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 available to you under a choice of one of two 8 * licenses. You may choose to be licensed under the terms of the GNU 9 * General Public License (GPL) Version 2, available from the file 10 * COPYING in the main directory of this source tree, or the 11 * OpenIB.org BSD license below: 12 * 13 * Redistribution and use in source and binary forms, with or 14 * without modification, are permitted provided that the following 15 * conditions are met: 16 * 17 * - Redistributions of source code must retain the above 18 * copyright notice, this list of conditions and the following 19 * disclaimer. 20 * 21 * - Redistributions in binary form must reproduce the above 22 * copyright notice, this list of conditions and the following 23 * disclaimer in the documentation and/or other materials 24 * provided with the distribution. 25 * 26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 33 * SOFTWARE. 34 */ 35 36 #include <linux/completion.h> 37 #include <linux/in.h> 38 #include <linux/in6.h> 39 #include <linux/mutex.h> 40 #include <linux/random.h> 41 #include <linux/idr.h> 42 #include <linux/inetdevice.h> 43 44 #include <net/tcp.h> 45 46 #include <rdma/rdma_cm.h> 47 #include <rdma/rdma_cm_ib.h> 48 #include <rdma/ib_cache.h> 49 #include <rdma/ib_cm.h> 50 #include <rdma/ib_sa.h> 51 #include <rdma/iw_cm.h> 52 53 MODULE_AUTHOR("Sean Hefty"); 54 MODULE_DESCRIPTION("Generic RDMA CM Agent"); 55 MODULE_LICENSE("Dual BSD/GPL"); 56 57 #define CMA_CM_RESPONSE_TIMEOUT 20 58 #define CMA_MAX_CM_RETRIES 15 59 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24) 60 61 static void cma_add_one(struct ib_device *device); 62 static void cma_remove_one(struct ib_device *device); 63 64 static struct ib_client cma_client = { 65 .name = "cma", 66 .add = cma_add_one, 67 .remove = cma_remove_one 68 }; 69 70 static struct ib_sa_client sa_client; 71 static struct rdma_addr_client addr_client; 72 static LIST_HEAD(dev_list); 73 static LIST_HEAD(listen_any_list); 74 static DEFINE_MUTEX(lock); 75 static struct workqueue_struct *cma_wq; 76 static DEFINE_IDR(sdp_ps); 77 static DEFINE_IDR(tcp_ps); 78 static DEFINE_IDR(udp_ps); 79 static DEFINE_IDR(ipoib_ps); 80 static int next_port; 81 82 struct cma_device { 83 struct list_head list; 84 struct ib_device *device; 85 struct completion comp; 86 atomic_t refcount; 87 struct list_head id_list; 88 }; 89 90 enum cma_state { 91 CMA_IDLE, 92 CMA_ADDR_QUERY, 93 CMA_ADDR_RESOLVED, 94 CMA_ROUTE_QUERY, 95 CMA_ROUTE_RESOLVED, 96 CMA_CONNECT, 97 CMA_DISCONNECT, 98 CMA_ADDR_BOUND, 99 CMA_LISTEN, 100 CMA_DEVICE_REMOVAL, 101 CMA_DESTROYING 102 }; 103 104 struct rdma_bind_list { 105 struct idr *ps; 106 struct hlist_head owners; 107 unsigned short port; 108 }; 109 110 /* 111 * Device removal can occur at anytime, so we need extra handling to 112 * serialize notifying the user of device removal with other callbacks. 113 * We do this by disabling removal notification while a callback is in process, 114 * and reporting it after the callback completes. 115 */ 116 struct rdma_id_private { 117 struct rdma_cm_id id; 118 119 struct rdma_bind_list *bind_list; 120 struct hlist_node node; 121 struct list_head list; /* listen_any_list or cma_device.list */ 122 struct list_head listen_list; /* per device listens */ 123 struct cma_device *cma_dev; 124 struct list_head mc_list; 125 126 int internal_id; 127 enum cma_state state; 128 spinlock_t lock; 129 struct mutex qp_mutex; 130 131 struct completion comp; 132 atomic_t refcount; 133 struct mutex handler_mutex; 134 135 int backlog; 136 int timeout_ms; 137 struct ib_sa_query *query; 138 int query_id; 139 union { 140 struct ib_cm_id *ib; 141 struct iw_cm_id *iw; 142 } cm_id; 143 144 u32 seq_num; 145 u32 qkey; 146 u32 qp_num; 147 u8 srq; 148 u8 tos; 149 }; 150 151 struct cma_multicast { 152 struct rdma_id_private *id_priv; 153 union { 154 struct ib_sa_multicast *ib; 155 } multicast; 156 struct list_head list; 157 void *context; 158 struct sockaddr addr; 159 u8 pad[sizeof(struct sockaddr_in6) - 160 sizeof(struct sockaddr)]; 161 }; 162 163 struct cma_work { 164 struct work_struct work; 165 struct rdma_id_private *id; 166 enum cma_state old_state; 167 enum cma_state new_state; 168 struct rdma_cm_event event; 169 }; 170 171 struct cma_ndev_work { 172 struct work_struct work; 173 struct rdma_id_private *id; 174 struct rdma_cm_event event; 175 }; 176 177 union cma_ip_addr { 178 struct in6_addr ip6; 179 struct { 180 __be32 pad[3]; 181 __be32 addr; 182 } ip4; 183 }; 184 185 struct cma_hdr { 186 u8 cma_version; 187 u8 ip_version; /* IP version: 7:4 */ 188 __be16 port; 189 union cma_ip_addr src_addr; 190 union cma_ip_addr dst_addr; 191 }; 192 193 struct sdp_hh { 194 u8 bsdh[16]; 195 u8 sdp_version; /* Major version: 7:4 */ 196 u8 ip_version; /* IP version: 7:4 */ 197 u8 sdp_specific1[10]; 198 __be16 port; 199 __be16 sdp_specific2; 200 union cma_ip_addr src_addr; 201 union cma_ip_addr dst_addr; 202 }; 203 204 struct sdp_hah { 205 u8 bsdh[16]; 206 u8 sdp_version; 207 }; 208 209 #define CMA_VERSION 0x00 210 #define SDP_MAJ_VERSION 0x2 211 212 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp) 213 { 214 unsigned long flags; 215 int ret; 216 217 spin_lock_irqsave(&id_priv->lock, flags); 218 ret = (id_priv->state == comp); 219 spin_unlock_irqrestore(&id_priv->lock, flags); 220 return ret; 221 } 222 223 static int cma_comp_exch(struct rdma_id_private *id_priv, 224 enum cma_state comp, enum cma_state exch) 225 { 226 unsigned long flags; 227 int ret; 228 229 spin_lock_irqsave(&id_priv->lock, flags); 230 if ((ret = (id_priv->state == comp))) 231 id_priv->state = exch; 232 spin_unlock_irqrestore(&id_priv->lock, flags); 233 return ret; 234 } 235 236 static enum cma_state cma_exch(struct rdma_id_private *id_priv, 237 enum cma_state exch) 238 { 239 unsigned long flags; 240 enum cma_state old; 241 242 spin_lock_irqsave(&id_priv->lock, flags); 243 old = id_priv->state; 244 id_priv->state = exch; 245 spin_unlock_irqrestore(&id_priv->lock, flags); 246 return old; 247 } 248 249 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr) 250 { 251 return hdr->ip_version >> 4; 252 } 253 254 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver) 255 { 256 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF); 257 } 258 259 static inline u8 sdp_get_majv(u8 sdp_version) 260 { 261 return sdp_version >> 4; 262 } 263 264 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh) 265 { 266 return hh->ip_version >> 4; 267 } 268 269 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver) 270 { 271 hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF); 272 } 273 274 static inline int cma_is_ud_ps(enum rdma_port_space ps) 275 { 276 return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB); 277 } 278 279 static void cma_attach_to_dev(struct rdma_id_private *id_priv, 280 struct cma_device *cma_dev) 281 { 282 atomic_inc(&cma_dev->refcount); 283 id_priv->cma_dev = cma_dev; 284 id_priv->id.device = cma_dev->device; 285 list_add_tail(&id_priv->list, &cma_dev->id_list); 286 } 287 288 static inline void cma_deref_dev(struct cma_device *cma_dev) 289 { 290 if (atomic_dec_and_test(&cma_dev->refcount)) 291 complete(&cma_dev->comp); 292 } 293 294 static void cma_detach_from_dev(struct rdma_id_private *id_priv) 295 { 296 list_del(&id_priv->list); 297 cma_deref_dev(id_priv->cma_dev); 298 id_priv->cma_dev = NULL; 299 } 300 301 static int cma_set_qkey(struct ib_device *device, u8 port_num, 302 enum rdma_port_space ps, 303 struct rdma_dev_addr *dev_addr, u32 *qkey) 304 { 305 struct ib_sa_mcmember_rec rec; 306 int ret = 0; 307 308 switch (ps) { 309 case RDMA_PS_UDP: 310 *qkey = RDMA_UDP_QKEY; 311 break; 312 case RDMA_PS_IPOIB: 313 ib_addr_get_mgid(dev_addr, &rec.mgid); 314 ret = ib_sa_get_mcmember_rec(device, port_num, &rec.mgid, &rec); 315 *qkey = be32_to_cpu(rec.qkey); 316 break; 317 default: 318 break; 319 } 320 return ret; 321 } 322 323 static int cma_acquire_dev(struct rdma_id_private *id_priv) 324 { 325 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 326 struct cma_device *cma_dev; 327 union ib_gid gid; 328 int ret = -ENODEV; 329 330 switch (rdma_node_get_transport(dev_addr->dev_type)) { 331 case RDMA_TRANSPORT_IB: 332 ib_addr_get_sgid(dev_addr, &gid); 333 break; 334 case RDMA_TRANSPORT_IWARP: 335 iw_addr_get_sgid(dev_addr, &gid); 336 break; 337 default: 338 return -ENODEV; 339 } 340 341 list_for_each_entry(cma_dev, &dev_list, list) { 342 ret = ib_find_cached_gid(cma_dev->device, &gid, 343 &id_priv->id.port_num, NULL); 344 if (!ret) { 345 ret = cma_set_qkey(cma_dev->device, 346 id_priv->id.port_num, 347 id_priv->id.ps, dev_addr, 348 &id_priv->qkey); 349 if (!ret) 350 cma_attach_to_dev(id_priv, cma_dev); 351 break; 352 } 353 } 354 return ret; 355 } 356 357 static void cma_deref_id(struct rdma_id_private *id_priv) 358 { 359 if (atomic_dec_and_test(&id_priv->refcount)) 360 complete(&id_priv->comp); 361 } 362 363 static int cma_disable_callback(struct rdma_id_private *id_priv, 364 enum cma_state state) 365 { 366 mutex_lock(&id_priv->handler_mutex); 367 if (id_priv->state != state) { 368 mutex_unlock(&id_priv->handler_mutex); 369 return -EINVAL; 370 } 371 return 0; 372 } 373 374 static int cma_has_cm_dev(struct rdma_id_private *id_priv) 375 { 376 return (id_priv->id.device && id_priv->cm_id.ib); 377 } 378 379 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler, 380 void *context, enum rdma_port_space ps) 381 { 382 struct rdma_id_private *id_priv; 383 384 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL); 385 if (!id_priv) 386 return ERR_PTR(-ENOMEM); 387 388 id_priv->state = CMA_IDLE; 389 id_priv->id.context = context; 390 id_priv->id.event_handler = event_handler; 391 id_priv->id.ps = ps; 392 spin_lock_init(&id_priv->lock); 393 mutex_init(&id_priv->qp_mutex); 394 init_completion(&id_priv->comp); 395 atomic_set(&id_priv->refcount, 1); 396 mutex_init(&id_priv->handler_mutex); 397 INIT_LIST_HEAD(&id_priv->listen_list); 398 INIT_LIST_HEAD(&id_priv->mc_list); 399 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num); 400 401 return &id_priv->id; 402 } 403 EXPORT_SYMBOL(rdma_create_id); 404 405 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp) 406 { 407 struct ib_qp_attr qp_attr; 408 int qp_attr_mask, ret; 409 410 qp_attr.qp_state = IB_QPS_INIT; 411 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 412 if (ret) 413 return ret; 414 415 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 416 if (ret) 417 return ret; 418 419 qp_attr.qp_state = IB_QPS_RTR; 420 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE); 421 if (ret) 422 return ret; 423 424 qp_attr.qp_state = IB_QPS_RTS; 425 qp_attr.sq_psn = 0; 426 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN); 427 428 return ret; 429 } 430 431 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp) 432 { 433 struct ib_qp_attr qp_attr; 434 int qp_attr_mask, ret; 435 436 qp_attr.qp_state = IB_QPS_INIT; 437 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 438 if (ret) 439 return ret; 440 441 return ib_modify_qp(qp, &qp_attr, qp_attr_mask); 442 } 443 444 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd, 445 struct ib_qp_init_attr *qp_init_attr) 446 { 447 struct rdma_id_private *id_priv; 448 struct ib_qp *qp; 449 int ret; 450 451 id_priv = container_of(id, struct rdma_id_private, id); 452 if (id->device != pd->device) 453 return -EINVAL; 454 455 qp = ib_create_qp(pd, qp_init_attr); 456 if (IS_ERR(qp)) 457 return PTR_ERR(qp); 458 459 if (cma_is_ud_ps(id_priv->id.ps)) 460 ret = cma_init_ud_qp(id_priv, qp); 461 else 462 ret = cma_init_conn_qp(id_priv, qp); 463 if (ret) 464 goto err; 465 466 id->qp = qp; 467 id_priv->qp_num = qp->qp_num; 468 id_priv->srq = (qp->srq != NULL); 469 return 0; 470 err: 471 ib_destroy_qp(qp); 472 return ret; 473 } 474 EXPORT_SYMBOL(rdma_create_qp); 475 476 void rdma_destroy_qp(struct rdma_cm_id *id) 477 { 478 struct rdma_id_private *id_priv; 479 480 id_priv = container_of(id, struct rdma_id_private, id); 481 mutex_lock(&id_priv->qp_mutex); 482 ib_destroy_qp(id_priv->id.qp); 483 id_priv->id.qp = NULL; 484 mutex_unlock(&id_priv->qp_mutex); 485 } 486 EXPORT_SYMBOL(rdma_destroy_qp); 487 488 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv, 489 struct rdma_conn_param *conn_param) 490 { 491 struct ib_qp_attr qp_attr; 492 int qp_attr_mask, ret; 493 494 mutex_lock(&id_priv->qp_mutex); 495 if (!id_priv->id.qp) { 496 ret = 0; 497 goto out; 498 } 499 500 /* Need to update QP attributes from default values. */ 501 qp_attr.qp_state = IB_QPS_INIT; 502 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 503 if (ret) 504 goto out; 505 506 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 507 if (ret) 508 goto out; 509 510 qp_attr.qp_state = IB_QPS_RTR; 511 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 512 if (ret) 513 goto out; 514 515 if (conn_param) 516 qp_attr.max_dest_rd_atomic = conn_param->responder_resources; 517 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 518 out: 519 mutex_unlock(&id_priv->qp_mutex); 520 return ret; 521 } 522 523 static int cma_modify_qp_rts(struct rdma_id_private *id_priv, 524 struct rdma_conn_param *conn_param) 525 { 526 struct ib_qp_attr qp_attr; 527 int qp_attr_mask, ret; 528 529 mutex_lock(&id_priv->qp_mutex); 530 if (!id_priv->id.qp) { 531 ret = 0; 532 goto out; 533 } 534 535 qp_attr.qp_state = IB_QPS_RTS; 536 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 537 if (ret) 538 goto out; 539 540 if (conn_param) 541 qp_attr.max_rd_atomic = conn_param->initiator_depth; 542 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 543 out: 544 mutex_unlock(&id_priv->qp_mutex); 545 return ret; 546 } 547 548 static int cma_modify_qp_err(struct rdma_id_private *id_priv) 549 { 550 struct ib_qp_attr qp_attr; 551 int ret; 552 553 mutex_lock(&id_priv->qp_mutex); 554 if (!id_priv->id.qp) { 555 ret = 0; 556 goto out; 557 } 558 559 qp_attr.qp_state = IB_QPS_ERR; 560 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE); 561 out: 562 mutex_unlock(&id_priv->qp_mutex); 563 return ret; 564 } 565 566 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv, 567 struct ib_qp_attr *qp_attr, int *qp_attr_mask) 568 { 569 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 570 int ret; 571 572 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num, 573 ib_addr_get_pkey(dev_addr), 574 &qp_attr->pkey_index); 575 if (ret) 576 return ret; 577 578 qp_attr->port_num = id_priv->id.port_num; 579 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT; 580 581 if (cma_is_ud_ps(id_priv->id.ps)) { 582 qp_attr->qkey = id_priv->qkey; 583 *qp_attr_mask |= IB_QP_QKEY; 584 } else { 585 qp_attr->qp_access_flags = 0; 586 *qp_attr_mask |= IB_QP_ACCESS_FLAGS; 587 } 588 return 0; 589 } 590 591 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr, 592 int *qp_attr_mask) 593 { 594 struct rdma_id_private *id_priv; 595 int ret = 0; 596 597 id_priv = container_of(id, struct rdma_id_private, id); 598 switch (rdma_node_get_transport(id_priv->id.device->node_type)) { 599 case RDMA_TRANSPORT_IB: 600 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps)) 601 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask); 602 else 603 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr, 604 qp_attr_mask); 605 if (qp_attr->qp_state == IB_QPS_RTR) 606 qp_attr->rq_psn = id_priv->seq_num; 607 break; 608 case RDMA_TRANSPORT_IWARP: 609 if (!id_priv->cm_id.iw) { 610 qp_attr->qp_access_flags = 0; 611 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS; 612 } else 613 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr, 614 qp_attr_mask); 615 break; 616 default: 617 ret = -ENOSYS; 618 break; 619 } 620 621 return ret; 622 } 623 EXPORT_SYMBOL(rdma_init_qp_attr); 624 625 static inline int cma_zero_addr(struct sockaddr *addr) 626 { 627 struct in6_addr *ip6; 628 629 if (addr->sa_family == AF_INET) 630 return ipv4_is_zeronet( 631 ((struct sockaddr_in *)addr)->sin_addr.s_addr); 632 else { 633 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr; 634 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] | 635 ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0; 636 } 637 } 638 639 static inline int cma_loopback_addr(struct sockaddr *addr) 640 { 641 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr); 642 } 643 644 static inline int cma_any_addr(struct sockaddr *addr) 645 { 646 return cma_zero_addr(addr) || cma_loopback_addr(addr); 647 } 648 649 static inline __be16 cma_port(struct sockaddr *addr) 650 { 651 if (addr->sa_family == AF_INET) 652 return ((struct sockaddr_in *) addr)->sin_port; 653 else 654 return ((struct sockaddr_in6 *) addr)->sin6_port; 655 } 656 657 static inline int cma_any_port(struct sockaddr *addr) 658 { 659 return !cma_port(addr); 660 } 661 662 static int cma_get_net_info(void *hdr, enum rdma_port_space ps, 663 u8 *ip_ver, __be16 *port, 664 union cma_ip_addr **src, union cma_ip_addr **dst) 665 { 666 switch (ps) { 667 case RDMA_PS_SDP: 668 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) != 669 SDP_MAJ_VERSION) 670 return -EINVAL; 671 672 *ip_ver = sdp_get_ip_ver(hdr); 673 *port = ((struct sdp_hh *) hdr)->port; 674 *src = &((struct sdp_hh *) hdr)->src_addr; 675 *dst = &((struct sdp_hh *) hdr)->dst_addr; 676 break; 677 default: 678 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION) 679 return -EINVAL; 680 681 *ip_ver = cma_get_ip_ver(hdr); 682 *port = ((struct cma_hdr *) hdr)->port; 683 *src = &((struct cma_hdr *) hdr)->src_addr; 684 *dst = &((struct cma_hdr *) hdr)->dst_addr; 685 break; 686 } 687 688 if (*ip_ver != 4 && *ip_ver != 6) 689 return -EINVAL; 690 return 0; 691 } 692 693 static void cma_save_net_info(struct rdma_addr *addr, 694 struct rdma_addr *listen_addr, 695 u8 ip_ver, __be16 port, 696 union cma_ip_addr *src, union cma_ip_addr *dst) 697 { 698 struct sockaddr_in *listen4, *ip4; 699 struct sockaddr_in6 *listen6, *ip6; 700 701 switch (ip_ver) { 702 case 4: 703 listen4 = (struct sockaddr_in *) &listen_addr->src_addr; 704 ip4 = (struct sockaddr_in *) &addr->src_addr; 705 ip4->sin_family = listen4->sin_family; 706 ip4->sin_addr.s_addr = dst->ip4.addr; 707 ip4->sin_port = listen4->sin_port; 708 709 ip4 = (struct sockaddr_in *) &addr->dst_addr; 710 ip4->sin_family = listen4->sin_family; 711 ip4->sin_addr.s_addr = src->ip4.addr; 712 ip4->sin_port = port; 713 break; 714 case 6: 715 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr; 716 ip6 = (struct sockaddr_in6 *) &addr->src_addr; 717 ip6->sin6_family = listen6->sin6_family; 718 ip6->sin6_addr = dst->ip6; 719 ip6->sin6_port = listen6->sin6_port; 720 721 ip6 = (struct sockaddr_in6 *) &addr->dst_addr; 722 ip6->sin6_family = listen6->sin6_family; 723 ip6->sin6_addr = src->ip6; 724 ip6->sin6_port = port; 725 break; 726 default: 727 break; 728 } 729 } 730 731 static inline int cma_user_data_offset(enum rdma_port_space ps) 732 { 733 switch (ps) { 734 case RDMA_PS_SDP: 735 return 0; 736 default: 737 return sizeof(struct cma_hdr); 738 } 739 } 740 741 static void cma_cancel_route(struct rdma_id_private *id_priv) 742 { 743 switch (rdma_node_get_transport(id_priv->id.device->node_type)) { 744 case RDMA_TRANSPORT_IB: 745 if (id_priv->query) 746 ib_sa_cancel_query(id_priv->query_id, id_priv->query); 747 break; 748 default: 749 break; 750 } 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 /* 758 * Remove from listen_any_list to prevent added devices from spawning 759 * additional listen requests. 760 */ 761 mutex_lock(&lock); 762 list_del(&id_priv->list); 763 764 while (!list_empty(&id_priv->listen_list)) { 765 dev_id_priv = list_entry(id_priv->listen_list.next, 766 struct rdma_id_private, listen_list); 767 /* sync with device removal to avoid duplicate destruction */ 768 list_del_init(&dev_id_priv->list); 769 list_del(&dev_id_priv->listen_list); 770 mutex_unlock(&lock); 771 772 rdma_destroy_id(&dev_id_priv->id); 773 mutex_lock(&lock); 774 } 775 mutex_unlock(&lock); 776 } 777 778 static void cma_cancel_operation(struct rdma_id_private *id_priv, 779 enum cma_state state) 780 { 781 switch (state) { 782 case CMA_ADDR_QUERY: 783 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr); 784 break; 785 case CMA_ROUTE_QUERY: 786 cma_cancel_route(id_priv); 787 break; 788 case CMA_LISTEN: 789 if (cma_any_addr(&id_priv->id.route.addr.src_addr) && 790 !id_priv->cma_dev) 791 cma_cancel_listens(id_priv); 792 break; 793 default: 794 break; 795 } 796 } 797 798 static void cma_release_port(struct rdma_id_private *id_priv) 799 { 800 struct rdma_bind_list *bind_list = id_priv->bind_list; 801 802 if (!bind_list) 803 return; 804 805 mutex_lock(&lock); 806 hlist_del(&id_priv->node); 807 if (hlist_empty(&bind_list->owners)) { 808 idr_remove(bind_list->ps, bind_list->port); 809 kfree(bind_list); 810 } 811 mutex_unlock(&lock); 812 } 813 814 static void cma_leave_mc_groups(struct rdma_id_private *id_priv) 815 { 816 struct cma_multicast *mc; 817 818 while (!list_empty(&id_priv->mc_list)) { 819 mc = container_of(id_priv->mc_list.next, 820 struct cma_multicast, list); 821 list_del(&mc->list); 822 ib_sa_free_multicast(mc->multicast.ib); 823 kfree(mc); 824 } 825 } 826 827 void rdma_destroy_id(struct rdma_cm_id *id) 828 { 829 struct rdma_id_private *id_priv; 830 enum cma_state state; 831 832 id_priv = container_of(id, struct rdma_id_private, id); 833 state = cma_exch(id_priv, CMA_DESTROYING); 834 cma_cancel_operation(id_priv, state); 835 836 mutex_lock(&lock); 837 if (id_priv->cma_dev) { 838 mutex_unlock(&lock); 839 switch (rdma_node_get_transport(id->device->node_type)) { 840 case RDMA_TRANSPORT_IB: 841 if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib)) 842 ib_destroy_cm_id(id_priv->cm_id.ib); 843 break; 844 case RDMA_TRANSPORT_IWARP: 845 if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw)) 846 iw_destroy_cm_id(id_priv->cm_id.iw); 847 break; 848 default: 849 break; 850 } 851 cma_leave_mc_groups(id_priv); 852 mutex_lock(&lock); 853 cma_detach_from_dev(id_priv); 854 } 855 mutex_unlock(&lock); 856 857 cma_release_port(id_priv); 858 cma_deref_id(id_priv); 859 wait_for_completion(&id_priv->comp); 860 861 if (id_priv->internal_id) 862 cma_deref_id(id_priv->id.context); 863 864 kfree(id_priv->id.route.path_rec); 865 kfree(id_priv); 866 } 867 EXPORT_SYMBOL(rdma_destroy_id); 868 869 static int cma_rep_recv(struct rdma_id_private *id_priv) 870 { 871 int ret; 872 873 ret = cma_modify_qp_rtr(id_priv, NULL); 874 if (ret) 875 goto reject; 876 877 ret = cma_modify_qp_rts(id_priv, NULL); 878 if (ret) 879 goto reject; 880 881 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0); 882 if (ret) 883 goto reject; 884 885 return 0; 886 reject: 887 cma_modify_qp_err(id_priv); 888 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED, 889 NULL, 0, NULL, 0); 890 return ret; 891 } 892 893 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data) 894 { 895 if (id_priv->id.ps == RDMA_PS_SDP && 896 sdp_get_majv(((struct sdp_hah *) data)->sdp_version) != 897 SDP_MAJ_VERSION) 898 return -EINVAL; 899 900 return 0; 901 } 902 903 static void cma_set_rep_event_data(struct rdma_cm_event *event, 904 struct ib_cm_rep_event_param *rep_data, 905 void *private_data) 906 { 907 event->param.conn.private_data = private_data; 908 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE; 909 event->param.conn.responder_resources = rep_data->responder_resources; 910 event->param.conn.initiator_depth = rep_data->initiator_depth; 911 event->param.conn.flow_control = rep_data->flow_control; 912 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count; 913 event->param.conn.srq = rep_data->srq; 914 event->param.conn.qp_num = rep_data->remote_qpn; 915 } 916 917 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event) 918 { 919 struct rdma_id_private *id_priv = cm_id->context; 920 struct rdma_cm_event event; 921 int ret = 0; 922 923 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT && 924 cma_disable_callback(id_priv, CMA_CONNECT)) || 925 (ib_event->event == IB_CM_TIMEWAIT_EXIT && 926 cma_disable_callback(id_priv, CMA_DISCONNECT))) 927 return 0; 928 929 memset(&event, 0, sizeof event); 930 switch (ib_event->event) { 931 case IB_CM_REQ_ERROR: 932 case IB_CM_REP_ERROR: 933 event.event = RDMA_CM_EVENT_UNREACHABLE; 934 event.status = -ETIMEDOUT; 935 break; 936 case IB_CM_REP_RECEIVED: 937 event.status = cma_verify_rep(id_priv, ib_event->private_data); 938 if (event.status) 939 event.event = RDMA_CM_EVENT_CONNECT_ERROR; 940 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) { 941 event.status = cma_rep_recv(id_priv); 942 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR : 943 RDMA_CM_EVENT_ESTABLISHED; 944 } else 945 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE; 946 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd, 947 ib_event->private_data); 948 break; 949 case IB_CM_RTU_RECEIVED: 950 case IB_CM_USER_ESTABLISHED: 951 event.event = RDMA_CM_EVENT_ESTABLISHED; 952 break; 953 case IB_CM_DREQ_ERROR: 954 event.status = -ETIMEDOUT; /* fall through */ 955 case IB_CM_DREQ_RECEIVED: 956 case IB_CM_DREP_RECEIVED: 957 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT)) 958 goto out; 959 event.event = RDMA_CM_EVENT_DISCONNECTED; 960 break; 961 case IB_CM_TIMEWAIT_EXIT: 962 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT; 963 break; 964 case IB_CM_MRA_RECEIVED: 965 /* ignore event */ 966 goto out; 967 case IB_CM_REJ_RECEIVED: 968 cma_modify_qp_err(id_priv); 969 event.status = ib_event->param.rej_rcvd.reason; 970 event.event = RDMA_CM_EVENT_REJECTED; 971 event.param.conn.private_data = ib_event->private_data; 972 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE; 973 break; 974 default: 975 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n", 976 ib_event->event); 977 goto out; 978 } 979 980 ret = id_priv->id.event_handler(&id_priv->id, &event); 981 if (ret) { 982 /* Destroy the CM ID by returning a non-zero value. */ 983 id_priv->cm_id.ib = NULL; 984 cma_exch(id_priv, CMA_DESTROYING); 985 mutex_unlock(&id_priv->handler_mutex); 986 rdma_destroy_id(&id_priv->id); 987 return ret; 988 } 989 out: 990 mutex_unlock(&id_priv->handler_mutex); 991 return ret; 992 } 993 994 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id, 995 struct ib_cm_event *ib_event) 996 { 997 struct rdma_id_private *id_priv; 998 struct rdma_cm_id *id; 999 struct rdma_route *rt; 1000 union cma_ip_addr *src, *dst; 1001 __be16 port; 1002 u8 ip_ver; 1003 int ret; 1004 1005 if (cma_get_net_info(ib_event->private_data, listen_id->ps, 1006 &ip_ver, &port, &src, &dst)) 1007 goto err; 1008 1009 id = rdma_create_id(listen_id->event_handler, listen_id->context, 1010 listen_id->ps); 1011 if (IS_ERR(id)) 1012 goto err; 1013 1014 cma_save_net_info(&id->route.addr, &listen_id->route.addr, 1015 ip_ver, port, src, dst); 1016 1017 rt = &id->route; 1018 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1; 1019 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths, 1020 GFP_KERNEL); 1021 if (!rt->path_rec) 1022 goto destroy_id; 1023 1024 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path; 1025 if (rt->num_paths == 2) 1026 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path; 1027 1028 ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid); 1029 ret = rdma_translate_ip(&id->route.addr.src_addr, 1030 &id->route.addr.dev_addr); 1031 if (ret) 1032 goto destroy_id; 1033 1034 id_priv = container_of(id, struct rdma_id_private, id); 1035 id_priv->state = CMA_CONNECT; 1036 return id_priv; 1037 1038 destroy_id: 1039 rdma_destroy_id(id); 1040 err: 1041 return NULL; 1042 } 1043 1044 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id, 1045 struct ib_cm_event *ib_event) 1046 { 1047 struct rdma_id_private *id_priv; 1048 struct rdma_cm_id *id; 1049 union cma_ip_addr *src, *dst; 1050 __be16 port; 1051 u8 ip_ver; 1052 int ret; 1053 1054 id = rdma_create_id(listen_id->event_handler, listen_id->context, 1055 listen_id->ps); 1056 if (IS_ERR(id)) 1057 return NULL; 1058 1059 1060 if (cma_get_net_info(ib_event->private_data, listen_id->ps, 1061 &ip_ver, &port, &src, &dst)) 1062 goto err; 1063 1064 cma_save_net_info(&id->route.addr, &listen_id->route.addr, 1065 ip_ver, port, src, dst); 1066 1067 ret = rdma_translate_ip(&id->route.addr.src_addr, 1068 &id->route.addr.dev_addr); 1069 if (ret) 1070 goto err; 1071 1072 id_priv = container_of(id, struct rdma_id_private, id); 1073 id_priv->state = CMA_CONNECT; 1074 return id_priv; 1075 err: 1076 rdma_destroy_id(id); 1077 return NULL; 1078 } 1079 1080 static void cma_set_req_event_data(struct rdma_cm_event *event, 1081 struct ib_cm_req_event_param *req_data, 1082 void *private_data, int offset) 1083 { 1084 event->param.conn.private_data = private_data + offset; 1085 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset; 1086 event->param.conn.responder_resources = req_data->responder_resources; 1087 event->param.conn.initiator_depth = req_data->initiator_depth; 1088 event->param.conn.flow_control = req_data->flow_control; 1089 event->param.conn.retry_count = req_data->retry_count; 1090 event->param.conn.rnr_retry_count = req_data->rnr_retry_count; 1091 event->param.conn.srq = req_data->srq; 1092 event->param.conn.qp_num = req_data->remote_qpn; 1093 } 1094 1095 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event) 1096 { 1097 struct rdma_id_private *listen_id, *conn_id; 1098 struct rdma_cm_event event; 1099 int offset, ret; 1100 1101 listen_id = cm_id->context; 1102 if (cma_disable_callback(listen_id, CMA_LISTEN)) 1103 return -ECONNABORTED; 1104 1105 memset(&event, 0, sizeof event); 1106 offset = cma_user_data_offset(listen_id->id.ps); 1107 event.event = RDMA_CM_EVENT_CONNECT_REQUEST; 1108 if (cma_is_ud_ps(listen_id->id.ps)) { 1109 conn_id = cma_new_udp_id(&listen_id->id, ib_event); 1110 event.param.ud.private_data = ib_event->private_data + offset; 1111 event.param.ud.private_data_len = 1112 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset; 1113 } else { 1114 conn_id = cma_new_conn_id(&listen_id->id, ib_event); 1115 cma_set_req_event_data(&event, &ib_event->param.req_rcvd, 1116 ib_event->private_data, offset); 1117 } 1118 if (!conn_id) { 1119 ret = -ENOMEM; 1120 goto out; 1121 } 1122 1123 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING); 1124 mutex_lock(&lock); 1125 ret = cma_acquire_dev(conn_id); 1126 mutex_unlock(&lock); 1127 if (ret) 1128 goto release_conn_id; 1129 1130 conn_id->cm_id.ib = cm_id; 1131 cm_id->context = conn_id; 1132 cm_id->cm_handler = cma_ib_handler; 1133 1134 ret = conn_id->id.event_handler(&conn_id->id, &event); 1135 if (!ret) { 1136 /* 1137 * Acquire mutex to prevent user executing rdma_destroy_id() 1138 * while we're accessing the cm_id. 1139 */ 1140 mutex_lock(&lock); 1141 if (cma_comp(conn_id, CMA_CONNECT) && 1142 !cma_is_ud_ps(conn_id->id.ps)) 1143 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0); 1144 mutex_unlock(&lock); 1145 mutex_unlock(&conn_id->handler_mutex); 1146 goto out; 1147 } 1148 1149 /* Destroy the CM ID by returning a non-zero value. */ 1150 conn_id->cm_id.ib = NULL; 1151 1152 release_conn_id: 1153 cma_exch(conn_id, CMA_DESTROYING); 1154 mutex_unlock(&conn_id->handler_mutex); 1155 rdma_destroy_id(&conn_id->id); 1156 1157 out: 1158 mutex_unlock(&listen_id->handler_mutex); 1159 return ret; 1160 } 1161 1162 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr) 1163 { 1164 return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr))); 1165 } 1166 1167 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr, 1168 struct ib_cm_compare_data *compare) 1169 { 1170 struct cma_hdr *cma_data, *cma_mask; 1171 struct sdp_hh *sdp_data, *sdp_mask; 1172 __be32 ip4_addr; 1173 struct in6_addr ip6_addr; 1174 1175 memset(compare, 0, sizeof *compare); 1176 cma_data = (void *) compare->data; 1177 cma_mask = (void *) compare->mask; 1178 sdp_data = (void *) compare->data; 1179 sdp_mask = (void *) compare->mask; 1180 1181 switch (addr->sa_family) { 1182 case AF_INET: 1183 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr; 1184 if (ps == RDMA_PS_SDP) { 1185 sdp_set_ip_ver(sdp_data, 4); 1186 sdp_set_ip_ver(sdp_mask, 0xF); 1187 sdp_data->dst_addr.ip4.addr = ip4_addr; 1188 sdp_mask->dst_addr.ip4.addr = htonl(~0); 1189 } else { 1190 cma_set_ip_ver(cma_data, 4); 1191 cma_set_ip_ver(cma_mask, 0xF); 1192 cma_data->dst_addr.ip4.addr = ip4_addr; 1193 cma_mask->dst_addr.ip4.addr = htonl(~0); 1194 } 1195 break; 1196 case AF_INET6: 1197 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr; 1198 if (ps == RDMA_PS_SDP) { 1199 sdp_set_ip_ver(sdp_data, 6); 1200 sdp_set_ip_ver(sdp_mask, 0xF); 1201 sdp_data->dst_addr.ip6 = ip6_addr; 1202 memset(&sdp_mask->dst_addr.ip6, 0xFF, 1203 sizeof sdp_mask->dst_addr.ip6); 1204 } else { 1205 cma_set_ip_ver(cma_data, 6); 1206 cma_set_ip_ver(cma_mask, 0xF); 1207 cma_data->dst_addr.ip6 = ip6_addr; 1208 memset(&cma_mask->dst_addr.ip6, 0xFF, 1209 sizeof cma_mask->dst_addr.ip6); 1210 } 1211 break; 1212 default: 1213 break; 1214 } 1215 } 1216 1217 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event) 1218 { 1219 struct rdma_id_private *id_priv = iw_id->context; 1220 struct rdma_cm_event event; 1221 struct sockaddr_in *sin; 1222 int ret = 0; 1223 1224 if (cma_disable_callback(id_priv, CMA_CONNECT)) 1225 return 0; 1226 1227 memset(&event, 0, sizeof event); 1228 switch (iw_event->event) { 1229 case IW_CM_EVENT_CLOSE: 1230 event.event = RDMA_CM_EVENT_DISCONNECTED; 1231 break; 1232 case IW_CM_EVENT_CONNECT_REPLY: 1233 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr; 1234 *sin = iw_event->local_addr; 1235 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr; 1236 *sin = iw_event->remote_addr; 1237 switch (iw_event->status) { 1238 case 0: 1239 event.event = RDMA_CM_EVENT_ESTABLISHED; 1240 break; 1241 case -ECONNRESET: 1242 case -ECONNREFUSED: 1243 event.event = RDMA_CM_EVENT_REJECTED; 1244 break; 1245 case -ETIMEDOUT: 1246 event.event = RDMA_CM_EVENT_UNREACHABLE; 1247 break; 1248 default: 1249 event.event = RDMA_CM_EVENT_CONNECT_ERROR; 1250 break; 1251 } 1252 break; 1253 case IW_CM_EVENT_ESTABLISHED: 1254 event.event = RDMA_CM_EVENT_ESTABLISHED; 1255 break; 1256 default: 1257 BUG_ON(1); 1258 } 1259 1260 event.status = iw_event->status; 1261 event.param.conn.private_data = iw_event->private_data; 1262 event.param.conn.private_data_len = iw_event->private_data_len; 1263 ret = id_priv->id.event_handler(&id_priv->id, &event); 1264 if (ret) { 1265 /* Destroy the CM ID by returning a non-zero value. */ 1266 id_priv->cm_id.iw = NULL; 1267 cma_exch(id_priv, CMA_DESTROYING); 1268 mutex_unlock(&id_priv->handler_mutex); 1269 rdma_destroy_id(&id_priv->id); 1270 return ret; 1271 } 1272 1273 mutex_unlock(&id_priv->handler_mutex); 1274 return ret; 1275 } 1276 1277 static int iw_conn_req_handler(struct iw_cm_id *cm_id, 1278 struct iw_cm_event *iw_event) 1279 { 1280 struct rdma_cm_id *new_cm_id; 1281 struct rdma_id_private *listen_id, *conn_id; 1282 struct sockaddr_in *sin; 1283 struct net_device *dev = NULL; 1284 struct rdma_cm_event event; 1285 int ret; 1286 struct ib_device_attr attr; 1287 1288 listen_id = cm_id->context; 1289 if (cma_disable_callback(listen_id, CMA_LISTEN)) 1290 return -ECONNABORTED; 1291 1292 /* Create a new RDMA id for the new IW CM ID */ 1293 new_cm_id = rdma_create_id(listen_id->id.event_handler, 1294 listen_id->id.context, 1295 RDMA_PS_TCP); 1296 if (IS_ERR(new_cm_id)) { 1297 ret = -ENOMEM; 1298 goto out; 1299 } 1300 conn_id = container_of(new_cm_id, struct rdma_id_private, id); 1301 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING); 1302 conn_id->state = CMA_CONNECT; 1303 1304 dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr); 1305 if (!dev) { 1306 ret = -EADDRNOTAVAIL; 1307 mutex_unlock(&conn_id->handler_mutex); 1308 rdma_destroy_id(new_cm_id); 1309 goto out; 1310 } 1311 ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL); 1312 if (ret) { 1313 mutex_unlock(&conn_id->handler_mutex); 1314 rdma_destroy_id(new_cm_id); 1315 goto out; 1316 } 1317 1318 mutex_lock(&lock); 1319 ret = cma_acquire_dev(conn_id); 1320 mutex_unlock(&lock); 1321 if (ret) { 1322 mutex_unlock(&conn_id->handler_mutex); 1323 rdma_destroy_id(new_cm_id); 1324 goto out; 1325 } 1326 1327 conn_id->cm_id.iw = cm_id; 1328 cm_id->context = conn_id; 1329 cm_id->cm_handler = cma_iw_handler; 1330 1331 sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr; 1332 *sin = iw_event->local_addr; 1333 sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr; 1334 *sin = iw_event->remote_addr; 1335 1336 ret = ib_query_device(conn_id->id.device, &attr); 1337 if (ret) { 1338 mutex_unlock(&conn_id->handler_mutex); 1339 rdma_destroy_id(new_cm_id); 1340 goto out; 1341 } 1342 1343 memset(&event, 0, sizeof event); 1344 event.event = RDMA_CM_EVENT_CONNECT_REQUEST; 1345 event.param.conn.private_data = iw_event->private_data; 1346 event.param.conn.private_data_len = iw_event->private_data_len; 1347 event.param.conn.initiator_depth = attr.max_qp_init_rd_atom; 1348 event.param.conn.responder_resources = attr.max_qp_rd_atom; 1349 ret = conn_id->id.event_handler(&conn_id->id, &event); 1350 if (ret) { 1351 /* User wants to destroy the CM ID */ 1352 conn_id->cm_id.iw = NULL; 1353 cma_exch(conn_id, CMA_DESTROYING); 1354 mutex_unlock(&conn_id->handler_mutex); 1355 rdma_destroy_id(&conn_id->id); 1356 goto out; 1357 } 1358 1359 mutex_unlock(&conn_id->handler_mutex); 1360 1361 out: 1362 if (dev) 1363 dev_put(dev); 1364 mutex_unlock(&listen_id->handler_mutex); 1365 return ret; 1366 } 1367 1368 static int cma_ib_listen(struct rdma_id_private *id_priv) 1369 { 1370 struct ib_cm_compare_data compare_data; 1371 struct sockaddr *addr; 1372 __be64 svc_id; 1373 int ret; 1374 1375 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler, 1376 id_priv); 1377 if (IS_ERR(id_priv->cm_id.ib)) 1378 return PTR_ERR(id_priv->cm_id.ib); 1379 1380 addr = &id_priv->id.route.addr.src_addr; 1381 svc_id = cma_get_service_id(id_priv->id.ps, addr); 1382 if (cma_any_addr(addr)) 1383 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL); 1384 else { 1385 cma_set_compare_data(id_priv->id.ps, addr, &compare_data); 1386 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data); 1387 } 1388 1389 if (ret) { 1390 ib_destroy_cm_id(id_priv->cm_id.ib); 1391 id_priv->cm_id.ib = NULL; 1392 } 1393 1394 return ret; 1395 } 1396 1397 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog) 1398 { 1399 int ret; 1400 struct sockaddr_in *sin; 1401 1402 id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device, 1403 iw_conn_req_handler, 1404 id_priv); 1405 if (IS_ERR(id_priv->cm_id.iw)) 1406 return PTR_ERR(id_priv->cm_id.iw); 1407 1408 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr; 1409 id_priv->cm_id.iw->local_addr = *sin; 1410 1411 ret = iw_cm_listen(id_priv->cm_id.iw, backlog); 1412 1413 if (ret) { 1414 iw_destroy_cm_id(id_priv->cm_id.iw); 1415 id_priv->cm_id.iw = NULL; 1416 } 1417 1418 return ret; 1419 } 1420 1421 static int cma_listen_handler(struct rdma_cm_id *id, 1422 struct rdma_cm_event *event) 1423 { 1424 struct rdma_id_private *id_priv = id->context; 1425 1426 id->context = id_priv->id.context; 1427 id->event_handler = id_priv->id.event_handler; 1428 return id_priv->id.event_handler(id, event); 1429 } 1430 1431 static void cma_listen_on_dev(struct rdma_id_private *id_priv, 1432 struct cma_device *cma_dev) 1433 { 1434 struct rdma_id_private *dev_id_priv; 1435 struct rdma_cm_id *id; 1436 int ret; 1437 1438 id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps); 1439 if (IS_ERR(id)) 1440 return; 1441 1442 dev_id_priv = container_of(id, struct rdma_id_private, id); 1443 1444 dev_id_priv->state = CMA_ADDR_BOUND; 1445 memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr, 1446 ip_addr_size(&id_priv->id.route.addr.src_addr)); 1447 1448 cma_attach_to_dev(dev_id_priv, cma_dev); 1449 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list); 1450 atomic_inc(&id_priv->refcount); 1451 dev_id_priv->internal_id = 1; 1452 1453 ret = rdma_listen(id, id_priv->backlog); 1454 if (ret) 1455 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, " 1456 "listening on device %s\n", ret, cma_dev->device->name); 1457 } 1458 1459 static void cma_listen_on_all(struct rdma_id_private *id_priv) 1460 { 1461 struct cma_device *cma_dev; 1462 1463 mutex_lock(&lock); 1464 list_add_tail(&id_priv->list, &listen_any_list); 1465 list_for_each_entry(cma_dev, &dev_list, list) 1466 cma_listen_on_dev(id_priv, cma_dev); 1467 mutex_unlock(&lock); 1468 } 1469 1470 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af) 1471 { 1472 struct sockaddr_in addr_in; 1473 1474 memset(&addr_in, 0, sizeof addr_in); 1475 addr_in.sin_family = af; 1476 return rdma_bind_addr(id, (struct sockaddr *) &addr_in); 1477 } 1478 1479 int rdma_listen(struct rdma_cm_id *id, int backlog) 1480 { 1481 struct rdma_id_private *id_priv; 1482 int ret; 1483 1484 id_priv = container_of(id, struct rdma_id_private, id); 1485 if (id_priv->state == CMA_IDLE) { 1486 ret = cma_bind_any(id, AF_INET); 1487 if (ret) 1488 return ret; 1489 } 1490 1491 if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN)) 1492 return -EINVAL; 1493 1494 id_priv->backlog = backlog; 1495 if (id->device) { 1496 switch (rdma_node_get_transport(id->device->node_type)) { 1497 case RDMA_TRANSPORT_IB: 1498 ret = cma_ib_listen(id_priv); 1499 if (ret) 1500 goto err; 1501 break; 1502 case RDMA_TRANSPORT_IWARP: 1503 ret = cma_iw_listen(id_priv, backlog); 1504 if (ret) 1505 goto err; 1506 break; 1507 default: 1508 ret = -ENOSYS; 1509 goto err; 1510 } 1511 } else 1512 cma_listen_on_all(id_priv); 1513 1514 return 0; 1515 err: 1516 id_priv->backlog = 0; 1517 cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND); 1518 return ret; 1519 } 1520 EXPORT_SYMBOL(rdma_listen); 1521 1522 void rdma_set_service_type(struct rdma_cm_id *id, int tos) 1523 { 1524 struct rdma_id_private *id_priv; 1525 1526 id_priv = container_of(id, struct rdma_id_private, id); 1527 id_priv->tos = (u8) tos; 1528 } 1529 EXPORT_SYMBOL(rdma_set_service_type); 1530 1531 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec, 1532 void *context) 1533 { 1534 struct cma_work *work = context; 1535 struct rdma_route *route; 1536 1537 route = &work->id->id.route; 1538 1539 if (!status) { 1540 route->num_paths = 1; 1541 *route->path_rec = *path_rec; 1542 } else { 1543 work->old_state = CMA_ROUTE_QUERY; 1544 work->new_state = CMA_ADDR_RESOLVED; 1545 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR; 1546 work->event.status = status; 1547 } 1548 1549 queue_work(cma_wq, &work->work); 1550 } 1551 1552 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms, 1553 struct cma_work *work) 1554 { 1555 struct rdma_addr *addr = &id_priv->id.route.addr; 1556 struct ib_sa_path_rec path_rec; 1557 ib_sa_comp_mask comp_mask; 1558 struct sockaddr_in6 *sin6; 1559 1560 memset(&path_rec, 0, sizeof path_rec); 1561 ib_addr_get_sgid(&addr->dev_addr, &path_rec.sgid); 1562 ib_addr_get_dgid(&addr->dev_addr, &path_rec.dgid); 1563 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr)); 1564 path_rec.numb_path = 1; 1565 path_rec.reversible = 1; 1566 path_rec.service_id = cma_get_service_id(id_priv->id.ps, &addr->dst_addr); 1567 1568 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID | 1569 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH | 1570 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID; 1571 1572 if (addr->src_addr.sa_family == AF_INET) { 1573 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos); 1574 comp_mask |= IB_SA_PATH_REC_QOS_CLASS; 1575 } else { 1576 sin6 = (struct sockaddr_in6 *) &addr->src_addr; 1577 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20); 1578 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS; 1579 } 1580 1581 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device, 1582 id_priv->id.port_num, &path_rec, 1583 comp_mask, timeout_ms, 1584 GFP_KERNEL, cma_query_handler, 1585 work, &id_priv->query); 1586 1587 return (id_priv->query_id < 0) ? id_priv->query_id : 0; 1588 } 1589 1590 static void cma_work_handler(struct work_struct *_work) 1591 { 1592 struct cma_work *work = container_of(_work, struct cma_work, work); 1593 struct rdma_id_private *id_priv = work->id; 1594 int destroy = 0; 1595 1596 mutex_lock(&id_priv->handler_mutex); 1597 if (!cma_comp_exch(id_priv, work->old_state, work->new_state)) 1598 goto out; 1599 1600 if (id_priv->id.event_handler(&id_priv->id, &work->event)) { 1601 cma_exch(id_priv, CMA_DESTROYING); 1602 destroy = 1; 1603 } 1604 out: 1605 mutex_unlock(&id_priv->handler_mutex); 1606 cma_deref_id(id_priv); 1607 if (destroy) 1608 rdma_destroy_id(&id_priv->id); 1609 kfree(work); 1610 } 1611 1612 static void cma_ndev_work_handler(struct work_struct *_work) 1613 { 1614 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work); 1615 struct rdma_id_private *id_priv = work->id; 1616 int destroy = 0; 1617 1618 mutex_lock(&id_priv->handler_mutex); 1619 if (id_priv->state == CMA_DESTROYING || 1620 id_priv->state == CMA_DEVICE_REMOVAL) 1621 goto out; 1622 1623 if (id_priv->id.event_handler(&id_priv->id, &work->event)) { 1624 cma_exch(id_priv, CMA_DESTROYING); 1625 destroy = 1; 1626 } 1627 1628 out: 1629 mutex_unlock(&id_priv->handler_mutex); 1630 cma_deref_id(id_priv); 1631 if (destroy) 1632 rdma_destroy_id(&id_priv->id); 1633 kfree(work); 1634 } 1635 1636 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms) 1637 { 1638 struct rdma_route *route = &id_priv->id.route; 1639 struct cma_work *work; 1640 int ret; 1641 1642 work = kzalloc(sizeof *work, GFP_KERNEL); 1643 if (!work) 1644 return -ENOMEM; 1645 1646 work->id = id_priv; 1647 INIT_WORK(&work->work, cma_work_handler); 1648 work->old_state = CMA_ROUTE_QUERY; 1649 work->new_state = CMA_ROUTE_RESOLVED; 1650 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 1651 1652 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL); 1653 if (!route->path_rec) { 1654 ret = -ENOMEM; 1655 goto err1; 1656 } 1657 1658 ret = cma_query_ib_route(id_priv, timeout_ms, work); 1659 if (ret) 1660 goto err2; 1661 1662 return 0; 1663 err2: 1664 kfree(route->path_rec); 1665 route->path_rec = NULL; 1666 err1: 1667 kfree(work); 1668 return ret; 1669 } 1670 1671 int rdma_set_ib_paths(struct rdma_cm_id *id, 1672 struct ib_sa_path_rec *path_rec, int num_paths) 1673 { 1674 struct rdma_id_private *id_priv; 1675 int ret; 1676 1677 id_priv = container_of(id, struct rdma_id_private, id); 1678 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED)) 1679 return -EINVAL; 1680 1681 id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL); 1682 if (!id->route.path_rec) { 1683 ret = -ENOMEM; 1684 goto err; 1685 } 1686 1687 memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths); 1688 return 0; 1689 err: 1690 cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED); 1691 return ret; 1692 } 1693 EXPORT_SYMBOL(rdma_set_ib_paths); 1694 1695 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms) 1696 { 1697 struct cma_work *work; 1698 1699 work = kzalloc(sizeof *work, GFP_KERNEL); 1700 if (!work) 1701 return -ENOMEM; 1702 1703 work->id = id_priv; 1704 INIT_WORK(&work->work, cma_work_handler); 1705 work->old_state = CMA_ROUTE_QUERY; 1706 work->new_state = CMA_ROUTE_RESOLVED; 1707 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 1708 queue_work(cma_wq, &work->work); 1709 return 0; 1710 } 1711 1712 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms) 1713 { 1714 struct rdma_id_private *id_priv; 1715 int ret; 1716 1717 id_priv = container_of(id, struct rdma_id_private, id); 1718 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY)) 1719 return -EINVAL; 1720 1721 atomic_inc(&id_priv->refcount); 1722 switch (rdma_node_get_transport(id->device->node_type)) { 1723 case RDMA_TRANSPORT_IB: 1724 ret = cma_resolve_ib_route(id_priv, timeout_ms); 1725 break; 1726 case RDMA_TRANSPORT_IWARP: 1727 ret = cma_resolve_iw_route(id_priv, timeout_ms); 1728 break; 1729 default: 1730 ret = -ENOSYS; 1731 break; 1732 } 1733 if (ret) 1734 goto err; 1735 1736 return 0; 1737 err: 1738 cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED); 1739 cma_deref_id(id_priv); 1740 return ret; 1741 } 1742 EXPORT_SYMBOL(rdma_resolve_route); 1743 1744 static int cma_bind_loopback(struct rdma_id_private *id_priv) 1745 { 1746 struct cma_device *cma_dev; 1747 struct ib_port_attr port_attr; 1748 union ib_gid gid; 1749 u16 pkey; 1750 int ret; 1751 u8 p; 1752 1753 mutex_lock(&lock); 1754 if (list_empty(&dev_list)) { 1755 ret = -ENODEV; 1756 goto out; 1757 } 1758 list_for_each_entry(cma_dev, &dev_list, list) 1759 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p) 1760 if (!ib_query_port(cma_dev->device, p, &port_attr) && 1761 port_attr.state == IB_PORT_ACTIVE) 1762 goto port_found; 1763 1764 p = 1; 1765 cma_dev = list_entry(dev_list.next, struct cma_device, list); 1766 1767 port_found: 1768 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid); 1769 if (ret) 1770 goto out; 1771 1772 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey); 1773 if (ret) 1774 goto out; 1775 1776 ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid); 1777 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey); 1778 id_priv->id.port_num = p; 1779 cma_attach_to_dev(id_priv, cma_dev); 1780 out: 1781 mutex_unlock(&lock); 1782 return ret; 1783 } 1784 1785 static void addr_handler(int status, struct sockaddr *src_addr, 1786 struct rdma_dev_addr *dev_addr, void *context) 1787 { 1788 struct rdma_id_private *id_priv = context; 1789 struct rdma_cm_event event; 1790 1791 memset(&event, 0, sizeof event); 1792 mutex_lock(&id_priv->handler_mutex); 1793 1794 /* 1795 * Grab mutex to block rdma_destroy_id() from removing the device while 1796 * we're trying to acquire it. 1797 */ 1798 mutex_lock(&lock); 1799 if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) { 1800 mutex_unlock(&lock); 1801 goto out; 1802 } 1803 1804 if (!status && !id_priv->cma_dev) 1805 status = cma_acquire_dev(id_priv); 1806 mutex_unlock(&lock); 1807 1808 if (status) { 1809 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND)) 1810 goto out; 1811 event.event = RDMA_CM_EVENT_ADDR_ERROR; 1812 event.status = status; 1813 } else { 1814 memcpy(&id_priv->id.route.addr.src_addr, src_addr, 1815 ip_addr_size(src_addr)); 1816 event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 1817 } 1818 1819 if (id_priv->id.event_handler(&id_priv->id, &event)) { 1820 cma_exch(id_priv, CMA_DESTROYING); 1821 mutex_unlock(&id_priv->handler_mutex); 1822 cma_deref_id(id_priv); 1823 rdma_destroy_id(&id_priv->id); 1824 return; 1825 } 1826 out: 1827 mutex_unlock(&id_priv->handler_mutex); 1828 cma_deref_id(id_priv); 1829 } 1830 1831 static int cma_resolve_loopback(struct rdma_id_private *id_priv) 1832 { 1833 struct cma_work *work; 1834 struct sockaddr_in *src_in, *dst_in; 1835 union ib_gid gid; 1836 int ret; 1837 1838 work = kzalloc(sizeof *work, GFP_KERNEL); 1839 if (!work) 1840 return -ENOMEM; 1841 1842 if (!id_priv->cma_dev) { 1843 ret = cma_bind_loopback(id_priv); 1844 if (ret) 1845 goto err; 1846 } 1847 1848 ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid); 1849 ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid); 1850 1851 if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) { 1852 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr; 1853 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr; 1854 src_in->sin_family = dst_in->sin_family; 1855 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr; 1856 } 1857 1858 work->id = id_priv; 1859 INIT_WORK(&work->work, cma_work_handler); 1860 work->old_state = CMA_ADDR_QUERY; 1861 work->new_state = CMA_ADDR_RESOLVED; 1862 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 1863 queue_work(cma_wq, &work->work); 1864 return 0; 1865 err: 1866 kfree(work); 1867 return ret; 1868 } 1869 1870 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr, 1871 struct sockaddr *dst_addr) 1872 { 1873 if (src_addr && src_addr->sa_family) 1874 return rdma_bind_addr(id, src_addr); 1875 else 1876 return cma_bind_any(id, dst_addr->sa_family); 1877 } 1878 1879 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr, 1880 struct sockaddr *dst_addr, int timeout_ms) 1881 { 1882 struct rdma_id_private *id_priv; 1883 int ret; 1884 1885 id_priv = container_of(id, struct rdma_id_private, id); 1886 if (id_priv->state == CMA_IDLE) { 1887 ret = cma_bind_addr(id, src_addr, dst_addr); 1888 if (ret) 1889 return ret; 1890 } 1891 1892 if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY)) 1893 return -EINVAL; 1894 1895 atomic_inc(&id_priv->refcount); 1896 memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr)); 1897 if (cma_any_addr(dst_addr)) 1898 ret = cma_resolve_loopback(id_priv); 1899 else 1900 ret = rdma_resolve_ip(&addr_client, &id->route.addr.src_addr, 1901 dst_addr, &id->route.addr.dev_addr, 1902 timeout_ms, addr_handler, id_priv); 1903 if (ret) 1904 goto err; 1905 1906 return 0; 1907 err: 1908 cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND); 1909 cma_deref_id(id_priv); 1910 return ret; 1911 } 1912 EXPORT_SYMBOL(rdma_resolve_addr); 1913 1914 static void cma_bind_port(struct rdma_bind_list *bind_list, 1915 struct rdma_id_private *id_priv) 1916 { 1917 struct sockaddr_in *sin; 1918 1919 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr; 1920 sin->sin_port = htons(bind_list->port); 1921 id_priv->bind_list = bind_list; 1922 hlist_add_head(&id_priv->node, &bind_list->owners); 1923 } 1924 1925 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv, 1926 unsigned short snum) 1927 { 1928 struct rdma_bind_list *bind_list; 1929 int port, ret; 1930 1931 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL); 1932 if (!bind_list) 1933 return -ENOMEM; 1934 1935 do { 1936 ret = idr_get_new_above(ps, bind_list, snum, &port); 1937 } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL)); 1938 1939 if (ret) 1940 goto err1; 1941 1942 if (port != snum) { 1943 ret = -EADDRNOTAVAIL; 1944 goto err2; 1945 } 1946 1947 bind_list->ps = ps; 1948 bind_list->port = (unsigned short) port; 1949 cma_bind_port(bind_list, id_priv); 1950 return 0; 1951 err2: 1952 idr_remove(ps, port); 1953 err1: 1954 kfree(bind_list); 1955 return ret; 1956 } 1957 1958 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv) 1959 { 1960 struct rdma_bind_list *bind_list; 1961 int port, ret, low, high; 1962 1963 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL); 1964 if (!bind_list) 1965 return -ENOMEM; 1966 1967 retry: 1968 /* FIXME: add proper port randomization per like inet_csk_get_port */ 1969 do { 1970 ret = idr_get_new_above(ps, bind_list, next_port, &port); 1971 } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL)); 1972 1973 if (ret) 1974 goto err1; 1975 1976 inet_get_local_port_range(&low, &high); 1977 if (port > high) { 1978 if (next_port != low) { 1979 idr_remove(ps, port); 1980 next_port = low; 1981 goto retry; 1982 } 1983 ret = -EADDRNOTAVAIL; 1984 goto err2; 1985 } 1986 1987 if (port == high) 1988 next_port = low; 1989 else 1990 next_port = port + 1; 1991 1992 bind_list->ps = ps; 1993 bind_list->port = (unsigned short) port; 1994 cma_bind_port(bind_list, id_priv); 1995 return 0; 1996 err2: 1997 idr_remove(ps, port); 1998 err1: 1999 kfree(bind_list); 2000 return ret; 2001 } 2002 2003 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv) 2004 { 2005 struct rdma_id_private *cur_id; 2006 struct sockaddr_in *sin, *cur_sin; 2007 struct rdma_bind_list *bind_list; 2008 struct hlist_node *node; 2009 unsigned short snum; 2010 2011 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr; 2012 snum = ntohs(sin->sin_port); 2013 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE)) 2014 return -EACCES; 2015 2016 bind_list = idr_find(ps, snum); 2017 if (!bind_list) 2018 return cma_alloc_port(ps, id_priv, snum); 2019 2020 /* 2021 * We don't support binding to any address if anyone is bound to 2022 * a specific address on the same port. 2023 */ 2024 if (cma_any_addr(&id_priv->id.route.addr.src_addr)) 2025 return -EADDRNOTAVAIL; 2026 2027 hlist_for_each_entry(cur_id, node, &bind_list->owners, node) { 2028 if (cma_any_addr(&cur_id->id.route.addr.src_addr)) 2029 return -EADDRNOTAVAIL; 2030 2031 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr; 2032 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr) 2033 return -EADDRINUSE; 2034 } 2035 2036 cma_bind_port(bind_list, id_priv); 2037 return 0; 2038 } 2039 2040 static int cma_get_port(struct rdma_id_private *id_priv) 2041 { 2042 struct idr *ps; 2043 int ret; 2044 2045 switch (id_priv->id.ps) { 2046 case RDMA_PS_SDP: 2047 ps = &sdp_ps; 2048 break; 2049 case RDMA_PS_TCP: 2050 ps = &tcp_ps; 2051 break; 2052 case RDMA_PS_UDP: 2053 ps = &udp_ps; 2054 break; 2055 case RDMA_PS_IPOIB: 2056 ps = &ipoib_ps; 2057 break; 2058 default: 2059 return -EPROTONOSUPPORT; 2060 } 2061 2062 mutex_lock(&lock); 2063 if (cma_any_port(&id_priv->id.route.addr.src_addr)) 2064 ret = cma_alloc_any_port(ps, id_priv); 2065 else 2066 ret = cma_use_port(ps, id_priv); 2067 mutex_unlock(&lock); 2068 2069 return ret; 2070 } 2071 2072 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr) 2073 { 2074 struct rdma_id_private *id_priv; 2075 int ret; 2076 2077 if (addr->sa_family != AF_INET) 2078 return -EAFNOSUPPORT; 2079 2080 id_priv = container_of(id, struct rdma_id_private, id); 2081 if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND)) 2082 return -EINVAL; 2083 2084 if (!cma_any_addr(addr)) { 2085 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr); 2086 if (ret) 2087 goto err1; 2088 2089 mutex_lock(&lock); 2090 ret = cma_acquire_dev(id_priv); 2091 mutex_unlock(&lock); 2092 if (ret) 2093 goto err1; 2094 } 2095 2096 memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr)); 2097 ret = cma_get_port(id_priv); 2098 if (ret) 2099 goto err2; 2100 2101 return 0; 2102 err2: 2103 if (!cma_any_addr(addr)) { 2104 mutex_lock(&lock); 2105 cma_detach_from_dev(id_priv); 2106 mutex_unlock(&lock); 2107 } 2108 err1: 2109 cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE); 2110 return ret; 2111 } 2112 EXPORT_SYMBOL(rdma_bind_addr); 2113 2114 static int cma_format_hdr(void *hdr, enum rdma_port_space ps, 2115 struct rdma_route *route) 2116 { 2117 struct sockaddr_in *src4, *dst4; 2118 struct cma_hdr *cma_hdr; 2119 struct sdp_hh *sdp_hdr; 2120 2121 src4 = (struct sockaddr_in *) &route->addr.src_addr; 2122 dst4 = (struct sockaddr_in *) &route->addr.dst_addr; 2123 2124 switch (ps) { 2125 case RDMA_PS_SDP: 2126 sdp_hdr = hdr; 2127 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION) 2128 return -EINVAL; 2129 sdp_set_ip_ver(sdp_hdr, 4); 2130 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr; 2131 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr; 2132 sdp_hdr->port = src4->sin_port; 2133 break; 2134 default: 2135 cma_hdr = hdr; 2136 cma_hdr->cma_version = CMA_VERSION; 2137 cma_set_ip_ver(cma_hdr, 4); 2138 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr; 2139 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr; 2140 cma_hdr->port = src4->sin_port; 2141 break; 2142 } 2143 return 0; 2144 } 2145 2146 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id, 2147 struct ib_cm_event *ib_event) 2148 { 2149 struct rdma_id_private *id_priv = cm_id->context; 2150 struct rdma_cm_event event; 2151 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd; 2152 int ret = 0; 2153 2154 if (cma_disable_callback(id_priv, CMA_CONNECT)) 2155 return 0; 2156 2157 memset(&event, 0, sizeof event); 2158 switch (ib_event->event) { 2159 case IB_CM_SIDR_REQ_ERROR: 2160 event.event = RDMA_CM_EVENT_UNREACHABLE; 2161 event.status = -ETIMEDOUT; 2162 break; 2163 case IB_CM_SIDR_REP_RECEIVED: 2164 event.param.ud.private_data = ib_event->private_data; 2165 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE; 2166 if (rep->status != IB_SIDR_SUCCESS) { 2167 event.event = RDMA_CM_EVENT_UNREACHABLE; 2168 event.status = ib_event->param.sidr_rep_rcvd.status; 2169 break; 2170 } 2171 if (id_priv->qkey != rep->qkey) { 2172 event.event = RDMA_CM_EVENT_UNREACHABLE; 2173 event.status = -EINVAL; 2174 break; 2175 } 2176 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num, 2177 id_priv->id.route.path_rec, 2178 &event.param.ud.ah_attr); 2179 event.param.ud.qp_num = rep->qpn; 2180 event.param.ud.qkey = rep->qkey; 2181 event.event = RDMA_CM_EVENT_ESTABLISHED; 2182 event.status = 0; 2183 break; 2184 default: 2185 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n", 2186 ib_event->event); 2187 goto out; 2188 } 2189 2190 ret = id_priv->id.event_handler(&id_priv->id, &event); 2191 if (ret) { 2192 /* Destroy the CM ID by returning a non-zero value. */ 2193 id_priv->cm_id.ib = NULL; 2194 cma_exch(id_priv, CMA_DESTROYING); 2195 mutex_unlock(&id_priv->handler_mutex); 2196 rdma_destroy_id(&id_priv->id); 2197 return ret; 2198 } 2199 out: 2200 mutex_unlock(&id_priv->handler_mutex); 2201 return ret; 2202 } 2203 2204 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv, 2205 struct rdma_conn_param *conn_param) 2206 { 2207 struct ib_cm_sidr_req_param req; 2208 struct rdma_route *route; 2209 int ret; 2210 2211 req.private_data_len = sizeof(struct cma_hdr) + 2212 conn_param->private_data_len; 2213 req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC); 2214 if (!req.private_data) 2215 return -ENOMEM; 2216 2217 if (conn_param->private_data && conn_param->private_data_len) 2218 memcpy((void *) req.private_data + sizeof(struct cma_hdr), 2219 conn_param->private_data, conn_param->private_data_len); 2220 2221 route = &id_priv->id.route; 2222 ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route); 2223 if (ret) 2224 goto out; 2225 2226 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, 2227 cma_sidr_rep_handler, id_priv); 2228 if (IS_ERR(id_priv->cm_id.ib)) { 2229 ret = PTR_ERR(id_priv->cm_id.ib); 2230 goto out; 2231 } 2232 2233 req.path = route->path_rec; 2234 req.service_id = cma_get_service_id(id_priv->id.ps, 2235 &route->addr.dst_addr); 2236 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8); 2237 req.max_cm_retries = CMA_MAX_CM_RETRIES; 2238 2239 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req); 2240 if (ret) { 2241 ib_destroy_cm_id(id_priv->cm_id.ib); 2242 id_priv->cm_id.ib = NULL; 2243 } 2244 out: 2245 kfree(req.private_data); 2246 return ret; 2247 } 2248 2249 static int cma_connect_ib(struct rdma_id_private *id_priv, 2250 struct rdma_conn_param *conn_param) 2251 { 2252 struct ib_cm_req_param req; 2253 struct rdma_route *route; 2254 void *private_data; 2255 int offset, ret; 2256 2257 memset(&req, 0, sizeof req); 2258 offset = cma_user_data_offset(id_priv->id.ps); 2259 req.private_data_len = offset + conn_param->private_data_len; 2260 private_data = kzalloc(req.private_data_len, GFP_ATOMIC); 2261 if (!private_data) 2262 return -ENOMEM; 2263 2264 if (conn_param->private_data && conn_param->private_data_len) 2265 memcpy(private_data + offset, conn_param->private_data, 2266 conn_param->private_data_len); 2267 2268 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler, 2269 id_priv); 2270 if (IS_ERR(id_priv->cm_id.ib)) { 2271 ret = PTR_ERR(id_priv->cm_id.ib); 2272 goto out; 2273 } 2274 2275 route = &id_priv->id.route; 2276 ret = cma_format_hdr(private_data, id_priv->id.ps, route); 2277 if (ret) 2278 goto out; 2279 req.private_data = private_data; 2280 2281 req.primary_path = &route->path_rec[0]; 2282 if (route->num_paths == 2) 2283 req.alternate_path = &route->path_rec[1]; 2284 2285 req.service_id = cma_get_service_id(id_priv->id.ps, 2286 &route->addr.dst_addr); 2287 req.qp_num = id_priv->qp_num; 2288 req.qp_type = IB_QPT_RC; 2289 req.starting_psn = id_priv->seq_num; 2290 req.responder_resources = conn_param->responder_resources; 2291 req.initiator_depth = conn_param->initiator_depth; 2292 req.flow_control = conn_param->flow_control; 2293 req.retry_count = conn_param->retry_count; 2294 req.rnr_retry_count = conn_param->rnr_retry_count; 2295 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT; 2296 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT; 2297 req.max_cm_retries = CMA_MAX_CM_RETRIES; 2298 req.srq = id_priv->srq ? 1 : 0; 2299 2300 ret = ib_send_cm_req(id_priv->cm_id.ib, &req); 2301 out: 2302 if (ret && !IS_ERR(id_priv->cm_id.ib)) { 2303 ib_destroy_cm_id(id_priv->cm_id.ib); 2304 id_priv->cm_id.ib = NULL; 2305 } 2306 2307 kfree(private_data); 2308 return ret; 2309 } 2310 2311 static int cma_connect_iw(struct rdma_id_private *id_priv, 2312 struct rdma_conn_param *conn_param) 2313 { 2314 struct iw_cm_id *cm_id; 2315 struct sockaddr_in* sin; 2316 int ret; 2317 struct iw_cm_conn_param iw_param; 2318 2319 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv); 2320 if (IS_ERR(cm_id)) { 2321 ret = PTR_ERR(cm_id); 2322 goto out; 2323 } 2324 2325 id_priv->cm_id.iw = cm_id; 2326 2327 sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr; 2328 cm_id->local_addr = *sin; 2329 2330 sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr; 2331 cm_id->remote_addr = *sin; 2332 2333 ret = cma_modify_qp_rtr(id_priv, conn_param); 2334 if (ret) 2335 goto out; 2336 2337 iw_param.ord = conn_param->initiator_depth; 2338 iw_param.ird = conn_param->responder_resources; 2339 iw_param.private_data = conn_param->private_data; 2340 iw_param.private_data_len = conn_param->private_data_len; 2341 if (id_priv->id.qp) 2342 iw_param.qpn = id_priv->qp_num; 2343 else 2344 iw_param.qpn = conn_param->qp_num; 2345 ret = iw_cm_connect(cm_id, &iw_param); 2346 out: 2347 if (ret && !IS_ERR(cm_id)) { 2348 iw_destroy_cm_id(cm_id); 2349 id_priv->cm_id.iw = NULL; 2350 } 2351 return ret; 2352 } 2353 2354 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param) 2355 { 2356 struct rdma_id_private *id_priv; 2357 int ret; 2358 2359 id_priv = container_of(id, struct rdma_id_private, id); 2360 if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT)) 2361 return -EINVAL; 2362 2363 if (!id->qp) { 2364 id_priv->qp_num = conn_param->qp_num; 2365 id_priv->srq = conn_param->srq; 2366 } 2367 2368 switch (rdma_node_get_transport(id->device->node_type)) { 2369 case RDMA_TRANSPORT_IB: 2370 if (cma_is_ud_ps(id->ps)) 2371 ret = cma_resolve_ib_udp(id_priv, conn_param); 2372 else 2373 ret = cma_connect_ib(id_priv, conn_param); 2374 break; 2375 case RDMA_TRANSPORT_IWARP: 2376 ret = cma_connect_iw(id_priv, conn_param); 2377 break; 2378 default: 2379 ret = -ENOSYS; 2380 break; 2381 } 2382 if (ret) 2383 goto err; 2384 2385 return 0; 2386 err: 2387 cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED); 2388 return ret; 2389 } 2390 EXPORT_SYMBOL(rdma_connect); 2391 2392 static int cma_accept_ib(struct rdma_id_private *id_priv, 2393 struct rdma_conn_param *conn_param) 2394 { 2395 struct ib_cm_rep_param rep; 2396 int ret; 2397 2398 ret = cma_modify_qp_rtr(id_priv, conn_param); 2399 if (ret) 2400 goto out; 2401 2402 ret = cma_modify_qp_rts(id_priv, conn_param); 2403 if (ret) 2404 goto out; 2405 2406 memset(&rep, 0, sizeof rep); 2407 rep.qp_num = id_priv->qp_num; 2408 rep.starting_psn = id_priv->seq_num; 2409 rep.private_data = conn_param->private_data; 2410 rep.private_data_len = conn_param->private_data_len; 2411 rep.responder_resources = conn_param->responder_resources; 2412 rep.initiator_depth = conn_param->initiator_depth; 2413 rep.failover_accepted = 0; 2414 rep.flow_control = conn_param->flow_control; 2415 rep.rnr_retry_count = conn_param->rnr_retry_count; 2416 rep.srq = id_priv->srq ? 1 : 0; 2417 2418 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep); 2419 out: 2420 return ret; 2421 } 2422 2423 static int cma_accept_iw(struct rdma_id_private *id_priv, 2424 struct rdma_conn_param *conn_param) 2425 { 2426 struct iw_cm_conn_param iw_param; 2427 int ret; 2428 2429 ret = cma_modify_qp_rtr(id_priv, conn_param); 2430 if (ret) 2431 return ret; 2432 2433 iw_param.ord = conn_param->initiator_depth; 2434 iw_param.ird = conn_param->responder_resources; 2435 iw_param.private_data = conn_param->private_data; 2436 iw_param.private_data_len = conn_param->private_data_len; 2437 if (id_priv->id.qp) { 2438 iw_param.qpn = id_priv->qp_num; 2439 } else 2440 iw_param.qpn = conn_param->qp_num; 2441 2442 return iw_cm_accept(id_priv->cm_id.iw, &iw_param); 2443 } 2444 2445 static int cma_send_sidr_rep(struct rdma_id_private *id_priv, 2446 enum ib_cm_sidr_status status, 2447 const void *private_data, int private_data_len) 2448 { 2449 struct ib_cm_sidr_rep_param rep; 2450 2451 memset(&rep, 0, sizeof rep); 2452 rep.status = status; 2453 if (status == IB_SIDR_SUCCESS) { 2454 rep.qp_num = id_priv->qp_num; 2455 rep.qkey = id_priv->qkey; 2456 } 2457 rep.private_data = private_data; 2458 rep.private_data_len = private_data_len; 2459 2460 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep); 2461 } 2462 2463 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param) 2464 { 2465 struct rdma_id_private *id_priv; 2466 int ret; 2467 2468 id_priv = container_of(id, struct rdma_id_private, id); 2469 if (!cma_comp(id_priv, CMA_CONNECT)) 2470 return -EINVAL; 2471 2472 if (!id->qp && conn_param) { 2473 id_priv->qp_num = conn_param->qp_num; 2474 id_priv->srq = conn_param->srq; 2475 } 2476 2477 switch (rdma_node_get_transport(id->device->node_type)) { 2478 case RDMA_TRANSPORT_IB: 2479 if (cma_is_ud_ps(id->ps)) 2480 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS, 2481 conn_param->private_data, 2482 conn_param->private_data_len); 2483 else if (conn_param) 2484 ret = cma_accept_ib(id_priv, conn_param); 2485 else 2486 ret = cma_rep_recv(id_priv); 2487 break; 2488 case RDMA_TRANSPORT_IWARP: 2489 ret = cma_accept_iw(id_priv, conn_param); 2490 break; 2491 default: 2492 ret = -ENOSYS; 2493 break; 2494 } 2495 2496 if (ret) 2497 goto reject; 2498 2499 return 0; 2500 reject: 2501 cma_modify_qp_err(id_priv); 2502 rdma_reject(id, NULL, 0); 2503 return ret; 2504 } 2505 EXPORT_SYMBOL(rdma_accept); 2506 2507 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event) 2508 { 2509 struct rdma_id_private *id_priv; 2510 int ret; 2511 2512 id_priv = container_of(id, struct rdma_id_private, id); 2513 if (!cma_has_cm_dev(id_priv)) 2514 return -EINVAL; 2515 2516 switch (id->device->node_type) { 2517 case RDMA_NODE_IB_CA: 2518 ret = ib_cm_notify(id_priv->cm_id.ib, event); 2519 break; 2520 default: 2521 ret = 0; 2522 break; 2523 } 2524 return ret; 2525 } 2526 EXPORT_SYMBOL(rdma_notify); 2527 2528 int rdma_reject(struct rdma_cm_id *id, const void *private_data, 2529 u8 private_data_len) 2530 { 2531 struct rdma_id_private *id_priv; 2532 int ret; 2533 2534 id_priv = container_of(id, struct rdma_id_private, id); 2535 if (!cma_has_cm_dev(id_priv)) 2536 return -EINVAL; 2537 2538 switch (rdma_node_get_transport(id->device->node_type)) { 2539 case RDMA_TRANSPORT_IB: 2540 if (cma_is_ud_ps(id->ps)) 2541 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 2542 private_data, private_data_len); 2543 else 2544 ret = ib_send_cm_rej(id_priv->cm_id.ib, 2545 IB_CM_REJ_CONSUMER_DEFINED, NULL, 2546 0, private_data, private_data_len); 2547 break; 2548 case RDMA_TRANSPORT_IWARP: 2549 ret = iw_cm_reject(id_priv->cm_id.iw, 2550 private_data, private_data_len); 2551 break; 2552 default: 2553 ret = -ENOSYS; 2554 break; 2555 } 2556 return ret; 2557 } 2558 EXPORT_SYMBOL(rdma_reject); 2559 2560 int rdma_disconnect(struct rdma_cm_id *id) 2561 { 2562 struct rdma_id_private *id_priv; 2563 int ret; 2564 2565 id_priv = container_of(id, struct rdma_id_private, id); 2566 if (!cma_has_cm_dev(id_priv)) 2567 return -EINVAL; 2568 2569 switch (rdma_node_get_transport(id->device->node_type)) { 2570 case RDMA_TRANSPORT_IB: 2571 ret = cma_modify_qp_err(id_priv); 2572 if (ret) 2573 goto out; 2574 /* Initiate or respond to a disconnect. */ 2575 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) 2576 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0); 2577 break; 2578 case RDMA_TRANSPORT_IWARP: 2579 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0); 2580 break; 2581 default: 2582 ret = -EINVAL; 2583 break; 2584 } 2585 out: 2586 return ret; 2587 } 2588 EXPORT_SYMBOL(rdma_disconnect); 2589 2590 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast) 2591 { 2592 struct rdma_id_private *id_priv; 2593 struct cma_multicast *mc = multicast->context; 2594 struct rdma_cm_event event; 2595 int ret; 2596 2597 id_priv = mc->id_priv; 2598 if (cma_disable_callback(id_priv, CMA_ADDR_BOUND) && 2599 cma_disable_callback(id_priv, CMA_ADDR_RESOLVED)) 2600 return 0; 2601 2602 mutex_lock(&id_priv->qp_mutex); 2603 if (!status && id_priv->id.qp) 2604 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid, 2605 multicast->rec.mlid); 2606 mutex_unlock(&id_priv->qp_mutex); 2607 2608 memset(&event, 0, sizeof event); 2609 event.status = status; 2610 event.param.ud.private_data = mc->context; 2611 if (!status) { 2612 event.event = RDMA_CM_EVENT_MULTICAST_JOIN; 2613 ib_init_ah_from_mcmember(id_priv->id.device, 2614 id_priv->id.port_num, &multicast->rec, 2615 &event.param.ud.ah_attr); 2616 event.param.ud.qp_num = 0xFFFFFF; 2617 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey); 2618 } else 2619 event.event = RDMA_CM_EVENT_MULTICAST_ERROR; 2620 2621 ret = id_priv->id.event_handler(&id_priv->id, &event); 2622 if (ret) { 2623 cma_exch(id_priv, CMA_DESTROYING); 2624 mutex_unlock(&id_priv->handler_mutex); 2625 rdma_destroy_id(&id_priv->id); 2626 return 0; 2627 } 2628 2629 mutex_unlock(&id_priv->handler_mutex); 2630 return 0; 2631 } 2632 2633 static void cma_set_mgid(struct rdma_id_private *id_priv, 2634 struct sockaddr *addr, union ib_gid *mgid) 2635 { 2636 unsigned char mc_map[MAX_ADDR_LEN]; 2637 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 2638 struct sockaddr_in *sin = (struct sockaddr_in *) addr; 2639 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr; 2640 2641 if (cma_any_addr(addr)) { 2642 memset(mgid, 0, sizeof *mgid); 2643 } else if ((addr->sa_family == AF_INET6) && 2644 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFF10A01B) == 2645 0xFF10A01B)) { 2646 /* IPv6 address is an SA assigned MGID. */ 2647 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid); 2648 } else { 2649 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map); 2650 if (id_priv->id.ps == RDMA_PS_UDP) 2651 mc_map[7] = 0x01; /* Use RDMA CM signature */ 2652 *mgid = *(union ib_gid *) (mc_map + 4); 2653 } 2654 } 2655 2656 static int cma_join_ib_multicast(struct rdma_id_private *id_priv, 2657 struct cma_multicast *mc) 2658 { 2659 struct ib_sa_mcmember_rec rec; 2660 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 2661 ib_sa_comp_mask comp_mask; 2662 int ret; 2663 2664 ib_addr_get_mgid(dev_addr, &rec.mgid); 2665 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num, 2666 &rec.mgid, &rec); 2667 if (ret) 2668 return ret; 2669 2670 cma_set_mgid(id_priv, &mc->addr, &rec.mgid); 2671 if (id_priv->id.ps == RDMA_PS_UDP) 2672 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY); 2673 ib_addr_get_sgid(dev_addr, &rec.port_gid); 2674 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr)); 2675 rec.join_state = 1; 2676 2677 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID | 2678 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE | 2679 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL | 2680 IB_SA_MCMEMBER_REC_FLOW_LABEL | 2681 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS; 2682 2683 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device, 2684 id_priv->id.port_num, &rec, 2685 comp_mask, GFP_KERNEL, 2686 cma_ib_mc_handler, mc); 2687 if (IS_ERR(mc->multicast.ib)) 2688 return PTR_ERR(mc->multicast.ib); 2689 2690 return 0; 2691 } 2692 2693 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr, 2694 void *context) 2695 { 2696 struct rdma_id_private *id_priv; 2697 struct cma_multicast *mc; 2698 int ret; 2699 2700 id_priv = container_of(id, struct rdma_id_private, id); 2701 if (!cma_comp(id_priv, CMA_ADDR_BOUND) && 2702 !cma_comp(id_priv, CMA_ADDR_RESOLVED)) 2703 return -EINVAL; 2704 2705 mc = kmalloc(sizeof *mc, GFP_KERNEL); 2706 if (!mc) 2707 return -ENOMEM; 2708 2709 memcpy(&mc->addr, addr, ip_addr_size(addr)); 2710 mc->context = context; 2711 mc->id_priv = id_priv; 2712 2713 spin_lock(&id_priv->lock); 2714 list_add(&mc->list, &id_priv->mc_list); 2715 spin_unlock(&id_priv->lock); 2716 2717 switch (rdma_node_get_transport(id->device->node_type)) { 2718 case RDMA_TRANSPORT_IB: 2719 ret = cma_join_ib_multicast(id_priv, mc); 2720 break; 2721 default: 2722 ret = -ENOSYS; 2723 break; 2724 } 2725 2726 if (ret) { 2727 spin_lock_irq(&id_priv->lock); 2728 list_del(&mc->list); 2729 spin_unlock_irq(&id_priv->lock); 2730 kfree(mc); 2731 } 2732 return ret; 2733 } 2734 EXPORT_SYMBOL(rdma_join_multicast); 2735 2736 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr) 2737 { 2738 struct rdma_id_private *id_priv; 2739 struct cma_multicast *mc; 2740 2741 id_priv = container_of(id, struct rdma_id_private, id); 2742 spin_lock_irq(&id_priv->lock); 2743 list_for_each_entry(mc, &id_priv->mc_list, list) { 2744 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) { 2745 list_del(&mc->list); 2746 spin_unlock_irq(&id_priv->lock); 2747 2748 if (id->qp) 2749 ib_detach_mcast(id->qp, 2750 &mc->multicast.ib->rec.mgid, 2751 mc->multicast.ib->rec.mlid); 2752 ib_sa_free_multicast(mc->multicast.ib); 2753 kfree(mc); 2754 return; 2755 } 2756 } 2757 spin_unlock_irq(&id_priv->lock); 2758 } 2759 EXPORT_SYMBOL(rdma_leave_multicast); 2760 2761 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv) 2762 { 2763 struct rdma_dev_addr *dev_addr; 2764 struct cma_ndev_work *work; 2765 2766 dev_addr = &id_priv->id.route.addr.dev_addr; 2767 2768 if ((dev_addr->src_dev == ndev) && 2769 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) { 2770 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n", 2771 ndev->name, &id_priv->id); 2772 work = kzalloc(sizeof *work, GFP_KERNEL); 2773 if (!work) 2774 return -ENOMEM; 2775 2776 INIT_WORK(&work->work, cma_ndev_work_handler); 2777 work->id = id_priv; 2778 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE; 2779 atomic_inc(&id_priv->refcount); 2780 queue_work(cma_wq, &work->work); 2781 } 2782 2783 return 0; 2784 } 2785 2786 static int cma_netdev_callback(struct notifier_block *self, unsigned long event, 2787 void *ctx) 2788 { 2789 struct net_device *ndev = (struct net_device *)ctx; 2790 struct cma_device *cma_dev; 2791 struct rdma_id_private *id_priv; 2792 int ret = NOTIFY_DONE; 2793 2794 if (dev_net(ndev) != &init_net) 2795 return NOTIFY_DONE; 2796 2797 if (event != NETDEV_BONDING_FAILOVER) 2798 return NOTIFY_DONE; 2799 2800 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING)) 2801 return NOTIFY_DONE; 2802 2803 mutex_lock(&lock); 2804 list_for_each_entry(cma_dev, &dev_list, list) 2805 list_for_each_entry(id_priv, &cma_dev->id_list, list) { 2806 ret = cma_netdev_change(ndev, id_priv); 2807 if (ret) 2808 goto out; 2809 } 2810 2811 out: 2812 mutex_unlock(&lock); 2813 return ret; 2814 } 2815 2816 static struct notifier_block cma_nb = { 2817 .notifier_call = cma_netdev_callback 2818 }; 2819 2820 static void cma_add_one(struct ib_device *device) 2821 { 2822 struct cma_device *cma_dev; 2823 struct rdma_id_private *id_priv; 2824 2825 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL); 2826 if (!cma_dev) 2827 return; 2828 2829 cma_dev->device = device; 2830 2831 init_completion(&cma_dev->comp); 2832 atomic_set(&cma_dev->refcount, 1); 2833 INIT_LIST_HEAD(&cma_dev->id_list); 2834 ib_set_client_data(device, &cma_client, cma_dev); 2835 2836 mutex_lock(&lock); 2837 list_add_tail(&cma_dev->list, &dev_list); 2838 list_for_each_entry(id_priv, &listen_any_list, list) 2839 cma_listen_on_dev(id_priv, cma_dev); 2840 mutex_unlock(&lock); 2841 } 2842 2843 static int cma_remove_id_dev(struct rdma_id_private *id_priv) 2844 { 2845 struct rdma_cm_event event; 2846 enum cma_state state; 2847 int ret = 0; 2848 2849 /* Record that we want to remove the device */ 2850 state = cma_exch(id_priv, CMA_DEVICE_REMOVAL); 2851 if (state == CMA_DESTROYING) 2852 return 0; 2853 2854 cma_cancel_operation(id_priv, state); 2855 mutex_lock(&id_priv->handler_mutex); 2856 2857 /* Check for destruction from another callback. */ 2858 if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL)) 2859 goto out; 2860 2861 memset(&event, 0, sizeof event); 2862 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL; 2863 ret = id_priv->id.event_handler(&id_priv->id, &event); 2864 out: 2865 mutex_unlock(&id_priv->handler_mutex); 2866 return ret; 2867 } 2868 2869 static void cma_process_remove(struct cma_device *cma_dev) 2870 { 2871 struct rdma_id_private *id_priv; 2872 int ret; 2873 2874 mutex_lock(&lock); 2875 while (!list_empty(&cma_dev->id_list)) { 2876 id_priv = list_entry(cma_dev->id_list.next, 2877 struct rdma_id_private, list); 2878 2879 list_del(&id_priv->listen_list); 2880 list_del_init(&id_priv->list); 2881 atomic_inc(&id_priv->refcount); 2882 mutex_unlock(&lock); 2883 2884 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv); 2885 cma_deref_id(id_priv); 2886 if (ret) 2887 rdma_destroy_id(&id_priv->id); 2888 2889 mutex_lock(&lock); 2890 } 2891 mutex_unlock(&lock); 2892 2893 cma_deref_dev(cma_dev); 2894 wait_for_completion(&cma_dev->comp); 2895 } 2896 2897 static void cma_remove_one(struct ib_device *device) 2898 { 2899 struct cma_device *cma_dev; 2900 2901 cma_dev = ib_get_client_data(device, &cma_client); 2902 if (!cma_dev) 2903 return; 2904 2905 mutex_lock(&lock); 2906 list_del(&cma_dev->list); 2907 mutex_unlock(&lock); 2908 2909 cma_process_remove(cma_dev); 2910 kfree(cma_dev); 2911 } 2912 2913 static int cma_init(void) 2914 { 2915 int ret, low, high, remaining; 2916 2917 get_random_bytes(&next_port, sizeof next_port); 2918 inet_get_local_port_range(&low, &high); 2919 remaining = (high - low) + 1; 2920 next_port = ((unsigned int) next_port % remaining) + low; 2921 2922 cma_wq = create_singlethread_workqueue("rdma_cm"); 2923 if (!cma_wq) 2924 return -ENOMEM; 2925 2926 ib_sa_register_client(&sa_client); 2927 rdma_addr_register_client(&addr_client); 2928 register_netdevice_notifier(&cma_nb); 2929 2930 ret = ib_register_client(&cma_client); 2931 if (ret) 2932 goto err; 2933 return 0; 2934 2935 err: 2936 unregister_netdevice_notifier(&cma_nb); 2937 rdma_addr_unregister_client(&addr_client); 2938 ib_sa_unregister_client(&sa_client); 2939 destroy_workqueue(cma_wq); 2940 return ret; 2941 } 2942 2943 static void cma_cleanup(void) 2944 { 2945 ib_unregister_client(&cma_client); 2946 unregister_netdevice_notifier(&cma_nb); 2947 rdma_addr_unregister_client(&addr_client); 2948 ib_sa_unregister_client(&sa_client); 2949 destroy_workqueue(cma_wq); 2950 idr_destroy(&sdp_ps); 2951 idr_destroy(&tcp_ps); 2952 idr_destroy(&udp_ps); 2953 idr_destroy(&ipoib_ps); 2954 } 2955 2956 module_init(cma_init); 2957 module_exit(cma_cleanup); 2958