irq_size = cep->ird; 866 qp_attrs.orq_size = cep->ord; 867 qp_attrs.sk = cep->sock; 868 qp_attrs.state = SIW_QP_STATE_RTS; 869 870 qp_attr_mask = SIW_QP_ATTR_STATE | SIW_QP_ATTR_LLP_HANDLE | 871 SIW_QP_ATTR_ORD | SIW_QP_ATTR_IRD | SIW_QP_ATTR_MPA; 872 873 /* Move socket RX/TX under QP control */ 874 down_write(&qp->state_lock); 875 if (qp->attrs.state > SIW_QP_STATE_RTR) { 876 rv = -EINVAL; 877 up_write(&qp->state_lock); 878 goto out_err; 879 } 880 rv = siw_qp_modify(qp, &qp_attrs, qp_attr_mask); 881 882 siw_qp_socket_assoc(cep, qp); 883 884 up_write(&qp->state_lock); 885 886 /* Send extra RDMA frame to trigger peer RTS if negotiated */ 887 if (mpa_p2p_mode != MPA_V2_RDMA_NO_RTR) { 888 rv = siw_qp_mpa_rts(qp, mpa_p2p_mode); 889 if (rv) 890 goto out_err; 891 } 892 if (!rv) { 893 rv = siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, 0); 894 if (!rv) 895 cep->state = SIW_EPSTATE_RDMA_MODE; 896 897 return 0; 898 } 899 900 out_err: 901 if (rv != -EAGAIN) 902 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, -EINVAL); 903 904 return rv; 905 } 906 907 /* 908 * siw_accept_newconn - accept an incoming pending connection 909 * 910 */ 911 static void siw_accept_newconn(struct siw_cep *cep) 912 { 913 struct socket *s = cep->sock; 914 struct socket *new_s = NULL; 915 struct siw_cep *new_cep = NULL; 916 int rv = 0; /* debug only. should disappear */ 917 918 if (cep->state != SIW_EPSTATE_LISTENING) 919 goto error; 920 921 new_cep = siw_cep_alloc(cep->sdev); 922 if (!new_cep) 923 goto error; 924 925 /* 926 * 4: Allocate a sufficient number of work elements 927 * to allow concurrent handling of local + peer close 928 * events, MPA header processing + MPA timeout. 929 */ 930 if (siw_cm_alloc_work(new_cep, 4) != 0) 931 goto error; 932 933 /* 934 * Copy saved socket callbacks from listening CEP 935 * and assign new socket with new CEP 936 */ 937 new_cep->sk_state_change = cep->sk_state_change; 938 new_cep->sk_data_ready = cep->sk_data_ready; 939 new_cep->sk_write_space = cep->sk_write_space; 940 new_cep->sk_error_report = cep->sk_error_report; 941 942 rv = kernel_accept(s, &new_s, O_NONBLOCK); 943 if (rv != 0) { 944 /* 945 * Connection already aborted by peer..? 946 */ 947 siw_dbg_cep(cep, "kernel_accept() error: %d\n", rv); 948 goto error; 949 } 950 new_cep->sock = new_s; 951 siw_cep_get(new_cep); 952 new_s->sk->sk_user_data = new_cep; 953 954 if (siw_tcp_nagle == false) 955 tcp_sock_set_nodelay(new_s->sk); 956 new_cep->state = SIW_EPSTATE_AWAIT_MPAREQ; 957 958 rv = siw_cm_queue_work(new_cep, SIW_CM_WORK_MPATIMEOUT); 959 if (rv) 960 goto error; 961 /* 962 * See siw_proc_mpareq() etc. for the use of new_cep->listen_cep. 963 */ 964 new_cep->listen_cep = cep; 965 siw_cep_get(cep); 966 967 if (atomic_read(&new_s->sk->sk_rmem_alloc)) { 968 /* 969 * MPA REQ already queued 970 */ 971 siw_dbg_cep(cep, "immediate mpa request\n"); 972 973 siw_cep_set_inuse(new_cep); 974 rv = siw_proc_mpareq(new_cep); 975 if (rv != -EAGAIN) { 976 siw_cep_put(cep); 977 new_cep->listen_cep = NULL; 978 if (rv) { 979 siw_cancel_mpatimer(new_cep); 980 siw_cep_set_free(new_cep); 981 goto error; 982 } 983 } 984 siw_cep_set_free(new_cep); 985 } 986 return; 987 988 error: 989 if (new_cep) 990 siw_cep_put(new_cep); 991 992 if (new_s) { 993 siw_socket_disassoc(new_s); 994 sock_release(new_s); 995 new_cep->sock = NULL; 996 } 997 siw_dbg_cep(cep, "error %d\n", rv); 998 } 999 1000 static void siw_cm_work_handler(struct work_struct *w) 1001 { 1002 struct siw_cm_work *work; 1003 struct siw_cep *cep; 1004 int release_cep = 0, rv = 0; 1005 1006 work = container_of(w, struct siw_cm_work, work.work); 1007 cep = work->cep; 1008 1009 siw_dbg_cep(cep, "[QP %u]: work type: %d, state %d\n", 1010 cep->qp ? qp_id(cep->qp) : UINT_MAX, 1011 work->type, cep->state); 1012 1013 siw_cep_set_inuse(cep); 1014 1015 switch (work->type) { 1016 case SIW_CM_WORK_ACCEPT: 1017 siw_accept_newconn(cep); 1018 break; 1019 1020 case SIW_CM_WORK_READ_MPAHDR: 1021 if (cep->state == SIW_EPSTATE_AWAIT_MPAREQ) { 1022 if (cep->listen_cep) { 1023 siw_cep_set_inuse(cep->listen_cep); 1024 1025 if (cep->listen_cep->state == 1026 SIW_EPSTATE_LISTENING) 1027 rv = siw_proc_mpareq(cep); 1028 else 1029 rv = -EFAULT; 1030 1031 siw_cep_set_free(cep->listen_cep); 1032 1033 if (rv != -EAGAIN) { 1034 siw_cep_put(cep->listen_cep); 1035 cep->listen_cep = NULL; 1036 if (rv) 1037 siw_cep_put(cep); 1038 } 1039 } 1040 } else if (cep->state == SIW_EPSTATE_AWAIT_MPAREP) { 1041 rv = siw_proc_mpareply(cep); 1042 } else { 1043 /* 1044 * CEP already moved out of MPA handshake. 1045 * any connection management already done. 1046 * silently ignore the mpa packet. 1047 */ 1048 if (cep->state == SIW_EPSTATE_RDMA_MODE) { 1049 cep->sock->sk->sk_data_ready(cep->sock->sk); 1050 siw_dbg_cep(cep, "already in RDMA mode"); 1051 } else { 1052 siw_dbg_cep(cep, "out of state: %d\n", 1053 cep->state); 1054 } 1055 } 1056 if (rv && rv != -EAGAIN) 1057 release_cep = 1; 1058 break; 1059 1060 case SIW_CM_WORK_CLOSE_LLP: 1061 /* 1062 * QP scheduled LLP close 1063 */ 1064 if (cep->qp && cep->qp->term_info.valid) 1065 siw_send_terminate(cep->qp); 1066 1067 if (cep->cm_id) 1068 siw_cm_upcall(cep, IW_CM_EVENT_CLOSE, 0); 1069 1070 release_cep = 1; 1071 break; 1072 1073 case SIW_CM_WORK_PEER_CLOSE: 1074 if (cep->cm_id) { 1075 if (cep->state == SIW_EPSTATE_AWAIT_MPAREP) { 1076 /* 1077 * MPA reply not received, but connection drop 1078 */ 1079 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, 1080 -ECONNRESET); 1081 } else if (cep->state == SIW_EPSTATE_RDMA_MODE) { 1082 /* 1083 * NOTE: IW_CM_EVENT_DISCONNECT is given just 1084 * to transition IWCM into CLOSING. 1085 */ 1086 siw_cm_upcall(cep, IW_CM_EVENT_DISCONNECT, 0); 1087 siw_cm_upcall(cep, IW_CM_EVENT_CLOSE, 0); 1088 } 1089 /* 1090 * for other states there is no connection 1091 * known to the IWCM. 1092 */ 1093 } else { 1094 if (cep->state == SIW_EPSTATE_RECVD_MPAREQ) { 1095 /* 1096 * Wait for the ulp/CM to call accept/reject 1097 */ 1098 siw_dbg_cep(cep, 1099 "mpa req recvd, wait for ULP\n"); 1100 } else if (cep->state == SIW_EPSTATE_AWAIT_MPAREQ) { 1101 /* 1102 * Socket close before MPA request received. 1103 */ 1104 if (cep->listen_cep) { 1105 siw_dbg_cep(cep, 1106 "no mpareq: drop listener\n"); 1107 siw_cep_put(cep->listen_cep); 1108 cep->listen_cep = NULL; 1109 } 1110 } 1111 } 1112 release_cep = 1; 1113 break; 1114 1115 case SIW_CM_WORK_MPATIMEOUT: 1116 cep->mpa_timer = NULL; 1117 1118 if (cep->state == SIW_EPSTATE_AWAIT_MPAREP) { 1119 /* 1120 * MPA request timed out: 1121 * Hide any partially received private data and signal 1122 * timeout 1123 */ 1124 cep->mpa.hdr.params.pd_len = 0; 1125 1126 if (cep->cm_id) 1127 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, 1128 -ETIMEDOUT); 1129 release_cep = 1; 1130 1131 } else if (cep->state == SIW_EPSTATE_AWAIT_MPAREQ) { 1132 /* 1133 * No MPA request received after peer TCP stream setup. 1134 */ 1135 if (cep->listen_cep) { 1136 siw_cep_put(cep->listen_cep); 1137 cep->listen_cep = NULL; 1138 } 1139 release_cep = 1; 1140 } 1141 break; 1142 1143 default: 1144 WARN(1, "Undefined CM work type: %d\n", work->type); 1145 } 1146 if (release_cep) { 1147 siw_dbg_cep(cep, 1148 "release: timer=%s, QP[%u]\n", 1149 cep->mpa_timer ? "y" : "n", 1150 cep->qp ? qp_id(cep->qp) : UINT_MAX); 1151 1152 siw_cancel_mpatimer(cep); 1153 1154 cep->state = SIW_EPSTATE_CLOSED; 1155 1156 if (cep->qp) { 1157 struct siw_qp *qp = cep->qp; 1158 /* 1159 * Serialize a potential race with application 1160 * closing the QP and calling siw_qp_cm_drop() 1161 */ 1162 siw_qp_get(qp); 1163 siw_cep_set_free(cep); 1164 1165 siw_qp_llp_close(qp); 1166 siw_qp_put(qp); 1167 1168 siw_cep_set_inuse(cep); 1169 cep->qp = NULL; 1170 siw_qp_put(qp); 1171 } 1172 if (cep->sock) { 1173 siw_socket_disassoc(cep->sock); 1174 sock_release(cep->sock); 1175 cep->sock = NULL; 1176 } 1177 if (cep->cm_id) { 1178 cep->cm_id->rem_ref(cep->cm_id); 1179 cep->cm_id = NULL; 1180 siw_cep_put(cep); 1181 } 1182 } 1183 siw_cep_set_free(cep); 1184 siw_put_work(work); 1185 siw_cep_put(cep); 1186 } 1187 1188 static struct workqueue_struct *siw_cm_wq; 1189 1190 int siw_cm_queue_work(struct siw_cep *cep, enum siw_work_type type) 1191 { 1192 struct siw_cm_work *work = siw_get_work(cep); 1193 unsigned long delay = 0; 1194 1195 if (!work) { 1196 siw_dbg_cep(cep, "failed with no work available\n"); 1197 return -ENOMEM; 1198 } 1199 work->type = type; 1200 work->cep = cep; 1201 1202 siw_cep_get(cep); 1203 1204 INIT_DELAYED_WORK(&work->work, siw_cm_work_handler); 1205 1206 if (type == SIW_CM_WORK_MPATIMEOUT) { 1207 cep->mpa_timer = work; 1208 1209 if (cep->state == SIW_EPSTATE_AWAIT_MPAREP) 1210 delay = MPAREQ_TIMEOUT; 1211 else 1212 delay = MPAREP_TIMEOUT; 1213 } 1214 siw_dbg_cep(cep, "[QP %u]: work type: %d, timeout %lu\n", 1215 cep->qp ? qp_id(cep->qp) : -1, type, delay); 1216 1217 queue_delayed_work(siw_cm_wq, &work->work, delay); 1218 1219 return 0; 1220 } 1221 1222 static void siw_cm_llp_data_ready(struct sock *sk) 1223 { 1224 struct siw_cep *cep; 1225 1226 trace_sk_data_ready(sk); 1227 1228 read_lock(&sk->sk_callback_lock); 1229 1230 cep = sk_to_cep(sk); 1231 if (!cep) 1232 goto out; 1233 1234 siw_dbg_cep(cep, "cep state: %d, socket state %d\n", 1235 cep->state, sk->sk_state); 1236 1237 if (sk->sk_state != TCP_ESTABLISHED) 1238 goto out; 1239 1240 switch (cep->state) { 1241 case SIW_EPSTATE_RDMA_MODE: 1242 case SIW_EPSTATE_LISTENING: 1243 break; 1244 1245 case SIW_EPSTATE_AWAIT_MPAREQ: 1246 case SIW_EPSTATE_AWAIT_MPAREP: 1247 siw_cm_queue_work(cep, SIW_CM_WORK_READ_MPAHDR); 1248 break; 1249 1250 default: 1251 siw_dbg_cep(cep, "unexpected data, state %d\n", cep->state); 1252 break; 1253 } 1254 out: 1255 read_unlock(&sk->sk_callback_lock); 1256 } 1257 1258 static void siw_cm_llp_write_space(struct sock *sk) 1259 { 1260 struct siw_cep *cep = sk_to_cep(sk); 1261 1262 if (cep) 1263 siw_dbg_cep(cep, "state: %d\n", cep->state); 1264 } 1265 1266 static void siw_cm_llp_error_report(struct sock *sk) 1267 { 1268 struct siw_cep *cep = sk_to_cep(sk); 1269 1270 if (cep) { 1271 siw_dbg_cep(cep, "error %d, socket state: %d, cep state: %d\n", 1272 sk->sk_err, sk->sk_state, cep->state); 1273 cep->sk_error_report(sk); 1274 } 1275 } 1276 1277 static void siw_cm_llp_state_change(struct sock *sk) 1278 { 1279 struct siw_cep *cep; 1280 void (*orig_state_change)(struct sock *s); 1281 1282 read_lock(&sk->sk_callback_lock); 1283 1284 cep = sk_to_cep(sk); 1285 if (!cep) { 1286 /* endpoint already disassociated */ 1287 read_unlock(&sk->sk_callback_lock); 1288 return; 1289 } 1290 orig_state_change = cep->sk_state_change; 1291 1292 siw_dbg_cep(cep, "state: %d\n", cep->state); 1293 1294 switch (sk->sk_state) { 1295 case TCP_ESTABLISHED: 1296 /* 1297 * handle accepting socket as special case where only 1298 * new connection is possible 1299 */ 1300 siw_cm_queue_work(cep, SIW_CM_WORK_ACCEPT); 1301 break; 1302 1303 case TCP_CLOSE: 1304 case TCP_CLOSE_WAIT: 1305 if (cep->qp) 1306 cep->qp->tx_ctx.tx_suspend = 1; 1307 siw_cm_queue_work(cep, SIW_CM_WORK_PEER_CLOSE); 1308 break; 1309 1310 default: 1311 siw_dbg_cep(cep, "unexpected socket state %d\n", sk->sk_state); 1312 } 1313 read_unlock(&sk->sk_callback_lock); 1314 orig_state_change(sk); 1315 } 1316 1317 static int kernel_bindconnect(struct socket *s, struct sockaddr *laddr, 1318 struct sockaddr *raddr, bool afonly) 1319 { 1320 int rv, flags = 0; 1321 size_t size = laddr->sa_family == AF_INET ? 1322 sizeof(struct sockaddr_in) : sizeof(struct sockaddr_in6); 1323 1324 /* 1325 * Make address available again asap. 1326 */ 1327 sock_set_reuseaddr(s->sk); 1328 1329 if (afonly) { 1330 rv = ip6_sock_set_v6only(s->sk); 1331 if (rv) 1332 return rv; 1333 } 1334 1335 rv = s->ops->bind(s, laddr, size); 1336 if (rv < 0) 1337 return rv; 1338 1339 rv = s->ops->connect(s, raddr, size, flags); 1340 1341 return rv < 0 ? rv : 0; 1342 } 1343 1344 int siw_connect(struct iw_cm_id *id, struct iw_cm_conn_param *params) 1345 { 1346 struct siw_device *sdev = to_siw_dev(id->device); 1347 struct siw_qp *qp; 1348 struct siw_cep *cep = NULL; 1349 struct socket *s = NULL; 1350 struct sockaddr *laddr = (struct sockaddr *)&id->local_addr, 1351 *raddr = (struct sockaddr *)&id->remote_addr; 1352 bool p2p_mode = peer_to_peer, v4 = true; 1353 u16 pd_len = params->private_data_len; 1354 int version = mpa_version, rv; 1355 1356 if (pd_len > MPA_MAX_PRIVDATA) 1357 return -EINVAL; 1358 1359 if (params->ird > sdev->attrs.max_ird || 1360 params->ord > sdev->attrs.max_ord) 1361 return -ENOMEM; 1362 1363 if (laddr->sa_family == AF_INET6) 1364 v4 = false; 1365 else if (laddr->sa_family != AF_INET) 1366 return -EAFNOSUPPORT; 1367 1368 /* 1369 * Respect any iwarp port mapping: Use mapped remote address 1370 * if valid. Local address must not be mapped, since siw 1371 * uses kernel TCP stack. 1372 */ 1373 if ((v4 && to_sockaddr_in(id->remote_addr).sin_port != 0) || 1374 to_sockaddr_in6(id->remote_addr).sin6_port != 0) 1375 raddr = (struct sockaddr *)&id->m_remote_addr; 1376 1377 qp = siw_qp_id2obj(sdev, params->qpn); 1378 if (!qp) { 1379 WARN(1, "[QP %u] does not exist\n", params->qpn); 1380 rv = -EINVAL; 1381 goto error; 1382 } 1383 siw_dbg_qp(qp, "pd_len %d, laddr %pISp, raddr %pISp\n", pd_len, laddr, 1384 raddr); 1385 1386 rv = sock_create(v4 ? AF_INET : AF_INET6, SOCK_STREAM, IPPROTO_TCP, &s); 1387 if (rv < 0) 1388 goto error; 1389 1390 /* 1391 * NOTE: For simplification, connect() is called in blocking 1392 * mode. Might be reconsidered for async connection setup at 1393 * TCP level. 1394 */ 1395 rv = kernel_bindconnect(s, laddr, raddr, id->afonly); 1396 if (rv != 0) { 1397 siw_dbg_qp(qp, "kernel_bindconnect: error %d\n", rv); 1398 goto error; 1399 } 1400 if (siw_tcp_nagle == false) 1401 tcp_sock_set_nodelay(s->sk); 1402 cep = siw_cep_alloc(sdev); 1403 if (!cep) { 1404 rv = -ENOMEM; 1405 goto error; 1406 } 1407 siw_cep_set_inuse(cep); 1408 1409 /* Associate QP with CEP */ 1410 siw_cep_get(cep); 1411 qp->cep = cep; 1412 1413 /* siw_qp_get(qp) already done by QP lookup */ 1414 cep->qp = qp; 1415 1416 id->add_ref(id); 1417 cep->cm_id = id; 1418 1419 /* 1420 * 4: Allocate a sufficient number of work elements 1421 * to allow concurrent handling of local + peer close 1422 * events, MPA header processing + MPA timeout. 1423 */ 1424 rv = siw_cm_alloc_work(cep, 4); 1425 if (rv != 0) { 1426 rv = -ENOMEM; 1427 goto error; 1428 } 1429 cep->ird = params->ird; 1430 cep->ord = params->ord; 1431 1432 if (p2p_mode && cep->ord == 0) 1433 cep->ord = 1; 1434 1435 cep->state = SIW_EPSTATE_CONNECTING; 1436 1437 /* 1438 * Associate CEP with socket 1439 */ 1440 siw_cep_socket_assoc(cep, s); 1441 1442 cep->state = SIW_EPSTATE_AWAIT_MPAREP; 1443 1444 /* 1445 * Set MPA Request bits: CRC if required, no MPA Markers, 1446 * MPA Rev. according to module parameter 'mpa_version', Key 'Request'. 1447 */ 1448 cep->mpa.hdr.params.bits = 0; 1449 if (version > MPA_REVISION_2) { 1450 pr_warn("Setting MPA version to %u\n", MPA_REVISION_2); 1451 version = MPA_REVISION_2; 1452 /* Adjust also module parameter */ 1453 mpa_version = MPA_REVISION_2; 1454 } 1455 __mpa_rr_set_revision(&cep->mpa.hdr.params.bits, version); 1456 1457 if (try_gso) 1458 cep->mpa.hdr.params.bits |= MPA_RR_FLAG_GSO_EXP; 1459 1460 if (mpa_crc_required) 1461 cep->mpa.hdr.params.bits |= MPA_RR_FLAG_CRC; 1462 1463 /* 1464 * If MPA version == 2: 1465 * o Include ORD and IRD. 1466 * o Indicate peer-to-peer mode, if required by module 1467 * parameter 'peer_to_peer'. 1468 */ 1469 if (version == MPA_REVISION_2) { 1470 cep->enhanced_rdma_conn_est = true; 1471 cep->mpa.hdr.params.bits |= MPA_RR_FLAG_ENHANCED; 1472 1473 cep->mpa.v2_ctrl.ird = htons(cep->ird); 1474 cep->mpa.v2_ctrl.ord = htons(cep->ord); 1475 1476 if (p2p_mode) { 1477 cep->mpa.v2_ctrl.ird |= MPA_V2_PEER_TO_PEER; 1478 cep->mpa.v2_ctrl.ord |= rtr_type; 1479 } 1480 /* Remember own P2P mode requested */ 1481 cep->mpa.v2_ctrl_req.ird = cep->mpa.v2_ctrl.ird; 1482 cep->mpa.v2_ctrl_req.ord = cep->mpa.v2_ctrl.ord; 1483 } 1484 memcpy(cep->mpa.hdr.key, MPA_KEY_REQ, 16); 1485 1486 rv = siw_send_mpareqrep(cep, params->private_data, pd_len); 1487 /* 1488 * Reset private data. 1489 */ 1490 cep->mpa.hdr.params.pd_len = 0; 1491 1492 if (rv >= 0) { 1493 rv = siw_cm_queue_work(cep, SIW_CM_WORK_MPATIMEOUT); 1494 if (!rv) { 1495 siw_dbg_cep(cep, "[QP %u]: exit\n", qp_id(qp)); 1496 siw_cep_set_free(cep); 1497 return 0; 1498 } 1499 } 1500 error: 1501 siw_dbg(id->device, "failed: %d\n", rv); 1502 1503 if (cep) { 1504 siw_socket_disassoc(s); 1505 sock_release(s); 1506 cep->sock = NULL; 1507 1508 cep->qp = NULL; 1509 1510 cep->cm_id = NULL; 1511 id->rem_ref(id); 1512 1513 qp->cep = NULL; 1514 siw_cep_put(cep); 1515 1516 cep->state = SIW_EPSTATE_CLOSED; 1517 1518 siw_cep_set_free(cep); 1519 1520 siw_cep_put(cep); 1521 1522 } else if (s) { 1523 sock_release(s); 1524 } 1525 if (qp) 1526 siw_qp_put(qp); 1527 1528 return rv; 1529 } 1530 1531 /* 1532 * siw_accept - Let SoftiWARP accept an RDMA connection request 1533 * 1534 * @id: New connection management id to be used for accepted 1535 * connection request 1536 * @params: Connection parameters provided by ULP for accepting connection 1537 * 1538 * Transition QP to RTS state, associate new CM id @id with accepted CEP 1539 * and get prepared for TCP input by installing socket callbacks. 1540 * Then send MPA Reply and generate the "connection established" event. 1541 * Socket callbacks must be installed before sending MPA Reply, because 1542 * the latter may cause a first RDMA message to arrive from the RDMA Initiator 1543 * side very quickly, at which time the socket callbacks must be ready. 1544 */ 1545 int siw_accept(struct iw_cm_id *id, struct iw_cm_conn_param *params) 1546 { 1547 struct siw_device *sdev = to_siw_dev(id->device); 1548 struct siw_cep *cep = (struct siw_cep *)id->provider_data; 1549 struct siw_qp *qp; 1550 struct siw_qp_attrs qp_attrs; 1551 int rv, max_priv_data = MPA_MAX_PRIVDATA; 1552 bool wait_for_peer_rts = false; 1553 1554 siw_cep_set_inuse(cep); 1555 siw_cep_put(cep); 1556 1557 /* Free lingering inbound private data */ 1558 if (cep->mpa.hdr.params.pd_len) { 1559 cep->mpa.hdr.params.pd_len = 0; 1560 kfree(cep->mpa.pdata); 1561 cep->mpa.pdata = NULL; 1562 } 1563 siw_cancel_mpatimer(cep); 1564 1565 if (cep->state != SIW_EPSTATE_RECVD_MPAREQ) { 1566 siw_dbg_cep(cep, "out of state\n"); 1567 1568 siw_cep_set_free(cep); 1569 siw_cep_put(cep); 1570 1571 return -ECONNRESET; 1572 } 1573 qp = siw_qp_id2obj(sdev, params->qpn); 1574 if (!qp) { 1575 WARN(1, "[QP %d] does not exist\n", params->qpn); 1576 siw_cep_set_free(cep); 1577 siw_cep_put(cep); 1578 1579 return -EINVAL; 1580 } 1581 down_write(&qp->state_lock); 1582 if (qp->attrs.state > SIW_QP_STATE_RTR) { 1583 rv = -EINVAL; 1584 up_write(&qp->state_lock); 1585 goto error; 1586 } 1587 siw_dbg_cep(cep, "[QP %d]\n", params->qpn); 1588 1589 if (try_gso && cep->mpa.hdr.params.bits & MPA_RR_FLAG_GSO_EXP) { 1590 siw_dbg_cep(cep, "peer allows GSO on TX\n"); 1591 qp->tx_ctx.gso_seg_limit = 0; 1592 } 1593 if (params->ord > sdev->attrs.max_ord || 1594 params->ird > sdev->attrs.max_ird) { 1595 siw_dbg_cep( 1596 cep, 1597 "[QP %u]: ord %d (max %d), ird %d (max %d)\n", 1598 qp_id(qp), params->ord, sdev->attrs.max_ord, 1599 params->ird, sdev->attrs.max_ird); 1600 rv = -EINVAL; 1601 up_write(&qp->state_lock); 1602 goto error; 1603 } 1604 if (cep->enhanced_rdma_conn_est) 1605 max_priv_data -= sizeof(struct mpa_v2_data); 1606 1607 if (params->private_data_len > max_priv_data) { 1608 siw_dbg_cep( 1609 cep, 1610 "[QP %u]: private data length: %d (max %d)\n", 1611 qp_id(qp), params->private_data_len, max_priv_data); 1612 rv = -EINVAL; 1613 up_write(&qp->state_lock); 1614 goto error; 1615 } 1616 if (cep->enhanced_rdma_conn_est) { 1617 if (params->ord > cep->ord) { 1618 if (relaxed_ird_negotiation) { 1619 params->ord = cep->ord; 1620 } else { 1621 cep->ird = params->ird; 1622 cep->ord = params->ord; 1623 rv = -EINVAL; 1624 up_write(&qp->state_lock); 1625 goto error; 1626 } 1627 } 1628 if (params->ird < cep->ird) { 1629 if (relaxed_ird_negotiation && 1630 cep->ird <= sdev->attrs.max_ird) 1631 params->ird = cep->ird; 1632 else { 1633 rv = -ENOMEM; 1634 up_write(&qp->state_lock); 1635 goto error; 1636 } 1637 } 1638 if (cep->mpa.v2_ctrl.ord & 1639 (MPA_V2_RDMA_WRITE_RTR | MPA_V2_RDMA_READ_RTR)) 1640 wait_for_peer_rts = true; 1641 /* 1642 * Signal back negotiated IRD and ORD values 1643 */ 1644 cep->mpa.v2_ctrl.ord = 1645 htons(params->ord & MPA_IRD_ORD_MASK) | 1646 (cep->mpa.v2_ctrl.ord & ~MPA_V2_MASK_IRD_ORD); 1647 cep->mpa.v2_ctrl.ird = 1648 htons(params->ird & MPA_IRD_ORD_MASK) | 1649 (cep->mpa.v2_ctrl.ird & ~MPA_V2_MASK_IRD_ORD); 1650 } 1651 cep->ird = params->ird; 1652 cep->ord = params->ord; 1653 1654 cep->cm_id = id; 1655 id->add_ref(id); 1656 1657 memset(&qp_attrs, 0, sizeof(qp_attrs)); 1658 qp_attrs.orq_size = cep->ord; 1659 qp_attrs.irq_size = cep->ird; 1660 qp_attrs.sk = cep->sock; 1661 if (cep->mpa.hdr.params.bits & MPA_RR_FLAG_CRC) 1662 qp_attrs.flags = SIW_MPA_CRC; 1663 qp_attrs.state = SIW_QP_STATE_RTS; 1664 1665 siw_dbg_cep(cep, "[QP%u]: moving to rts\n", qp_id(qp)); 1666 1667 /* Associate QP with CEP */ 1668 siw_cep_get(cep); 1669 qp->cep = cep; 1670 1671 /* siw_qp_get(qp) already done by QP lookup */ 1672 cep->qp = qp; 1673 1674 cep->state = SIW_EPSTATE_RDMA_MODE; 1675 1676 /* Move socket RX/TX under QP control */ 1677 rv = siw_qp_modify(qp, &qp_attrs, 1678 SIW_QP_ATTR_STATE | SIW_QP_ATTR_LLP_HANDLE | 1679 SIW_QP_ATTR_ORD | SIW_QP_ATTR_IRD | 1680 SIW_QP_ATTR_MPA); 1681 up_write(&qp->state_lock); 1682 1683 if (rv) 1684 goto error; 1685 1686 siw_dbg_cep(cep, "[QP %u]: send mpa reply, %d byte pdata\n", 1687 qp_id(qp), params->private_data_len); 1688 1689 rv = siw_send_mpareqrep(cep, params->private_data, 1690 params->private_data_len); 1691 if (rv != 0) 1692 goto error; 1693 1694 if (wait_for_peer_rts) { 1695 siw_sk_assign_rtr_upcalls(cep); 1696 } else { 1697 siw_qp_socket_assoc(cep, qp); 1698 rv = siw_cm_upcall(cep, IW_CM_EVENT_ESTABLISHED, 0); 1699 if (rv) 1700 goto error; 1701 } 1702 siw_cep_set_free(cep); 1703 1704 return 0; 1705 error: 1706 siw_socket_disassoc(cep->sock); 1707 sock_release(cep->sock); 1708 cep->sock = NULL; 1709 1710 cep->state = SIW_EPSTATE_CLOSED; 1711 1712 if (cep->cm_id) { 1713 cep->cm_id->rem_ref(id); 1714 cep->cm_id = NULL; 1715 } 1716 if (qp->cep) { 1717 siw_cep_put(cep); 1718 qp->cep = NULL; 1719 } 1720 cep->qp = NULL; 1721 siw_qp_put(qp); 1722 1723 siw_cep_set_free(cep); 1724 siw_cep_put(cep); 1725 1726 return rv; 1727 } 1728 1729 /* 1730 * siw_reject() 1731 * 1732 * Local connection reject case. Send private data back to peer, 1733 * close connection and dereference connection id. 1734 */ 1735 int siw_reject(struct iw_cm_id *id, const void *pdata, u8 pd_len) 1736 { 1737 struct siw_cep *cep = (struct siw_cep *)id->provider_data; 1738 1739 siw_cep_set_inuse(cep); 1740 siw_cep_put(cep); 1741 1742 siw_cancel_mpatimer(cep); 1743 1744 if (cep->state != SIW_EPSTATE_RECVD_MPAREQ) { 1745 siw_dbg_cep(cep, "out of state\n"); 1746 1747 siw_cep_set_free(cep); 1748 siw_cep_put(cep); /* put last reference */ 1749 1750 return -ECONNRESET; 1751 } 1752 siw_dbg_cep(cep, "cep->state %d, pd_len %d\n", cep->state, 1753 pd_len); 1754 1755 if (__mpa_rr_revision(cep->mpa.hdr.params.bits) >= MPA_REVISION_1) { 1756 cep->mpa.hdr.params.bits |= MPA_RR_FLAG_REJECT; /* reject */ 1757 siw_send_mpareqrep(cep, pdata, pd_len); 1758 } 1759 siw_socket_disassoc(cep->sock); 1760 sock_release(cep->sock); 1761 cep->sock = NULL; 1762 1763 cep->state = SIW_EPSTATE_CLOSED; 1764 1765 siw_cep_set_free(cep); 1766 siw_cep_put(cep); 1767 1768 return 0; 1769 } 1770 1771 /* 1772 * siw_create_listen - Create resources for a listener's IWCM ID @id 1773 * 1774 * Starts listen on the socket address id->local_addr. 1775 * 1776 */ 1777 int siw_create_listen(struct iw_cm_id *id, int backlog) 1778 { 1779 struct socket *s; 1780 struct siw_cep *cep = NULL; 1781 struct siw_device *sdev = to_siw_dev(id->device); 1782 int addr_family = id->local_addr.ss_family; 1783 int rv = 0; 1784 1785 if (addr_family != AF_INET && addr_family != AF_INET6) 1786 return -EAFNOSUPPORT; 1787 1788 rv = sock_create(addr_family, SOCK_STREAM, IPPROTO_TCP, &s); 1789 if (rv < 0) 1790 return rv; 1791 1792 /* 1793 * Allow binding local port when still in TIME_WAIT from last close. 1794 */ 1795 sock_set_reuseaddr(s->sk); 1796 1797 if (addr_family == AF_INET) { 1798 struct sockaddr_in *laddr = &to_sockaddr_in(id->local_addr); 1799 1800 /* For wildcard addr, limit binding to current device only */ 1801 if (ipv4_is_zeronet(laddr->sin_addr.s_addr)) 1802 s->sk->sk_bound_dev_if = sdev->netdev->ifindex; 1803 1804 rv = s->ops->bind(s, (struct sockaddr *)laddr, 1805 sizeof(struct sockaddr_in)); 1806 } else { 1807 struct sockaddr_in6 *laddr = &to_sockaddr_in6(id->local_addr); 1808 1809 if (id->afonly) { 1810 rv = ip6_sock_set_v6only(s->sk); 1811 if (rv) { 1812 siw_dbg(id->device, 1813 "ip6_sock_set_v6only erro: %d\n", rv); 1814 goto error; 1815 } 1816 } 1817 1818 /* For wildcard addr, limit binding to current device only */ 1819 if (ipv6_addr_any(&laddr->sin6_addr)) 1820 s->sk->sk_bound_dev_if = sdev->netdev->ifindex; 1821 1822 rv = s->ops->bind(s, (struct sockaddr *)laddr, 1823 sizeof(struct sockaddr_in6)); 1824 } 1825 if (rv) { 1826 siw_dbg(id->device, "socket bind error: %d\n", rv); 1827 goto error; 1828 } 1829 cep = siw_cep_alloc(sdev); 1830 if (!cep) { 1831 rv = -ENOMEM; 1832 goto error; 1833 } 1834 siw_cep_socket_assoc(cep, s); 1835 1836 rv = siw_cm_alloc_work(cep, backlog); 1837 if (rv) { 1838 siw_dbg(id->device, 1839 "alloc_work error %d, backlog %d\n", 1840 rv, backlog); 1841 goto error; 1842 } 1843 rv = s->ops->listen(s, backlog); 1844 if (rv) { 1845 siw_dbg(id->device, "listen error %d\n", rv); 1846 goto error; 1847 } 1848 cep->cm_id = id; 1849 id->add_ref(id); 1850 1851 /* 1852 * In case of a wildcard rdma_listen on a multi-homed device, 1853 * a listener's IWCM id is associated with more than one listening CEP. 1854 * 1855 * We currently use id->provider_data in three different ways: 1856 * 1857 * o For a listener's IWCM id, id->provider_data points to 1858 * the list_head of the list of listening CEPs. 1859 * Uses: siw_create_listen(), siw_destroy_listen() 1860 * 1861 * o For each accepted passive-side IWCM id, id->provider_data 1862 * points to the CEP itself. This is a consequence of 1863 * - siw_cm_upcall() setting event.provider_data = cep and 1864 * - the IWCM's cm_conn_req_handler() setting provider_data of the 1865 * new passive-side IWCM id equal to event.provider_data 1866 * Uses: siw_accept(), siw_reject() 1867 * 1868 * o For an active-side IWCM id, id->provider_data is not used at all. 1869 * 1870 */ 1871 if (!id->provider_data) { 1872 id->provider_data = 1873 kmalloc(sizeof(struct list_head), GFP_KERNEL); 1874 if (!id->provider_data) { 1875 rv = -ENOMEM; 1876 goto error; 1877 } 1878 INIT_LIST_HEAD((struct list_head *)id->provider_data); 1879 } 1880 list_add_tail(&cep->listenq, (struct list_head *)id->provider_data); 1881 cep->state = SIW_EPSTATE_LISTENING; 1882 1883 siw_dbg(id->device, "Listen at laddr %pISp\n", &id->local_addr); 1884 1885 return 0; 1886 1887 error: 1888 siw_dbg(id->device, "failed: %d\n", rv); 1889 1890 if (cep) { 1891 siw_cep_set_inuse(cep); 1892 1893 if (cep->cm_id) { 1894 cep->cm_id->rem_ref(cep->cm_id); 1895 cep->cm_id = NULL; 1896 } 1897 cep->sock = NULL; 1898 siw_socket_disassoc(s); 1899 cep->state = SIW_EPSTATE_CLOSED; 1900 1901 siw_cep_set_free(cep); 1902 siw_cep_put(cep); 1903 } 1904 sock_release(s); 1905 1906 return rv; 1907 } 1908 1909 static void siw_drop_listeners(struct iw_cm_id *id) 1910 { 1911 struct list_head *p, *tmp; 1912 1913 /* 1914 * In case of a wildcard rdma_listen on a multi-homed device, 1915 * a listener's IWCM id is associated with more than one listening CEP. 1916 */ 1917 list_for_each_safe(p, tmp, (struct list_head *)id->provider_data) { 1918 struct siw_cep *cep = list_entry(p, struct siw_cep, listenq); 1919 1920 list_del(p); 1921 1922 siw_dbg_cep(cep, "drop cep, state %d\n", cep->state); 1923 1924 siw_cep_set_inuse(cep); 1925 1926 if (cep->cm_id) { 1927 cep->cm_id->rem_ref(cep->cm_id); 1928 cep->cm_id = NULL; 1929 } 1930 if (cep->sock) { 1931 siw_socket_disassoc(cep->sock); 1932 sock_release(cep->sock); 1933 cep->sock = NULL; 1934 } 1935 cep->state = SIW_EPSTATE_CLOSED; 1936 siw_cep_set_free(cep); 1937 siw_cep_put(cep); 1938 } 1939 } 1940 1941 int siw_destroy_listen(struct iw_cm_id *id) 1942 { 1943 if (!id->provider_data) { 1944 siw_dbg(id->device, "no cep(s)\n"); 1945 return 0; 1946 } 1947 siw_drop_listeners(id); 1948 kfree(id->provider_data); 1949 id->provider_data = NULL; 1950 1951 return 0; 1952 } 1953 1954 int siw_cm_init(void) 1955 { 1956 /* 1957 * create_single_workqueue for strict ordering 1958 */ 1959 siw_cm_wq = create_singlethread_workqueue("siw_cm_wq"); 1960 if (!siw_cm_wq) 1961 return -ENOMEM; 1962 1963 return 0; 1964 } 1965 1966 void siw_cm_exit(void) 1967 { 1968 if (siw_cm_wq) 1969 destroy_workqueue(siw_cm_wq); 1970 } 1971