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