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 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24) 56 57 static void cma_add_one(struct ib_device *device); 58 static void cma_remove_one(struct ib_device *device); 59 60 static struct ib_client cma_client = { 61 .name = "cma", 62 .add = cma_add_one, 63 .remove = cma_remove_one 64 }; 65 66 static struct ib_sa_client sa_client; 67 static struct rdma_addr_client addr_client; 68 static LIST_HEAD(dev_list); 69 static LIST_HEAD(listen_any_list); 70 static DEFINE_MUTEX(lock); 71 static struct workqueue_struct *cma_wq; 72 static DEFINE_IDR(sdp_ps); 73 static DEFINE_IDR(tcp_ps); 74 static DEFINE_IDR(udp_ps); 75 static DEFINE_IDR(ipoib_ps); 76 static int next_port; 77 78 struct cma_device { 79 struct list_head list; 80 struct ib_device *device; 81 struct completion comp; 82 atomic_t refcount; 83 struct list_head id_list; 84 }; 85 86 enum cma_state { 87 CMA_IDLE, 88 CMA_ADDR_QUERY, 89 CMA_ADDR_RESOLVED, 90 CMA_ROUTE_QUERY, 91 CMA_ROUTE_RESOLVED, 92 CMA_CONNECT, 93 CMA_DISCONNECT, 94 CMA_ADDR_BOUND, 95 CMA_LISTEN, 96 CMA_DEVICE_REMOVAL, 97 CMA_DESTROYING 98 }; 99 100 struct rdma_bind_list { 101 struct idr *ps; 102 struct hlist_head owners; 103 unsigned short port; 104 }; 105 106 /* 107 * Device removal can occur at anytime, so we need extra handling to 108 * serialize notifying the user of device removal with other callbacks. 109 * We do this by disabling removal notification while a callback is in process, 110 * and reporting it after the callback completes. 111 */ 112 struct rdma_id_private { 113 struct rdma_cm_id id; 114 115 struct rdma_bind_list *bind_list; 116 struct hlist_node node; 117 struct list_head list; 118 struct list_head listen_list; 119 struct cma_device *cma_dev; 120 struct list_head mc_list; 121 122 enum cma_state state; 123 spinlock_t lock; 124 struct completion comp; 125 atomic_t refcount; 126 wait_queue_head_t wait_remove; 127 atomic_t dev_remove; 128 129 int backlog; 130 int timeout_ms; 131 struct ib_sa_query *query; 132 int query_id; 133 union { 134 struct ib_cm_id *ib; 135 struct iw_cm_id *iw; 136 } cm_id; 137 138 u32 seq_num; 139 u32 qkey; 140 u32 qp_num; 141 u8 srq; 142 u8 tos; 143 }; 144 145 struct cma_multicast { 146 struct rdma_id_private *id_priv; 147 union { 148 struct ib_sa_multicast *ib; 149 } multicast; 150 struct list_head list; 151 void *context; 152 struct sockaddr addr; 153 u8 pad[sizeof(struct sockaddr_in6) - 154 sizeof(struct sockaddr)]; 155 }; 156 157 struct cma_work { 158 struct work_struct work; 159 struct rdma_id_private *id; 160 enum cma_state old_state; 161 enum cma_state new_state; 162 struct rdma_cm_event event; 163 }; 164 165 union cma_ip_addr { 166 struct in6_addr ip6; 167 struct { 168 __u32 pad[3]; 169 __u32 addr; 170 } ip4; 171 }; 172 173 struct cma_hdr { 174 u8 cma_version; 175 u8 ip_version; /* IP version: 7:4 */ 176 __u16 port; 177 union cma_ip_addr src_addr; 178 union cma_ip_addr dst_addr; 179 }; 180 181 struct sdp_hh { 182 u8 bsdh[16]; 183 u8 sdp_version; /* Major version: 7:4 */ 184 u8 ip_version; /* IP version: 7:4 */ 185 u8 sdp_specific1[10]; 186 __u16 port; 187 __u16 sdp_specific2; 188 union cma_ip_addr src_addr; 189 union cma_ip_addr dst_addr; 190 }; 191 192 struct sdp_hah { 193 u8 bsdh[16]; 194 u8 sdp_version; 195 }; 196 197 #define CMA_VERSION 0x00 198 #define SDP_MAJ_VERSION 0x2 199 200 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp) 201 { 202 unsigned long flags; 203 int ret; 204 205 spin_lock_irqsave(&id_priv->lock, flags); 206 ret = (id_priv->state == comp); 207 spin_unlock_irqrestore(&id_priv->lock, flags); 208 return ret; 209 } 210 211 static int cma_comp_exch(struct rdma_id_private *id_priv, 212 enum cma_state comp, enum cma_state exch) 213 { 214 unsigned long flags; 215 int ret; 216 217 spin_lock_irqsave(&id_priv->lock, flags); 218 if ((ret = (id_priv->state == comp))) 219 id_priv->state = exch; 220 spin_unlock_irqrestore(&id_priv->lock, flags); 221 return ret; 222 } 223 224 static enum cma_state cma_exch(struct rdma_id_private *id_priv, 225 enum cma_state exch) 226 { 227 unsigned long flags; 228 enum cma_state old; 229 230 spin_lock_irqsave(&id_priv->lock, flags); 231 old = id_priv->state; 232 id_priv->state = exch; 233 spin_unlock_irqrestore(&id_priv->lock, flags); 234 return old; 235 } 236 237 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr) 238 { 239 return hdr->ip_version >> 4; 240 } 241 242 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver) 243 { 244 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF); 245 } 246 247 static inline u8 sdp_get_majv(u8 sdp_version) 248 { 249 return sdp_version >> 4; 250 } 251 252 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh) 253 { 254 return hh->ip_version >> 4; 255 } 256 257 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver) 258 { 259 hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF); 260 } 261 262 static inline int cma_is_ud_ps(enum rdma_port_space ps) 263 { 264 return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB); 265 } 266 267 static void cma_attach_to_dev(struct rdma_id_private *id_priv, 268 struct cma_device *cma_dev) 269 { 270 atomic_inc(&cma_dev->refcount); 271 id_priv->cma_dev = cma_dev; 272 id_priv->id.device = cma_dev->device; 273 list_add_tail(&id_priv->list, &cma_dev->id_list); 274 } 275 276 static inline void cma_deref_dev(struct cma_device *cma_dev) 277 { 278 if (atomic_dec_and_test(&cma_dev->refcount)) 279 complete(&cma_dev->comp); 280 } 281 282 static void cma_detach_from_dev(struct rdma_id_private *id_priv) 283 { 284 list_del(&id_priv->list); 285 cma_deref_dev(id_priv->cma_dev); 286 id_priv->cma_dev = NULL; 287 } 288 289 static int cma_set_qkey(struct ib_device *device, u8 port_num, 290 enum rdma_port_space ps, 291 struct rdma_dev_addr *dev_addr, u32 *qkey) 292 { 293 struct ib_sa_mcmember_rec rec; 294 int ret = 0; 295 296 switch (ps) { 297 case RDMA_PS_UDP: 298 *qkey = RDMA_UDP_QKEY; 299 break; 300 case RDMA_PS_IPOIB: 301 ib_addr_get_mgid(dev_addr, &rec.mgid); 302 ret = ib_sa_get_mcmember_rec(device, port_num, &rec.mgid, &rec); 303 *qkey = be32_to_cpu(rec.qkey); 304 break; 305 default: 306 break; 307 } 308 return ret; 309 } 310 311 static int cma_acquire_dev(struct rdma_id_private *id_priv) 312 { 313 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 314 struct cma_device *cma_dev; 315 union ib_gid gid; 316 int ret = -ENODEV; 317 318 switch (rdma_node_get_transport(dev_addr->dev_type)) { 319 case RDMA_TRANSPORT_IB: 320 ib_addr_get_sgid(dev_addr, &gid); 321 break; 322 case RDMA_TRANSPORT_IWARP: 323 iw_addr_get_sgid(dev_addr, &gid); 324 break; 325 default: 326 return -ENODEV; 327 } 328 329 list_for_each_entry(cma_dev, &dev_list, list) { 330 ret = ib_find_cached_gid(cma_dev->device, &gid, 331 &id_priv->id.port_num, NULL); 332 if (!ret) { 333 ret = cma_set_qkey(cma_dev->device, 334 id_priv->id.port_num, 335 id_priv->id.ps, dev_addr, 336 &id_priv->qkey); 337 if (!ret) 338 cma_attach_to_dev(id_priv, cma_dev); 339 break; 340 } 341 } 342 return ret; 343 } 344 345 static void cma_deref_id(struct rdma_id_private *id_priv) 346 { 347 if (atomic_dec_and_test(&id_priv->refcount)) 348 complete(&id_priv->comp); 349 } 350 351 static int cma_disable_remove(struct rdma_id_private *id_priv, 352 enum cma_state state) 353 { 354 unsigned long flags; 355 int ret; 356 357 spin_lock_irqsave(&id_priv->lock, flags); 358 if (id_priv->state == state) { 359 atomic_inc(&id_priv->dev_remove); 360 ret = 0; 361 } else 362 ret = -EINVAL; 363 spin_unlock_irqrestore(&id_priv->lock, flags); 364 return ret; 365 } 366 367 static void cma_enable_remove(struct rdma_id_private *id_priv) 368 { 369 if (atomic_dec_and_test(&id_priv->dev_remove)) 370 wake_up(&id_priv->wait_remove); 371 } 372 373 static int cma_has_cm_dev(struct rdma_id_private *id_priv) 374 { 375 return (id_priv->id.device && id_priv->cm_id.ib); 376 } 377 378 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler, 379 void *context, enum rdma_port_space ps) 380 { 381 struct rdma_id_private *id_priv; 382 383 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL); 384 if (!id_priv) 385 return ERR_PTR(-ENOMEM); 386 387 id_priv->state = CMA_IDLE; 388 id_priv->id.context = context; 389 id_priv->id.event_handler = event_handler; 390 id_priv->id.ps = ps; 391 spin_lock_init(&id_priv->lock); 392 init_completion(&id_priv->comp); 393 atomic_set(&id_priv->refcount, 1); 394 init_waitqueue_head(&id_priv->wait_remove); 395 atomic_set(&id_priv->dev_remove, 0); 396 INIT_LIST_HEAD(&id_priv->listen_list); 397 INIT_LIST_HEAD(&id_priv->mc_list); 398 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num); 399 400 return &id_priv->id; 401 } 402 EXPORT_SYMBOL(rdma_create_id); 403 404 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp) 405 { 406 struct ib_qp_attr qp_attr; 407 int qp_attr_mask, ret; 408 409 qp_attr.qp_state = IB_QPS_INIT; 410 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 411 if (ret) 412 return ret; 413 414 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 415 if (ret) 416 return ret; 417 418 qp_attr.qp_state = IB_QPS_RTR; 419 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE); 420 if (ret) 421 return ret; 422 423 qp_attr.qp_state = IB_QPS_RTS; 424 qp_attr.sq_psn = 0; 425 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN); 426 427 return ret; 428 } 429 430 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp) 431 { 432 struct ib_qp_attr qp_attr; 433 int qp_attr_mask, ret; 434 435 qp_attr.qp_state = IB_QPS_INIT; 436 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 437 if (ret) 438 return ret; 439 440 return ib_modify_qp(qp, &qp_attr, qp_attr_mask); 441 } 442 443 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd, 444 struct ib_qp_init_attr *qp_init_attr) 445 { 446 struct rdma_id_private *id_priv; 447 struct ib_qp *qp; 448 int ret; 449 450 id_priv = container_of(id, struct rdma_id_private, id); 451 if (id->device != pd->device) 452 return -EINVAL; 453 454 qp = ib_create_qp(pd, qp_init_attr); 455 if (IS_ERR(qp)) 456 return PTR_ERR(qp); 457 458 if (cma_is_ud_ps(id_priv->id.ps)) 459 ret = cma_init_ud_qp(id_priv, qp); 460 else 461 ret = cma_init_conn_qp(id_priv, qp); 462 if (ret) 463 goto err; 464 465 id->qp = qp; 466 id_priv->qp_num = qp->qp_num; 467 id_priv->srq = (qp->srq != NULL); 468 return 0; 469 err: 470 ib_destroy_qp(qp); 471 return ret; 472 } 473 EXPORT_SYMBOL(rdma_create_qp); 474 475 void rdma_destroy_qp(struct rdma_cm_id *id) 476 { 477 ib_destroy_qp(id->qp); 478 } 479 EXPORT_SYMBOL(rdma_destroy_qp); 480 481 static int cma_modify_qp_rtr(struct rdma_cm_id *id) 482 { 483 struct ib_qp_attr qp_attr; 484 int qp_attr_mask, ret; 485 486 if (!id->qp) 487 return 0; 488 489 /* Need to update QP attributes from default values. */ 490 qp_attr.qp_state = IB_QPS_INIT; 491 ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask); 492 if (ret) 493 return ret; 494 495 ret = ib_modify_qp(id->qp, &qp_attr, qp_attr_mask); 496 if (ret) 497 return ret; 498 499 qp_attr.qp_state = IB_QPS_RTR; 500 ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask); 501 if (ret) 502 return ret; 503 504 return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask); 505 } 506 507 static int cma_modify_qp_rts(struct rdma_cm_id *id) 508 { 509 struct ib_qp_attr qp_attr; 510 int qp_attr_mask, ret; 511 512 if (!id->qp) 513 return 0; 514 515 qp_attr.qp_state = IB_QPS_RTS; 516 ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask); 517 if (ret) 518 return ret; 519 520 return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask); 521 } 522 523 static int cma_modify_qp_err(struct rdma_cm_id *id) 524 { 525 struct ib_qp_attr qp_attr; 526 527 if (!id->qp) 528 return 0; 529 530 qp_attr.qp_state = IB_QPS_ERR; 531 return ib_modify_qp(id->qp, &qp_attr, IB_QP_STATE); 532 } 533 534 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv, 535 struct ib_qp_attr *qp_attr, int *qp_attr_mask) 536 { 537 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 538 int ret; 539 540 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num, 541 ib_addr_get_pkey(dev_addr), 542 &qp_attr->pkey_index); 543 if (ret) 544 return ret; 545 546 qp_attr->port_num = id_priv->id.port_num; 547 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT; 548 549 if (cma_is_ud_ps(id_priv->id.ps)) { 550 qp_attr->qkey = id_priv->qkey; 551 *qp_attr_mask |= IB_QP_QKEY; 552 } else { 553 qp_attr->qp_access_flags = 0; 554 *qp_attr_mask |= IB_QP_ACCESS_FLAGS; 555 } 556 return 0; 557 } 558 559 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr, 560 int *qp_attr_mask) 561 { 562 struct rdma_id_private *id_priv; 563 int ret = 0; 564 565 id_priv = container_of(id, struct rdma_id_private, id); 566 switch (rdma_node_get_transport(id_priv->id.device->node_type)) { 567 case RDMA_TRANSPORT_IB: 568 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps)) 569 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask); 570 else 571 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr, 572 qp_attr_mask); 573 if (qp_attr->qp_state == IB_QPS_RTR) 574 qp_attr->rq_psn = id_priv->seq_num; 575 break; 576 case RDMA_TRANSPORT_IWARP: 577 if (!id_priv->cm_id.iw) { 578 qp_attr->qp_access_flags = 0; 579 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS; 580 } else 581 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr, 582 qp_attr_mask); 583 break; 584 default: 585 ret = -ENOSYS; 586 break; 587 } 588 589 return ret; 590 } 591 EXPORT_SYMBOL(rdma_init_qp_attr); 592 593 static inline int cma_zero_addr(struct sockaddr *addr) 594 { 595 struct in6_addr *ip6; 596 597 if (addr->sa_family == AF_INET) 598 return ZERONET(((struct sockaddr_in *) addr)->sin_addr.s_addr); 599 else { 600 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr; 601 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] | 602 ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0; 603 } 604 } 605 606 static inline int cma_loopback_addr(struct sockaddr *addr) 607 { 608 return LOOPBACK(((struct sockaddr_in *) addr)->sin_addr.s_addr); 609 } 610 611 static inline int cma_any_addr(struct sockaddr *addr) 612 { 613 return cma_zero_addr(addr) || cma_loopback_addr(addr); 614 } 615 616 static inline __be16 cma_port(struct sockaddr *addr) 617 { 618 if (addr->sa_family == AF_INET) 619 return ((struct sockaddr_in *) addr)->sin_port; 620 else 621 return ((struct sockaddr_in6 *) addr)->sin6_port; 622 } 623 624 static inline int cma_any_port(struct sockaddr *addr) 625 { 626 return !cma_port(addr); 627 } 628 629 static int cma_get_net_info(void *hdr, enum rdma_port_space ps, 630 u8 *ip_ver, __u16 *port, 631 union cma_ip_addr **src, union cma_ip_addr **dst) 632 { 633 switch (ps) { 634 case RDMA_PS_SDP: 635 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) != 636 SDP_MAJ_VERSION) 637 return -EINVAL; 638 639 *ip_ver = sdp_get_ip_ver(hdr); 640 *port = ((struct sdp_hh *) hdr)->port; 641 *src = &((struct sdp_hh *) hdr)->src_addr; 642 *dst = &((struct sdp_hh *) hdr)->dst_addr; 643 break; 644 default: 645 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION) 646 return -EINVAL; 647 648 *ip_ver = cma_get_ip_ver(hdr); 649 *port = ((struct cma_hdr *) hdr)->port; 650 *src = &((struct cma_hdr *) hdr)->src_addr; 651 *dst = &((struct cma_hdr *) hdr)->dst_addr; 652 break; 653 } 654 655 if (*ip_ver != 4 && *ip_ver != 6) 656 return -EINVAL; 657 return 0; 658 } 659 660 static void cma_save_net_info(struct rdma_addr *addr, 661 struct rdma_addr *listen_addr, 662 u8 ip_ver, __u16 port, 663 union cma_ip_addr *src, union cma_ip_addr *dst) 664 { 665 struct sockaddr_in *listen4, *ip4; 666 struct sockaddr_in6 *listen6, *ip6; 667 668 switch (ip_ver) { 669 case 4: 670 listen4 = (struct sockaddr_in *) &listen_addr->src_addr; 671 ip4 = (struct sockaddr_in *) &addr->src_addr; 672 ip4->sin_family = listen4->sin_family; 673 ip4->sin_addr.s_addr = dst->ip4.addr; 674 ip4->sin_port = listen4->sin_port; 675 676 ip4 = (struct sockaddr_in *) &addr->dst_addr; 677 ip4->sin_family = listen4->sin_family; 678 ip4->sin_addr.s_addr = src->ip4.addr; 679 ip4->sin_port = port; 680 break; 681 case 6: 682 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr; 683 ip6 = (struct sockaddr_in6 *) &addr->src_addr; 684 ip6->sin6_family = listen6->sin6_family; 685 ip6->sin6_addr = dst->ip6; 686 ip6->sin6_port = listen6->sin6_port; 687 688 ip6 = (struct sockaddr_in6 *) &addr->dst_addr; 689 ip6->sin6_family = listen6->sin6_family; 690 ip6->sin6_addr = src->ip6; 691 ip6->sin6_port = port; 692 break; 693 default: 694 break; 695 } 696 } 697 698 static inline int cma_user_data_offset(enum rdma_port_space ps) 699 { 700 switch (ps) { 701 case RDMA_PS_SDP: 702 return 0; 703 default: 704 return sizeof(struct cma_hdr); 705 } 706 } 707 708 static void cma_cancel_route(struct rdma_id_private *id_priv) 709 { 710 switch (rdma_node_get_transport(id_priv->id.device->node_type)) { 711 case RDMA_TRANSPORT_IB: 712 if (id_priv->query) 713 ib_sa_cancel_query(id_priv->query_id, id_priv->query); 714 break; 715 default: 716 break; 717 } 718 } 719 720 static inline int cma_internal_listen(struct rdma_id_private *id_priv) 721 { 722 return (id_priv->state == CMA_LISTEN) && id_priv->cma_dev && 723 cma_any_addr(&id_priv->id.route.addr.src_addr); 724 } 725 726 static void cma_destroy_listen(struct rdma_id_private *id_priv) 727 { 728 cma_exch(id_priv, CMA_DESTROYING); 729 730 if (id_priv->cma_dev) { 731 switch (rdma_node_get_transport(id_priv->id.device->node_type)) { 732 case RDMA_TRANSPORT_IB: 733 if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib)) 734 ib_destroy_cm_id(id_priv->cm_id.ib); 735 break; 736 case RDMA_TRANSPORT_IWARP: 737 if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw)) 738 iw_destroy_cm_id(id_priv->cm_id.iw); 739 break; 740 default: 741 break; 742 } 743 cma_detach_from_dev(id_priv); 744 } 745 list_del(&id_priv->listen_list); 746 747 cma_deref_id(id_priv); 748 wait_for_completion(&id_priv->comp); 749 750 kfree(id_priv); 751 } 752 753 static void cma_cancel_listens(struct rdma_id_private *id_priv) 754 { 755 struct rdma_id_private *dev_id_priv; 756 757 mutex_lock(&lock); 758 list_del(&id_priv->list); 759 760 while (!list_empty(&id_priv->listen_list)) { 761 dev_id_priv = list_entry(id_priv->listen_list.next, 762 struct rdma_id_private, listen_list); 763 cma_destroy_listen(dev_id_priv); 764 } 765 mutex_unlock(&lock); 766 } 767 768 static void cma_cancel_operation(struct rdma_id_private *id_priv, 769 enum cma_state state) 770 { 771 switch (state) { 772 case CMA_ADDR_QUERY: 773 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr); 774 break; 775 case CMA_ROUTE_QUERY: 776 cma_cancel_route(id_priv); 777 break; 778 case CMA_LISTEN: 779 if (cma_any_addr(&id_priv->id.route.addr.src_addr) && 780 !id_priv->cma_dev) 781 cma_cancel_listens(id_priv); 782 break; 783 default: 784 break; 785 } 786 } 787 788 static void cma_release_port(struct rdma_id_private *id_priv) 789 { 790 struct rdma_bind_list *bind_list = id_priv->bind_list; 791 792 if (!bind_list) 793 return; 794 795 mutex_lock(&lock); 796 hlist_del(&id_priv->node); 797 if (hlist_empty(&bind_list->owners)) { 798 idr_remove(bind_list->ps, bind_list->port); 799 kfree(bind_list); 800 } 801 mutex_unlock(&lock); 802 } 803 804 static void cma_leave_mc_groups(struct rdma_id_private *id_priv) 805 { 806 struct cma_multicast *mc; 807 808 while (!list_empty(&id_priv->mc_list)) { 809 mc = container_of(id_priv->mc_list.next, 810 struct cma_multicast, list); 811 list_del(&mc->list); 812 ib_sa_free_multicast(mc->multicast.ib); 813 kfree(mc); 814 } 815 } 816 817 void rdma_destroy_id(struct rdma_cm_id *id) 818 { 819 struct rdma_id_private *id_priv; 820 enum cma_state state; 821 822 id_priv = container_of(id, struct rdma_id_private, id); 823 state = cma_exch(id_priv, CMA_DESTROYING); 824 cma_cancel_operation(id_priv, state); 825 826 mutex_lock(&lock); 827 if (id_priv->cma_dev) { 828 mutex_unlock(&lock); 829 switch (rdma_node_get_transport(id->device->node_type)) { 830 case RDMA_TRANSPORT_IB: 831 if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib)) 832 ib_destroy_cm_id(id_priv->cm_id.ib); 833 break; 834 case RDMA_TRANSPORT_IWARP: 835 if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw)) 836 iw_destroy_cm_id(id_priv->cm_id.iw); 837 break; 838 default: 839 break; 840 } 841 cma_leave_mc_groups(id_priv); 842 mutex_lock(&lock); 843 cma_detach_from_dev(id_priv); 844 } 845 mutex_unlock(&lock); 846 847 cma_release_port(id_priv); 848 cma_deref_id(id_priv); 849 wait_for_completion(&id_priv->comp); 850 851 kfree(id_priv->id.route.path_rec); 852 kfree(id_priv); 853 } 854 EXPORT_SYMBOL(rdma_destroy_id); 855 856 static int cma_rep_recv(struct rdma_id_private *id_priv) 857 { 858 int ret; 859 860 ret = cma_modify_qp_rtr(&id_priv->id); 861 if (ret) 862 goto reject; 863 864 ret = cma_modify_qp_rts(&id_priv->id); 865 if (ret) 866 goto reject; 867 868 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0); 869 if (ret) 870 goto reject; 871 872 return 0; 873 reject: 874 cma_modify_qp_err(&id_priv->id); 875 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED, 876 NULL, 0, NULL, 0); 877 return ret; 878 } 879 880 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data) 881 { 882 if (id_priv->id.ps == RDMA_PS_SDP && 883 sdp_get_majv(((struct sdp_hah *) data)->sdp_version) != 884 SDP_MAJ_VERSION) 885 return -EINVAL; 886 887 return 0; 888 } 889 890 static void cma_set_rep_event_data(struct rdma_cm_event *event, 891 struct ib_cm_rep_event_param *rep_data, 892 void *private_data) 893 { 894 event->param.conn.private_data = private_data; 895 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE; 896 event->param.conn.responder_resources = rep_data->responder_resources; 897 event->param.conn.initiator_depth = rep_data->initiator_depth; 898 event->param.conn.flow_control = rep_data->flow_control; 899 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count; 900 event->param.conn.srq = rep_data->srq; 901 event->param.conn.qp_num = rep_data->remote_qpn; 902 } 903 904 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event) 905 { 906 struct rdma_id_private *id_priv = cm_id->context; 907 struct rdma_cm_event event; 908 int ret = 0; 909 910 if (cma_disable_remove(id_priv, CMA_CONNECT)) 911 return 0; 912 913 memset(&event, 0, sizeof event); 914 switch (ib_event->event) { 915 case IB_CM_REQ_ERROR: 916 case IB_CM_REP_ERROR: 917 event.event = RDMA_CM_EVENT_UNREACHABLE; 918 event.status = -ETIMEDOUT; 919 break; 920 case IB_CM_REP_RECEIVED: 921 event.status = cma_verify_rep(id_priv, ib_event->private_data); 922 if (event.status) 923 event.event = RDMA_CM_EVENT_CONNECT_ERROR; 924 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) { 925 event.status = cma_rep_recv(id_priv); 926 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR : 927 RDMA_CM_EVENT_ESTABLISHED; 928 } else 929 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE; 930 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd, 931 ib_event->private_data); 932 break; 933 case IB_CM_RTU_RECEIVED: 934 case IB_CM_USER_ESTABLISHED: 935 event.event = RDMA_CM_EVENT_ESTABLISHED; 936 break; 937 case IB_CM_DREQ_ERROR: 938 event.status = -ETIMEDOUT; /* fall through */ 939 case IB_CM_DREQ_RECEIVED: 940 case IB_CM_DREP_RECEIVED: 941 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT)) 942 goto out; 943 event.event = RDMA_CM_EVENT_DISCONNECTED; 944 break; 945 case IB_CM_TIMEWAIT_EXIT: 946 case IB_CM_MRA_RECEIVED: 947 /* ignore event */ 948 goto out; 949 case IB_CM_REJ_RECEIVED: 950 cma_modify_qp_err(&id_priv->id); 951 event.status = ib_event->param.rej_rcvd.reason; 952 event.event = RDMA_CM_EVENT_REJECTED; 953 event.param.conn.private_data = ib_event->private_data; 954 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE; 955 break; 956 default: 957 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d", 958 ib_event->event); 959 goto out; 960 } 961 962 ret = id_priv->id.event_handler(&id_priv->id, &event); 963 if (ret) { 964 /* Destroy the CM ID by returning a non-zero value. */ 965 id_priv->cm_id.ib = NULL; 966 cma_exch(id_priv, CMA_DESTROYING); 967 cma_enable_remove(id_priv); 968 rdma_destroy_id(&id_priv->id); 969 return ret; 970 } 971 out: 972 cma_enable_remove(id_priv); 973 return ret; 974 } 975 976 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id, 977 struct ib_cm_event *ib_event) 978 { 979 struct rdma_id_private *id_priv; 980 struct rdma_cm_id *id; 981 struct rdma_route *rt; 982 union cma_ip_addr *src, *dst; 983 __u16 port; 984 u8 ip_ver; 985 986 if (cma_get_net_info(ib_event->private_data, listen_id->ps, 987 &ip_ver, &port, &src, &dst)) 988 goto err; 989 990 id = rdma_create_id(listen_id->event_handler, listen_id->context, 991 listen_id->ps); 992 if (IS_ERR(id)) 993 goto err; 994 995 cma_save_net_info(&id->route.addr, &listen_id->route.addr, 996 ip_ver, port, src, dst); 997 998 rt = &id->route; 999 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1; 1000 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths, 1001 GFP_KERNEL); 1002 if (!rt->path_rec) 1003 goto destroy_id; 1004 1005 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path; 1006 if (rt->num_paths == 2) 1007 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path; 1008 1009 ib_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid); 1010 ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid); 1011 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey)); 1012 rt->addr.dev_addr.dev_type = RDMA_NODE_IB_CA; 1013 1014 id_priv = container_of(id, struct rdma_id_private, id); 1015 id_priv->state = CMA_CONNECT; 1016 return id_priv; 1017 1018 destroy_id: 1019 rdma_destroy_id(id); 1020 err: 1021 return NULL; 1022 } 1023 1024 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id, 1025 struct ib_cm_event *ib_event) 1026 { 1027 struct rdma_id_private *id_priv; 1028 struct rdma_cm_id *id; 1029 union cma_ip_addr *src, *dst; 1030 __u16 port; 1031 u8 ip_ver; 1032 int ret; 1033 1034 id = rdma_create_id(listen_id->event_handler, listen_id->context, 1035 listen_id->ps); 1036 if (IS_ERR(id)) 1037 return NULL; 1038 1039 1040 if (cma_get_net_info(ib_event->private_data, listen_id->ps, 1041 &ip_ver, &port, &src, &dst)) 1042 goto err; 1043 1044 cma_save_net_info(&id->route.addr, &listen_id->route.addr, 1045 ip_ver, port, src, dst); 1046 1047 ret = rdma_translate_ip(&id->route.addr.src_addr, 1048 &id->route.addr.dev_addr); 1049 if (ret) 1050 goto err; 1051 1052 id_priv = container_of(id, struct rdma_id_private, id); 1053 id_priv->state = CMA_CONNECT; 1054 return id_priv; 1055 err: 1056 rdma_destroy_id(id); 1057 return NULL; 1058 } 1059 1060 static void cma_set_req_event_data(struct rdma_cm_event *event, 1061 struct ib_cm_req_event_param *req_data, 1062 void *private_data, int offset) 1063 { 1064 event->param.conn.private_data = private_data + offset; 1065 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset; 1066 event->param.conn.responder_resources = req_data->responder_resources; 1067 event->param.conn.initiator_depth = req_data->initiator_depth; 1068 event->param.conn.flow_control = req_data->flow_control; 1069 event->param.conn.retry_count = req_data->retry_count; 1070 event->param.conn.rnr_retry_count = req_data->rnr_retry_count; 1071 event->param.conn.srq = req_data->srq; 1072 event->param.conn.qp_num = req_data->remote_qpn; 1073 } 1074 1075 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event) 1076 { 1077 struct rdma_id_private *listen_id, *conn_id; 1078 struct rdma_cm_event event; 1079 int offset, ret; 1080 1081 listen_id = cm_id->context; 1082 if (cma_disable_remove(listen_id, CMA_LISTEN)) 1083 return -ECONNABORTED; 1084 1085 memset(&event, 0, sizeof event); 1086 offset = cma_user_data_offset(listen_id->id.ps); 1087 event.event = RDMA_CM_EVENT_CONNECT_REQUEST; 1088 if (cma_is_ud_ps(listen_id->id.ps)) { 1089 conn_id = cma_new_udp_id(&listen_id->id, ib_event); 1090 event.param.ud.private_data = ib_event->private_data + offset; 1091 event.param.ud.private_data_len = 1092 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset; 1093 } else { 1094 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0); 1095 conn_id = cma_new_conn_id(&listen_id->id, ib_event); 1096 cma_set_req_event_data(&event, &ib_event->param.req_rcvd, 1097 ib_event->private_data, offset); 1098 } 1099 if (!conn_id) { 1100 ret = -ENOMEM; 1101 goto out; 1102 } 1103 1104 atomic_inc(&conn_id->dev_remove); 1105 mutex_lock(&lock); 1106 ret = cma_acquire_dev(conn_id); 1107 mutex_unlock(&lock); 1108 if (ret) 1109 goto release_conn_id; 1110 1111 conn_id->cm_id.ib = cm_id; 1112 cm_id->context = conn_id; 1113 cm_id->cm_handler = cma_ib_handler; 1114 1115 ret = conn_id->id.event_handler(&conn_id->id, &event); 1116 if (!ret) 1117 goto out; 1118 1119 /* Destroy the CM ID by returning a non-zero value. */ 1120 conn_id->cm_id.ib = NULL; 1121 1122 release_conn_id: 1123 cma_exch(conn_id, CMA_DESTROYING); 1124 cma_enable_remove(conn_id); 1125 rdma_destroy_id(&conn_id->id); 1126 1127 out: 1128 cma_enable_remove(listen_id); 1129 return ret; 1130 } 1131 1132 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr) 1133 { 1134 return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr))); 1135 } 1136 1137 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr, 1138 struct ib_cm_compare_data *compare) 1139 { 1140 struct cma_hdr *cma_data, *cma_mask; 1141 struct sdp_hh *sdp_data, *sdp_mask; 1142 __u32 ip4_addr; 1143 struct in6_addr ip6_addr; 1144 1145 memset(compare, 0, sizeof *compare); 1146 cma_data = (void *) compare->data; 1147 cma_mask = (void *) compare->mask; 1148 sdp_data = (void *) compare->data; 1149 sdp_mask = (void *) compare->mask; 1150 1151 switch (addr->sa_family) { 1152 case AF_INET: 1153 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr; 1154 if (ps == RDMA_PS_SDP) { 1155 sdp_set_ip_ver(sdp_data, 4); 1156 sdp_set_ip_ver(sdp_mask, 0xF); 1157 sdp_data->dst_addr.ip4.addr = ip4_addr; 1158 sdp_mask->dst_addr.ip4.addr = ~0; 1159 } else { 1160 cma_set_ip_ver(cma_data, 4); 1161 cma_set_ip_ver(cma_mask, 0xF); 1162 cma_data->dst_addr.ip4.addr = ip4_addr; 1163 cma_mask->dst_addr.ip4.addr = ~0; 1164 } 1165 break; 1166 case AF_INET6: 1167 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr; 1168 if (ps == RDMA_PS_SDP) { 1169 sdp_set_ip_ver(sdp_data, 6); 1170 sdp_set_ip_ver(sdp_mask, 0xF); 1171 sdp_data->dst_addr.ip6 = ip6_addr; 1172 memset(&sdp_mask->dst_addr.ip6, 0xFF, 1173 sizeof sdp_mask->dst_addr.ip6); 1174 } else { 1175 cma_set_ip_ver(cma_data, 6); 1176 cma_set_ip_ver(cma_mask, 0xF); 1177 cma_data->dst_addr.ip6 = ip6_addr; 1178 memset(&cma_mask->dst_addr.ip6, 0xFF, 1179 sizeof cma_mask->dst_addr.ip6); 1180 } 1181 break; 1182 default: 1183 break; 1184 } 1185 } 1186 1187 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event) 1188 { 1189 struct rdma_id_private *id_priv = iw_id->context; 1190 struct rdma_cm_event event; 1191 struct sockaddr_in *sin; 1192 int ret = 0; 1193 1194 if (cma_disable_remove(id_priv, CMA_CONNECT)) 1195 return 0; 1196 1197 memset(&event, 0, sizeof event); 1198 switch (iw_event->event) { 1199 case IW_CM_EVENT_CLOSE: 1200 event.event = RDMA_CM_EVENT_DISCONNECTED; 1201 break; 1202 case IW_CM_EVENT_CONNECT_REPLY: 1203 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr; 1204 *sin = iw_event->local_addr; 1205 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr; 1206 *sin = iw_event->remote_addr; 1207 switch (iw_event->status) { 1208 case 0: 1209 event.event = RDMA_CM_EVENT_ESTABLISHED; 1210 break; 1211 case -ECONNRESET: 1212 case -ECONNREFUSED: 1213 event.event = RDMA_CM_EVENT_REJECTED; 1214 break; 1215 case -ETIMEDOUT: 1216 event.event = RDMA_CM_EVENT_UNREACHABLE; 1217 break; 1218 default: 1219 event.event = RDMA_CM_EVENT_CONNECT_ERROR; 1220 break; 1221 } 1222 break; 1223 case IW_CM_EVENT_ESTABLISHED: 1224 event.event = RDMA_CM_EVENT_ESTABLISHED; 1225 break; 1226 default: 1227 BUG_ON(1); 1228 } 1229 1230 event.status = iw_event->status; 1231 event.param.conn.private_data = iw_event->private_data; 1232 event.param.conn.private_data_len = iw_event->private_data_len; 1233 ret = id_priv->id.event_handler(&id_priv->id, &event); 1234 if (ret) { 1235 /* Destroy the CM ID by returning a non-zero value. */ 1236 id_priv->cm_id.iw = NULL; 1237 cma_exch(id_priv, CMA_DESTROYING); 1238 cma_enable_remove(id_priv); 1239 rdma_destroy_id(&id_priv->id); 1240 return ret; 1241 } 1242 1243 cma_enable_remove(id_priv); 1244 return ret; 1245 } 1246 1247 static int iw_conn_req_handler(struct iw_cm_id *cm_id, 1248 struct iw_cm_event *iw_event) 1249 { 1250 struct rdma_cm_id *new_cm_id; 1251 struct rdma_id_private *listen_id, *conn_id; 1252 struct sockaddr_in *sin; 1253 struct net_device *dev = NULL; 1254 struct rdma_cm_event event; 1255 int ret; 1256 1257 listen_id = cm_id->context; 1258 if (cma_disable_remove(listen_id, CMA_LISTEN)) 1259 return -ECONNABORTED; 1260 1261 /* Create a new RDMA id for the new IW CM ID */ 1262 new_cm_id = rdma_create_id(listen_id->id.event_handler, 1263 listen_id->id.context, 1264 RDMA_PS_TCP); 1265 if (!new_cm_id) { 1266 ret = -ENOMEM; 1267 goto out; 1268 } 1269 conn_id = container_of(new_cm_id, struct rdma_id_private, id); 1270 atomic_inc(&conn_id->dev_remove); 1271 conn_id->state = CMA_CONNECT; 1272 1273 dev = ip_dev_find(iw_event->local_addr.sin_addr.s_addr); 1274 if (!dev) { 1275 ret = -EADDRNOTAVAIL; 1276 cma_enable_remove(conn_id); 1277 rdma_destroy_id(new_cm_id); 1278 goto out; 1279 } 1280 ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL); 1281 if (ret) { 1282 cma_enable_remove(conn_id); 1283 rdma_destroy_id(new_cm_id); 1284 goto out; 1285 } 1286 1287 mutex_lock(&lock); 1288 ret = cma_acquire_dev(conn_id); 1289 mutex_unlock(&lock); 1290 if (ret) { 1291 cma_enable_remove(conn_id); 1292 rdma_destroy_id(new_cm_id); 1293 goto out; 1294 } 1295 1296 conn_id->cm_id.iw = cm_id; 1297 cm_id->context = conn_id; 1298 cm_id->cm_handler = cma_iw_handler; 1299 1300 sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr; 1301 *sin = iw_event->local_addr; 1302 sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr; 1303 *sin = iw_event->remote_addr; 1304 1305 memset(&event, 0, sizeof event); 1306 event.event = RDMA_CM_EVENT_CONNECT_REQUEST; 1307 event.param.conn.private_data = iw_event->private_data; 1308 event.param.conn.private_data_len = iw_event->private_data_len; 1309 ret = conn_id->id.event_handler(&conn_id->id, &event); 1310 if (ret) { 1311 /* User wants to destroy the CM ID */ 1312 conn_id->cm_id.iw = NULL; 1313 cma_exch(conn_id, CMA_DESTROYING); 1314 cma_enable_remove(conn_id); 1315 rdma_destroy_id(&conn_id->id); 1316 } 1317 1318 out: 1319 if (dev) 1320 dev_put(dev); 1321 cma_enable_remove(listen_id); 1322 return ret; 1323 } 1324 1325 static int cma_ib_listen(struct rdma_id_private *id_priv) 1326 { 1327 struct ib_cm_compare_data compare_data; 1328 struct sockaddr *addr; 1329 __be64 svc_id; 1330 int ret; 1331 1332 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler, 1333 id_priv); 1334 if (IS_ERR(id_priv->cm_id.ib)) 1335 return PTR_ERR(id_priv->cm_id.ib); 1336 1337 addr = &id_priv->id.route.addr.src_addr; 1338 svc_id = cma_get_service_id(id_priv->id.ps, addr); 1339 if (cma_any_addr(addr)) 1340 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL); 1341 else { 1342 cma_set_compare_data(id_priv->id.ps, addr, &compare_data); 1343 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data); 1344 } 1345 1346 if (ret) { 1347 ib_destroy_cm_id(id_priv->cm_id.ib); 1348 id_priv->cm_id.ib = NULL; 1349 } 1350 1351 return ret; 1352 } 1353 1354 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog) 1355 { 1356 int ret; 1357 struct sockaddr_in *sin; 1358 1359 id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device, 1360 iw_conn_req_handler, 1361 id_priv); 1362 if (IS_ERR(id_priv->cm_id.iw)) 1363 return PTR_ERR(id_priv->cm_id.iw); 1364 1365 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr; 1366 id_priv->cm_id.iw->local_addr = *sin; 1367 1368 ret = iw_cm_listen(id_priv->cm_id.iw, backlog); 1369 1370 if (ret) { 1371 iw_destroy_cm_id(id_priv->cm_id.iw); 1372 id_priv->cm_id.iw = NULL; 1373 } 1374 1375 return ret; 1376 } 1377 1378 static int cma_listen_handler(struct rdma_cm_id *id, 1379 struct rdma_cm_event *event) 1380 { 1381 struct rdma_id_private *id_priv = id->context; 1382 1383 id->context = id_priv->id.context; 1384 id->event_handler = id_priv->id.event_handler; 1385 return id_priv->id.event_handler(id, event); 1386 } 1387 1388 static void cma_listen_on_dev(struct rdma_id_private *id_priv, 1389 struct cma_device *cma_dev) 1390 { 1391 struct rdma_id_private *dev_id_priv; 1392 struct rdma_cm_id *id; 1393 int ret; 1394 1395 id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps); 1396 if (IS_ERR(id)) 1397 return; 1398 1399 dev_id_priv = container_of(id, struct rdma_id_private, id); 1400 1401 dev_id_priv->state = CMA_ADDR_BOUND; 1402 memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr, 1403 ip_addr_size(&id_priv->id.route.addr.src_addr)); 1404 1405 cma_attach_to_dev(dev_id_priv, cma_dev); 1406 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list); 1407 1408 ret = rdma_listen(id, id_priv->backlog); 1409 if (ret) 1410 goto err; 1411 1412 return; 1413 err: 1414 cma_destroy_listen(dev_id_priv); 1415 } 1416 1417 static void cma_listen_on_all(struct rdma_id_private *id_priv) 1418 { 1419 struct cma_device *cma_dev; 1420 1421 mutex_lock(&lock); 1422 list_add_tail(&id_priv->list, &listen_any_list); 1423 list_for_each_entry(cma_dev, &dev_list, list) 1424 cma_listen_on_dev(id_priv, cma_dev); 1425 mutex_unlock(&lock); 1426 } 1427 1428 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af) 1429 { 1430 struct sockaddr_in addr_in; 1431 1432 memset(&addr_in, 0, sizeof addr_in); 1433 addr_in.sin_family = af; 1434 return rdma_bind_addr(id, (struct sockaddr *) &addr_in); 1435 } 1436 1437 int rdma_listen(struct rdma_cm_id *id, int backlog) 1438 { 1439 struct rdma_id_private *id_priv; 1440 int ret; 1441 1442 id_priv = container_of(id, struct rdma_id_private, id); 1443 if (id_priv->state == CMA_IDLE) { 1444 ret = cma_bind_any(id, AF_INET); 1445 if (ret) 1446 return ret; 1447 } 1448 1449 if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN)) 1450 return -EINVAL; 1451 1452 id_priv->backlog = backlog; 1453 if (id->device) { 1454 switch (rdma_node_get_transport(id->device->node_type)) { 1455 case RDMA_TRANSPORT_IB: 1456 ret = cma_ib_listen(id_priv); 1457 if (ret) 1458 goto err; 1459 break; 1460 case RDMA_TRANSPORT_IWARP: 1461 ret = cma_iw_listen(id_priv, backlog); 1462 if (ret) 1463 goto err; 1464 break; 1465 default: 1466 ret = -ENOSYS; 1467 goto err; 1468 } 1469 } else 1470 cma_listen_on_all(id_priv); 1471 1472 return 0; 1473 err: 1474 id_priv->backlog = 0; 1475 cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND); 1476 return ret; 1477 } 1478 EXPORT_SYMBOL(rdma_listen); 1479 1480 void rdma_set_service_type(struct rdma_cm_id *id, int tos) 1481 { 1482 struct rdma_id_private *id_priv; 1483 1484 id_priv = container_of(id, struct rdma_id_private, id); 1485 id_priv->tos = (u8) tos; 1486 } 1487 EXPORT_SYMBOL(rdma_set_service_type); 1488 1489 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec, 1490 void *context) 1491 { 1492 struct cma_work *work = context; 1493 struct rdma_route *route; 1494 1495 route = &work->id->id.route; 1496 1497 if (!status) { 1498 route->num_paths = 1; 1499 *route->path_rec = *path_rec; 1500 } else { 1501 work->old_state = CMA_ROUTE_QUERY; 1502 work->new_state = CMA_ADDR_RESOLVED; 1503 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR; 1504 work->event.status = status; 1505 } 1506 1507 queue_work(cma_wq, &work->work); 1508 } 1509 1510 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms, 1511 struct cma_work *work) 1512 { 1513 struct rdma_addr *addr = &id_priv->id.route.addr; 1514 struct ib_sa_path_rec path_rec; 1515 ib_sa_comp_mask comp_mask; 1516 struct sockaddr_in6 *sin6; 1517 1518 memset(&path_rec, 0, sizeof path_rec); 1519 ib_addr_get_sgid(&addr->dev_addr, &path_rec.sgid); 1520 ib_addr_get_dgid(&addr->dev_addr, &path_rec.dgid); 1521 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr)); 1522 path_rec.numb_path = 1; 1523 path_rec.reversible = 1; 1524 path_rec.service_id = cma_get_service_id(id_priv->id.ps, &addr->dst_addr); 1525 1526 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID | 1527 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH | 1528 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID; 1529 1530 if (addr->src_addr.sa_family == AF_INET) { 1531 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos); 1532 comp_mask |= IB_SA_PATH_REC_QOS_CLASS; 1533 } else { 1534 sin6 = (struct sockaddr_in6 *) &addr->src_addr; 1535 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20); 1536 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS; 1537 } 1538 1539 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device, 1540 id_priv->id.port_num, &path_rec, 1541 comp_mask, timeout_ms, 1542 GFP_KERNEL, cma_query_handler, 1543 work, &id_priv->query); 1544 1545 return (id_priv->query_id < 0) ? id_priv->query_id : 0; 1546 } 1547 1548 static void cma_work_handler(struct work_struct *_work) 1549 { 1550 struct cma_work *work = container_of(_work, struct cma_work, work); 1551 struct rdma_id_private *id_priv = work->id; 1552 int destroy = 0; 1553 1554 atomic_inc(&id_priv->dev_remove); 1555 if (!cma_comp_exch(id_priv, work->old_state, work->new_state)) 1556 goto out; 1557 1558 if (id_priv->id.event_handler(&id_priv->id, &work->event)) { 1559 cma_exch(id_priv, CMA_DESTROYING); 1560 destroy = 1; 1561 } 1562 out: 1563 cma_enable_remove(id_priv); 1564 cma_deref_id(id_priv); 1565 if (destroy) 1566 rdma_destroy_id(&id_priv->id); 1567 kfree(work); 1568 } 1569 1570 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms) 1571 { 1572 struct rdma_route *route = &id_priv->id.route; 1573 struct cma_work *work; 1574 int ret; 1575 1576 work = kzalloc(sizeof *work, GFP_KERNEL); 1577 if (!work) 1578 return -ENOMEM; 1579 1580 work->id = id_priv; 1581 INIT_WORK(&work->work, cma_work_handler); 1582 work->old_state = CMA_ROUTE_QUERY; 1583 work->new_state = CMA_ROUTE_RESOLVED; 1584 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 1585 1586 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL); 1587 if (!route->path_rec) { 1588 ret = -ENOMEM; 1589 goto err1; 1590 } 1591 1592 ret = cma_query_ib_route(id_priv, timeout_ms, work); 1593 if (ret) 1594 goto err2; 1595 1596 return 0; 1597 err2: 1598 kfree(route->path_rec); 1599 route->path_rec = NULL; 1600 err1: 1601 kfree(work); 1602 return ret; 1603 } 1604 1605 int rdma_set_ib_paths(struct rdma_cm_id *id, 1606 struct ib_sa_path_rec *path_rec, int num_paths) 1607 { 1608 struct rdma_id_private *id_priv; 1609 int ret; 1610 1611 id_priv = container_of(id, struct rdma_id_private, id); 1612 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED)) 1613 return -EINVAL; 1614 1615 id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL); 1616 if (!id->route.path_rec) { 1617 ret = -ENOMEM; 1618 goto err; 1619 } 1620 1621 memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths); 1622 return 0; 1623 err: 1624 cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED); 1625 return ret; 1626 } 1627 EXPORT_SYMBOL(rdma_set_ib_paths); 1628 1629 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms) 1630 { 1631 struct cma_work *work; 1632 1633 work = kzalloc(sizeof *work, GFP_KERNEL); 1634 if (!work) 1635 return -ENOMEM; 1636 1637 work->id = id_priv; 1638 INIT_WORK(&work->work, cma_work_handler); 1639 work->old_state = CMA_ROUTE_QUERY; 1640 work->new_state = CMA_ROUTE_RESOLVED; 1641 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 1642 queue_work(cma_wq, &work->work); 1643 return 0; 1644 } 1645 1646 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms) 1647 { 1648 struct rdma_id_private *id_priv; 1649 int ret; 1650 1651 id_priv = container_of(id, struct rdma_id_private, id); 1652 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY)) 1653 return -EINVAL; 1654 1655 atomic_inc(&id_priv->refcount); 1656 switch (rdma_node_get_transport(id->device->node_type)) { 1657 case RDMA_TRANSPORT_IB: 1658 ret = cma_resolve_ib_route(id_priv, timeout_ms); 1659 break; 1660 case RDMA_TRANSPORT_IWARP: 1661 ret = cma_resolve_iw_route(id_priv, timeout_ms); 1662 break; 1663 default: 1664 ret = -ENOSYS; 1665 break; 1666 } 1667 if (ret) 1668 goto err; 1669 1670 return 0; 1671 err: 1672 cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED); 1673 cma_deref_id(id_priv); 1674 return ret; 1675 } 1676 EXPORT_SYMBOL(rdma_resolve_route); 1677 1678 static int cma_bind_loopback(struct rdma_id_private *id_priv) 1679 { 1680 struct cma_device *cma_dev; 1681 struct ib_port_attr port_attr; 1682 union ib_gid gid; 1683 u16 pkey; 1684 int ret; 1685 u8 p; 1686 1687 mutex_lock(&lock); 1688 if (list_empty(&dev_list)) { 1689 ret = -ENODEV; 1690 goto out; 1691 } 1692 list_for_each_entry(cma_dev, &dev_list, list) 1693 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p) 1694 if (!ib_query_port(cma_dev->device, p, &port_attr) && 1695 port_attr.state == IB_PORT_ACTIVE) 1696 goto port_found; 1697 1698 p = 1; 1699 cma_dev = list_entry(dev_list.next, struct cma_device, list); 1700 1701 port_found: 1702 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid); 1703 if (ret) 1704 goto out; 1705 1706 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey); 1707 if (ret) 1708 goto out; 1709 1710 ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid); 1711 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey); 1712 id_priv->id.port_num = p; 1713 cma_attach_to_dev(id_priv, cma_dev); 1714 out: 1715 mutex_unlock(&lock); 1716 return ret; 1717 } 1718 1719 static void addr_handler(int status, struct sockaddr *src_addr, 1720 struct rdma_dev_addr *dev_addr, void *context) 1721 { 1722 struct rdma_id_private *id_priv = context; 1723 struct rdma_cm_event event; 1724 1725 memset(&event, 0, sizeof event); 1726 atomic_inc(&id_priv->dev_remove); 1727 1728 /* 1729 * Grab mutex to block rdma_destroy_id() from removing the device while 1730 * we're trying to acquire it. 1731 */ 1732 mutex_lock(&lock); 1733 if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) { 1734 mutex_unlock(&lock); 1735 goto out; 1736 } 1737 1738 if (!status && !id_priv->cma_dev) 1739 status = cma_acquire_dev(id_priv); 1740 mutex_unlock(&lock); 1741 1742 if (status) { 1743 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND)) 1744 goto out; 1745 event.event = RDMA_CM_EVENT_ADDR_ERROR; 1746 event.status = status; 1747 } else { 1748 memcpy(&id_priv->id.route.addr.src_addr, src_addr, 1749 ip_addr_size(src_addr)); 1750 event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 1751 } 1752 1753 if (id_priv->id.event_handler(&id_priv->id, &event)) { 1754 cma_exch(id_priv, CMA_DESTROYING); 1755 cma_enable_remove(id_priv); 1756 cma_deref_id(id_priv); 1757 rdma_destroy_id(&id_priv->id); 1758 return; 1759 } 1760 out: 1761 cma_enable_remove(id_priv); 1762 cma_deref_id(id_priv); 1763 } 1764 1765 static int cma_resolve_loopback(struct rdma_id_private *id_priv) 1766 { 1767 struct cma_work *work; 1768 struct sockaddr_in *src_in, *dst_in; 1769 union ib_gid gid; 1770 int ret; 1771 1772 work = kzalloc(sizeof *work, GFP_KERNEL); 1773 if (!work) 1774 return -ENOMEM; 1775 1776 if (!id_priv->cma_dev) { 1777 ret = cma_bind_loopback(id_priv); 1778 if (ret) 1779 goto err; 1780 } 1781 1782 ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid); 1783 ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid); 1784 1785 if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) { 1786 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr; 1787 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr; 1788 src_in->sin_family = dst_in->sin_family; 1789 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr; 1790 } 1791 1792 work->id = id_priv; 1793 INIT_WORK(&work->work, cma_work_handler); 1794 work->old_state = CMA_ADDR_QUERY; 1795 work->new_state = CMA_ADDR_RESOLVED; 1796 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 1797 queue_work(cma_wq, &work->work); 1798 return 0; 1799 err: 1800 kfree(work); 1801 return ret; 1802 } 1803 1804 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr, 1805 struct sockaddr *dst_addr) 1806 { 1807 if (src_addr && src_addr->sa_family) 1808 return rdma_bind_addr(id, src_addr); 1809 else 1810 return cma_bind_any(id, dst_addr->sa_family); 1811 } 1812 1813 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr, 1814 struct sockaddr *dst_addr, int timeout_ms) 1815 { 1816 struct rdma_id_private *id_priv; 1817 int ret; 1818 1819 id_priv = container_of(id, struct rdma_id_private, id); 1820 if (id_priv->state == CMA_IDLE) { 1821 ret = cma_bind_addr(id, src_addr, dst_addr); 1822 if (ret) 1823 return ret; 1824 } 1825 1826 if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY)) 1827 return -EINVAL; 1828 1829 atomic_inc(&id_priv->refcount); 1830 memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr)); 1831 if (cma_any_addr(dst_addr)) 1832 ret = cma_resolve_loopback(id_priv); 1833 else 1834 ret = rdma_resolve_ip(&addr_client, &id->route.addr.src_addr, 1835 dst_addr, &id->route.addr.dev_addr, 1836 timeout_ms, addr_handler, id_priv); 1837 if (ret) 1838 goto err; 1839 1840 return 0; 1841 err: 1842 cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND); 1843 cma_deref_id(id_priv); 1844 return ret; 1845 } 1846 EXPORT_SYMBOL(rdma_resolve_addr); 1847 1848 static void cma_bind_port(struct rdma_bind_list *bind_list, 1849 struct rdma_id_private *id_priv) 1850 { 1851 struct sockaddr_in *sin; 1852 1853 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr; 1854 sin->sin_port = htons(bind_list->port); 1855 id_priv->bind_list = bind_list; 1856 hlist_add_head(&id_priv->node, &bind_list->owners); 1857 } 1858 1859 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv, 1860 unsigned short snum) 1861 { 1862 struct rdma_bind_list *bind_list; 1863 int port, ret; 1864 1865 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL); 1866 if (!bind_list) 1867 return -ENOMEM; 1868 1869 do { 1870 ret = idr_get_new_above(ps, bind_list, snum, &port); 1871 } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL)); 1872 1873 if (ret) 1874 goto err1; 1875 1876 if (port != snum) { 1877 ret = -EADDRNOTAVAIL; 1878 goto err2; 1879 } 1880 1881 bind_list->ps = ps; 1882 bind_list->port = (unsigned short) port; 1883 cma_bind_port(bind_list, id_priv); 1884 return 0; 1885 err2: 1886 idr_remove(ps, port); 1887 err1: 1888 kfree(bind_list); 1889 return ret; 1890 } 1891 1892 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv) 1893 { 1894 struct rdma_bind_list *bind_list; 1895 int port, ret, low, high; 1896 1897 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL); 1898 if (!bind_list) 1899 return -ENOMEM; 1900 1901 retry: 1902 /* FIXME: add proper port randomization per like inet_csk_get_port */ 1903 do { 1904 ret = idr_get_new_above(ps, bind_list, next_port, &port); 1905 } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL)); 1906 1907 if (ret) 1908 goto err1; 1909 1910 inet_get_local_port_range(&low, &high); 1911 if (port > high) { 1912 if (next_port != low) { 1913 idr_remove(ps, port); 1914 next_port = low; 1915 goto retry; 1916 } 1917 ret = -EADDRNOTAVAIL; 1918 goto err2; 1919 } 1920 1921 if (port == high) 1922 next_port = low; 1923 else 1924 next_port = port + 1; 1925 1926 bind_list->ps = ps; 1927 bind_list->port = (unsigned short) port; 1928 cma_bind_port(bind_list, id_priv); 1929 return 0; 1930 err2: 1931 idr_remove(ps, port); 1932 err1: 1933 kfree(bind_list); 1934 return ret; 1935 } 1936 1937 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv) 1938 { 1939 struct rdma_id_private *cur_id; 1940 struct sockaddr_in *sin, *cur_sin; 1941 struct rdma_bind_list *bind_list; 1942 struct hlist_node *node; 1943 unsigned short snum; 1944 1945 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr; 1946 snum = ntohs(sin->sin_port); 1947 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE)) 1948 return -EACCES; 1949 1950 bind_list = idr_find(ps, snum); 1951 if (!bind_list) 1952 return cma_alloc_port(ps, id_priv, snum); 1953 1954 /* 1955 * We don't support binding to any address if anyone is bound to 1956 * a specific address on the same port. 1957 */ 1958 if (cma_any_addr(&id_priv->id.route.addr.src_addr)) 1959 return -EADDRNOTAVAIL; 1960 1961 hlist_for_each_entry(cur_id, node, &bind_list->owners, node) { 1962 if (cma_any_addr(&cur_id->id.route.addr.src_addr)) 1963 return -EADDRNOTAVAIL; 1964 1965 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr; 1966 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr) 1967 return -EADDRINUSE; 1968 } 1969 1970 cma_bind_port(bind_list, id_priv); 1971 return 0; 1972 } 1973 1974 static int cma_get_port(struct rdma_id_private *id_priv) 1975 { 1976 struct idr *ps; 1977 int ret; 1978 1979 switch (id_priv->id.ps) { 1980 case RDMA_PS_SDP: 1981 ps = &sdp_ps; 1982 break; 1983 case RDMA_PS_TCP: 1984 ps = &tcp_ps; 1985 break; 1986 case RDMA_PS_UDP: 1987 ps = &udp_ps; 1988 break; 1989 case RDMA_PS_IPOIB: 1990 ps = &ipoib_ps; 1991 break; 1992 default: 1993 return -EPROTONOSUPPORT; 1994 } 1995 1996 mutex_lock(&lock); 1997 if (cma_any_port(&id_priv->id.route.addr.src_addr)) 1998 ret = cma_alloc_any_port(ps, id_priv); 1999 else 2000 ret = cma_use_port(ps, id_priv); 2001 mutex_unlock(&lock); 2002 2003 return ret; 2004 } 2005 2006 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr) 2007 { 2008 struct rdma_id_private *id_priv; 2009 int ret; 2010 2011 if (addr->sa_family != AF_INET) 2012 return -EAFNOSUPPORT; 2013 2014 id_priv = container_of(id, struct rdma_id_private, id); 2015 if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND)) 2016 return -EINVAL; 2017 2018 if (!cma_any_addr(addr)) { 2019 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr); 2020 if (ret) 2021 goto err1; 2022 2023 mutex_lock(&lock); 2024 ret = cma_acquire_dev(id_priv); 2025 mutex_unlock(&lock); 2026 if (ret) 2027 goto err1; 2028 } 2029 2030 memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr)); 2031 ret = cma_get_port(id_priv); 2032 if (ret) 2033 goto err2; 2034 2035 return 0; 2036 err2: 2037 if (!cma_any_addr(addr)) { 2038 mutex_lock(&lock); 2039 cma_detach_from_dev(id_priv); 2040 mutex_unlock(&lock); 2041 } 2042 err1: 2043 cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE); 2044 return ret; 2045 } 2046 EXPORT_SYMBOL(rdma_bind_addr); 2047 2048 static int cma_format_hdr(void *hdr, enum rdma_port_space ps, 2049 struct rdma_route *route) 2050 { 2051 struct sockaddr_in *src4, *dst4; 2052 struct cma_hdr *cma_hdr; 2053 struct sdp_hh *sdp_hdr; 2054 2055 src4 = (struct sockaddr_in *) &route->addr.src_addr; 2056 dst4 = (struct sockaddr_in *) &route->addr.dst_addr; 2057 2058 switch (ps) { 2059 case RDMA_PS_SDP: 2060 sdp_hdr = hdr; 2061 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION) 2062 return -EINVAL; 2063 sdp_set_ip_ver(sdp_hdr, 4); 2064 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr; 2065 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr; 2066 sdp_hdr->port = src4->sin_port; 2067 break; 2068 default: 2069 cma_hdr = hdr; 2070 cma_hdr->cma_version = CMA_VERSION; 2071 cma_set_ip_ver(cma_hdr, 4); 2072 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr; 2073 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr; 2074 cma_hdr->port = src4->sin_port; 2075 break; 2076 } 2077 return 0; 2078 } 2079 2080 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id, 2081 struct ib_cm_event *ib_event) 2082 { 2083 struct rdma_id_private *id_priv = cm_id->context; 2084 struct rdma_cm_event event; 2085 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd; 2086 int ret = 0; 2087 2088 if (cma_disable_remove(id_priv, CMA_CONNECT)) 2089 return 0; 2090 2091 memset(&event, 0, sizeof event); 2092 switch (ib_event->event) { 2093 case IB_CM_SIDR_REQ_ERROR: 2094 event.event = RDMA_CM_EVENT_UNREACHABLE; 2095 event.status = -ETIMEDOUT; 2096 break; 2097 case IB_CM_SIDR_REP_RECEIVED: 2098 event.param.ud.private_data = ib_event->private_data; 2099 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE; 2100 if (rep->status != IB_SIDR_SUCCESS) { 2101 event.event = RDMA_CM_EVENT_UNREACHABLE; 2102 event.status = ib_event->param.sidr_rep_rcvd.status; 2103 break; 2104 } 2105 if (id_priv->qkey != rep->qkey) { 2106 event.event = RDMA_CM_EVENT_UNREACHABLE; 2107 event.status = -EINVAL; 2108 break; 2109 } 2110 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num, 2111 id_priv->id.route.path_rec, 2112 &event.param.ud.ah_attr); 2113 event.param.ud.qp_num = rep->qpn; 2114 event.param.ud.qkey = rep->qkey; 2115 event.event = RDMA_CM_EVENT_ESTABLISHED; 2116 event.status = 0; 2117 break; 2118 default: 2119 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d", 2120 ib_event->event); 2121 goto out; 2122 } 2123 2124 ret = id_priv->id.event_handler(&id_priv->id, &event); 2125 if (ret) { 2126 /* Destroy the CM ID by returning a non-zero value. */ 2127 id_priv->cm_id.ib = NULL; 2128 cma_exch(id_priv, CMA_DESTROYING); 2129 cma_enable_remove(id_priv); 2130 rdma_destroy_id(&id_priv->id); 2131 return ret; 2132 } 2133 out: 2134 cma_enable_remove(id_priv); 2135 return ret; 2136 } 2137 2138 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv, 2139 struct rdma_conn_param *conn_param) 2140 { 2141 struct ib_cm_sidr_req_param req; 2142 struct rdma_route *route; 2143 int ret; 2144 2145 req.private_data_len = sizeof(struct cma_hdr) + 2146 conn_param->private_data_len; 2147 req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC); 2148 if (!req.private_data) 2149 return -ENOMEM; 2150 2151 if (conn_param->private_data && conn_param->private_data_len) 2152 memcpy((void *) req.private_data + sizeof(struct cma_hdr), 2153 conn_param->private_data, conn_param->private_data_len); 2154 2155 route = &id_priv->id.route; 2156 ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route); 2157 if (ret) 2158 goto out; 2159 2160 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, 2161 cma_sidr_rep_handler, id_priv); 2162 if (IS_ERR(id_priv->cm_id.ib)) { 2163 ret = PTR_ERR(id_priv->cm_id.ib); 2164 goto out; 2165 } 2166 2167 req.path = route->path_rec; 2168 req.service_id = cma_get_service_id(id_priv->id.ps, 2169 &route->addr.dst_addr); 2170 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8); 2171 req.max_cm_retries = CMA_MAX_CM_RETRIES; 2172 2173 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req); 2174 if (ret) { 2175 ib_destroy_cm_id(id_priv->cm_id.ib); 2176 id_priv->cm_id.ib = NULL; 2177 } 2178 out: 2179 kfree(req.private_data); 2180 return ret; 2181 } 2182 2183 static int cma_connect_ib(struct rdma_id_private *id_priv, 2184 struct rdma_conn_param *conn_param) 2185 { 2186 struct ib_cm_req_param req; 2187 struct rdma_route *route; 2188 void *private_data; 2189 int offset, ret; 2190 2191 memset(&req, 0, sizeof req); 2192 offset = cma_user_data_offset(id_priv->id.ps); 2193 req.private_data_len = offset + conn_param->private_data_len; 2194 private_data = kzalloc(req.private_data_len, GFP_ATOMIC); 2195 if (!private_data) 2196 return -ENOMEM; 2197 2198 if (conn_param->private_data && conn_param->private_data_len) 2199 memcpy(private_data + offset, conn_param->private_data, 2200 conn_param->private_data_len); 2201 2202 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler, 2203 id_priv); 2204 if (IS_ERR(id_priv->cm_id.ib)) { 2205 ret = PTR_ERR(id_priv->cm_id.ib); 2206 goto out; 2207 } 2208 2209 route = &id_priv->id.route; 2210 ret = cma_format_hdr(private_data, id_priv->id.ps, route); 2211 if (ret) 2212 goto out; 2213 req.private_data = private_data; 2214 2215 req.primary_path = &route->path_rec[0]; 2216 if (route->num_paths == 2) 2217 req.alternate_path = &route->path_rec[1]; 2218 2219 req.service_id = cma_get_service_id(id_priv->id.ps, 2220 &route->addr.dst_addr); 2221 req.qp_num = id_priv->qp_num; 2222 req.qp_type = IB_QPT_RC; 2223 req.starting_psn = id_priv->seq_num; 2224 req.responder_resources = conn_param->responder_resources; 2225 req.initiator_depth = conn_param->initiator_depth; 2226 req.flow_control = conn_param->flow_control; 2227 req.retry_count = conn_param->retry_count; 2228 req.rnr_retry_count = conn_param->rnr_retry_count; 2229 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT; 2230 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT; 2231 req.max_cm_retries = CMA_MAX_CM_RETRIES; 2232 req.srq = id_priv->srq ? 1 : 0; 2233 2234 ret = ib_send_cm_req(id_priv->cm_id.ib, &req); 2235 out: 2236 if (ret && !IS_ERR(id_priv->cm_id.ib)) { 2237 ib_destroy_cm_id(id_priv->cm_id.ib); 2238 id_priv->cm_id.ib = NULL; 2239 } 2240 2241 kfree(private_data); 2242 return ret; 2243 } 2244 2245 static int cma_connect_iw(struct rdma_id_private *id_priv, 2246 struct rdma_conn_param *conn_param) 2247 { 2248 struct iw_cm_id *cm_id; 2249 struct sockaddr_in* sin; 2250 int ret; 2251 struct iw_cm_conn_param iw_param; 2252 2253 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv); 2254 if (IS_ERR(cm_id)) { 2255 ret = PTR_ERR(cm_id); 2256 goto out; 2257 } 2258 2259 id_priv->cm_id.iw = cm_id; 2260 2261 sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr; 2262 cm_id->local_addr = *sin; 2263 2264 sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr; 2265 cm_id->remote_addr = *sin; 2266 2267 ret = cma_modify_qp_rtr(&id_priv->id); 2268 if (ret) 2269 goto out; 2270 2271 iw_param.ord = conn_param->initiator_depth; 2272 iw_param.ird = conn_param->responder_resources; 2273 iw_param.private_data = conn_param->private_data; 2274 iw_param.private_data_len = conn_param->private_data_len; 2275 if (id_priv->id.qp) 2276 iw_param.qpn = id_priv->qp_num; 2277 else 2278 iw_param.qpn = conn_param->qp_num; 2279 ret = iw_cm_connect(cm_id, &iw_param); 2280 out: 2281 if (ret && !IS_ERR(cm_id)) { 2282 iw_destroy_cm_id(cm_id); 2283 id_priv->cm_id.iw = NULL; 2284 } 2285 return ret; 2286 } 2287 2288 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param) 2289 { 2290 struct rdma_id_private *id_priv; 2291 int ret; 2292 2293 id_priv = container_of(id, struct rdma_id_private, id); 2294 if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT)) 2295 return -EINVAL; 2296 2297 if (!id->qp) { 2298 id_priv->qp_num = conn_param->qp_num; 2299 id_priv->srq = conn_param->srq; 2300 } 2301 2302 switch (rdma_node_get_transport(id->device->node_type)) { 2303 case RDMA_TRANSPORT_IB: 2304 if (cma_is_ud_ps(id->ps)) 2305 ret = cma_resolve_ib_udp(id_priv, conn_param); 2306 else 2307 ret = cma_connect_ib(id_priv, conn_param); 2308 break; 2309 case RDMA_TRANSPORT_IWARP: 2310 ret = cma_connect_iw(id_priv, conn_param); 2311 break; 2312 default: 2313 ret = -ENOSYS; 2314 break; 2315 } 2316 if (ret) 2317 goto err; 2318 2319 return 0; 2320 err: 2321 cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED); 2322 return ret; 2323 } 2324 EXPORT_SYMBOL(rdma_connect); 2325 2326 static int cma_accept_ib(struct rdma_id_private *id_priv, 2327 struct rdma_conn_param *conn_param) 2328 { 2329 struct ib_cm_rep_param rep; 2330 struct ib_qp_attr qp_attr; 2331 int qp_attr_mask, ret; 2332 2333 if (id_priv->id.qp) { 2334 ret = cma_modify_qp_rtr(&id_priv->id); 2335 if (ret) 2336 goto out; 2337 2338 qp_attr.qp_state = IB_QPS_RTS; 2339 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, &qp_attr, 2340 &qp_attr_mask); 2341 if (ret) 2342 goto out; 2343 2344 qp_attr.max_rd_atomic = conn_param->initiator_depth; 2345 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 2346 if (ret) 2347 goto out; 2348 } 2349 2350 memset(&rep, 0, sizeof rep); 2351 rep.qp_num = id_priv->qp_num; 2352 rep.starting_psn = id_priv->seq_num; 2353 rep.private_data = conn_param->private_data; 2354 rep.private_data_len = conn_param->private_data_len; 2355 rep.responder_resources = conn_param->responder_resources; 2356 rep.initiator_depth = conn_param->initiator_depth; 2357 rep.failover_accepted = 0; 2358 rep.flow_control = conn_param->flow_control; 2359 rep.rnr_retry_count = conn_param->rnr_retry_count; 2360 rep.srq = id_priv->srq ? 1 : 0; 2361 2362 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep); 2363 out: 2364 return ret; 2365 } 2366 2367 static int cma_accept_iw(struct rdma_id_private *id_priv, 2368 struct rdma_conn_param *conn_param) 2369 { 2370 struct iw_cm_conn_param iw_param; 2371 int ret; 2372 2373 ret = cma_modify_qp_rtr(&id_priv->id); 2374 if (ret) 2375 return ret; 2376 2377 iw_param.ord = conn_param->initiator_depth; 2378 iw_param.ird = conn_param->responder_resources; 2379 iw_param.private_data = conn_param->private_data; 2380 iw_param.private_data_len = conn_param->private_data_len; 2381 if (id_priv->id.qp) { 2382 iw_param.qpn = id_priv->qp_num; 2383 } else 2384 iw_param.qpn = conn_param->qp_num; 2385 2386 return iw_cm_accept(id_priv->cm_id.iw, &iw_param); 2387 } 2388 2389 static int cma_send_sidr_rep(struct rdma_id_private *id_priv, 2390 enum ib_cm_sidr_status status, 2391 const void *private_data, int private_data_len) 2392 { 2393 struct ib_cm_sidr_rep_param rep; 2394 2395 memset(&rep, 0, sizeof rep); 2396 rep.status = status; 2397 if (status == IB_SIDR_SUCCESS) { 2398 rep.qp_num = id_priv->qp_num; 2399 rep.qkey = id_priv->qkey; 2400 } 2401 rep.private_data = private_data; 2402 rep.private_data_len = private_data_len; 2403 2404 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep); 2405 } 2406 2407 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param) 2408 { 2409 struct rdma_id_private *id_priv; 2410 int ret; 2411 2412 id_priv = container_of(id, struct rdma_id_private, id); 2413 if (!cma_comp(id_priv, CMA_CONNECT)) 2414 return -EINVAL; 2415 2416 if (!id->qp && conn_param) { 2417 id_priv->qp_num = conn_param->qp_num; 2418 id_priv->srq = conn_param->srq; 2419 } 2420 2421 switch (rdma_node_get_transport(id->device->node_type)) { 2422 case RDMA_TRANSPORT_IB: 2423 if (cma_is_ud_ps(id->ps)) 2424 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS, 2425 conn_param->private_data, 2426 conn_param->private_data_len); 2427 else if (conn_param) 2428 ret = cma_accept_ib(id_priv, conn_param); 2429 else 2430 ret = cma_rep_recv(id_priv); 2431 break; 2432 case RDMA_TRANSPORT_IWARP: 2433 ret = cma_accept_iw(id_priv, conn_param); 2434 break; 2435 default: 2436 ret = -ENOSYS; 2437 break; 2438 } 2439 2440 if (ret) 2441 goto reject; 2442 2443 return 0; 2444 reject: 2445 cma_modify_qp_err(id); 2446 rdma_reject(id, NULL, 0); 2447 return ret; 2448 } 2449 EXPORT_SYMBOL(rdma_accept); 2450 2451 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event) 2452 { 2453 struct rdma_id_private *id_priv; 2454 int ret; 2455 2456 id_priv = container_of(id, struct rdma_id_private, id); 2457 if (!cma_has_cm_dev(id_priv)) 2458 return -EINVAL; 2459 2460 switch (id->device->node_type) { 2461 case RDMA_NODE_IB_CA: 2462 ret = ib_cm_notify(id_priv->cm_id.ib, event); 2463 break; 2464 default: 2465 ret = 0; 2466 break; 2467 } 2468 return ret; 2469 } 2470 EXPORT_SYMBOL(rdma_notify); 2471 2472 int rdma_reject(struct rdma_cm_id *id, const void *private_data, 2473 u8 private_data_len) 2474 { 2475 struct rdma_id_private *id_priv; 2476 int ret; 2477 2478 id_priv = container_of(id, struct rdma_id_private, id); 2479 if (!cma_has_cm_dev(id_priv)) 2480 return -EINVAL; 2481 2482 switch (rdma_node_get_transport(id->device->node_type)) { 2483 case RDMA_TRANSPORT_IB: 2484 if (cma_is_ud_ps(id->ps)) 2485 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 2486 private_data, private_data_len); 2487 else 2488 ret = ib_send_cm_rej(id_priv->cm_id.ib, 2489 IB_CM_REJ_CONSUMER_DEFINED, NULL, 2490 0, private_data, private_data_len); 2491 break; 2492 case RDMA_TRANSPORT_IWARP: 2493 ret = iw_cm_reject(id_priv->cm_id.iw, 2494 private_data, private_data_len); 2495 break; 2496 default: 2497 ret = -ENOSYS; 2498 break; 2499 } 2500 return ret; 2501 } 2502 EXPORT_SYMBOL(rdma_reject); 2503 2504 int rdma_disconnect(struct rdma_cm_id *id) 2505 { 2506 struct rdma_id_private *id_priv; 2507 int ret; 2508 2509 id_priv = container_of(id, struct rdma_id_private, id); 2510 if (!cma_has_cm_dev(id_priv)) 2511 return -EINVAL; 2512 2513 switch (rdma_node_get_transport(id->device->node_type)) { 2514 case RDMA_TRANSPORT_IB: 2515 ret = cma_modify_qp_err(id); 2516 if (ret) 2517 goto out; 2518 /* Initiate or respond to a disconnect. */ 2519 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) 2520 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0); 2521 break; 2522 case RDMA_TRANSPORT_IWARP: 2523 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0); 2524 break; 2525 default: 2526 ret = -EINVAL; 2527 break; 2528 } 2529 out: 2530 return ret; 2531 } 2532 EXPORT_SYMBOL(rdma_disconnect); 2533 2534 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast) 2535 { 2536 struct rdma_id_private *id_priv; 2537 struct cma_multicast *mc = multicast->context; 2538 struct rdma_cm_event event; 2539 int ret; 2540 2541 id_priv = mc->id_priv; 2542 if (cma_disable_remove(id_priv, CMA_ADDR_BOUND) && 2543 cma_disable_remove(id_priv, CMA_ADDR_RESOLVED)) 2544 return 0; 2545 2546 if (!status && id_priv->id.qp) 2547 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid, 2548 multicast->rec.mlid); 2549 2550 memset(&event, 0, sizeof event); 2551 event.status = status; 2552 event.param.ud.private_data = mc->context; 2553 if (!status) { 2554 event.event = RDMA_CM_EVENT_MULTICAST_JOIN; 2555 ib_init_ah_from_mcmember(id_priv->id.device, 2556 id_priv->id.port_num, &multicast->rec, 2557 &event.param.ud.ah_attr); 2558 event.param.ud.qp_num = 0xFFFFFF; 2559 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey); 2560 } else 2561 event.event = RDMA_CM_EVENT_MULTICAST_ERROR; 2562 2563 ret = id_priv->id.event_handler(&id_priv->id, &event); 2564 if (ret) { 2565 cma_exch(id_priv, CMA_DESTROYING); 2566 cma_enable_remove(id_priv); 2567 rdma_destroy_id(&id_priv->id); 2568 return 0; 2569 } 2570 2571 cma_enable_remove(id_priv); 2572 return 0; 2573 } 2574 2575 static void cma_set_mgid(struct rdma_id_private *id_priv, 2576 struct sockaddr *addr, union ib_gid *mgid) 2577 { 2578 unsigned char mc_map[MAX_ADDR_LEN]; 2579 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 2580 struct sockaddr_in *sin = (struct sockaddr_in *) addr; 2581 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr; 2582 2583 if (cma_any_addr(addr)) { 2584 memset(mgid, 0, sizeof *mgid); 2585 } else if ((addr->sa_family == AF_INET6) && 2586 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFF10A01B) == 2587 0xFF10A01B)) { 2588 /* IPv6 address is an SA assigned MGID. */ 2589 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid); 2590 } else { 2591 ip_ib_mc_map(sin->sin_addr.s_addr, mc_map); 2592 if (id_priv->id.ps == RDMA_PS_UDP) 2593 mc_map[7] = 0x01; /* Use RDMA CM signature */ 2594 mc_map[8] = ib_addr_get_pkey(dev_addr) >> 8; 2595 mc_map[9] = (unsigned char) ib_addr_get_pkey(dev_addr); 2596 *mgid = *(union ib_gid *) (mc_map + 4); 2597 } 2598 } 2599 2600 static int cma_join_ib_multicast(struct rdma_id_private *id_priv, 2601 struct cma_multicast *mc) 2602 { 2603 struct ib_sa_mcmember_rec rec; 2604 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 2605 ib_sa_comp_mask comp_mask; 2606 int ret; 2607 2608 ib_addr_get_mgid(dev_addr, &rec.mgid); 2609 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num, 2610 &rec.mgid, &rec); 2611 if (ret) 2612 return ret; 2613 2614 cma_set_mgid(id_priv, &mc->addr, &rec.mgid); 2615 if (id_priv->id.ps == RDMA_PS_UDP) 2616 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY); 2617 ib_addr_get_sgid(dev_addr, &rec.port_gid); 2618 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr)); 2619 rec.join_state = 1; 2620 2621 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID | 2622 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE | 2623 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL | 2624 IB_SA_MCMEMBER_REC_FLOW_LABEL | 2625 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS; 2626 2627 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device, 2628 id_priv->id.port_num, &rec, 2629 comp_mask, GFP_KERNEL, 2630 cma_ib_mc_handler, mc); 2631 if (IS_ERR(mc->multicast.ib)) 2632 return PTR_ERR(mc->multicast.ib); 2633 2634 return 0; 2635 } 2636 2637 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr, 2638 void *context) 2639 { 2640 struct rdma_id_private *id_priv; 2641 struct cma_multicast *mc; 2642 int ret; 2643 2644 id_priv = container_of(id, struct rdma_id_private, id); 2645 if (!cma_comp(id_priv, CMA_ADDR_BOUND) && 2646 !cma_comp(id_priv, CMA_ADDR_RESOLVED)) 2647 return -EINVAL; 2648 2649 mc = kmalloc(sizeof *mc, GFP_KERNEL); 2650 if (!mc) 2651 return -ENOMEM; 2652 2653 memcpy(&mc->addr, addr, ip_addr_size(addr)); 2654 mc->context = context; 2655 mc->id_priv = id_priv; 2656 2657 spin_lock(&id_priv->lock); 2658 list_add(&mc->list, &id_priv->mc_list); 2659 spin_unlock(&id_priv->lock); 2660 2661 switch (rdma_node_get_transport(id->device->node_type)) { 2662 case RDMA_TRANSPORT_IB: 2663 ret = cma_join_ib_multicast(id_priv, mc); 2664 break; 2665 default: 2666 ret = -ENOSYS; 2667 break; 2668 } 2669 2670 if (ret) { 2671 spin_lock_irq(&id_priv->lock); 2672 list_del(&mc->list); 2673 spin_unlock_irq(&id_priv->lock); 2674 kfree(mc); 2675 } 2676 return ret; 2677 } 2678 EXPORT_SYMBOL(rdma_join_multicast); 2679 2680 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr) 2681 { 2682 struct rdma_id_private *id_priv; 2683 struct cma_multicast *mc; 2684 2685 id_priv = container_of(id, struct rdma_id_private, id); 2686 spin_lock_irq(&id_priv->lock); 2687 list_for_each_entry(mc, &id_priv->mc_list, list) { 2688 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) { 2689 list_del(&mc->list); 2690 spin_unlock_irq(&id_priv->lock); 2691 2692 if (id->qp) 2693 ib_detach_mcast(id->qp, 2694 &mc->multicast.ib->rec.mgid, 2695 mc->multicast.ib->rec.mlid); 2696 ib_sa_free_multicast(mc->multicast.ib); 2697 kfree(mc); 2698 return; 2699 } 2700 } 2701 spin_unlock_irq(&id_priv->lock); 2702 } 2703 EXPORT_SYMBOL(rdma_leave_multicast); 2704 2705 static void cma_add_one(struct ib_device *device) 2706 { 2707 struct cma_device *cma_dev; 2708 struct rdma_id_private *id_priv; 2709 2710 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL); 2711 if (!cma_dev) 2712 return; 2713 2714 cma_dev->device = device; 2715 2716 init_completion(&cma_dev->comp); 2717 atomic_set(&cma_dev->refcount, 1); 2718 INIT_LIST_HEAD(&cma_dev->id_list); 2719 ib_set_client_data(device, &cma_client, cma_dev); 2720 2721 mutex_lock(&lock); 2722 list_add_tail(&cma_dev->list, &dev_list); 2723 list_for_each_entry(id_priv, &listen_any_list, list) 2724 cma_listen_on_dev(id_priv, cma_dev); 2725 mutex_unlock(&lock); 2726 } 2727 2728 static int cma_remove_id_dev(struct rdma_id_private *id_priv) 2729 { 2730 struct rdma_cm_event event; 2731 enum cma_state state; 2732 2733 /* Record that we want to remove the device */ 2734 state = cma_exch(id_priv, CMA_DEVICE_REMOVAL); 2735 if (state == CMA_DESTROYING) 2736 return 0; 2737 2738 cma_cancel_operation(id_priv, state); 2739 wait_event(id_priv->wait_remove, !atomic_read(&id_priv->dev_remove)); 2740 2741 /* Check for destruction from another callback. */ 2742 if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL)) 2743 return 0; 2744 2745 memset(&event, 0, sizeof event); 2746 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL; 2747 return id_priv->id.event_handler(&id_priv->id, &event); 2748 } 2749 2750 static void cma_process_remove(struct cma_device *cma_dev) 2751 { 2752 struct rdma_id_private *id_priv; 2753 int ret; 2754 2755 mutex_lock(&lock); 2756 while (!list_empty(&cma_dev->id_list)) { 2757 id_priv = list_entry(cma_dev->id_list.next, 2758 struct rdma_id_private, list); 2759 2760 if (cma_internal_listen(id_priv)) { 2761 cma_destroy_listen(id_priv); 2762 continue; 2763 } 2764 2765 list_del_init(&id_priv->list); 2766 atomic_inc(&id_priv->refcount); 2767 mutex_unlock(&lock); 2768 2769 ret = cma_remove_id_dev(id_priv); 2770 cma_deref_id(id_priv); 2771 if (ret) 2772 rdma_destroy_id(&id_priv->id); 2773 2774 mutex_lock(&lock); 2775 } 2776 mutex_unlock(&lock); 2777 2778 cma_deref_dev(cma_dev); 2779 wait_for_completion(&cma_dev->comp); 2780 } 2781 2782 static void cma_remove_one(struct ib_device *device) 2783 { 2784 struct cma_device *cma_dev; 2785 2786 cma_dev = ib_get_client_data(device, &cma_client); 2787 if (!cma_dev) 2788 return; 2789 2790 mutex_lock(&lock); 2791 list_del(&cma_dev->list); 2792 mutex_unlock(&lock); 2793 2794 cma_process_remove(cma_dev); 2795 kfree(cma_dev); 2796 } 2797 2798 static int cma_init(void) 2799 { 2800 int ret, low, high; 2801 2802 get_random_bytes(&next_port, sizeof next_port); 2803 inet_get_local_port_range(&low, &high); 2804 next_port = ((unsigned int) next_port % (high - low)) + low; 2805 2806 cma_wq = create_singlethread_workqueue("rdma_cm"); 2807 if (!cma_wq) 2808 return -ENOMEM; 2809 2810 ib_sa_register_client(&sa_client); 2811 rdma_addr_register_client(&addr_client); 2812 2813 ret = ib_register_client(&cma_client); 2814 if (ret) 2815 goto err; 2816 return 0; 2817 2818 err: 2819 rdma_addr_unregister_client(&addr_client); 2820 ib_sa_unregister_client(&sa_client); 2821 destroy_workqueue(cma_wq); 2822 return ret; 2823 } 2824 2825 static void cma_cleanup(void) 2826 { 2827 ib_unregister_client(&cma_client); 2828 rdma_addr_unregister_client(&addr_client); 2829 ib_sa_unregister_client(&sa_client); 2830 destroy_workqueue(cma_wq); 2831 idr_destroy(&sdp_ps); 2832 idr_destroy(&tcp_ps); 2833 idr_destroy(&udp_ps); 2834 idr_destroy(&ipoib_ps); 2835 } 2836 2837 module_init(cma_init); 2838 module_exit(cma_cleanup); 2839