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_cep_set_free(new_cep); 980 goto error; 981 } 982 } 983 siw_cep_set_free(new_cep); 984 } 985 return; 986 987 error: 988 if (new_cep) 989 siw_cep_put(new_cep); 990 991 if (new_s) { 992 siw_socket_disassoc(new_s); 993 sock_release(new_s); 994 new_cep->sock = NULL; 995 } 996 siw_dbg_cep(cep, "error %d\n", rv); 997 } 998 999 static void siw_cm_work_handler(struct work_struct *w) 1000 { 1001 struct siw_cm_work *work; 1002 struct siw_cep *cep; 1003 int release_cep = 0, rv = 0; 1004 1005 work = container_of(w, struct siw_cm_work, work.work); 1006 cep = work->cep; 1007 1008 siw_dbg_cep(cep, "[QP %u]: work type: %d, state %d\n", 1009 cep->qp ? qp_id(cep->qp) : UINT_MAX, 1010 work->type, cep->state); 1011 1012 siw_cep_set_inuse(cep); 1013 1014 switch (work->type) { 1015 case SIW_CM_WORK_ACCEPT: 1016 siw_accept_newconn(cep); 1017 break; 1018 1019 case SIW_CM_WORK_READ_MPAHDR: 1020 if (cep->state == SIW_EPSTATE_AWAIT_MPAREQ) { 1021 if (cep->listen_cep) { 1022 siw_cep_set_inuse(cep->listen_cep); 1023 1024 if (cep->listen_cep->state == 1025 SIW_EPSTATE_LISTENING) 1026 rv = siw_proc_mpareq(cep); 1027 else 1028 rv = -EFAULT; 1029 1030 siw_cep_set_free(cep->listen_cep); 1031 1032 if (rv != -EAGAIN) { 1033 siw_cep_put(cep->listen_cep); 1034 cep->listen_cep = NULL; 1035 if (rv) 1036 siw_cep_put(cep); 1037 } 1038 } 1039 } else if (cep->state == SIW_EPSTATE_AWAIT_MPAREP) { 1040 rv = siw_proc_mpareply(cep); 1041 } else { 1042 /* 1043 * CEP already moved out of MPA handshake. 1044 * any connection management already done. 1045 * silently ignore the mpa packet. 1046 */ 1047 if (cep->state == SIW_EPSTATE_RDMA_MODE) { 1048 cep->sock->sk->sk_data_ready(cep->sock->sk); 1049 siw_dbg_cep(cep, "already in RDMA mode"); 1050 } else { 1051 siw_dbg_cep(cep, "out of state: %d\n", 1052 cep->state); 1053 } 1054 } 1055 if (rv && rv != -EAGAIN) 1056 release_cep = 1; 1057 break; 1058 1059 case SIW_CM_WORK_CLOSE_LLP: 1060 /* 1061 * QP scheduled LLP close 1062 */ 1063 if (cep->qp && cep->qp->term_info.valid) 1064 siw_send_terminate(cep->qp); 1065 1066 if (cep->cm_id) 1067 siw_cm_upcall(cep, IW_CM_EVENT_CLOSE, 0); 1068 1069 release_cep = 1; 1070 break; 1071 1072 case SIW_CM_WORK_PEER_CLOSE: 1073 if (cep->cm_id) { 1074 if (cep->state == SIW_EPSTATE_AWAIT_MPAREP) { 1075 /* 1076 * MPA reply not received, but connection drop 1077 */ 1078 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, 1079 -ECONNRESET); 1080 } else if (cep->state == SIW_EPSTATE_RDMA_MODE) { 1081 /* 1082 * NOTE: IW_CM_EVENT_DISCONNECT is given just 1083 * to transition IWCM into CLOSING. 1084 */ 1085 siw_cm_upcall(cep, IW_CM_EVENT_DISCONNECT, 0); 1086 siw_cm_upcall(cep, IW_CM_EVENT_CLOSE, 0); 1087 } 1088 /* 1089 * for other states there is no connection 1090 * known to the IWCM. 1091 */ 1092 } else { 1093 if (cep->state == SIW_EPSTATE_RECVD_MPAREQ) { 1094 /* 1095 * Wait for the ulp/CM to call accept/reject 1096 */ 1097 siw_dbg_cep(cep, 1098 "mpa req recvd, wait for ULP\n"); 1099 } else if (cep->state == SIW_EPSTATE_AWAIT_MPAREQ) { 1100 /* 1101 * Socket close before MPA request received. 1102 */ 1103 siw_dbg_cep(cep, "no mpareq: drop listener\n"); 1104 siw_cep_put(cep->listen_cep); 1105 cep->listen_cep = NULL; 1106 } 1107 } 1108 release_cep = 1; 1109 break; 1110 1111 case SIW_CM_WORK_MPATIMEOUT: 1112 cep->mpa_timer = NULL; 1113 1114 if (cep->state == SIW_EPSTATE_AWAIT_MPAREP) { 1115 /* 1116 * MPA request timed out: 1117 * Hide any partially received private data and signal 1118 * timeout 1119 */ 1120 cep->mpa.hdr.params.pd_len = 0; 1121 1122 if (cep->cm_id) 1123 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, 1124 -ETIMEDOUT); 1125 release_cep = 1; 1126 1127 } else if (cep->state == SIW_EPSTATE_AWAIT_MPAREQ) { 1128 /* 1129 * No MPA request received after peer TCP stream setup. 1130 */ 1131 if (cep->listen_cep) { 1132 siw_cep_put(cep->listen_cep); 1133 cep->listen_cep = NULL; 1134 } 1135 release_cep = 1; 1136 } 1137 break; 1138 1139 default: 1140 WARN(1, "Undefined CM work type: %d\n", work->type); 1141 } 1142 if (release_cep) { 1143 siw_dbg_cep(cep, 1144 "release: timer=%s, QP[%u]\n", 1145 cep->mpa_timer ? "y" : "n", 1146 cep->qp ? qp_id(cep->qp) : UINT_MAX); 1147 1148 siw_cancel_mpatimer(cep); 1149 1150 cep->state = SIW_EPSTATE_CLOSED; 1151 1152 if (cep->qp) { 1153 struct siw_qp *qp = cep->qp; 1154 /* 1155 * Serialize a potential race with application 1156 * closing the QP and calling siw_qp_cm_drop() 1157 */ 1158 siw_qp_get(qp); 1159 siw_cep_set_free(cep); 1160 1161 siw_qp_llp_close(qp); 1162 siw_qp_put(qp); 1163 1164 siw_cep_set_inuse(cep); 1165 cep->qp = NULL; 1166 siw_qp_put(qp); 1167 } 1168 if (cep->sock) { 1169 siw_socket_disassoc(cep->sock); 1170 sock_release(cep->sock); 1171 cep->sock = NULL; 1172 } 1173 if (cep->cm_id) { 1174 cep->cm_id->rem_ref(cep->cm_id); 1175 cep->cm_id = NULL; 1176 siw_cep_put(cep); 1177 } 1178 } 1179 siw_cep_set_free(cep); 1180 siw_put_work(work); 1181 siw_cep_put(cep); 1182 } 1183 1184 static struct workqueue_struct *siw_cm_wq; 1185 1186 int siw_cm_queue_work(struct siw_cep *cep, enum siw_work_type type) 1187 { 1188 struct siw_cm_work *work = siw_get_work(cep); 1189 unsigned long delay = 0; 1190 1191 if (!work) { 1192 siw_dbg_cep(cep, "failed with no work available\n"); 1193 return -ENOMEM; 1194 } 1195 work->type = type; 1196 work->cep = cep; 1197 1198 siw_cep_get(cep); 1199 1200 INIT_DELAYED_WORK(&work->work, siw_cm_work_handler); 1201 1202 if (type == SIW_CM_WORK_MPATIMEOUT) { 1203 cep->mpa_timer = work; 1204 1205 if (cep->state == SIW_EPSTATE_AWAIT_MPAREP) 1206 delay = MPAREQ_TIMEOUT; 1207 else 1208 delay = MPAREP_TIMEOUT; 1209 } 1210 siw_dbg_cep(cep, "[QP %u]: work type: %d, timeout %lu\n", 1211 cep->qp ? qp_id(cep->qp) : -1, type, delay); 1212 1213 queue_delayed_work(siw_cm_wq, &work->work, delay); 1214 1215 return 0; 1216 } 1217 1218 static void siw_cm_llp_data_ready(struct sock *sk) 1219 { 1220 struct siw_cep *cep; 1221 1222 trace_sk_data_ready(sk); 1223 1224 read_lock(&sk->sk_callback_lock); 1225 1226 cep = sk_to_cep(sk); 1227 if (!cep) 1228 goto out; 1229 1230 siw_dbg_cep(cep, "state: %d\n", cep->state); 1231 1232 switch (cep->state) { 1233 case SIW_EPSTATE_RDMA_MODE: 1234 case SIW_EPSTATE_LISTENING: 1235 break; 1236 1237 case SIW_EPSTATE_AWAIT_MPAREQ: 1238 case SIW_EPSTATE_AWAIT_MPAREP: 1239 siw_cm_queue_work(cep, SIW_CM_WORK_READ_MPAHDR); 1240 break; 1241 1242 default: 1243 siw_dbg_cep(cep, "unexpected data, state %d\n", cep->state); 1244 break; 1245 } 1246 out: 1247 read_unlock(&sk->sk_callback_lock); 1248 } 1249 1250 static void siw_cm_llp_write_space(struct sock *sk) 1251 { 1252 struct siw_cep *cep = sk_to_cep(sk); 1253 1254 if (cep) 1255 siw_dbg_cep(cep, "state: %d\n", cep->state); 1256 } 1257 1258 static void siw_cm_llp_error_report(struct sock *sk) 1259 { 1260 struct siw_cep *cep = sk_to_cep(sk); 1261 1262 if (cep) { 1263 siw_dbg_cep(cep, "error %d, socket state: %d, cep state: %d\n", 1264 sk->sk_err, sk->sk_state, cep->state); 1265 cep->sk_error_report(sk); 1266 } 1267 } 1268 1269 static void siw_cm_llp_state_change(struct sock *sk) 1270 { 1271 struct siw_cep *cep; 1272 void (*orig_state_change)(struct sock *s); 1273 1274 read_lock(&sk->sk_callback_lock); 1275 1276 cep = sk_to_cep(sk); 1277 if (!cep) { 1278 /* endpoint already disassociated */ 1279 read_unlock(&sk->sk_callback_lock); 1280 return; 1281 } 1282 orig_state_change = cep->sk_state_change; 1283 1284 siw_dbg_cep(cep, "state: %d\n", cep->state); 1285 1286 switch (sk->sk_state) { 1287 case TCP_ESTABLISHED: 1288 /* 1289 * handle accepting socket as special case where only 1290 * new connection is possible 1291 */ 1292 siw_cm_queue_work(cep, SIW_CM_WORK_ACCEPT); 1293 break; 1294 1295 case TCP_CLOSE: 1296 case TCP_CLOSE_WAIT: 1297 if (cep->qp) 1298 cep->qp->tx_ctx.tx_suspend = 1; 1299 siw_cm_queue_work(cep, SIW_CM_WORK_PEER_CLOSE); 1300 break; 1301 1302 default: 1303 siw_dbg_cep(cep, "unexpected socket state %d\n", sk->sk_state); 1304 } 1305 read_unlock(&sk->sk_callback_lock); 1306 orig_state_change(sk); 1307 } 1308 1309 static int kernel_bindconnect(struct socket *s, struct sockaddr *laddr, 1310 struct sockaddr *raddr, bool afonly) 1311 { 1312 int rv, flags = 0; 1313 size_t size = laddr->sa_family == AF_INET ? 1314 sizeof(struct sockaddr_in) : sizeof(struct sockaddr_in6); 1315 1316 /* 1317 * Make address available again asap. 1318 */ 1319 sock_set_reuseaddr(s->sk); 1320 1321 if (afonly) { 1322 rv = ip6_sock_set_v6only(s->sk); 1323 if (rv) 1324 return rv; 1325 } 1326 1327 rv = s->ops->bind(s, laddr, size); 1328 if (rv < 0) 1329 return rv; 1330 1331 rv = s->ops->connect(s, raddr, size, flags); 1332 1333 return rv < 0 ? rv : 0; 1334 } 1335 1336 int siw_connect(struct iw_cm_id *id, struct iw_cm_conn_param *params) 1337 { 1338 struct siw_device *sdev = to_siw_dev(id->device); 1339 struct siw_qp *qp; 1340 struct siw_cep *cep = NULL; 1341 struct socket *s = NULL; 1342 struct sockaddr *laddr = (struct sockaddr *)&id->local_addr, 1343 *raddr = (struct sockaddr *)&id->remote_addr; 1344 bool p2p_mode = peer_to_peer, v4 = true; 1345 u16 pd_len = params->private_data_len; 1346 int version = mpa_version, rv; 1347 1348 if (pd_len > MPA_MAX_PRIVDATA) 1349 return -EINVAL; 1350 1351 if (params->ird > sdev->attrs.max_ird || 1352 params->ord > sdev->attrs.max_ord) 1353 return -ENOMEM; 1354 1355 if (laddr->sa_family == AF_INET6) 1356 v4 = false; 1357 else if (laddr->sa_family != AF_INET) 1358 return -EAFNOSUPPORT; 1359 1360 /* 1361 * Respect any iwarp port mapping: Use mapped remote address 1362 * if valid. Local address must not be mapped, since siw 1363 * uses kernel TCP stack. 1364 */ 1365 if ((v4 && to_sockaddr_in(id->remote_addr).sin_port != 0) || 1366 to_sockaddr_in6(id->remote_addr).sin6_port != 0) 1367 raddr = (struct sockaddr *)&id->m_remote_addr; 1368 1369 qp = siw_qp_id2obj(sdev, params->qpn); 1370 if (!qp) { 1371 WARN(1, "[QP %u] does not exist\n", params->qpn); 1372 rv = -EINVAL; 1373 goto error; 1374 } 1375 siw_dbg_qp(qp, "pd_len %d, laddr %pISp, raddr %pISp\n", pd_len, laddr, 1376 raddr); 1377 1378 rv = sock_create(v4 ? AF_INET : AF_INET6, SOCK_STREAM, IPPROTO_TCP, &s); 1379 if (rv < 0) 1380 goto error; 1381 1382 /* 1383 * NOTE: For simplification, connect() is called in blocking 1384 * mode. Might be reconsidered for async connection setup at 1385 * TCP level. 1386 */ 1387 rv = kernel_bindconnect(s, laddr, raddr, id->afonly); 1388 if (rv != 0) { 1389 siw_dbg_qp(qp, "kernel_bindconnect: error %d\n", rv); 1390 goto error; 1391 } 1392 if (siw_tcp_nagle == false) 1393 tcp_sock_set_nodelay(s->sk); 1394 cep = siw_cep_alloc(sdev); 1395 if (!cep) { 1396 rv = -ENOMEM; 1397 goto error; 1398 } 1399 siw_cep_set_inuse(cep); 1400 1401 /* Associate QP with CEP */ 1402 siw_cep_get(cep); 1403 qp->cep = cep; 1404 1405 /* siw_qp_get(qp) already done by QP lookup */ 1406 cep->qp = qp; 1407 1408 id->add_ref(id); 1409 cep->cm_id = id; 1410 1411 /* 1412 * 4: Allocate a sufficient number of work elements 1413 * to allow concurrent handling of local + peer close 1414 * events, MPA header processing + MPA timeout. 1415 */ 1416 rv = siw_cm_alloc_work(cep, 4); 1417 if (rv != 0) { 1418 rv = -ENOMEM; 1419 goto error; 1420 } 1421 cep->ird = params->ird; 1422 cep->ord = params->ord; 1423 1424 if (p2p_mode && cep->ord == 0) 1425 cep->ord = 1; 1426 1427 cep->state = SIW_EPSTATE_CONNECTING; 1428 1429 /* 1430 * Associate CEP with socket 1431 */ 1432 siw_cep_socket_assoc(cep, s); 1433 1434 cep->state = SIW_EPSTATE_AWAIT_MPAREP; 1435 1436 /* 1437 * Set MPA Request bits: CRC if required, no MPA Markers, 1438 * MPA Rev. according to module parameter 'mpa_version', Key 'Request'. 1439 */ 1440 cep->mpa.hdr.params.bits = 0; 1441 if (version > MPA_REVISION_2) { 1442 pr_warn("Setting MPA version to %u\n", MPA_REVISION_2); 1443 version = MPA_REVISION_2; 1444 /* Adjust also module parameter */ 1445 mpa_version = MPA_REVISION_2; 1446 } 1447 __mpa_rr_set_revision(&cep->mpa.hdr.params.bits, version); 1448 1449 if (try_gso) 1450 cep->mpa.hdr.params.bits |= MPA_RR_FLAG_GSO_EXP; 1451 1452 if (mpa_crc_required) 1453 cep->mpa.hdr.params.bits |= MPA_RR_FLAG_CRC; 1454 1455 /* 1456 * If MPA version == 2: 1457 * o Include ORD and IRD. 1458 * o Indicate peer-to-peer mode, if required by module 1459 * parameter 'peer_to_peer'. 1460 */ 1461 if (version == MPA_REVISION_2) { 1462 cep->enhanced_rdma_conn_est = true; 1463 cep->mpa.hdr.params.bits |= MPA_RR_FLAG_ENHANCED; 1464 1465 cep->mpa.v2_ctrl.ird = htons(cep->ird); 1466 cep->mpa.v2_ctrl.ord = htons(cep->ord); 1467 1468 if (p2p_mode) { 1469 cep->mpa.v2_ctrl.ird |= MPA_V2_PEER_TO_PEER; 1470 cep->mpa.v2_ctrl.ord |= rtr_type; 1471 } 1472 /* Remember own P2P mode requested */ 1473 cep->mpa.v2_ctrl_req.ird = cep->mpa.v2_ctrl.ird; 1474 cep->mpa.v2_ctrl_req.ord = cep->mpa.v2_ctrl.ord; 1475 } 1476 memcpy(cep->mpa.hdr.key, MPA_KEY_REQ, 16); 1477 1478 rv = siw_send_mpareqrep(cep, params->private_data, pd_len); 1479 /* 1480 * Reset private data. 1481 */ 1482 cep->mpa.hdr.params.pd_len = 0; 1483 1484 if (rv >= 0) { 1485 rv = siw_cm_queue_work(cep, SIW_CM_WORK_MPATIMEOUT); 1486 if (!rv) { 1487 siw_dbg_cep(cep, "[QP %u]: exit\n", qp_id(qp)); 1488 siw_cep_set_free(cep); 1489 return 0; 1490 } 1491 } 1492 error: 1493 siw_dbg(id->device, "failed: %d\n", rv); 1494 1495 if (cep) { 1496 siw_socket_disassoc(s); 1497 sock_release(s); 1498 cep->sock = NULL; 1499 1500 cep->qp = NULL; 1501 1502 cep->cm_id = NULL; 1503 id->rem_ref(id); 1504 siw_cep_put(cep); 1505 1506 qp->cep = NULL; 1507 siw_cep_put(cep); 1508 1509 cep->state = SIW_EPSTATE_CLOSED; 1510 1511 siw_cep_set_free(cep); 1512 1513 siw_cep_put(cep); 1514 1515 } else if (s) { 1516 sock_release(s); 1517 } 1518 if (qp) 1519 siw_qp_put(qp); 1520 1521 return rv; 1522 } 1523 1524 /* 1525 * siw_accept - Let SoftiWARP accept an RDMA connection request 1526 * 1527 * @id: New connection management id to be used for accepted 1528 * connection request 1529 * @params: Connection parameters provided by ULP for accepting connection 1530 * 1531 * Transition QP to RTS state, associate new CM id @id with accepted CEP 1532 * and get prepared for TCP input by installing socket callbacks. 1533 * Then send MPA Reply and generate the "connection established" event. 1534 * Socket callbacks must be installed before sending MPA Reply, because 1535 * the latter may cause a first RDMA message to arrive from the RDMA Initiator 1536 * side very quickly, at which time the socket callbacks must be ready. 1537 */ 1538 int siw_accept(struct iw_cm_id *id, struct iw_cm_conn_param *params) 1539 { 1540 struct siw_device *sdev = to_siw_dev(id->device); 1541 struct siw_cep *cep = (struct siw_cep *)id->provider_data; 1542 struct siw_qp *qp; 1543 struct siw_qp_attrs qp_attrs; 1544 int rv, max_priv_data = MPA_MAX_PRIVDATA; 1545 bool wait_for_peer_rts = false; 1546 1547 siw_cep_set_inuse(cep); 1548 siw_cep_put(cep); 1549 1550 /* Free lingering inbound private data */ 1551 if (cep->mpa.hdr.params.pd_len) { 1552 cep->mpa.hdr.params.pd_len = 0; 1553 kfree(cep->mpa.pdata); 1554 cep->mpa.pdata = NULL; 1555 } 1556 siw_cancel_mpatimer(cep); 1557 1558 if (cep->state != SIW_EPSTATE_RECVD_MPAREQ) { 1559 siw_dbg_cep(cep, "out of state\n"); 1560 1561 siw_cep_set_free(cep); 1562 siw_cep_put(cep); 1563 1564 return -ECONNRESET; 1565 } 1566 qp = siw_qp_id2obj(sdev, params->qpn); 1567 if (!qp) { 1568 WARN(1, "[QP %d] does not exist\n", params->qpn); 1569 siw_cep_set_free(cep); 1570 siw_cep_put(cep); 1571 1572 return -EINVAL; 1573 } 1574 down_write(&qp->state_lock); 1575 if (qp->attrs.state > SIW_QP_STATE_RTR) { 1576 rv = -EINVAL; 1577 up_write(&qp->state_lock); 1578 goto error; 1579 } 1580 siw_dbg_cep(cep, "[QP %d]\n", params->qpn); 1581 1582 if (try_gso && cep->mpa.hdr.params.bits & MPA_RR_FLAG_GSO_EXP) { 1583 siw_dbg_cep(cep, "peer allows GSO on TX\n"); 1584 qp->tx_ctx.gso_seg_limit = 0; 1585 } 1586 if (params->ord > sdev->attrs.max_ord || 1587 params->ird > sdev->attrs.max_ird) { 1588 siw_dbg_cep( 1589 cep, 1590 "[QP %u]: ord %d (max %d), ird %d (max %d)\n", 1591 qp_id(qp), params->ord, sdev->attrs.max_ord, 1592 params->ird, sdev->attrs.max_ird); 1593 rv = -EINVAL; 1594 up_write(&qp->state_lock); 1595 goto error; 1596 } 1597 if (cep->enhanced_rdma_conn_est) 1598 max_priv_data -= sizeof(struct mpa_v2_data); 1599 1600 if (params->private_data_len > max_priv_data) { 1601 siw_dbg_cep( 1602 cep, 1603 "[QP %u]: private data length: %d (max %d)\n", 1604 qp_id(qp), params->private_data_len, max_priv_data); 1605 rv = -EINVAL; 1606 up_write(&qp->state_lock); 1607 goto error; 1608 } 1609 if (cep->enhanced_rdma_conn_est) { 1610 if (params->ord > cep->ord) { 1611 if (relaxed_ird_negotiation) { 1612 params->ord = cep->ord; 1613 } else { 1614 cep->ird = params->ird; 1615 cep->ord = params->ord; 1616 rv = -EINVAL; 1617 up_write(&qp->state_lock); 1618 goto error; 1619 } 1620 } 1621 if (params->ird < cep->ird) { 1622 if (relaxed_ird_negotiation && 1623 cep->ird <= sdev->attrs.max_ird) 1624 params->ird = cep->ird; 1625 else { 1626 rv = -ENOMEM; 1627 up_write(&qp->state_lock); 1628 goto error; 1629 } 1630 } 1631 if (cep->mpa.v2_ctrl.ord & 1632 (MPA_V2_RDMA_WRITE_RTR | MPA_V2_RDMA_READ_RTR)) 1633 wait_for_peer_rts = true; 1634 /* 1635 * Signal back negotiated IRD and ORD values 1636 */ 1637 cep->mpa.v2_ctrl.ord = 1638 htons(params->ord & MPA_IRD_ORD_MASK) | 1639 (cep->mpa.v2_ctrl.ord & ~MPA_V2_MASK_IRD_ORD); 1640 cep->mpa.v2_ctrl.ird = 1641 htons(params->ird & MPA_IRD_ORD_MASK) | 1642 (cep->mpa.v2_ctrl.ird & ~MPA_V2_MASK_IRD_ORD); 1643 } 1644 cep->ird = params->ird; 1645 cep->ord = params->ord; 1646 1647 cep->cm_id = id; 1648 id->add_ref(id); 1649 1650 memset(&qp_attrs, 0, sizeof(qp_attrs)); 1651 qp_attrs.orq_size = cep->ord; 1652 qp_attrs.irq_size = cep->ird; 1653 qp_attrs.sk = cep->sock; 1654 if (cep->mpa.hdr.params.bits & MPA_RR_FLAG_CRC) 1655 qp_attrs.flags = SIW_MPA_CRC; 1656 qp_attrs.state = SIW_QP_STATE_RTS; 1657 1658 siw_dbg_cep(cep, "[QP%u]: moving to rts\n", qp_id(qp)); 1659 1660 /* Associate QP with CEP */ 1661 siw_cep_get(cep); 1662 qp->cep = cep; 1663 1664 /* siw_qp_get(qp) already done by QP lookup */ 1665 cep->qp = qp; 1666 1667 cep->state = SIW_EPSTATE_RDMA_MODE; 1668 1669 /* Move socket RX/TX under QP control */ 1670 rv = siw_qp_modify(qp, &qp_attrs, 1671 SIW_QP_ATTR_STATE | SIW_QP_ATTR_LLP_HANDLE | 1672 SIW_QP_ATTR_ORD | SIW_QP_ATTR_IRD | 1673 SIW_QP_ATTR_MPA); 1674 up_write(&qp->state_lock); 1675 1676 if (rv) 1677 goto error; 1678 1679 siw_dbg_cep(cep, "[QP %u]: send mpa reply, %d byte pdata\n", 1680 qp_id(qp), params->private_data_len); 1681 1682 rv = siw_send_mpareqrep(cep, params->private_data, 1683 params->private_data_len); 1684 if (rv != 0) 1685 goto error; 1686 1687 if (wait_for_peer_rts) { 1688 siw_sk_assign_rtr_upcalls(cep); 1689 } else { 1690 siw_qp_socket_assoc(cep, qp); 1691 rv = siw_cm_upcall(cep, IW_CM_EVENT_ESTABLISHED, 0); 1692 if (rv) 1693 goto error; 1694 } 1695 siw_cep_set_free(cep); 1696 1697 return 0; 1698 error: 1699 siw_socket_disassoc(cep->sock); 1700 sock_release(cep->sock); 1701 cep->sock = NULL; 1702 1703 cep->state = SIW_EPSTATE_CLOSED; 1704 1705 if (cep->cm_id) { 1706 cep->cm_id->rem_ref(id); 1707 cep->cm_id = NULL; 1708 } 1709 if (qp->cep) { 1710 siw_cep_put(cep); 1711 qp->cep = NULL; 1712 } 1713 cep->qp = NULL; 1714 siw_qp_put(qp); 1715 1716 siw_cep_set_free(cep); 1717 siw_cep_put(cep); 1718 1719 return rv; 1720 } 1721 1722 /* 1723 * siw_reject() 1724 * 1725 * Local connection reject case. Send private data back to peer, 1726 * close connection and dereference connection id. 1727 */ 1728 int siw_reject(struct iw_cm_id *id, const void *pdata, u8 pd_len) 1729 { 1730 struct siw_cep *cep = (struct siw_cep *)id->provider_data; 1731 1732 siw_cep_set_inuse(cep); 1733 siw_cep_put(cep); 1734 1735 siw_cancel_mpatimer(cep); 1736 1737 if (cep->state != SIW_EPSTATE_RECVD_MPAREQ) { 1738 siw_dbg_cep(cep, "out of state\n"); 1739 1740 siw_cep_set_free(cep); 1741 siw_cep_put(cep); /* put last reference */ 1742 1743 return -ECONNRESET; 1744 } 1745 siw_dbg_cep(cep, "cep->state %d, pd_len %d\n", cep->state, 1746 pd_len); 1747 1748 if (__mpa_rr_revision(cep->mpa.hdr.params.bits) >= MPA_REVISION_1) { 1749 cep->mpa.hdr.params.bits |= MPA_RR_FLAG_REJECT; /* reject */ 1750 siw_send_mpareqrep(cep, pdata, pd_len); 1751 } 1752 siw_socket_disassoc(cep->sock); 1753 sock_release(cep->sock); 1754 cep->sock = NULL; 1755 1756 cep->state = SIW_EPSTATE_CLOSED; 1757 1758 siw_cep_set_free(cep); 1759 siw_cep_put(cep); 1760 1761 return 0; 1762 } 1763 1764 /* 1765 * siw_create_listen - Create resources for a listener's IWCM ID @id 1766 * 1767 * Starts listen on the socket address id->local_addr. 1768 * 1769 */ 1770 int siw_create_listen(struct iw_cm_id *id, int backlog) 1771 { 1772 struct socket *s; 1773 struct siw_cep *cep = NULL; 1774 struct siw_device *sdev = to_siw_dev(id->device); 1775 int addr_family = id->local_addr.ss_family; 1776 int rv = 0; 1777 1778 if (addr_family != AF_INET && addr_family != AF_INET6) 1779 return -EAFNOSUPPORT; 1780 1781 rv = sock_create(addr_family, SOCK_STREAM, IPPROTO_TCP, &s); 1782 if (rv < 0) 1783 return rv; 1784 1785 /* 1786 * Allow binding local port when still in TIME_WAIT from last close. 1787 */ 1788 sock_set_reuseaddr(s->sk); 1789 1790 if (addr_family == AF_INET) { 1791 struct sockaddr_in *laddr = &to_sockaddr_in(id->local_addr); 1792 1793 /* For wildcard addr, limit binding to current device only */ 1794 if (ipv4_is_zeronet(laddr->sin_addr.s_addr)) 1795 s->sk->sk_bound_dev_if = sdev->netdev->ifindex; 1796 1797 rv = s->ops->bind(s, (struct sockaddr *)laddr, 1798 sizeof(struct sockaddr_in)); 1799 } else { 1800 struct sockaddr_in6 *laddr = &to_sockaddr_in6(id->local_addr); 1801 1802 if (id->afonly) { 1803 rv = ip6_sock_set_v6only(s->sk); 1804 if (rv) { 1805 siw_dbg(id->device, 1806 "ip6_sock_set_v6only erro: %d\n", rv); 1807 goto error; 1808 } 1809 } 1810 1811 /* For wildcard addr, limit binding to current device only */ 1812 if (ipv6_addr_any(&laddr->sin6_addr)) 1813 s->sk->sk_bound_dev_if = sdev->netdev->ifindex; 1814 1815 rv = s->ops->bind(s, (struct sockaddr *)laddr, 1816 sizeof(struct sockaddr_in6)); 1817 } 1818 if (rv) { 1819 siw_dbg(id->device, "socket bind error: %d\n", rv); 1820 goto error; 1821 } 1822 cep = siw_cep_alloc(sdev); 1823 if (!cep) { 1824 rv = -ENOMEM; 1825 goto error; 1826 } 1827 siw_cep_socket_assoc(cep, s); 1828 1829 rv = siw_cm_alloc_work(cep, backlog); 1830 if (rv) { 1831 siw_dbg(id->device, 1832 "alloc_work error %d, backlog %d\n", 1833 rv, backlog); 1834 goto error; 1835 } 1836 rv = s->ops->listen(s, backlog); 1837 if (rv) { 1838 siw_dbg(id->device, "listen error %d\n", rv); 1839 goto error; 1840 } 1841 cep->cm_id = id; 1842 id->add_ref(id); 1843 1844 /* 1845 * In case of a wildcard rdma_listen on a multi-homed device, 1846 * a listener's IWCM id is associated with more than one listening CEP. 1847 * 1848 * We currently use id->provider_data in three different ways: 1849 * 1850 * o For a listener's IWCM id, id->provider_data points to 1851 * the list_head of the list of listening CEPs. 1852 * Uses: siw_create_listen(), siw_destroy_listen() 1853 * 1854 * o For each accepted passive-side IWCM id, id->provider_data 1855 * points to the CEP itself. This is a consequence of 1856 * - siw_cm_upcall() setting event.provider_data = cep and 1857 * - the IWCM's cm_conn_req_handler() setting provider_data of the 1858 * new passive-side IWCM id equal to event.provider_data 1859 * Uses: siw_accept(), siw_reject() 1860 * 1861 * o For an active-side IWCM id, id->provider_data is not used at all. 1862 * 1863 */ 1864 if (!id->provider_data) { 1865 id->provider_data = 1866 kmalloc(sizeof(struct list_head), GFP_KERNEL); 1867 if (!id->provider_data) { 1868 rv = -ENOMEM; 1869 goto error; 1870 } 1871 INIT_LIST_HEAD((struct list_head *)id->provider_data); 1872 } 1873 list_add_tail(&cep->listenq, (struct list_head *)id->provider_data); 1874 cep->state = SIW_EPSTATE_LISTENING; 1875 1876 siw_dbg(id->device, "Listen at laddr %pISp\n", &id->local_addr); 1877 1878 return 0; 1879 1880 error: 1881 siw_dbg(id->device, "failed: %d\n", rv); 1882 1883 if (cep) { 1884 siw_cep_set_inuse(cep); 1885 1886 if (cep->cm_id) { 1887 cep->cm_id->rem_ref(cep->cm_id); 1888 cep->cm_id = NULL; 1889 } 1890 cep->sock = NULL; 1891 siw_socket_disassoc(s); 1892 cep->state = SIW_EPSTATE_CLOSED; 1893 1894 siw_cep_set_free(cep); 1895 siw_cep_put(cep); 1896 } 1897 sock_release(s); 1898 1899 return rv; 1900 } 1901 1902 static void siw_drop_listeners(struct iw_cm_id *id) 1903 { 1904 struct list_head *p, *tmp; 1905 1906 /* 1907 * In case of a wildcard rdma_listen on a multi-homed device, 1908 * a listener's IWCM id is associated with more than one listening CEP. 1909 */ 1910 list_for_each_safe(p, tmp, (struct list_head *)id->provider_data) { 1911 struct siw_cep *cep = list_entry(p, struct siw_cep, listenq); 1912 1913 list_del(p); 1914 1915 siw_dbg_cep(cep, "drop cep, state %d\n", cep->state); 1916 1917 siw_cep_set_inuse(cep); 1918 1919 if (cep->cm_id) { 1920 cep->cm_id->rem_ref(cep->cm_id); 1921 cep->cm_id = NULL; 1922 } 1923 if (cep->sock) { 1924 siw_socket_disassoc(cep->sock); 1925 sock_release(cep->sock); 1926 cep->sock = NULL; 1927 } 1928 cep->state = SIW_EPSTATE_CLOSED; 1929 siw_cep_set_free(cep); 1930 siw_cep_put(cep); 1931 } 1932 } 1933 1934 int siw_destroy_listen(struct iw_cm_id *id) 1935 { 1936 if (!id->provider_data) { 1937 siw_dbg(id->device, "no cep(s)\n"); 1938 return 0; 1939 } 1940 siw_drop_listeners(id); 1941 kfree(id->provider_data); 1942 id->provider_data = NULL; 1943 1944 return 0; 1945 } 1946 1947 int siw_cm_init(void) 1948 { 1949 /* 1950 * create_single_workqueue for strict ordering 1951 */ 1952 siw_cm_wq = create_singlethread_workqueue("siw_cm_wq"); 1953 if (!siw_cm_wq) 1954 return -ENOMEM; 1955 1956 return 0; 1957 } 1958 1959 void siw_cm_exit(void) 1960 { 1961 if (siw_cm_wq) 1962 destroy_workqueue(siw_cm_wq); 1963 } 1964