1 /* 2 * Copyright (c) 2005 Voltaire Inc. All rights reserved. 3 * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved. 4 * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved. 5 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved. 6 * 7 * This Software is licensed under one of the following licenses: 8 * 9 * 1) under the terms of the "Common Public License 1.0" a copy of which is 10 * available from the Open Source Initiative, see 11 * http://www.opensource.org/licenses/cpl.php. 12 * 13 * 2) under the terms of the "The BSD License" a copy of which is 14 * available from the Open Source Initiative, see 15 * http://www.opensource.org/licenses/bsd-license.php. 16 * 17 * 3) under the terms of the "GNU General Public License (GPL) Version 2" a 18 * copy of which is available from the Open Source Initiative, see 19 * http://www.opensource.org/licenses/gpl-license.php. 20 * 21 * Licensee has the right to choose one of the above licenses. 22 * 23 * Redistributions of source code must retain the above copyright 24 * notice and one of the license notices. 25 * 26 * Redistributions in binary form must reproduce both the above copyright 27 * notice, one of the license notices in the documentation 28 * and/or other materials provided with the distribution. 29 * 30 */ 31 32 #include <linux/completion.h> 33 #include <linux/in.h> 34 #include <linux/in6.h> 35 #include <linux/mutex.h> 36 #include <linux/random.h> 37 #include <linux/idr.h> 38 #include <linux/inetdevice.h> 39 40 #include <net/tcp.h> 41 42 #include <rdma/rdma_cm.h> 43 #include <rdma/rdma_cm_ib.h> 44 #include <rdma/ib_cache.h> 45 #include <rdma/ib_cm.h> 46 #include <rdma/ib_sa.h> 47 #include <rdma/iw_cm.h> 48 49 MODULE_AUTHOR("Sean Hefty"); 50 MODULE_DESCRIPTION("Generic RDMA CM Agent"); 51 MODULE_LICENSE("Dual BSD/GPL"); 52 53 #define CMA_CM_RESPONSE_TIMEOUT 20 54 #define CMA_MAX_CM_RETRIES 15 55 56 static void cma_add_one(struct ib_device *device); 57 static void cma_remove_one(struct ib_device *device); 58 59 static struct ib_client cma_client = { 60 .name = "cma", 61 .add = cma_add_one, 62 .remove = cma_remove_one 63 }; 64 65 static struct ib_sa_client sa_client; 66 static struct rdma_addr_client addr_client; 67 static LIST_HEAD(dev_list); 68 static LIST_HEAD(listen_any_list); 69 static DEFINE_MUTEX(lock); 70 static struct workqueue_struct *cma_wq; 71 static DEFINE_IDR(sdp_ps); 72 static DEFINE_IDR(tcp_ps); 73 static DEFINE_IDR(udp_ps); 74 static int next_port; 75 76 struct cma_device { 77 struct list_head list; 78 struct ib_device *device; 79 __be64 node_guid; 80 struct completion comp; 81 atomic_t refcount; 82 struct list_head id_list; 83 }; 84 85 enum cma_state { 86 CMA_IDLE, 87 CMA_ADDR_QUERY, 88 CMA_ADDR_RESOLVED, 89 CMA_ROUTE_QUERY, 90 CMA_ROUTE_RESOLVED, 91 CMA_CONNECT, 92 CMA_DISCONNECT, 93 CMA_ADDR_BOUND, 94 CMA_LISTEN, 95 CMA_DEVICE_REMOVAL, 96 CMA_DESTROYING 97 }; 98 99 struct rdma_bind_list { 100 struct idr *ps; 101 struct hlist_head owners; 102 unsigned short port; 103 }; 104 105 /* 106 * Device removal can occur at anytime, so we need extra handling to 107 * serialize notifying the user of device removal with other callbacks. 108 * We do this by disabling removal notification while a callback is in process, 109 * and reporting it after the callback completes. 110 */ 111 struct rdma_id_private { 112 struct rdma_cm_id id; 113 114 struct rdma_bind_list *bind_list; 115 struct hlist_node node; 116 struct list_head list; 117 struct list_head listen_list; 118 struct cma_device *cma_dev; 119 120 enum cma_state state; 121 spinlock_t lock; 122 struct completion comp; 123 atomic_t refcount; 124 wait_queue_head_t wait_remove; 125 atomic_t dev_remove; 126 127 int backlog; 128 int timeout_ms; 129 struct ib_sa_query *query; 130 int query_id; 131 union { 132 struct ib_cm_id *ib; 133 struct iw_cm_id *iw; 134 } cm_id; 135 136 u32 seq_num; 137 u32 qp_num; 138 u8 srq; 139 }; 140 141 struct cma_work { 142 struct work_struct work; 143 struct rdma_id_private *id; 144 enum cma_state old_state; 145 enum cma_state new_state; 146 struct rdma_cm_event event; 147 }; 148 149 union cma_ip_addr { 150 struct in6_addr ip6; 151 struct { 152 __u32 pad[3]; 153 __u32 addr; 154 } ip4; 155 }; 156 157 struct cma_hdr { 158 u8 cma_version; 159 u8 ip_version; /* IP version: 7:4 */ 160 __u16 port; 161 union cma_ip_addr src_addr; 162 union cma_ip_addr dst_addr; 163 }; 164 165 struct sdp_hh { 166 u8 bsdh[16]; 167 u8 sdp_version; /* Major version: 7:4 */ 168 u8 ip_version; /* IP version: 7:4 */ 169 u8 sdp_specific1[10]; 170 __u16 port; 171 __u16 sdp_specific2; 172 union cma_ip_addr src_addr; 173 union cma_ip_addr dst_addr; 174 }; 175 176 struct sdp_hah { 177 u8 bsdh[16]; 178 u8 sdp_version; 179 }; 180 181 #define CMA_VERSION 0x00 182 #define SDP_MAJ_VERSION 0x2 183 184 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp) 185 { 186 unsigned long flags; 187 int ret; 188 189 spin_lock_irqsave(&id_priv->lock, flags); 190 ret = (id_priv->state == comp); 191 spin_unlock_irqrestore(&id_priv->lock, flags); 192 return ret; 193 } 194 195 static int cma_comp_exch(struct rdma_id_private *id_priv, 196 enum cma_state comp, enum cma_state exch) 197 { 198 unsigned long flags; 199 int ret; 200 201 spin_lock_irqsave(&id_priv->lock, flags); 202 if ((ret = (id_priv->state == comp))) 203 id_priv->state = exch; 204 spin_unlock_irqrestore(&id_priv->lock, flags); 205 return ret; 206 } 207 208 static enum cma_state cma_exch(struct rdma_id_private *id_priv, 209 enum cma_state exch) 210 { 211 unsigned long flags; 212 enum cma_state old; 213 214 spin_lock_irqsave(&id_priv->lock, flags); 215 old = id_priv->state; 216 id_priv->state = exch; 217 spin_unlock_irqrestore(&id_priv->lock, flags); 218 return old; 219 } 220 221 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr) 222 { 223 return hdr->ip_version >> 4; 224 } 225 226 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver) 227 { 228 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF); 229 } 230 231 static inline u8 sdp_get_majv(u8 sdp_version) 232 { 233 return sdp_version >> 4; 234 } 235 236 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh) 237 { 238 return hh->ip_version >> 4; 239 } 240 241 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver) 242 { 243 hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF); 244 } 245 246 static void cma_attach_to_dev(struct rdma_id_private *id_priv, 247 struct cma_device *cma_dev) 248 { 249 atomic_inc(&cma_dev->refcount); 250 id_priv->cma_dev = cma_dev; 251 id_priv->id.device = cma_dev->device; 252 list_add_tail(&id_priv->list, &cma_dev->id_list); 253 } 254 255 static inline void cma_deref_dev(struct cma_device *cma_dev) 256 { 257 if (atomic_dec_and_test(&cma_dev->refcount)) 258 complete(&cma_dev->comp); 259 } 260 261 static void cma_detach_from_dev(struct rdma_id_private *id_priv) 262 { 263 list_del(&id_priv->list); 264 cma_deref_dev(id_priv->cma_dev); 265 id_priv->cma_dev = NULL; 266 } 267 268 static int cma_acquire_dev(struct rdma_id_private *id_priv) 269 { 270 enum rdma_node_type dev_type = id_priv->id.route.addr.dev_addr.dev_type; 271 struct cma_device *cma_dev; 272 union ib_gid gid; 273 int ret = -ENODEV; 274 275 switch (rdma_node_get_transport(dev_type)) { 276 case RDMA_TRANSPORT_IB: 277 ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid); 278 break; 279 case RDMA_TRANSPORT_IWARP: 280 iw_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid); 281 break; 282 default: 283 return -ENODEV; 284 } 285 286 list_for_each_entry(cma_dev, &dev_list, list) { 287 ret = ib_find_cached_gid(cma_dev->device, &gid, 288 &id_priv->id.port_num, NULL); 289 if (!ret) { 290 cma_attach_to_dev(id_priv, cma_dev); 291 break; 292 } 293 } 294 return ret; 295 } 296 297 static void cma_deref_id(struct rdma_id_private *id_priv) 298 { 299 if (atomic_dec_and_test(&id_priv->refcount)) 300 complete(&id_priv->comp); 301 } 302 303 static void cma_release_remove(struct rdma_id_private *id_priv) 304 { 305 if (atomic_dec_and_test(&id_priv->dev_remove)) 306 wake_up(&id_priv->wait_remove); 307 } 308 309 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler, 310 void *context, enum rdma_port_space ps) 311 { 312 struct rdma_id_private *id_priv; 313 314 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL); 315 if (!id_priv) 316 return ERR_PTR(-ENOMEM); 317 318 id_priv->state = CMA_IDLE; 319 id_priv->id.context = context; 320 id_priv->id.event_handler = event_handler; 321 id_priv->id.ps = ps; 322 spin_lock_init(&id_priv->lock); 323 init_completion(&id_priv->comp); 324 atomic_set(&id_priv->refcount, 1); 325 init_waitqueue_head(&id_priv->wait_remove); 326 atomic_set(&id_priv->dev_remove, 0); 327 INIT_LIST_HEAD(&id_priv->listen_list); 328 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num); 329 330 return &id_priv->id; 331 } 332 EXPORT_SYMBOL(rdma_create_id); 333 334 static int cma_init_ib_qp(struct rdma_id_private *id_priv, struct ib_qp *qp) 335 { 336 struct ib_qp_attr qp_attr; 337 struct rdma_dev_addr *dev_addr; 338 int ret; 339 340 dev_addr = &id_priv->id.route.addr.dev_addr; 341 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num, 342 ib_addr_get_pkey(dev_addr), 343 &qp_attr.pkey_index); 344 if (ret) 345 return ret; 346 347 qp_attr.qp_state = IB_QPS_INIT; 348 qp_attr.qp_access_flags = 0; 349 qp_attr.port_num = id_priv->id.port_num; 350 return ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_ACCESS_FLAGS | 351 IB_QP_PKEY_INDEX | IB_QP_PORT); 352 } 353 354 static int cma_init_iw_qp(struct rdma_id_private *id_priv, struct ib_qp *qp) 355 { 356 struct ib_qp_attr qp_attr; 357 358 qp_attr.qp_state = IB_QPS_INIT; 359 qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE; 360 361 return ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_ACCESS_FLAGS); 362 } 363 364 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd, 365 struct ib_qp_init_attr *qp_init_attr) 366 { 367 struct rdma_id_private *id_priv; 368 struct ib_qp *qp; 369 int ret; 370 371 id_priv = container_of(id, struct rdma_id_private, id); 372 if (id->device != pd->device) 373 return -EINVAL; 374 375 qp = ib_create_qp(pd, qp_init_attr); 376 if (IS_ERR(qp)) 377 return PTR_ERR(qp); 378 379 switch (rdma_node_get_transport(id->device->node_type)) { 380 case RDMA_TRANSPORT_IB: 381 ret = cma_init_ib_qp(id_priv, qp); 382 break; 383 case RDMA_TRANSPORT_IWARP: 384 ret = cma_init_iw_qp(id_priv, qp); 385 break; 386 default: 387 ret = -ENOSYS; 388 break; 389 } 390 391 if (ret) 392 goto err; 393 394 id->qp = qp; 395 id_priv->qp_num = qp->qp_num; 396 id_priv->srq = (qp->srq != NULL); 397 return 0; 398 err: 399 ib_destroy_qp(qp); 400 return ret; 401 } 402 EXPORT_SYMBOL(rdma_create_qp); 403 404 void rdma_destroy_qp(struct rdma_cm_id *id) 405 { 406 ib_destroy_qp(id->qp); 407 } 408 EXPORT_SYMBOL(rdma_destroy_qp); 409 410 static int cma_modify_qp_rtr(struct rdma_cm_id *id) 411 { 412 struct ib_qp_attr qp_attr; 413 int qp_attr_mask, ret; 414 415 if (!id->qp) 416 return 0; 417 418 /* Need to update QP attributes from default values. */ 419 qp_attr.qp_state = IB_QPS_INIT; 420 ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask); 421 if (ret) 422 return ret; 423 424 ret = ib_modify_qp(id->qp, &qp_attr, qp_attr_mask); 425 if (ret) 426 return ret; 427 428 qp_attr.qp_state = IB_QPS_RTR; 429 ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask); 430 if (ret) 431 return ret; 432 433 return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask); 434 } 435 436 static int cma_modify_qp_rts(struct rdma_cm_id *id) 437 { 438 struct ib_qp_attr qp_attr; 439 int qp_attr_mask, ret; 440 441 if (!id->qp) 442 return 0; 443 444 qp_attr.qp_state = IB_QPS_RTS; 445 ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask); 446 if (ret) 447 return ret; 448 449 return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask); 450 } 451 452 static int cma_modify_qp_err(struct rdma_cm_id *id) 453 { 454 struct ib_qp_attr qp_attr; 455 456 if (!id->qp) 457 return 0; 458 459 qp_attr.qp_state = IB_QPS_ERR; 460 return ib_modify_qp(id->qp, &qp_attr, IB_QP_STATE); 461 } 462 463 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr, 464 int *qp_attr_mask) 465 { 466 struct rdma_id_private *id_priv; 467 int ret; 468 469 id_priv = container_of(id, struct rdma_id_private, id); 470 switch (rdma_node_get_transport(id_priv->id.device->node_type)) { 471 case RDMA_TRANSPORT_IB: 472 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr, 473 qp_attr_mask); 474 if (qp_attr->qp_state == IB_QPS_RTR) 475 qp_attr->rq_psn = id_priv->seq_num; 476 break; 477 case RDMA_TRANSPORT_IWARP: 478 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr, 479 qp_attr_mask); 480 break; 481 default: 482 ret = -ENOSYS; 483 break; 484 } 485 486 return ret; 487 } 488 EXPORT_SYMBOL(rdma_init_qp_attr); 489 490 static inline int cma_zero_addr(struct sockaddr *addr) 491 { 492 struct in6_addr *ip6; 493 494 if (addr->sa_family == AF_INET) 495 return ZERONET(((struct sockaddr_in *) addr)->sin_addr.s_addr); 496 else { 497 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr; 498 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] | 499 ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0; 500 } 501 } 502 503 static inline int cma_loopback_addr(struct sockaddr *addr) 504 { 505 return LOOPBACK(((struct sockaddr_in *) addr)->sin_addr.s_addr); 506 } 507 508 static inline int cma_any_addr(struct sockaddr *addr) 509 { 510 return cma_zero_addr(addr) || cma_loopback_addr(addr); 511 } 512 513 static inline __be16 cma_port(struct sockaddr *addr) 514 { 515 if (addr->sa_family == AF_INET) 516 return ((struct sockaddr_in *) addr)->sin_port; 517 else 518 return ((struct sockaddr_in6 *) addr)->sin6_port; 519 } 520 521 static inline int cma_any_port(struct sockaddr *addr) 522 { 523 return !cma_port(addr); 524 } 525 526 static int cma_get_net_info(void *hdr, enum rdma_port_space ps, 527 u8 *ip_ver, __u16 *port, 528 union cma_ip_addr **src, union cma_ip_addr **dst) 529 { 530 switch (ps) { 531 case RDMA_PS_SDP: 532 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) != 533 SDP_MAJ_VERSION) 534 return -EINVAL; 535 536 *ip_ver = sdp_get_ip_ver(hdr); 537 *port = ((struct sdp_hh *) hdr)->port; 538 *src = &((struct sdp_hh *) hdr)->src_addr; 539 *dst = &((struct sdp_hh *) hdr)->dst_addr; 540 break; 541 default: 542 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION) 543 return -EINVAL; 544 545 *ip_ver = cma_get_ip_ver(hdr); 546 *port = ((struct cma_hdr *) hdr)->port; 547 *src = &((struct cma_hdr *) hdr)->src_addr; 548 *dst = &((struct cma_hdr *) hdr)->dst_addr; 549 break; 550 } 551 552 if (*ip_ver != 4 && *ip_ver != 6) 553 return -EINVAL; 554 return 0; 555 } 556 557 static void cma_save_net_info(struct rdma_addr *addr, 558 struct rdma_addr *listen_addr, 559 u8 ip_ver, __u16 port, 560 union cma_ip_addr *src, union cma_ip_addr *dst) 561 { 562 struct sockaddr_in *listen4, *ip4; 563 struct sockaddr_in6 *listen6, *ip6; 564 565 switch (ip_ver) { 566 case 4: 567 listen4 = (struct sockaddr_in *) &listen_addr->src_addr; 568 ip4 = (struct sockaddr_in *) &addr->src_addr; 569 ip4->sin_family = listen4->sin_family; 570 ip4->sin_addr.s_addr = dst->ip4.addr; 571 ip4->sin_port = listen4->sin_port; 572 573 ip4 = (struct sockaddr_in *) &addr->dst_addr; 574 ip4->sin_family = listen4->sin_family; 575 ip4->sin_addr.s_addr = src->ip4.addr; 576 ip4->sin_port = port; 577 break; 578 case 6: 579 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr; 580 ip6 = (struct sockaddr_in6 *) &addr->src_addr; 581 ip6->sin6_family = listen6->sin6_family; 582 ip6->sin6_addr = dst->ip6; 583 ip6->sin6_port = listen6->sin6_port; 584 585 ip6 = (struct sockaddr_in6 *) &addr->dst_addr; 586 ip6->sin6_family = listen6->sin6_family; 587 ip6->sin6_addr = src->ip6; 588 ip6->sin6_port = port; 589 break; 590 default: 591 break; 592 } 593 } 594 595 static inline int cma_user_data_offset(enum rdma_port_space ps) 596 { 597 switch (ps) { 598 case RDMA_PS_SDP: 599 return 0; 600 default: 601 return sizeof(struct cma_hdr); 602 } 603 } 604 605 static void cma_cancel_route(struct rdma_id_private *id_priv) 606 { 607 switch (rdma_node_get_transport(id_priv->id.device->node_type)) { 608 case RDMA_TRANSPORT_IB: 609 if (id_priv->query) 610 ib_sa_cancel_query(id_priv->query_id, id_priv->query); 611 break; 612 default: 613 break; 614 } 615 } 616 617 static inline int cma_internal_listen(struct rdma_id_private *id_priv) 618 { 619 return (id_priv->state == CMA_LISTEN) && id_priv->cma_dev && 620 cma_any_addr(&id_priv->id.route.addr.src_addr); 621 } 622 623 static void cma_destroy_listen(struct rdma_id_private *id_priv) 624 { 625 cma_exch(id_priv, CMA_DESTROYING); 626 627 if (id_priv->cma_dev) { 628 switch (rdma_node_get_transport(id_priv->id.device->node_type)) { 629 case RDMA_TRANSPORT_IB: 630 if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib)) 631 ib_destroy_cm_id(id_priv->cm_id.ib); 632 break; 633 case RDMA_TRANSPORT_IWARP: 634 if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw)) 635 iw_destroy_cm_id(id_priv->cm_id.iw); 636 break; 637 default: 638 break; 639 } 640 cma_detach_from_dev(id_priv); 641 } 642 list_del(&id_priv->listen_list); 643 644 cma_deref_id(id_priv); 645 wait_for_completion(&id_priv->comp); 646 647 kfree(id_priv); 648 } 649 650 static void cma_cancel_listens(struct rdma_id_private *id_priv) 651 { 652 struct rdma_id_private *dev_id_priv; 653 654 mutex_lock(&lock); 655 list_del(&id_priv->list); 656 657 while (!list_empty(&id_priv->listen_list)) { 658 dev_id_priv = list_entry(id_priv->listen_list.next, 659 struct rdma_id_private, listen_list); 660 cma_destroy_listen(dev_id_priv); 661 } 662 mutex_unlock(&lock); 663 } 664 665 static void cma_cancel_operation(struct rdma_id_private *id_priv, 666 enum cma_state state) 667 { 668 switch (state) { 669 case CMA_ADDR_QUERY: 670 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr); 671 break; 672 case CMA_ROUTE_QUERY: 673 cma_cancel_route(id_priv); 674 break; 675 case CMA_LISTEN: 676 if (cma_any_addr(&id_priv->id.route.addr.src_addr) && 677 !id_priv->cma_dev) 678 cma_cancel_listens(id_priv); 679 break; 680 default: 681 break; 682 } 683 } 684 685 static void cma_release_port(struct rdma_id_private *id_priv) 686 { 687 struct rdma_bind_list *bind_list = id_priv->bind_list; 688 689 if (!bind_list) 690 return; 691 692 mutex_lock(&lock); 693 hlist_del(&id_priv->node); 694 if (hlist_empty(&bind_list->owners)) { 695 idr_remove(bind_list->ps, bind_list->port); 696 kfree(bind_list); 697 } 698 mutex_unlock(&lock); 699 } 700 701 void rdma_destroy_id(struct rdma_cm_id *id) 702 { 703 struct rdma_id_private *id_priv; 704 enum cma_state state; 705 706 id_priv = container_of(id, struct rdma_id_private, id); 707 state = cma_exch(id_priv, CMA_DESTROYING); 708 cma_cancel_operation(id_priv, state); 709 710 mutex_lock(&lock); 711 if (id_priv->cma_dev) { 712 mutex_unlock(&lock); 713 switch (rdma_node_get_transport(id->device->node_type)) { 714 case RDMA_TRANSPORT_IB: 715 if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib)) 716 ib_destroy_cm_id(id_priv->cm_id.ib); 717 break; 718 case RDMA_TRANSPORT_IWARP: 719 if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw)) 720 iw_destroy_cm_id(id_priv->cm_id.iw); 721 break; 722 default: 723 break; 724 } 725 mutex_lock(&lock); 726 cma_detach_from_dev(id_priv); 727 } 728 mutex_unlock(&lock); 729 730 cma_release_port(id_priv); 731 cma_deref_id(id_priv); 732 wait_for_completion(&id_priv->comp); 733 734 kfree(id_priv->id.route.path_rec); 735 kfree(id_priv); 736 } 737 EXPORT_SYMBOL(rdma_destroy_id); 738 739 static int cma_rep_recv(struct rdma_id_private *id_priv) 740 { 741 int ret; 742 743 ret = cma_modify_qp_rtr(&id_priv->id); 744 if (ret) 745 goto reject; 746 747 ret = cma_modify_qp_rts(&id_priv->id); 748 if (ret) 749 goto reject; 750 751 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0); 752 if (ret) 753 goto reject; 754 755 return 0; 756 reject: 757 cma_modify_qp_err(&id_priv->id); 758 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED, 759 NULL, 0, NULL, 0); 760 return ret; 761 } 762 763 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data) 764 { 765 if (id_priv->id.ps == RDMA_PS_SDP && 766 sdp_get_majv(((struct sdp_hah *) data)->sdp_version) != 767 SDP_MAJ_VERSION) 768 return -EINVAL; 769 770 return 0; 771 } 772 773 static void cma_set_rep_event_data(struct rdma_cm_event *event, 774 struct ib_cm_rep_event_param *rep_data, 775 void *private_data) 776 { 777 event->param.conn.private_data = private_data; 778 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE; 779 event->param.conn.responder_resources = rep_data->responder_resources; 780 event->param.conn.initiator_depth = rep_data->initiator_depth; 781 event->param.conn.flow_control = rep_data->flow_control; 782 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count; 783 event->param.conn.srq = rep_data->srq; 784 event->param.conn.qp_num = rep_data->remote_qpn; 785 } 786 787 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event) 788 { 789 struct rdma_id_private *id_priv = cm_id->context; 790 struct rdma_cm_event event; 791 int ret = 0; 792 793 atomic_inc(&id_priv->dev_remove); 794 if (!cma_comp(id_priv, CMA_CONNECT)) 795 goto out; 796 797 memset(&event, 0, sizeof event); 798 switch (ib_event->event) { 799 case IB_CM_REQ_ERROR: 800 case IB_CM_REP_ERROR: 801 event.event = RDMA_CM_EVENT_UNREACHABLE; 802 event.status = -ETIMEDOUT; 803 break; 804 case IB_CM_REP_RECEIVED: 805 event.status = cma_verify_rep(id_priv, ib_event->private_data); 806 if (event.status) 807 event.event = RDMA_CM_EVENT_CONNECT_ERROR; 808 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) { 809 event.status = cma_rep_recv(id_priv); 810 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR : 811 RDMA_CM_EVENT_ESTABLISHED; 812 } else 813 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE; 814 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd, 815 ib_event->private_data); 816 break; 817 case IB_CM_RTU_RECEIVED: 818 case IB_CM_USER_ESTABLISHED: 819 event.event = RDMA_CM_EVENT_ESTABLISHED; 820 break; 821 case IB_CM_DREQ_ERROR: 822 event.status = -ETIMEDOUT; /* fall through */ 823 case IB_CM_DREQ_RECEIVED: 824 case IB_CM_DREP_RECEIVED: 825 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT)) 826 goto out; 827 event.event = RDMA_CM_EVENT_DISCONNECTED; 828 break; 829 case IB_CM_TIMEWAIT_EXIT: 830 case IB_CM_MRA_RECEIVED: 831 /* ignore event */ 832 goto out; 833 case IB_CM_REJ_RECEIVED: 834 cma_modify_qp_err(&id_priv->id); 835 event.status = ib_event->param.rej_rcvd.reason; 836 event.event = RDMA_CM_EVENT_REJECTED; 837 event.param.conn.private_data = ib_event->private_data; 838 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE; 839 break; 840 default: 841 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d", 842 ib_event->event); 843 goto out; 844 } 845 846 ret = id_priv->id.event_handler(&id_priv->id, &event); 847 if (ret) { 848 /* Destroy the CM ID by returning a non-zero value. */ 849 id_priv->cm_id.ib = NULL; 850 cma_exch(id_priv, CMA_DESTROYING); 851 cma_release_remove(id_priv); 852 rdma_destroy_id(&id_priv->id); 853 return ret; 854 } 855 out: 856 cma_release_remove(id_priv); 857 return ret; 858 } 859 860 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id, 861 struct ib_cm_event *ib_event) 862 { 863 struct rdma_id_private *id_priv; 864 struct rdma_cm_id *id; 865 struct rdma_route *rt; 866 union cma_ip_addr *src, *dst; 867 __u16 port; 868 u8 ip_ver; 869 870 if (cma_get_net_info(ib_event->private_data, listen_id->ps, 871 &ip_ver, &port, &src, &dst)) 872 goto err; 873 874 id = rdma_create_id(listen_id->event_handler, listen_id->context, 875 listen_id->ps); 876 if (IS_ERR(id)) 877 goto err; 878 879 cma_save_net_info(&id->route.addr, &listen_id->route.addr, 880 ip_ver, port, src, dst); 881 882 rt = &id->route; 883 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1; 884 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths, 885 GFP_KERNEL); 886 if (!rt->path_rec) 887 goto destroy_id; 888 889 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path; 890 if (rt->num_paths == 2) 891 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path; 892 893 ib_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid); 894 ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid); 895 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey)); 896 rt->addr.dev_addr.dev_type = RDMA_NODE_IB_CA; 897 898 id_priv = container_of(id, struct rdma_id_private, id); 899 id_priv->state = CMA_CONNECT; 900 return id_priv; 901 902 destroy_id: 903 rdma_destroy_id(id); 904 err: 905 return NULL; 906 } 907 908 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id, 909 struct ib_cm_event *ib_event) 910 { 911 struct rdma_id_private *id_priv; 912 struct rdma_cm_id *id; 913 union cma_ip_addr *src, *dst; 914 __u16 port; 915 u8 ip_ver; 916 int ret; 917 918 id = rdma_create_id(listen_id->event_handler, listen_id->context, 919 listen_id->ps); 920 if (IS_ERR(id)) 921 return NULL; 922 923 924 if (cma_get_net_info(ib_event->private_data, listen_id->ps, 925 &ip_ver, &port, &src, &dst)) 926 goto err; 927 928 cma_save_net_info(&id->route.addr, &listen_id->route.addr, 929 ip_ver, port, src, dst); 930 931 ret = rdma_translate_ip(&id->route.addr.src_addr, 932 &id->route.addr.dev_addr); 933 if (ret) 934 goto err; 935 936 id_priv = container_of(id, struct rdma_id_private, id); 937 id_priv->state = CMA_CONNECT; 938 return id_priv; 939 err: 940 rdma_destroy_id(id); 941 return NULL; 942 } 943 944 static void cma_set_req_event_data(struct rdma_cm_event *event, 945 struct ib_cm_req_event_param *req_data, 946 void *private_data, int offset) 947 { 948 event->param.conn.private_data = private_data + offset; 949 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset; 950 event->param.conn.responder_resources = req_data->responder_resources; 951 event->param.conn.initiator_depth = req_data->initiator_depth; 952 event->param.conn.flow_control = req_data->flow_control; 953 event->param.conn.retry_count = req_data->retry_count; 954 event->param.conn.rnr_retry_count = req_data->rnr_retry_count; 955 event->param.conn.srq = req_data->srq; 956 event->param.conn.qp_num = req_data->remote_qpn; 957 } 958 959 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event) 960 { 961 struct rdma_id_private *listen_id, *conn_id; 962 struct rdma_cm_event event; 963 int offset, ret; 964 965 listen_id = cm_id->context; 966 atomic_inc(&listen_id->dev_remove); 967 if (!cma_comp(listen_id, CMA_LISTEN)) { 968 ret = -ECONNABORTED; 969 goto out; 970 } 971 972 memset(&event, 0, sizeof event); 973 offset = cma_user_data_offset(listen_id->id.ps); 974 event.event = RDMA_CM_EVENT_CONNECT_REQUEST; 975 if (listen_id->id.ps == RDMA_PS_UDP) { 976 conn_id = cma_new_udp_id(&listen_id->id, ib_event); 977 event.param.ud.private_data = ib_event->private_data + offset; 978 event.param.ud.private_data_len = 979 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset; 980 } else { 981 conn_id = cma_new_conn_id(&listen_id->id, ib_event); 982 cma_set_req_event_data(&event, &ib_event->param.req_rcvd, 983 ib_event->private_data, offset); 984 } 985 if (!conn_id) { 986 ret = -ENOMEM; 987 goto out; 988 } 989 990 atomic_inc(&conn_id->dev_remove); 991 mutex_lock(&lock); 992 ret = cma_acquire_dev(conn_id); 993 mutex_unlock(&lock); 994 if (ret) 995 goto release_conn_id; 996 997 conn_id->cm_id.ib = cm_id; 998 cm_id->context = conn_id; 999 cm_id->cm_handler = cma_ib_handler; 1000 1001 ret = conn_id->id.event_handler(&conn_id->id, &event); 1002 if (!ret) 1003 goto out; 1004 1005 /* Destroy the CM ID by returning a non-zero value. */ 1006 conn_id->cm_id.ib = NULL; 1007 1008 release_conn_id: 1009 cma_exch(conn_id, CMA_DESTROYING); 1010 cma_release_remove(conn_id); 1011 rdma_destroy_id(&conn_id->id); 1012 1013 out: 1014 cma_release_remove(listen_id); 1015 return ret; 1016 } 1017 1018 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr) 1019 { 1020 return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr))); 1021 } 1022 1023 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr, 1024 struct ib_cm_compare_data *compare) 1025 { 1026 struct cma_hdr *cma_data, *cma_mask; 1027 struct sdp_hh *sdp_data, *sdp_mask; 1028 __u32 ip4_addr; 1029 struct in6_addr ip6_addr; 1030 1031 memset(compare, 0, sizeof *compare); 1032 cma_data = (void *) compare->data; 1033 cma_mask = (void *) compare->mask; 1034 sdp_data = (void *) compare->data; 1035 sdp_mask = (void *) compare->mask; 1036 1037 switch (addr->sa_family) { 1038 case AF_INET: 1039 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr; 1040 if (ps == RDMA_PS_SDP) { 1041 sdp_set_ip_ver(sdp_data, 4); 1042 sdp_set_ip_ver(sdp_mask, 0xF); 1043 sdp_data->dst_addr.ip4.addr = ip4_addr; 1044 sdp_mask->dst_addr.ip4.addr = ~0; 1045 } else { 1046 cma_set_ip_ver(cma_data, 4); 1047 cma_set_ip_ver(cma_mask, 0xF); 1048 cma_data->dst_addr.ip4.addr = ip4_addr; 1049 cma_mask->dst_addr.ip4.addr = ~0; 1050 } 1051 break; 1052 case AF_INET6: 1053 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr; 1054 if (ps == RDMA_PS_SDP) { 1055 sdp_set_ip_ver(sdp_data, 6); 1056 sdp_set_ip_ver(sdp_mask, 0xF); 1057 sdp_data->dst_addr.ip6 = ip6_addr; 1058 memset(&sdp_mask->dst_addr.ip6, 0xFF, 1059 sizeof sdp_mask->dst_addr.ip6); 1060 } else { 1061 cma_set_ip_ver(cma_data, 6); 1062 cma_set_ip_ver(cma_mask, 0xF); 1063 cma_data->dst_addr.ip6 = ip6_addr; 1064 memset(&cma_mask->dst_addr.ip6, 0xFF, 1065 sizeof cma_mask->dst_addr.ip6); 1066 } 1067 break; 1068 default: 1069 break; 1070 } 1071 } 1072 1073 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event) 1074 { 1075 struct rdma_id_private *id_priv = iw_id->context; 1076 struct rdma_cm_event event; 1077 struct sockaddr_in *sin; 1078 int ret = 0; 1079 1080 memset(&event, 0, sizeof event); 1081 atomic_inc(&id_priv->dev_remove); 1082 1083 switch (iw_event->event) { 1084 case IW_CM_EVENT_CLOSE: 1085 event.event = RDMA_CM_EVENT_DISCONNECTED; 1086 break; 1087 case IW_CM_EVENT_CONNECT_REPLY: 1088 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr; 1089 *sin = iw_event->local_addr; 1090 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr; 1091 *sin = iw_event->remote_addr; 1092 switch (iw_event->status) { 1093 case 0: 1094 event.event = RDMA_CM_EVENT_ESTABLISHED; 1095 break; 1096 case -ECONNRESET: 1097 case -ECONNREFUSED: 1098 event.event = RDMA_CM_EVENT_REJECTED; 1099 break; 1100 case -ETIMEDOUT: 1101 event.event = RDMA_CM_EVENT_UNREACHABLE; 1102 break; 1103 default: 1104 event.event = RDMA_CM_EVENT_CONNECT_ERROR; 1105 break; 1106 } 1107 break; 1108 case IW_CM_EVENT_ESTABLISHED: 1109 event.event = RDMA_CM_EVENT_ESTABLISHED; 1110 break; 1111 default: 1112 BUG_ON(1); 1113 } 1114 1115 event.status = iw_event->status; 1116 event.param.conn.private_data = iw_event->private_data; 1117 event.param.conn.private_data_len = iw_event->private_data_len; 1118 ret = id_priv->id.event_handler(&id_priv->id, &event); 1119 if (ret) { 1120 /* Destroy the CM ID by returning a non-zero value. */ 1121 id_priv->cm_id.iw = NULL; 1122 cma_exch(id_priv, CMA_DESTROYING); 1123 cma_release_remove(id_priv); 1124 rdma_destroy_id(&id_priv->id); 1125 return ret; 1126 } 1127 1128 cma_release_remove(id_priv); 1129 return ret; 1130 } 1131 1132 static int iw_conn_req_handler(struct iw_cm_id *cm_id, 1133 struct iw_cm_event *iw_event) 1134 { 1135 struct rdma_cm_id *new_cm_id; 1136 struct rdma_id_private *listen_id, *conn_id; 1137 struct sockaddr_in *sin; 1138 struct net_device *dev = NULL; 1139 struct rdma_cm_event event; 1140 int ret; 1141 1142 listen_id = cm_id->context; 1143 atomic_inc(&listen_id->dev_remove); 1144 if (!cma_comp(listen_id, CMA_LISTEN)) { 1145 ret = -ECONNABORTED; 1146 goto out; 1147 } 1148 1149 /* Create a new RDMA id for the new IW CM ID */ 1150 new_cm_id = rdma_create_id(listen_id->id.event_handler, 1151 listen_id->id.context, 1152 RDMA_PS_TCP); 1153 if (!new_cm_id) { 1154 ret = -ENOMEM; 1155 goto out; 1156 } 1157 conn_id = container_of(new_cm_id, struct rdma_id_private, id); 1158 atomic_inc(&conn_id->dev_remove); 1159 conn_id->state = CMA_CONNECT; 1160 1161 dev = ip_dev_find(iw_event->local_addr.sin_addr.s_addr); 1162 if (!dev) { 1163 ret = -EADDRNOTAVAIL; 1164 cma_release_remove(conn_id); 1165 rdma_destroy_id(new_cm_id); 1166 goto out; 1167 } 1168 ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL); 1169 if (ret) { 1170 cma_release_remove(conn_id); 1171 rdma_destroy_id(new_cm_id); 1172 goto out; 1173 } 1174 1175 mutex_lock(&lock); 1176 ret = cma_acquire_dev(conn_id); 1177 mutex_unlock(&lock); 1178 if (ret) { 1179 cma_release_remove(conn_id); 1180 rdma_destroy_id(new_cm_id); 1181 goto out; 1182 } 1183 1184 conn_id->cm_id.iw = cm_id; 1185 cm_id->context = conn_id; 1186 cm_id->cm_handler = cma_iw_handler; 1187 1188 sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr; 1189 *sin = iw_event->local_addr; 1190 sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr; 1191 *sin = iw_event->remote_addr; 1192 1193 memset(&event, 0, sizeof event); 1194 event.event = RDMA_CM_EVENT_CONNECT_REQUEST; 1195 event.param.conn.private_data = iw_event->private_data; 1196 event.param.conn.private_data_len = iw_event->private_data_len; 1197 ret = conn_id->id.event_handler(&conn_id->id, &event); 1198 if (ret) { 1199 /* User wants to destroy the CM ID */ 1200 conn_id->cm_id.iw = NULL; 1201 cma_exch(conn_id, CMA_DESTROYING); 1202 cma_release_remove(conn_id); 1203 rdma_destroy_id(&conn_id->id); 1204 } 1205 1206 out: 1207 if (dev) 1208 dev_put(dev); 1209 cma_release_remove(listen_id); 1210 return ret; 1211 } 1212 1213 static int cma_ib_listen(struct rdma_id_private *id_priv) 1214 { 1215 struct ib_cm_compare_data compare_data; 1216 struct sockaddr *addr; 1217 __be64 svc_id; 1218 int ret; 1219 1220 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler, 1221 id_priv); 1222 if (IS_ERR(id_priv->cm_id.ib)) 1223 return PTR_ERR(id_priv->cm_id.ib); 1224 1225 addr = &id_priv->id.route.addr.src_addr; 1226 svc_id = cma_get_service_id(id_priv->id.ps, addr); 1227 if (cma_any_addr(addr)) 1228 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL); 1229 else { 1230 cma_set_compare_data(id_priv->id.ps, addr, &compare_data); 1231 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data); 1232 } 1233 1234 if (ret) { 1235 ib_destroy_cm_id(id_priv->cm_id.ib); 1236 id_priv->cm_id.ib = NULL; 1237 } 1238 1239 return ret; 1240 } 1241 1242 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog) 1243 { 1244 int ret; 1245 struct sockaddr_in *sin; 1246 1247 id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device, 1248 iw_conn_req_handler, 1249 id_priv); 1250 if (IS_ERR(id_priv->cm_id.iw)) 1251 return PTR_ERR(id_priv->cm_id.iw); 1252 1253 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr; 1254 id_priv->cm_id.iw->local_addr = *sin; 1255 1256 ret = iw_cm_listen(id_priv->cm_id.iw, backlog); 1257 1258 if (ret) { 1259 iw_destroy_cm_id(id_priv->cm_id.iw); 1260 id_priv->cm_id.iw = NULL; 1261 } 1262 1263 return ret; 1264 } 1265 1266 static int cma_listen_handler(struct rdma_cm_id *id, 1267 struct rdma_cm_event *event) 1268 { 1269 struct rdma_id_private *id_priv = id->context; 1270 1271 id->context = id_priv->id.context; 1272 id->event_handler = id_priv->id.event_handler; 1273 return id_priv->id.event_handler(id, event); 1274 } 1275 1276 static void cma_listen_on_dev(struct rdma_id_private *id_priv, 1277 struct cma_device *cma_dev) 1278 { 1279 struct rdma_id_private *dev_id_priv; 1280 struct rdma_cm_id *id; 1281 int ret; 1282 1283 id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps); 1284 if (IS_ERR(id)) 1285 return; 1286 1287 dev_id_priv = container_of(id, struct rdma_id_private, id); 1288 1289 dev_id_priv->state = CMA_ADDR_BOUND; 1290 memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr, 1291 ip_addr_size(&id_priv->id.route.addr.src_addr)); 1292 1293 cma_attach_to_dev(dev_id_priv, cma_dev); 1294 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list); 1295 1296 ret = rdma_listen(id, id_priv->backlog); 1297 if (ret) 1298 goto err; 1299 1300 return; 1301 err: 1302 cma_destroy_listen(dev_id_priv); 1303 } 1304 1305 static void cma_listen_on_all(struct rdma_id_private *id_priv) 1306 { 1307 struct cma_device *cma_dev; 1308 1309 mutex_lock(&lock); 1310 list_add_tail(&id_priv->list, &listen_any_list); 1311 list_for_each_entry(cma_dev, &dev_list, list) 1312 cma_listen_on_dev(id_priv, cma_dev); 1313 mutex_unlock(&lock); 1314 } 1315 1316 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af) 1317 { 1318 struct sockaddr_in addr_in; 1319 1320 memset(&addr_in, 0, sizeof addr_in); 1321 addr_in.sin_family = af; 1322 return rdma_bind_addr(id, (struct sockaddr *) &addr_in); 1323 } 1324 1325 int rdma_listen(struct rdma_cm_id *id, int backlog) 1326 { 1327 struct rdma_id_private *id_priv; 1328 int ret; 1329 1330 id_priv = container_of(id, struct rdma_id_private, id); 1331 if (id_priv->state == CMA_IDLE) { 1332 ret = cma_bind_any(id, AF_INET); 1333 if (ret) 1334 return ret; 1335 } 1336 1337 if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN)) 1338 return -EINVAL; 1339 1340 id_priv->backlog = backlog; 1341 if (id->device) { 1342 switch (rdma_node_get_transport(id->device->node_type)) { 1343 case RDMA_TRANSPORT_IB: 1344 ret = cma_ib_listen(id_priv); 1345 if (ret) 1346 goto err; 1347 break; 1348 case RDMA_TRANSPORT_IWARP: 1349 ret = cma_iw_listen(id_priv, backlog); 1350 if (ret) 1351 goto err; 1352 break; 1353 default: 1354 ret = -ENOSYS; 1355 goto err; 1356 } 1357 } else 1358 cma_listen_on_all(id_priv); 1359 1360 return 0; 1361 err: 1362 id_priv->backlog = 0; 1363 cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND); 1364 return ret; 1365 } 1366 EXPORT_SYMBOL(rdma_listen); 1367 1368 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec, 1369 void *context) 1370 { 1371 struct cma_work *work = context; 1372 struct rdma_route *route; 1373 1374 route = &work->id->id.route; 1375 1376 if (!status) { 1377 route->num_paths = 1; 1378 *route->path_rec = *path_rec; 1379 } else { 1380 work->old_state = CMA_ROUTE_QUERY; 1381 work->new_state = CMA_ADDR_RESOLVED; 1382 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR; 1383 work->event.status = status; 1384 } 1385 1386 queue_work(cma_wq, &work->work); 1387 } 1388 1389 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms, 1390 struct cma_work *work) 1391 { 1392 struct rdma_dev_addr *addr = &id_priv->id.route.addr.dev_addr; 1393 struct ib_sa_path_rec path_rec; 1394 1395 memset(&path_rec, 0, sizeof path_rec); 1396 ib_addr_get_sgid(addr, &path_rec.sgid); 1397 ib_addr_get_dgid(addr, &path_rec.dgid); 1398 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(addr)); 1399 path_rec.numb_path = 1; 1400 1401 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device, 1402 id_priv->id.port_num, &path_rec, 1403 IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID | 1404 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH, 1405 timeout_ms, GFP_KERNEL, 1406 cma_query_handler, work, &id_priv->query); 1407 1408 return (id_priv->query_id < 0) ? id_priv->query_id : 0; 1409 } 1410 1411 static void cma_work_handler(struct work_struct *_work) 1412 { 1413 struct cma_work *work = container_of(_work, struct cma_work, work); 1414 struct rdma_id_private *id_priv = work->id; 1415 int destroy = 0; 1416 1417 atomic_inc(&id_priv->dev_remove); 1418 if (!cma_comp_exch(id_priv, work->old_state, work->new_state)) 1419 goto out; 1420 1421 if (id_priv->id.event_handler(&id_priv->id, &work->event)) { 1422 cma_exch(id_priv, CMA_DESTROYING); 1423 destroy = 1; 1424 } 1425 out: 1426 cma_release_remove(id_priv); 1427 cma_deref_id(id_priv); 1428 if (destroy) 1429 rdma_destroy_id(&id_priv->id); 1430 kfree(work); 1431 } 1432 1433 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms) 1434 { 1435 struct rdma_route *route = &id_priv->id.route; 1436 struct cma_work *work; 1437 int ret; 1438 1439 work = kzalloc(sizeof *work, GFP_KERNEL); 1440 if (!work) 1441 return -ENOMEM; 1442 1443 work->id = id_priv; 1444 INIT_WORK(&work->work, cma_work_handler); 1445 work->old_state = CMA_ROUTE_QUERY; 1446 work->new_state = CMA_ROUTE_RESOLVED; 1447 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 1448 1449 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL); 1450 if (!route->path_rec) { 1451 ret = -ENOMEM; 1452 goto err1; 1453 } 1454 1455 ret = cma_query_ib_route(id_priv, timeout_ms, work); 1456 if (ret) 1457 goto err2; 1458 1459 return 0; 1460 err2: 1461 kfree(route->path_rec); 1462 route->path_rec = NULL; 1463 err1: 1464 kfree(work); 1465 return ret; 1466 } 1467 1468 int rdma_set_ib_paths(struct rdma_cm_id *id, 1469 struct ib_sa_path_rec *path_rec, int num_paths) 1470 { 1471 struct rdma_id_private *id_priv; 1472 int ret; 1473 1474 id_priv = container_of(id, struct rdma_id_private, id); 1475 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED)) 1476 return -EINVAL; 1477 1478 id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL); 1479 if (!id->route.path_rec) { 1480 ret = -ENOMEM; 1481 goto err; 1482 } 1483 1484 memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths); 1485 return 0; 1486 err: 1487 cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED); 1488 return ret; 1489 } 1490 EXPORT_SYMBOL(rdma_set_ib_paths); 1491 1492 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms) 1493 { 1494 struct cma_work *work; 1495 1496 work = kzalloc(sizeof *work, GFP_KERNEL); 1497 if (!work) 1498 return -ENOMEM; 1499 1500 work->id = id_priv; 1501 INIT_WORK(&work->work, cma_work_handler); 1502 work->old_state = CMA_ROUTE_QUERY; 1503 work->new_state = CMA_ROUTE_RESOLVED; 1504 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 1505 queue_work(cma_wq, &work->work); 1506 return 0; 1507 } 1508 1509 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms) 1510 { 1511 struct rdma_id_private *id_priv; 1512 int ret; 1513 1514 id_priv = container_of(id, struct rdma_id_private, id); 1515 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY)) 1516 return -EINVAL; 1517 1518 atomic_inc(&id_priv->refcount); 1519 switch (rdma_node_get_transport(id->device->node_type)) { 1520 case RDMA_TRANSPORT_IB: 1521 ret = cma_resolve_ib_route(id_priv, timeout_ms); 1522 break; 1523 case RDMA_TRANSPORT_IWARP: 1524 ret = cma_resolve_iw_route(id_priv, timeout_ms); 1525 break; 1526 default: 1527 ret = -ENOSYS; 1528 break; 1529 } 1530 if (ret) 1531 goto err; 1532 1533 return 0; 1534 err: 1535 cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED); 1536 cma_deref_id(id_priv); 1537 return ret; 1538 } 1539 EXPORT_SYMBOL(rdma_resolve_route); 1540 1541 static int cma_bind_loopback(struct rdma_id_private *id_priv) 1542 { 1543 struct cma_device *cma_dev; 1544 struct ib_port_attr port_attr; 1545 union ib_gid gid; 1546 u16 pkey; 1547 int ret; 1548 u8 p; 1549 1550 mutex_lock(&lock); 1551 if (list_empty(&dev_list)) { 1552 ret = -ENODEV; 1553 goto out; 1554 } 1555 list_for_each_entry(cma_dev, &dev_list, list) 1556 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p) 1557 if (!ib_query_port(cma_dev->device, p, &port_attr) && 1558 port_attr.state == IB_PORT_ACTIVE) 1559 goto port_found; 1560 1561 p = 1; 1562 cma_dev = list_entry(dev_list.next, struct cma_device, list); 1563 1564 port_found: 1565 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid); 1566 if (ret) 1567 goto out; 1568 1569 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey); 1570 if (ret) 1571 goto out; 1572 1573 ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid); 1574 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey); 1575 id_priv->id.port_num = p; 1576 cma_attach_to_dev(id_priv, cma_dev); 1577 out: 1578 mutex_unlock(&lock); 1579 return ret; 1580 } 1581 1582 static void addr_handler(int status, struct sockaddr *src_addr, 1583 struct rdma_dev_addr *dev_addr, void *context) 1584 { 1585 struct rdma_id_private *id_priv = context; 1586 struct rdma_cm_event event; 1587 1588 memset(&event, 0, sizeof event); 1589 atomic_inc(&id_priv->dev_remove); 1590 1591 /* 1592 * Grab mutex to block rdma_destroy_id() from removing the device while 1593 * we're trying to acquire it. 1594 */ 1595 mutex_lock(&lock); 1596 if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) { 1597 mutex_unlock(&lock); 1598 goto out; 1599 } 1600 1601 if (!status && !id_priv->cma_dev) 1602 status = cma_acquire_dev(id_priv); 1603 mutex_unlock(&lock); 1604 1605 if (status) { 1606 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND)) 1607 goto out; 1608 event.event = RDMA_CM_EVENT_ADDR_ERROR; 1609 event.status = status; 1610 } else { 1611 memcpy(&id_priv->id.route.addr.src_addr, src_addr, 1612 ip_addr_size(src_addr)); 1613 event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 1614 } 1615 1616 if (id_priv->id.event_handler(&id_priv->id, &event)) { 1617 cma_exch(id_priv, CMA_DESTROYING); 1618 cma_release_remove(id_priv); 1619 cma_deref_id(id_priv); 1620 rdma_destroy_id(&id_priv->id); 1621 return; 1622 } 1623 out: 1624 cma_release_remove(id_priv); 1625 cma_deref_id(id_priv); 1626 } 1627 1628 static int cma_resolve_loopback(struct rdma_id_private *id_priv) 1629 { 1630 struct cma_work *work; 1631 struct sockaddr_in *src_in, *dst_in; 1632 union ib_gid gid; 1633 int ret; 1634 1635 work = kzalloc(sizeof *work, GFP_KERNEL); 1636 if (!work) 1637 return -ENOMEM; 1638 1639 if (!id_priv->cma_dev) { 1640 ret = cma_bind_loopback(id_priv); 1641 if (ret) 1642 goto err; 1643 } 1644 1645 ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid); 1646 ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid); 1647 1648 if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) { 1649 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr; 1650 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr; 1651 src_in->sin_family = dst_in->sin_family; 1652 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr; 1653 } 1654 1655 work->id = id_priv; 1656 INIT_WORK(&work->work, cma_work_handler); 1657 work->old_state = CMA_ADDR_QUERY; 1658 work->new_state = CMA_ADDR_RESOLVED; 1659 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 1660 queue_work(cma_wq, &work->work); 1661 return 0; 1662 err: 1663 kfree(work); 1664 return ret; 1665 } 1666 1667 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr, 1668 struct sockaddr *dst_addr) 1669 { 1670 if (src_addr && src_addr->sa_family) 1671 return rdma_bind_addr(id, src_addr); 1672 else 1673 return cma_bind_any(id, dst_addr->sa_family); 1674 } 1675 1676 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr, 1677 struct sockaddr *dst_addr, int timeout_ms) 1678 { 1679 struct rdma_id_private *id_priv; 1680 int ret; 1681 1682 id_priv = container_of(id, struct rdma_id_private, id); 1683 if (id_priv->state == CMA_IDLE) { 1684 ret = cma_bind_addr(id, src_addr, dst_addr); 1685 if (ret) 1686 return ret; 1687 } 1688 1689 if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY)) 1690 return -EINVAL; 1691 1692 atomic_inc(&id_priv->refcount); 1693 memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr)); 1694 if (cma_any_addr(dst_addr)) 1695 ret = cma_resolve_loopback(id_priv); 1696 else 1697 ret = rdma_resolve_ip(&addr_client, &id->route.addr.src_addr, 1698 dst_addr, &id->route.addr.dev_addr, 1699 timeout_ms, addr_handler, id_priv); 1700 if (ret) 1701 goto err; 1702 1703 return 0; 1704 err: 1705 cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND); 1706 cma_deref_id(id_priv); 1707 return ret; 1708 } 1709 EXPORT_SYMBOL(rdma_resolve_addr); 1710 1711 static void cma_bind_port(struct rdma_bind_list *bind_list, 1712 struct rdma_id_private *id_priv) 1713 { 1714 struct sockaddr_in *sin; 1715 1716 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr; 1717 sin->sin_port = htons(bind_list->port); 1718 id_priv->bind_list = bind_list; 1719 hlist_add_head(&id_priv->node, &bind_list->owners); 1720 } 1721 1722 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv, 1723 unsigned short snum) 1724 { 1725 struct rdma_bind_list *bind_list; 1726 int port, ret; 1727 1728 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL); 1729 if (!bind_list) 1730 return -ENOMEM; 1731 1732 do { 1733 ret = idr_get_new_above(ps, bind_list, snum, &port); 1734 } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL)); 1735 1736 if (ret) 1737 goto err1; 1738 1739 if (port != snum) { 1740 ret = -EADDRNOTAVAIL; 1741 goto err2; 1742 } 1743 1744 bind_list->ps = ps; 1745 bind_list->port = (unsigned short) port; 1746 cma_bind_port(bind_list, id_priv); 1747 return 0; 1748 err2: 1749 idr_remove(ps, port); 1750 err1: 1751 kfree(bind_list); 1752 return ret; 1753 } 1754 1755 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv) 1756 { 1757 struct rdma_bind_list *bind_list; 1758 int port, ret; 1759 1760 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL); 1761 if (!bind_list) 1762 return -ENOMEM; 1763 1764 retry: 1765 do { 1766 ret = idr_get_new_above(ps, bind_list, next_port, &port); 1767 } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL)); 1768 1769 if (ret) 1770 goto err1; 1771 1772 if (port > sysctl_local_port_range[1]) { 1773 if (next_port != sysctl_local_port_range[0]) { 1774 idr_remove(ps, port); 1775 next_port = sysctl_local_port_range[0]; 1776 goto retry; 1777 } 1778 ret = -EADDRNOTAVAIL; 1779 goto err2; 1780 } 1781 1782 if (port == sysctl_local_port_range[1]) 1783 next_port = sysctl_local_port_range[0]; 1784 else 1785 next_port = port + 1; 1786 1787 bind_list->ps = ps; 1788 bind_list->port = (unsigned short) port; 1789 cma_bind_port(bind_list, id_priv); 1790 return 0; 1791 err2: 1792 idr_remove(ps, port); 1793 err1: 1794 kfree(bind_list); 1795 return ret; 1796 } 1797 1798 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv) 1799 { 1800 struct rdma_id_private *cur_id; 1801 struct sockaddr_in *sin, *cur_sin; 1802 struct rdma_bind_list *bind_list; 1803 struct hlist_node *node; 1804 unsigned short snum; 1805 1806 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr; 1807 snum = ntohs(sin->sin_port); 1808 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE)) 1809 return -EACCES; 1810 1811 bind_list = idr_find(ps, snum); 1812 if (!bind_list) 1813 return cma_alloc_port(ps, id_priv, snum); 1814 1815 /* 1816 * We don't support binding to any address if anyone is bound to 1817 * a specific address on the same port. 1818 */ 1819 if (cma_any_addr(&id_priv->id.route.addr.src_addr)) 1820 return -EADDRNOTAVAIL; 1821 1822 hlist_for_each_entry(cur_id, node, &bind_list->owners, node) { 1823 if (cma_any_addr(&cur_id->id.route.addr.src_addr)) 1824 return -EADDRNOTAVAIL; 1825 1826 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr; 1827 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr) 1828 return -EADDRINUSE; 1829 } 1830 1831 cma_bind_port(bind_list, id_priv); 1832 return 0; 1833 } 1834 1835 static int cma_get_port(struct rdma_id_private *id_priv) 1836 { 1837 struct idr *ps; 1838 int ret; 1839 1840 switch (id_priv->id.ps) { 1841 case RDMA_PS_SDP: 1842 ps = &sdp_ps; 1843 break; 1844 case RDMA_PS_TCP: 1845 ps = &tcp_ps; 1846 break; 1847 case RDMA_PS_UDP: 1848 ps = &udp_ps; 1849 break; 1850 default: 1851 return -EPROTONOSUPPORT; 1852 } 1853 1854 mutex_lock(&lock); 1855 if (cma_any_port(&id_priv->id.route.addr.src_addr)) 1856 ret = cma_alloc_any_port(ps, id_priv); 1857 else 1858 ret = cma_use_port(ps, id_priv); 1859 mutex_unlock(&lock); 1860 1861 return ret; 1862 } 1863 1864 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr) 1865 { 1866 struct rdma_id_private *id_priv; 1867 int ret; 1868 1869 if (addr->sa_family != AF_INET) 1870 return -EAFNOSUPPORT; 1871 1872 id_priv = container_of(id, struct rdma_id_private, id); 1873 if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND)) 1874 return -EINVAL; 1875 1876 if (!cma_any_addr(addr)) { 1877 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr); 1878 if (ret) 1879 goto err1; 1880 1881 mutex_lock(&lock); 1882 ret = cma_acquire_dev(id_priv); 1883 mutex_unlock(&lock); 1884 if (ret) 1885 goto err1; 1886 } 1887 1888 memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr)); 1889 ret = cma_get_port(id_priv); 1890 if (ret) 1891 goto err2; 1892 1893 return 0; 1894 err2: 1895 if (!cma_any_addr(addr)) { 1896 mutex_lock(&lock); 1897 cma_detach_from_dev(id_priv); 1898 mutex_unlock(&lock); 1899 } 1900 err1: 1901 cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE); 1902 return ret; 1903 } 1904 EXPORT_SYMBOL(rdma_bind_addr); 1905 1906 static int cma_format_hdr(void *hdr, enum rdma_port_space ps, 1907 struct rdma_route *route) 1908 { 1909 struct sockaddr_in *src4, *dst4; 1910 struct cma_hdr *cma_hdr; 1911 struct sdp_hh *sdp_hdr; 1912 1913 src4 = (struct sockaddr_in *) &route->addr.src_addr; 1914 dst4 = (struct sockaddr_in *) &route->addr.dst_addr; 1915 1916 switch (ps) { 1917 case RDMA_PS_SDP: 1918 sdp_hdr = hdr; 1919 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION) 1920 return -EINVAL; 1921 sdp_set_ip_ver(sdp_hdr, 4); 1922 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr; 1923 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr; 1924 sdp_hdr->port = src4->sin_port; 1925 break; 1926 default: 1927 cma_hdr = hdr; 1928 cma_hdr->cma_version = CMA_VERSION; 1929 cma_set_ip_ver(cma_hdr, 4); 1930 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr; 1931 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr; 1932 cma_hdr->port = src4->sin_port; 1933 break; 1934 } 1935 return 0; 1936 } 1937 1938 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id, 1939 struct ib_cm_event *ib_event) 1940 { 1941 struct rdma_id_private *id_priv = cm_id->context; 1942 struct rdma_cm_event event; 1943 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd; 1944 int ret = 0; 1945 1946 memset(&event, 0, sizeof event); 1947 atomic_inc(&id_priv->dev_remove); 1948 if (!cma_comp(id_priv, CMA_CONNECT)) 1949 goto out; 1950 1951 switch (ib_event->event) { 1952 case IB_CM_SIDR_REQ_ERROR: 1953 event.event = RDMA_CM_EVENT_UNREACHABLE; 1954 event.status = -ETIMEDOUT; 1955 break; 1956 case IB_CM_SIDR_REP_RECEIVED: 1957 event.param.ud.private_data = ib_event->private_data; 1958 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE; 1959 if (rep->status != IB_SIDR_SUCCESS) { 1960 event.event = RDMA_CM_EVENT_UNREACHABLE; 1961 event.status = ib_event->param.sidr_rep_rcvd.status; 1962 break; 1963 } 1964 if (rep->qkey != RDMA_UD_QKEY) { 1965 event.event = RDMA_CM_EVENT_UNREACHABLE; 1966 event.status = -EINVAL; 1967 break; 1968 } 1969 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num, 1970 id_priv->id.route.path_rec, 1971 &event.param.ud.ah_attr); 1972 event.param.ud.qp_num = rep->qpn; 1973 event.param.ud.qkey = rep->qkey; 1974 event.event = RDMA_CM_EVENT_ESTABLISHED; 1975 event.status = 0; 1976 break; 1977 default: 1978 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d", 1979 ib_event->event); 1980 goto out; 1981 } 1982 1983 ret = id_priv->id.event_handler(&id_priv->id, &event); 1984 if (ret) { 1985 /* Destroy the CM ID by returning a non-zero value. */ 1986 id_priv->cm_id.ib = NULL; 1987 cma_exch(id_priv, CMA_DESTROYING); 1988 cma_release_remove(id_priv); 1989 rdma_destroy_id(&id_priv->id); 1990 return ret; 1991 } 1992 out: 1993 cma_release_remove(id_priv); 1994 return ret; 1995 } 1996 1997 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv, 1998 struct rdma_conn_param *conn_param) 1999 { 2000 struct ib_cm_sidr_req_param req; 2001 struct rdma_route *route; 2002 int ret; 2003 2004 req.private_data_len = sizeof(struct cma_hdr) + 2005 conn_param->private_data_len; 2006 req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC); 2007 if (!req.private_data) 2008 return -ENOMEM; 2009 2010 if (conn_param->private_data && conn_param->private_data_len) 2011 memcpy((void *) req.private_data + sizeof(struct cma_hdr), 2012 conn_param->private_data, conn_param->private_data_len); 2013 2014 route = &id_priv->id.route; 2015 ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route); 2016 if (ret) 2017 goto out; 2018 2019 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, 2020 cma_sidr_rep_handler, id_priv); 2021 if (IS_ERR(id_priv->cm_id.ib)) { 2022 ret = PTR_ERR(id_priv->cm_id.ib); 2023 goto out; 2024 } 2025 2026 req.path = route->path_rec; 2027 req.service_id = cma_get_service_id(id_priv->id.ps, 2028 &route->addr.dst_addr); 2029 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8); 2030 req.max_cm_retries = CMA_MAX_CM_RETRIES; 2031 2032 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req); 2033 if (ret) { 2034 ib_destroy_cm_id(id_priv->cm_id.ib); 2035 id_priv->cm_id.ib = NULL; 2036 } 2037 out: 2038 kfree(req.private_data); 2039 return ret; 2040 } 2041 2042 static int cma_connect_ib(struct rdma_id_private *id_priv, 2043 struct rdma_conn_param *conn_param) 2044 { 2045 struct ib_cm_req_param req; 2046 struct rdma_route *route; 2047 void *private_data; 2048 int offset, ret; 2049 2050 memset(&req, 0, sizeof req); 2051 offset = cma_user_data_offset(id_priv->id.ps); 2052 req.private_data_len = offset + conn_param->private_data_len; 2053 private_data = kzalloc(req.private_data_len, GFP_ATOMIC); 2054 if (!private_data) 2055 return -ENOMEM; 2056 2057 if (conn_param->private_data && conn_param->private_data_len) 2058 memcpy(private_data + offset, conn_param->private_data, 2059 conn_param->private_data_len); 2060 2061 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler, 2062 id_priv); 2063 if (IS_ERR(id_priv->cm_id.ib)) { 2064 ret = PTR_ERR(id_priv->cm_id.ib); 2065 goto out; 2066 } 2067 2068 route = &id_priv->id.route; 2069 ret = cma_format_hdr(private_data, id_priv->id.ps, route); 2070 if (ret) 2071 goto out; 2072 req.private_data = private_data; 2073 2074 req.primary_path = &route->path_rec[0]; 2075 if (route->num_paths == 2) 2076 req.alternate_path = &route->path_rec[1]; 2077 2078 req.service_id = cma_get_service_id(id_priv->id.ps, 2079 &route->addr.dst_addr); 2080 req.qp_num = id_priv->qp_num; 2081 req.qp_type = IB_QPT_RC; 2082 req.starting_psn = id_priv->seq_num; 2083 req.responder_resources = conn_param->responder_resources; 2084 req.initiator_depth = conn_param->initiator_depth; 2085 req.flow_control = conn_param->flow_control; 2086 req.retry_count = conn_param->retry_count; 2087 req.rnr_retry_count = conn_param->rnr_retry_count; 2088 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT; 2089 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT; 2090 req.max_cm_retries = CMA_MAX_CM_RETRIES; 2091 req.srq = id_priv->srq ? 1 : 0; 2092 2093 ret = ib_send_cm_req(id_priv->cm_id.ib, &req); 2094 out: 2095 if (ret && !IS_ERR(id_priv->cm_id.ib)) { 2096 ib_destroy_cm_id(id_priv->cm_id.ib); 2097 id_priv->cm_id.ib = NULL; 2098 } 2099 2100 kfree(private_data); 2101 return ret; 2102 } 2103 2104 static int cma_connect_iw(struct rdma_id_private *id_priv, 2105 struct rdma_conn_param *conn_param) 2106 { 2107 struct iw_cm_id *cm_id; 2108 struct sockaddr_in* sin; 2109 int ret; 2110 struct iw_cm_conn_param iw_param; 2111 2112 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv); 2113 if (IS_ERR(cm_id)) { 2114 ret = PTR_ERR(cm_id); 2115 goto out; 2116 } 2117 2118 id_priv->cm_id.iw = cm_id; 2119 2120 sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr; 2121 cm_id->local_addr = *sin; 2122 2123 sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr; 2124 cm_id->remote_addr = *sin; 2125 2126 ret = cma_modify_qp_rtr(&id_priv->id); 2127 if (ret) 2128 goto out; 2129 2130 iw_param.ord = conn_param->initiator_depth; 2131 iw_param.ird = conn_param->responder_resources; 2132 iw_param.private_data = conn_param->private_data; 2133 iw_param.private_data_len = conn_param->private_data_len; 2134 if (id_priv->id.qp) 2135 iw_param.qpn = id_priv->qp_num; 2136 else 2137 iw_param.qpn = conn_param->qp_num; 2138 ret = iw_cm_connect(cm_id, &iw_param); 2139 out: 2140 if (ret && !IS_ERR(cm_id)) { 2141 iw_destroy_cm_id(cm_id); 2142 id_priv->cm_id.iw = NULL; 2143 } 2144 return ret; 2145 } 2146 2147 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param) 2148 { 2149 struct rdma_id_private *id_priv; 2150 int ret; 2151 2152 id_priv = container_of(id, struct rdma_id_private, id); 2153 if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT)) 2154 return -EINVAL; 2155 2156 if (!id->qp) { 2157 id_priv->qp_num = conn_param->qp_num; 2158 id_priv->srq = conn_param->srq; 2159 } 2160 2161 switch (rdma_node_get_transport(id->device->node_type)) { 2162 case RDMA_TRANSPORT_IB: 2163 if (id->ps == RDMA_PS_UDP) 2164 ret = cma_resolve_ib_udp(id_priv, conn_param); 2165 else 2166 ret = cma_connect_ib(id_priv, conn_param); 2167 break; 2168 case RDMA_TRANSPORT_IWARP: 2169 ret = cma_connect_iw(id_priv, conn_param); 2170 break; 2171 default: 2172 ret = -ENOSYS; 2173 break; 2174 } 2175 if (ret) 2176 goto err; 2177 2178 return 0; 2179 err: 2180 cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED); 2181 return ret; 2182 } 2183 EXPORT_SYMBOL(rdma_connect); 2184 2185 static int cma_accept_ib(struct rdma_id_private *id_priv, 2186 struct rdma_conn_param *conn_param) 2187 { 2188 struct ib_cm_rep_param rep; 2189 struct ib_qp_attr qp_attr; 2190 int qp_attr_mask, ret; 2191 2192 if (id_priv->id.qp) { 2193 ret = cma_modify_qp_rtr(&id_priv->id); 2194 if (ret) 2195 goto out; 2196 2197 qp_attr.qp_state = IB_QPS_RTS; 2198 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, &qp_attr, 2199 &qp_attr_mask); 2200 if (ret) 2201 goto out; 2202 2203 qp_attr.max_rd_atomic = conn_param->initiator_depth; 2204 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 2205 if (ret) 2206 goto out; 2207 } 2208 2209 memset(&rep, 0, sizeof rep); 2210 rep.qp_num = id_priv->qp_num; 2211 rep.starting_psn = id_priv->seq_num; 2212 rep.private_data = conn_param->private_data; 2213 rep.private_data_len = conn_param->private_data_len; 2214 rep.responder_resources = conn_param->responder_resources; 2215 rep.initiator_depth = conn_param->initiator_depth; 2216 rep.target_ack_delay = CMA_CM_RESPONSE_TIMEOUT; 2217 rep.failover_accepted = 0; 2218 rep.flow_control = conn_param->flow_control; 2219 rep.rnr_retry_count = conn_param->rnr_retry_count; 2220 rep.srq = id_priv->srq ? 1 : 0; 2221 2222 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep); 2223 out: 2224 return ret; 2225 } 2226 2227 static int cma_accept_iw(struct rdma_id_private *id_priv, 2228 struct rdma_conn_param *conn_param) 2229 { 2230 struct iw_cm_conn_param iw_param; 2231 int ret; 2232 2233 ret = cma_modify_qp_rtr(&id_priv->id); 2234 if (ret) 2235 return ret; 2236 2237 iw_param.ord = conn_param->initiator_depth; 2238 iw_param.ird = conn_param->responder_resources; 2239 iw_param.private_data = conn_param->private_data; 2240 iw_param.private_data_len = conn_param->private_data_len; 2241 if (id_priv->id.qp) { 2242 iw_param.qpn = id_priv->qp_num; 2243 } else 2244 iw_param.qpn = conn_param->qp_num; 2245 2246 return iw_cm_accept(id_priv->cm_id.iw, &iw_param); 2247 } 2248 2249 static int cma_send_sidr_rep(struct rdma_id_private *id_priv, 2250 enum ib_cm_sidr_status status, 2251 const void *private_data, int private_data_len) 2252 { 2253 struct ib_cm_sidr_rep_param rep; 2254 2255 memset(&rep, 0, sizeof rep); 2256 rep.status = status; 2257 if (status == IB_SIDR_SUCCESS) { 2258 rep.qp_num = id_priv->qp_num; 2259 rep.qkey = RDMA_UD_QKEY; 2260 } 2261 rep.private_data = private_data; 2262 rep.private_data_len = private_data_len; 2263 2264 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep); 2265 } 2266 2267 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param) 2268 { 2269 struct rdma_id_private *id_priv; 2270 int ret; 2271 2272 id_priv = container_of(id, struct rdma_id_private, id); 2273 if (!cma_comp(id_priv, CMA_CONNECT)) 2274 return -EINVAL; 2275 2276 if (!id->qp && conn_param) { 2277 id_priv->qp_num = conn_param->qp_num; 2278 id_priv->srq = conn_param->srq; 2279 } 2280 2281 switch (rdma_node_get_transport(id->device->node_type)) { 2282 case RDMA_TRANSPORT_IB: 2283 if (id->ps == RDMA_PS_UDP) 2284 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS, 2285 conn_param->private_data, 2286 conn_param->private_data_len); 2287 else if (conn_param) 2288 ret = cma_accept_ib(id_priv, conn_param); 2289 else 2290 ret = cma_rep_recv(id_priv); 2291 break; 2292 case RDMA_TRANSPORT_IWARP: 2293 ret = cma_accept_iw(id_priv, conn_param); 2294 break; 2295 default: 2296 ret = -ENOSYS; 2297 break; 2298 } 2299 2300 if (ret) 2301 goto reject; 2302 2303 return 0; 2304 reject: 2305 cma_modify_qp_err(id); 2306 rdma_reject(id, NULL, 0); 2307 return ret; 2308 } 2309 EXPORT_SYMBOL(rdma_accept); 2310 2311 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event) 2312 { 2313 struct rdma_id_private *id_priv; 2314 int ret; 2315 2316 id_priv = container_of(id, struct rdma_id_private, id); 2317 if (!cma_comp(id_priv, CMA_CONNECT)) 2318 return -EINVAL; 2319 2320 switch (id->device->node_type) { 2321 case RDMA_NODE_IB_CA: 2322 ret = ib_cm_notify(id_priv->cm_id.ib, event); 2323 break; 2324 default: 2325 ret = 0; 2326 break; 2327 } 2328 return ret; 2329 } 2330 EXPORT_SYMBOL(rdma_notify); 2331 2332 int rdma_reject(struct rdma_cm_id *id, const void *private_data, 2333 u8 private_data_len) 2334 { 2335 struct rdma_id_private *id_priv; 2336 int ret; 2337 2338 id_priv = container_of(id, struct rdma_id_private, id); 2339 if (!cma_comp(id_priv, CMA_CONNECT)) 2340 return -EINVAL; 2341 2342 switch (rdma_node_get_transport(id->device->node_type)) { 2343 case RDMA_TRANSPORT_IB: 2344 if (id->ps == RDMA_PS_UDP) 2345 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 2346 private_data, private_data_len); 2347 else 2348 ret = ib_send_cm_rej(id_priv->cm_id.ib, 2349 IB_CM_REJ_CONSUMER_DEFINED, NULL, 2350 0, private_data, private_data_len); 2351 break; 2352 case RDMA_TRANSPORT_IWARP: 2353 ret = iw_cm_reject(id_priv->cm_id.iw, 2354 private_data, private_data_len); 2355 break; 2356 default: 2357 ret = -ENOSYS; 2358 break; 2359 } 2360 return ret; 2361 } 2362 EXPORT_SYMBOL(rdma_reject); 2363 2364 int rdma_disconnect(struct rdma_cm_id *id) 2365 { 2366 struct rdma_id_private *id_priv; 2367 int ret; 2368 2369 id_priv = container_of(id, struct rdma_id_private, id); 2370 if (!cma_comp(id_priv, CMA_CONNECT) && 2371 !cma_comp(id_priv, CMA_DISCONNECT)) 2372 return -EINVAL; 2373 2374 switch (rdma_node_get_transport(id->device->node_type)) { 2375 case RDMA_TRANSPORT_IB: 2376 ret = cma_modify_qp_err(id); 2377 if (ret) 2378 goto out; 2379 /* Initiate or respond to a disconnect. */ 2380 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) 2381 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0); 2382 break; 2383 case RDMA_TRANSPORT_IWARP: 2384 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0); 2385 break; 2386 default: 2387 ret = -EINVAL; 2388 break; 2389 } 2390 out: 2391 return ret; 2392 } 2393 EXPORT_SYMBOL(rdma_disconnect); 2394 2395 static void cma_add_one(struct ib_device *device) 2396 { 2397 struct cma_device *cma_dev; 2398 struct rdma_id_private *id_priv; 2399 2400 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL); 2401 if (!cma_dev) 2402 return; 2403 2404 cma_dev->device = device; 2405 cma_dev->node_guid = device->node_guid; 2406 2407 init_completion(&cma_dev->comp); 2408 atomic_set(&cma_dev->refcount, 1); 2409 INIT_LIST_HEAD(&cma_dev->id_list); 2410 ib_set_client_data(device, &cma_client, cma_dev); 2411 2412 mutex_lock(&lock); 2413 list_add_tail(&cma_dev->list, &dev_list); 2414 list_for_each_entry(id_priv, &listen_any_list, list) 2415 cma_listen_on_dev(id_priv, cma_dev); 2416 mutex_unlock(&lock); 2417 } 2418 2419 static int cma_remove_id_dev(struct rdma_id_private *id_priv) 2420 { 2421 struct rdma_cm_event event; 2422 enum cma_state state; 2423 2424 /* Record that we want to remove the device */ 2425 state = cma_exch(id_priv, CMA_DEVICE_REMOVAL); 2426 if (state == CMA_DESTROYING) 2427 return 0; 2428 2429 cma_cancel_operation(id_priv, state); 2430 wait_event(id_priv->wait_remove, !atomic_read(&id_priv->dev_remove)); 2431 2432 /* Check for destruction from another callback. */ 2433 if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL)) 2434 return 0; 2435 2436 memset(&event, 0, sizeof event); 2437 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL; 2438 return id_priv->id.event_handler(&id_priv->id, &event); 2439 } 2440 2441 static void cma_process_remove(struct cma_device *cma_dev) 2442 { 2443 struct rdma_id_private *id_priv; 2444 int ret; 2445 2446 mutex_lock(&lock); 2447 while (!list_empty(&cma_dev->id_list)) { 2448 id_priv = list_entry(cma_dev->id_list.next, 2449 struct rdma_id_private, list); 2450 2451 if (cma_internal_listen(id_priv)) { 2452 cma_destroy_listen(id_priv); 2453 continue; 2454 } 2455 2456 list_del_init(&id_priv->list); 2457 atomic_inc(&id_priv->refcount); 2458 mutex_unlock(&lock); 2459 2460 ret = cma_remove_id_dev(id_priv); 2461 cma_deref_id(id_priv); 2462 if (ret) 2463 rdma_destroy_id(&id_priv->id); 2464 2465 mutex_lock(&lock); 2466 } 2467 mutex_unlock(&lock); 2468 2469 cma_deref_dev(cma_dev); 2470 wait_for_completion(&cma_dev->comp); 2471 } 2472 2473 static void cma_remove_one(struct ib_device *device) 2474 { 2475 struct cma_device *cma_dev; 2476 2477 cma_dev = ib_get_client_data(device, &cma_client); 2478 if (!cma_dev) 2479 return; 2480 2481 mutex_lock(&lock); 2482 list_del(&cma_dev->list); 2483 mutex_unlock(&lock); 2484 2485 cma_process_remove(cma_dev); 2486 kfree(cma_dev); 2487 } 2488 2489 static int cma_init(void) 2490 { 2491 int ret; 2492 2493 get_random_bytes(&next_port, sizeof next_port); 2494 next_port = (next_port % (sysctl_local_port_range[1] - 2495 sysctl_local_port_range[0])) + 2496 sysctl_local_port_range[0]; 2497 cma_wq = create_singlethread_workqueue("rdma_cm"); 2498 if (!cma_wq) 2499 return -ENOMEM; 2500 2501 ib_sa_register_client(&sa_client); 2502 rdma_addr_register_client(&addr_client); 2503 2504 ret = ib_register_client(&cma_client); 2505 if (ret) 2506 goto err; 2507 return 0; 2508 2509 err: 2510 rdma_addr_unregister_client(&addr_client); 2511 ib_sa_unregister_client(&sa_client); 2512 destroy_workqueue(cma_wq); 2513 return ret; 2514 } 2515 2516 static void cma_cleanup(void) 2517 { 2518 ib_unregister_client(&cma_client); 2519 rdma_addr_unregister_client(&addr_client); 2520 ib_sa_unregister_client(&sa_client); 2521 destroy_workqueue(cma_wq); 2522 idr_destroy(&sdp_ps); 2523 idr_destroy(&tcp_ps); 2524 idr_destroy(&udp_ps); 2525 } 2526 2527 module_init(cma_init); 2528 module_exit(cma_cleanup); 2529