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 #include <linux/slab.h> 44 #include <linux/module.h> 45 #include <net/route.h> 46 47 #include <net/tcp.h> 48 #include <net/ipv6.h> 49 50 #include <rdma/rdma_cm.h> 51 #include <rdma/rdma_cm_ib.h> 52 #include <rdma/rdma_netlink.h> 53 #include <rdma/ib.h> 54 #include <rdma/ib_cache.h> 55 #include <rdma/ib_cm.h> 56 #include <rdma/ib_sa.h> 57 #include <rdma/iw_cm.h> 58 59 MODULE_AUTHOR("Sean Hefty"); 60 MODULE_DESCRIPTION("Generic RDMA CM Agent"); 61 MODULE_LICENSE("Dual BSD/GPL"); 62 63 #define CMA_CM_RESPONSE_TIMEOUT 20 64 #define CMA_MAX_CM_RETRIES 15 65 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24) 66 #define CMA_IBOE_PACKET_LIFETIME 18 67 68 static void cma_add_one(struct ib_device *device); 69 static void cma_remove_one(struct ib_device *device); 70 71 static struct ib_client cma_client = { 72 .name = "cma", 73 .add = cma_add_one, 74 .remove = cma_remove_one 75 }; 76 77 static struct ib_sa_client sa_client; 78 static struct rdma_addr_client addr_client; 79 static LIST_HEAD(dev_list); 80 static LIST_HEAD(listen_any_list); 81 static DEFINE_MUTEX(lock); 82 static struct workqueue_struct *cma_wq; 83 static DEFINE_IDR(tcp_ps); 84 static DEFINE_IDR(udp_ps); 85 static DEFINE_IDR(ipoib_ps); 86 static DEFINE_IDR(ib_ps); 87 88 struct cma_device { 89 struct list_head list; 90 struct ib_device *device; 91 struct completion comp; 92 atomic_t refcount; 93 struct list_head id_list; 94 }; 95 96 struct rdma_bind_list { 97 struct idr *ps; 98 struct hlist_head owners; 99 unsigned short port; 100 }; 101 102 enum { 103 CMA_OPTION_AFONLY, 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; /* listen_any_list or cma_device.list */ 118 struct list_head listen_list; /* per device listens */ 119 struct cma_device *cma_dev; 120 struct list_head mc_list; 121 122 int internal_id; 123 enum rdma_cm_state state; 124 spinlock_t lock; 125 struct mutex qp_mutex; 126 127 struct completion comp; 128 atomic_t refcount; 129 struct mutex handler_mutex; 130 131 int backlog; 132 int timeout_ms; 133 struct ib_sa_query *query; 134 int query_id; 135 union { 136 struct ib_cm_id *ib; 137 struct iw_cm_id *iw; 138 } cm_id; 139 140 u32 seq_num; 141 u32 qkey; 142 u32 qp_num; 143 pid_t owner; 144 u32 options; 145 u8 srq; 146 u8 tos; 147 u8 reuseaddr; 148 u8 afonly; 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_storage addr; 159 struct kref mcref; 160 }; 161 162 struct cma_work { 163 struct work_struct work; 164 struct rdma_id_private *id; 165 enum rdma_cm_state old_state; 166 enum rdma_cm_state new_state; 167 struct rdma_cm_event event; 168 }; 169 170 struct cma_ndev_work { 171 struct work_struct work; 172 struct rdma_id_private *id; 173 struct rdma_cm_event event; 174 }; 175 176 struct iboe_mcast_work { 177 struct work_struct work; 178 struct rdma_id_private *id; 179 struct cma_multicast *mc; 180 }; 181 182 union cma_ip_addr { 183 struct in6_addr ip6; 184 struct { 185 __be32 pad[3]; 186 __be32 addr; 187 } ip4; 188 }; 189 190 struct cma_hdr { 191 u8 cma_version; 192 u8 ip_version; /* IP version: 7:4 */ 193 __be16 port; 194 union cma_ip_addr src_addr; 195 union cma_ip_addr dst_addr; 196 }; 197 198 #define CMA_VERSION 0x00 199 200 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_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 rdma_cm_state comp, enum rdma_cm_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 rdma_cm_state cma_exch(struct rdma_id_private *id_priv, 225 enum rdma_cm_state exch) 226 { 227 unsigned long flags; 228 enum rdma_cm_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 void cma_attach_to_dev(struct rdma_id_private *id_priv, 248 struct cma_device *cma_dev) 249 { 250 atomic_inc(&cma_dev->refcount); 251 id_priv->cma_dev = cma_dev; 252 id_priv->id.device = cma_dev->device; 253 id_priv->id.route.addr.dev_addr.transport = 254 rdma_node_get_transport(cma_dev->device->node_type); 255 list_add_tail(&id_priv->list, &cma_dev->id_list); 256 } 257 258 static inline void cma_deref_dev(struct cma_device *cma_dev) 259 { 260 if (atomic_dec_and_test(&cma_dev->refcount)) 261 complete(&cma_dev->comp); 262 } 263 264 static inline void release_mc(struct kref *kref) 265 { 266 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref); 267 268 kfree(mc->multicast.ib); 269 kfree(mc); 270 } 271 272 static void cma_release_dev(struct rdma_id_private *id_priv) 273 { 274 mutex_lock(&lock); 275 list_del(&id_priv->list); 276 cma_deref_dev(id_priv->cma_dev); 277 id_priv->cma_dev = NULL; 278 mutex_unlock(&lock); 279 } 280 281 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv) 282 { 283 return (struct sockaddr *) &id_priv->id.route.addr.src_addr; 284 } 285 286 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv) 287 { 288 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr; 289 } 290 291 static inline unsigned short cma_family(struct rdma_id_private *id_priv) 292 { 293 return id_priv->id.route.addr.src_addr.ss_family; 294 } 295 296 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey) 297 { 298 struct ib_sa_mcmember_rec rec; 299 int ret = 0; 300 301 if (id_priv->qkey) { 302 if (qkey && id_priv->qkey != qkey) 303 return -EINVAL; 304 return 0; 305 } 306 307 if (qkey) { 308 id_priv->qkey = qkey; 309 return 0; 310 } 311 312 switch (id_priv->id.ps) { 313 case RDMA_PS_UDP: 314 case RDMA_PS_IB: 315 id_priv->qkey = RDMA_UDP_QKEY; 316 break; 317 case RDMA_PS_IPOIB: 318 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid); 319 ret = ib_sa_get_mcmember_rec(id_priv->id.device, 320 id_priv->id.port_num, &rec.mgid, 321 &rec); 322 if (!ret) 323 id_priv->qkey = be32_to_cpu(rec.qkey); 324 break; 325 default: 326 break; 327 } 328 return ret; 329 } 330 331 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr) 332 { 333 dev_addr->dev_type = ARPHRD_INFINIBAND; 334 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr); 335 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey)); 336 } 337 338 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr) 339 { 340 int ret; 341 342 if (addr->sa_family != AF_IB) { 343 ret = rdma_translate_ip(addr, dev_addr, NULL); 344 } else { 345 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr); 346 ret = 0; 347 } 348 349 return ret; 350 } 351 352 static int cma_acquire_dev(struct rdma_id_private *id_priv, 353 struct rdma_id_private *listen_id_priv) 354 { 355 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 356 struct cma_device *cma_dev; 357 union ib_gid gid, iboe_gid; 358 int ret = -ENODEV; 359 u8 port, found_port; 360 enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ? 361 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET; 362 363 if (dev_ll != IB_LINK_LAYER_INFINIBAND && 364 id_priv->id.ps == RDMA_PS_IPOIB) 365 return -EINVAL; 366 367 mutex_lock(&lock); 368 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr, 369 &iboe_gid); 370 371 memcpy(&gid, dev_addr->src_dev_addr + 372 rdma_addr_gid_offset(dev_addr), sizeof gid); 373 if (listen_id_priv && 374 rdma_port_get_link_layer(listen_id_priv->id.device, 375 listen_id_priv->id.port_num) == dev_ll) { 376 cma_dev = listen_id_priv->cma_dev; 377 port = listen_id_priv->id.port_num; 378 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB && 379 rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET) 380 ret = ib_find_cached_gid(cma_dev->device, &iboe_gid, 381 &found_port, NULL); 382 else 383 ret = ib_find_cached_gid(cma_dev->device, &gid, 384 &found_port, NULL); 385 386 if (!ret && (port == found_port)) { 387 id_priv->id.port_num = found_port; 388 goto out; 389 } 390 } 391 list_for_each_entry(cma_dev, &dev_list, list) { 392 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) { 393 if (listen_id_priv && 394 listen_id_priv->cma_dev == cma_dev && 395 listen_id_priv->id.port_num == port) 396 continue; 397 if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) { 398 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB && 399 rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET) 400 ret = ib_find_cached_gid(cma_dev->device, &iboe_gid, &found_port, NULL); 401 else 402 ret = ib_find_cached_gid(cma_dev->device, &gid, &found_port, NULL); 403 404 if (!ret && (port == found_port)) { 405 id_priv->id.port_num = found_port; 406 goto out; 407 } 408 } 409 } 410 } 411 412 out: 413 if (!ret) 414 cma_attach_to_dev(id_priv, cma_dev); 415 416 mutex_unlock(&lock); 417 return ret; 418 } 419 420 /* 421 * Select the source IB device and address to reach the destination IB address. 422 */ 423 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv) 424 { 425 struct cma_device *cma_dev, *cur_dev; 426 struct sockaddr_ib *addr; 427 union ib_gid gid, sgid, *dgid; 428 u16 pkey, index; 429 u8 p; 430 int i; 431 432 cma_dev = NULL; 433 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv); 434 dgid = (union ib_gid *) &addr->sib_addr; 435 pkey = ntohs(addr->sib_pkey); 436 437 list_for_each_entry(cur_dev, &dev_list, list) { 438 if (rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB) 439 continue; 440 441 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) { 442 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index)) 443 continue; 444 445 for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i, &gid); i++) { 446 if (!memcmp(&gid, dgid, sizeof(gid))) { 447 cma_dev = cur_dev; 448 sgid = gid; 449 id_priv->id.port_num = p; 450 goto found; 451 } 452 453 if (!cma_dev && (gid.global.subnet_prefix == 454 dgid->global.subnet_prefix)) { 455 cma_dev = cur_dev; 456 sgid = gid; 457 id_priv->id.port_num = p; 458 } 459 } 460 } 461 } 462 463 if (!cma_dev) 464 return -ENODEV; 465 466 found: 467 cma_attach_to_dev(id_priv, cma_dev); 468 addr = (struct sockaddr_ib *) cma_src_addr(id_priv); 469 memcpy(&addr->sib_addr, &sgid, sizeof sgid); 470 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr); 471 return 0; 472 } 473 474 static void cma_deref_id(struct rdma_id_private *id_priv) 475 { 476 if (atomic_dec_and_test(&id_priv->refcount)) 477 complete(&id_priv->comp); 478 } 479 480 static int cma_disable_callback(struct rdma_id_private *id_priv, 481 enum rdma_cm_state state) 482 { 483 mutex_lock(&id_priv->handler_mutex); 484 if (id_priv->state != state) { 485 mutex_unlock(&id_priv->handler_mutex); 486 return -EINVAL; 487 } 488 return 0; 489 } 490 491 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler, 492 void *context, enum rdma_port_space ps, 493 enum ib_qp_type qp_type) 494 { 495 struct rdma_id_private *id_priv; 496 497 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL); 498 if (!id_priv) 499 return ERR_PTR(-ENOMEM); 500 501 id_priv->owner = task_pid_nr(current); 502 id_priv->state = RDMA_CM_IDLE; 503 id_priv->id.context = context; 504 id_priv->id.event_handler = event_handler; 505 id_priv->id.ps = ps; 506 id_priv->id.qp_type = qp_type; 507 spin_lock_init(&id_priv->lock); 508 mutex_init(&id_priv->qp_mutex); 509 init_completion(&id_priv->comp); 510 atomic_set(&id_priv->refcount, 1); 511 mutex_init(&id_priv->handler_mutex); 512 INIT_LIST_HEAD(&id_priv->listen_list); 513 INIT_LIST_HEAD(&id_priv->mc_list); 514 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num); 515 516 return &id_priv->id; 517 } 518 EXPORT_SYMBOL(rdma_create_id); 519 520 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp) 521 { 522 struct ib_qp_attr qp_attr; 523 int qp_attr_mask, ret; 524 525 qp_attr.qp_state = IB_QPS_INIT; 526 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 527 if (ret) 528 return ret; 529 530 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 531 if (ret) 532 return ret; 533 534 qp_attr.qp_state = IB_QPS_RTR; 535 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE); 536 if (ret) 537 return ret; 538 539 qp_attr.qp_state = IB_QPS_RTS; 540 qp_attr.sq_psn = 0; 541 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN); 542 543 return ret; 544 } 545 546 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp) 547 { 548 struct ib_qp_attr qp_attr; 549 int qp_attr_mask, ret; 550 551 qp_attr.qp_state = IB_QPS_INIT; 552 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 553 if (ret) 554 return ret; 555 556 return ib_modify_qp(qp, &qp_attr, qp_attr_mask); 557 } 558 559 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd, 560 struct ib_qp_init_attr *qp_init_attr) 561 { 562 struct rdma_id_private *id_priv; 563 struct ib_qp *qp; 564 int ret; 565 566 id_priv = container_of(id, struct rdma_id_private, id); 567 if (id->device != pd->device) 568 return -EINVAL; 569 570 qp = ib_create_qp(pd, qp_init_attr); 571 if (IS_ERR(qp)) 572 return PTR_ERR(qp); 573 574 if (id->qp_type == IB_QPT_UD) 575 ret = cma_init_ud_qp(id_priv, qp); 576 else 577 ret = cma_init_conn_qp(id_priv, qp); 578 if (ret) 579 goto err; 580 581 id->qp = qp; 582 id_priv->qp_num = qp->qp_num; 583 id_priv->srq = (qp->srq != NULL); 584 return 0; 585 err: 586 ib_destroy_qp(qp); 587 return ret; 588 } 589 EXPORT_SYMBOL(rdma_create_qp); 590 591 void rdma_destroy_qp(struct rdma_cm_id *id) 592 { 593 struct rdma_id_private *id_priv; 594 595 id_priv = container_of(id, struct rdma_id_private, id); 596 mutex_lock(&id_priv->qp_mutex); 597 ib_destroy_qp(id_priv->id.qp); 598 id_priv->id.qp = NULL; 599 mutex_unlock(&id_priv->qp_mutex); 600 } 601 EXPORT_SYMBOL(rdma_destroy_qp); 602 603 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv, 604 struct rdma_conn_param *conn_param) 605 { 606 struct ib_qp_attr qp_attr; 607 int qp_attr_mask, ret; 608 union ib_gid sgid; 609 610 mutex_lock(&id_priv->qp_mutex); 611 if (!id_priv->id.qp) { 612 ret = 0; 613 goto out; 614 } 615 616 /* Need to update QP attributes from default values. */ 617 qp_attr.qp_state = IB_QPS_INIT; 618 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 619 if (ret) 620 goto out; 621 622 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 623 if (ret) 624 goto out; 625 626 qp_attr.qp_state = IB_QPS_RTR; 627 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 628 if (ret) 629 goto out; 630 631 ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num, 632 qp_attr.ah_attr.grh.sgid_index, &sgid); 633 if (ret) 634 goto out; 635 636 if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) 637 == RDMA_TRANSPORT_IB && 638 rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) 639 == IB_LINK_LAYER_ETHERNET) { 640 ret = rdma_addr_find_smac_by_sgid(&sgid, qp_attr.smac, NULL); 641 642 if (ret) 643 goto out; 644 } 645 if (conn_param) 646 qp_attr.max_dest_rd_atomic = conn_param->responder_resources; 647 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 648 out: 649 mutex_unlock(&id_priv->qp_mutex); 650 return ret; 651 } 652 653 static int cma_modify_qp_rts(struct rdma_id_private *id_priv, 654 struct rdma_conn_param *conn_param) 655 { 656 struct ib_qp_attr qp_attr; 657 int qp_attr_mask, ret; 658 659 mutex_lock(&id_priv->qp_mutex); 660 if (!id_priv->id.qp) { 661 ret = 0; 662 goto out; 663 } 664 665 qp_attr.qp_state = IB_QPS_RTS; 666 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 667 if (ret) 668 goto out; 669 670 if (conn_param) 671 qp_attr.max_rd_atomic = conn_param->initiator_depth; 672 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 673 out: 674 mutex_unlock(&id_priv->qp_mutex); 675 return ret; 676 } 677 678 static int cma_modify_qp_err(struct rdma_id_private *id_priv) 679 { 680 struct ib_qp_attr qp_attr; 681 int ret; 682 683 mutex_lock(&id_priv->qp_mutex); 684 if (!id_priv->id.qp) { 685 ret = 0; 686 goto out; 687 } 688 689 qp_attr.qp_state = IB_QPS_ERR; 690 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE); 691 out: 692 mutex_unlock(&id_priv->qp_mutex); 693 return ret; 694 } 695 696 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv, 697 struct ib_qp_attr *qp_attr, int *qp_attr_mask) 698 { 699 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 700 int ret; 701 u16 pkey; 702 703 if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) == 704 IB_LINK_LAYER_INFINIBAND) 705 pkey = ib_addr_get_pkey(dev_addr); 706 else 707 pkey = 0xffff; 708 709 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num, 710 pkey, &qp_attr->pkey_index); 711 if (ret) 712 return ret; 713 714 qp_attr->port_num = id_priv->id.port_num; 715 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT; 716 717 if (id_priv->id.qp_type == IB_QPT_UD) { 718 ret = cma_set_qkey(id_priv, 0); 719 if (ret) 720 return ret; 721 722 qp_attr->qkey = id_priv->qkey; 723 *qp_attr_mask |= IB_QP_QKEY; 724 } else { 725 qp_attr->qp_access_flags = 0; 726 *qp_attr_mask |= IB_QP_ACCESS_FLAGS; 727 } 728 return 0; 729 } 730 731 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr, 732 int *qp_attr_mask) 733 { 734 struct rdma_id_private *id_priv; 735 int ret = 0; 736 737 id_priv = container_of(id, struct rdma_id_private, id); 738 switch (rdma_node_get_transport(id_priv->id.device->node_type)) { 739 case RDMA_TRANSPORT_IB: 740 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD)) 741 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask); 742 else 743 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr, 744 qp_attr_mask); 745 746 if (qp_attr->qp_state == IB_QPS_RTR) 747 qp_attr->rq_psn = id_priv->seq_num; 748 break; 749 case RDMA_TRANSPORT_IWARP: 750 if (!id_priv->cm_id.iw) { 751 qp_attr->qp_access_flags = 0; 752 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS; 753 } else 754 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr, 755 qp_attr_mask); 756 break; 757 default: 758 ret = -ENOSYS; 759 break; 760 } 761 762 return ret; 763 } 764 EXPORT_SYMBOL(rdma_init_qp_attr); 765 766 static inline int cma_zero_addr(struct sockaddr *addr) 767 { 768 switch (addr->sa_family) { 769 case AF_INET: 770 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr); 771 case AF_INET6: 772 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr); 773 case AF_IB: 774 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr); 775 default: 776 return 0; 777 } 778 } 779 780 static inline int cma_loopback_addr(struct sockaddr *addr) 781 { 782 switch (addr->sa_family) { 783 case AF_INET: 784 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr); 785 case AF_INET6: 786 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr); 787 case AF_IB: 788 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr); 789 default: 790 return 0; 791 } 792 } 793 794 static inline int cma_any_addr(struct sockaddr *addr) 795 { 796 return cma_zero_addr(addr) || cma_loopback_addr(addr); 797 } 798 799 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst) 800 { 801 if (src->sa_family != dst->sa_family) 802 return -1; 803 804 switch (src->sa_family) { 805 case AF_INET: 806 return ((struct sockaddr_in *) src)->sin_addr.s_addr != 807 ((struct sockaddr_in *) dst)->sin_addr.s_addr; 808 case AF_INET6: 809 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr, 810 &((struct sockaddr_in6 *) dst)->sin6_addr); 811 default: 812 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr, 813 &((struct sockaddr_ib *) dst)->sib_addr); 814 } 815 } 816 817 static __be16 cma_port(struct sockaddr *addr) 818 { 819 struct sockaddr_ib *sib; 820 821 switch (addr->sa_family) { 822 case AF_INET: 823 return ((struct sockaddr_in *) addr)->sin_port; 824 case AF_INET6: 825 return ((struct sockaddr_in6 *) addr)->sin6_port; 826 case AF_IB: 827 sib = (struct sockaddr_ib *) addr; 828 return htons((u16) (be64_to_cpu(sib->sib_sid) & 829 be64_to_cpu(sib->sib_sid_mask))); 830 default: 831 return 0; 832 } 833 } 834 835 static inline int cma_any_port(struct sockaddr *addr) 836 { 837 return !cma_port(addr); 838 } 839 840 static void cma_save_ib_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id, 841 struct ib_sa_path_rec *path) 842 { 843 struct sockaddr_ib *listen_ib, *ib; 844 845 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr; 846 ib = (struct sockaddr_ib *) &id->route.addr.src_addr; 847 ib->sib_family = listen_ib->sib_family; 848 ib->sib_pkey = path->pkey; 849 ib->sib_flowinfo = path->flow_label; 850 memcpy(&ib->sib_addr, &path->sgid, 16); 851 ib->sib_sid = listen_ib->sib_sid; 852 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL); 853 ib->sib_scope_id = listen_ib->sib_scope_id; 854 855 ib = (struct sockaddr_ib *) &id->route.addr.dst_addr; 856 ib->sib_family = listen_ib->sib_family; 857 ib->sib_pkey = path->pkey; 858 ib->sib_flowinfo = path->flow_label; 859 memcpy(&ib->sib_addr, &path->dgid, 16); 860 } 861 862 static void cma_save_ip4_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id, 863 struct cma_hdr *hdr) 864 { 865 struct sockaddr_in *listen4, *ip4; 866 867 listen4 = (struct sockaddr_in *) &listen_id->route.addr.src_addr; 868 ip4 = (struct sockaddr_in *) &id->route.addr.src_addr; 869 ip4->sin_family = listen4->sin_family; 870 ip4->sin_addr.s_addr = hdr->dst_addr.ip4.addr; 871 ip4->sin_port = listen4->sin_port; 872 873 ip4 = (struct sockaddr_in *) &id->route.addr.dst_addr; 874 ip4->sin_family = listen4->sin_family; 875 ip4->sin_addr.s_addr = hdr->src_addr.ip4.addr; 876 ip4->sin_port = hdr->port; 877 } 878 879 static void cma_save_ip6_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id, 880 struct cma_hdr *hdr) 881 { 882 struct sockaddr_in6 *listen6, *ip6; 883 884 listen6 = (struct sockaddr_in6 *) &listen_id->route.addr.src_addr; 885 ip6 = (struct sockaddr_in6 *) &id->route.addr.src_addr; 886 ip6->sin6_family = listen6->sin6_family; 887 ip6->sin6_addr = hdr->dst_addr.ip6; 888 ip6->sin6_port = listen6->sin6_port; 889 890 ip6 = (struct sockaddr_in6 *) &id->route.addr.dst_addr; 891 ip6->sin6_family = listen6->sin6_family; 892 ip6->sin6_addr = hdr->src_addr.ip6; 893 ip6->sin6_port = hdr->port; 894 } 895 896 static int cma_save_net_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id, 897 struct ib_cm_event *ib_event) 898 { 899 struct cma_hdr *hdr; 900 901 if ((listen_id->route.addr.src_addr.ss_family == AF_IB) && 902 (ib_event->event == IB_CM_REQ_RECEIVED)) { 903 cma_save_ib_info(id, listen_id, ib_event->param.req_rcvd.primary_path); 904 return 0; 905 } 906 907 hdr = ib_event->private_data; 908 if (hdr->cma_version != CMA_VERSION) 909 return -EINVAL; 910 911 switch (cma_get_ip_ver(hdr)) { 912 case 4: 913 cma_save_ip4_info(id, listen_id, hdr); 914 break; 915 case 6: 916 cma_save_ip6_info(id, listen_id, hdr); 917 break; 918 default: 919 return -EINVAL; 920 } 921 return 0; 922 } 923 924 static inline int cma_user_data_offset(struct rdma_id_private *id_priv) 925 { 926 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr); 927 } 928 929 static void cma_cancel_route(struct rdma_id_private *id_priv) 930 { 931 switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) { 932 case IB_LINK_LAYER_INFINIBAND: 933 if (id_priv->query) 934 ib_sa_cancel_query(id_priv->query_id, id_priv->query); 935 break; 936 default: 937 break; 938 } 939 } 940 941 static void cma_cancel_listens(struct rdma_id_private *id_priv) 942 { 943 struct rdma_id_private *dev_id_priv; 944 945 /* 946 * Remove from listen_any_list to prevent added devices from spawning 947 * additional listen requests. 948 */ 949 mutex_lock(&lock); 950 list_del(&id_priv->list); 951 952 while (!list_empty(&id_priv->listen_list)) { 953 dev_id_priv = list_entry(id_priv->listen_list.next, 954 struct rdma_id_private, listen_list); 955 /* sync with device removal to avoid duplicate destruction */ 956 list_del_init(&dev_id_priv->list); 957 list_del(&dev_id_priv->listen_list); 958 mutex_unlock(&lock); 959 960 rdma_destroy_id(&dev_id_priv->id); 961 mutex_lock(&lock); 962 } 963 mutex_unlock(&lock); 964 } 965 966 static void cma_cancel_operation(struct rdma_id_private *id_priv, 967 enum rdma_cm_state state) 968 { 969 switch (state) { 970 case RDMA_CM_ADDR_QUERY: 971 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr); 972 break; 973 case RDMA_CM_ROUTE_QUERY: 974 cma_cancel_route(id_priv); 975 break; 976 case RDMA_CM_LISTEN: 977 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev) 978 cma_cancel_listens(id_priv); 979 break; 980 default: 981 break; 982 } 983 } 984 985 static void cma_release_port(struct rdma_id_private *id_priv) 986 { 987 struct rdma_bind_list *bind_list = id_priv->bind_list; 988 989 if (!bind_list) 990 return; 991 992 mutex_lock(&lock); 993 hlist_del(&id_priv->node); 994 if (hlist_empty(&bind_list->owners)) { 995 idr_remove(bind_list->ps, bind_list->port); 996 kfree(bind_list); 997 } 998 mutex_unlock(&lock); 999 } 1000 1001 static void cma_leave_mc_groups(struct rdma_id_private *id_priv) 1002 { 1003 struct cma_multicast *mc; 1004 1005 while (!list_empty(&id_priv->mc_list)) { 1006 mc = container_of(id_priv->mc_list.next, 1007 struct cma_multicast, list); 1008 list_del(&mc->list); 1009 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) { 1010 case IB_LINK_LAYER_INFINIBAND: 1011 ib_sa_free_multicast(mc->multicast.ib); 1012 kfree(mc); 1013 break; 1014 case IB_LINK_LAYER_ETHERNET: 1015 kref_put(&mc->mcref, release_mc); 1016 break; 1017 default: 1018 break; 1019 } 1020 } 1021 } 1022 1023 void rdma_destroy_id(struct rdma_cm_id *id) 1024 { 1025 struct rdma_id_private *id_priv; 1026 enum rdma_cm_state state; 1027 1028 id_priv = container_of(id, struct rdma_id_private, id); 1029 state = cma_exch(id_priv, RDMA_CM_DESTROYING); 1030 cma_cancel_operation(id_priv, state); 1031 1032 /* 1033 * Wait for any active callback to finish. New callbacks will find 1034 * the id_priv state set to destroying and abort. 1035 */ 1036 mutex_lock(&id_priv->handler_mutex); 1037 mutex_unlock(&id_priv->handler_mutex); 1038 1039 if (id_priv->cma_dev) { 1040 switch (rdma_node_get_transport(id_priv->id.device->node_type)) { 1041 case RDMA_TRANSPORT_IB: 1042 if (id_priv->cm_id.ib) 1043 ib_destroy_cm_id(id_priv->cm_id.ib); 1044 break; 1045 case RDMA_TRANSPORT_IWARP: 1046 if (id_priv->cm_id.iw) 1047 iw_destroy_cm_id(id_priv->cm_id.iw); 1048 break; 1049 default: 1050 break; 1051 } 1052 cma_leave_mc_groups(id_priv); 1053 cma_release_dev(id_priv); 1054 } 1055 1056 cma_release_port(id_priv); 1057 cma_deref_id(id_priv); 1058 wait_for_completion(&id_priv->comp); 1059 1060 if (id_priv->internal_id) 1061 cma_deref_id(id_priv->id.context); 1062 1063 kfree(id_priv->id.route.path_rec); 1064 kfree(id_priv); 1065 } 1066 EXPORT_SYMBOL(rdma_destroy_id); 1067 1068 static int cma_rep_recv(struct rdma_id_private *id_priv) 1069 { 1070 int ret; 1071 1072 ret = cma_modify_qp_rtr(id_priv, NULL); 1073 if (ret) 1074 goto reject; 1075 1076 ret = cma_modify_qp_rts(id_priv, NULL); 1077 if (ret) 1078 goto reject; 1079 1080 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0); 1081 if (ret) 1082 goto reject; 1083 1084 return 0; 1085 reject: 1086 cma_modify_qp_err(id_priv); 1087 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED, 1088 NULL, 0, NULL, 0); 1089 return ret; 1090 } 1091 1092 static void cma_set_rep_event_data(struct rdma_cm_event *event, 1093 struct ib_cm_rep_event_param *rep_data, 1094 void *private_data) 1095 { 1096 event->param.conn.private_data = private_data; 1097 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE; 1098 event->param.conn.responder_resources = rep_data->responder_resources; 1099 event->param.conn.initiator_depth = rep_data->initiator_depth; 1100 event->param.conn.flow_control = rep_data->flow_control; 1101 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count; 1102 event->param.conn.srq = rep_data->srq; 1103 event->param.conn.qp_num = rep_data->remote_qpn; 1104 } 1105 1106 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event) 1107 { 1108 struct rdma_id_private *id_priv = cm_id->context; 1109 struct rdma_cm_event event; 1110 int ret = 0; 1111 1112 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT && 1113 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) || 1114 (ib_event->event == IB_CM_TIMEWAIT_EXIT && 1115 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT))) 1116 return 0; 1117 1118 memset(&event, 0, sizeof event); 1119 switch (ib_event->event) { 1120 case IB_CM_REQ_ERROR: 1121 case IB_CM_REP_ERROR: 1122 event.event = RDMA_CM_EVENT_UNREACHABLE; 1123 event.status = -ETIMEDOUT; 1124 break; 1125 case IB_CM_REP_RECEIVED: 1126 if (id_priv->id.qp) { 1127 event.status = cma_rep_recv(id_priv); 1128 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR : 1129 RDMA_CM_EVENT_ESTABLISHED; 1130 } else { 1131 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE; 1132 } 1133 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd, 1134 ib_event->private_data); 1135 break; 1136 case IB_CM_RTU_RECEIVED: 1137 case IB_CM_USER_ESTABLISHED: 1138 event.event = RDMA_CM_EVENT_ESTABLISHED; 1139 break; 1140 case IB_CM_DREQ_ERROR: 1141 event.status = -ETIMEDOUT; /* fall through */ 1142 case IB_CM_DREQ_RECEIVED: 1143 case IB_CM_DREP_RECEIVED: 1144 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT, 1145 RDMA_CM_DISCONNECT)) 1146 goto out; 1147 event.event = RDMA_CM_EVENT_DISCONNECTED; 1148 break; 1149 case IB_CM_TIMEWAIT_EXIT: 1150 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT; 1151 break; 1152 case IB_CM_MRA_RECEIVED: 1153 /* ignore event */ 1154 goto out; 1155 case IB_CM_REJ_RECEIVED: 1156 cma_modify_qp_err(id_priv); 1157 event.status = ib_event->param.rej_rcvd.reason; 1158 event.event = RDMA_CM_EVENT_REJECTED; 1159 event.param.conn.private_data = ib_event->private_data; 1160 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE; 1161 break; 1162 default: 1163 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n", 1164 ib_event->event); 1165 goto out; 1166 } 1167 1168 ret = id_priv->id.event_handler(&id_priv->id, &event); 1169 if (ret) { 1170 /* Destroy the CM ID by returning a non-zero value. */ 1171 id_priv->cm_id.ib = NULL; 1172 cma_exch(id_priv, RDMA_CM_DESTROYING); 1173 mutex_unlock(&id_priv->handler_mutex); 1174 rdma_destroy_id(&id_priv->id); 1175 return ret; 1176 } 1177 out: 1178 mutex_unlock(&id_priv->handler_mutex); 1179 return ret; 1180 } 1181 1182 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id, 1183 struct ib_cm_event *ib_event) 1184 { 1185 struct rdma_id_private *id_priv; 1186 struct rdma_cm_id *id; 1187 struct rdma_route *rt; 1188 int ret; 1189 1190 id = rdma_create_id(listen_id->event_handler, listen_id->context, 1191 listen_id->ps, ib_event->param.req_rcvd.qp_type); 1192 if (IS_ERR(id)) 1193 return NULL; 1194 1195 id_priv = container_of(id, struct rdma_id_private, id); 1196 if (cma_save_net_info(id, listen_id, ib_event)) 1197 goto err; 1198 1199 rt = &id->route; 1200 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1; 1201 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths, 1202 GFP_KERNEL); 1203 if (!rt->path_rec) 1204 goto err; 1205 1206 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path; 1207 if (rt->num_paths == 2) 1208 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path; 1209 1210 if (cma_any_addr(cma_src_addr(id_priv))) { 1211 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND; 1212 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid); 1213 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey)); 1214 } else { 1215 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr); 1216 if (ret) 1217 goto err; 1218 } 1219 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid); 1220 1221 id_priv->state = RDMA_CM_CONNECT; 1222 return id_priv; 1223 1224 err: 1225 rdma_destroy_id(id); 1226 return NULL; 1227 } 1228 1229 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id, 1230 struct ib_cm_event *ib_event) 1231 { 1232 struct rdma_id_private *id_priv; 1233 struct rdma_cm_id *id; 1234 int ret; 1235 1236 id = rdma_create_id(listen_id->event_handler, listen_id->context, 1237 listen_id->ps, IB_QPT_UD); 1238 if (IS_ERR(id)) 1239 return NULL; 1240 1241 id_priv = container_of(id, struct rdma_id_private, id); 1242 if (cma_save_net_info(id, listen_id, ib_event)) 1243 goto err; 1244 1245 if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) { 1246 ret = cma_translate_addr(cma_src_addr(id_priv), &id->route.addr.dev_addr); 1247 if (ret) 1248 goto err; 1249 } 1250 1251 id_priv->state = RDMA_CM_CONNECT; 1252 return id_priv; 1253 err: 1254 rdma_destroy_id(id); 1255 return NULL; 1256 } 1257 1258 static void cma_set_req_event_data(struct rdma_cm_event *event, 1259 struct ib_cm_req_event_param *req_data, 1260 void *private_data, int offset) 1261 { 1262 event->param.conn.private_data = private_data + offset; 1263 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset; 1264 event->param.conn.responder_resources = req_data->responder_resources; 1265 event->param.conn.initiator_depth = req_data->initiator_depth; 1266 event->param.conn.flow_control = req_data->flow_control; 1267 event->param.conn.retry_count = req_data->retry_count; 1268 event->param.conn.rnr_retry_count = req_data->rnr_retry_count; 1269 event->param.conn.srq = req_data->srq; 1270 event->param.conn.qp_num = req_data->remote_qpn; 1271 } 1272 1273 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event) 1274 { 1275 return (((ib_event->event == IB_CM_REQ_RECEIVED) && 1276 (ib_event->param.req_rcvd.qp_type == id->qp_type)) || 1277 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) && 1278 (id->qp_type == IB_QPT_UD)) || 1279 (!id->qp_type)); 1280 } 1281 1282 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event) 1283 { 1284 struct rdma_id_private *listen_id, *conn_id; 1285 struct rdma_cm_event event; 1286 int offset, ret; 1287 1288 listen_id = cm_id->context; 1289 if (!cma_check_req_qp_type(&listen_id->id, ib_event)) 1290 return -EINVAL; 1291 1292 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN)) 1293 return -ECONNABORTED; 1294 1295 memset(&event, 0, sizeof event); 1296 offset = cma_user_data_offset(listen_id); 1297 event.event = RDMA_CM_EVENT_CONNECT_REQUEST; 1298 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) { 1299 conn_id = cma_new_udp_id(&listen_id->id, ib_event); 1300 event.param.ud.private_data = ib_event->private_data + offset; 1301 event.param.ud.private_data_len = 1302 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset; 1303 } else { 1304 conn_id = cma_new_conn_id(&listen_id->id, ib_event); 1305 cma_set_req_event_data(&event, &ib_event->param.req_rcvd, 1306 ib_event->private_data, offset); 1307 } 1308 if (!conn_id) { 1309 ret = -ENOMEM; 1310 goto err1; 1311 } 1312 1313 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING); 1314 ret = cma_acquire_dev(conn_id, listen_id); 1315 if (ret) 1316 goto err2; 1317 1318 conn_id->cm_id.ib = cm_id; 1319 cm_id->context = conn_id; 1320 cm_id->cm_handler = cma_ib_handler; 1321 1322 /* 1323 * Protect against the user destroying conn_id from another thread 1324 * until we're done accessing it. 1325 */ 1326 atomic_inc(&conn_id->refcount); 1327 ret = conn_id->id.event_handler(&conn_id->id, &event); 1328 if (ret) 1329 goto err3; 1330 /* 1331 * Acquire mutex to prevent user executing rdma_destroy_id() 1332 * while we're accessing the cm_id. 1333 */ 1334 mutex_lock(&lock); 1335 if (cma_comp(conn_id, RDMA_CM_CONNECT) && 1336 (conn_id->id.qp_type != IB_QPT_UD)) 1337 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0); 1338 mutex_unlock(&lock); 1339 mutex_unlock(&conn_id->handler_mutex); 1340 mutex_unlock(&listen_id->handler_mutex); 1341 cma_deref_id(conn_id); 1342 return 0; 1343 1344 err3: 1345 cma_deref_id(conn_id); 1346 /* Destroy the CM ID by returning a non-zero value. */ 1347 conn_id->cm_id.ib = NULL; 1348 err2: 1349 cma_exch(conn_id, RDMA_CM_DESTROYING); 1350 mutex_unlock(&conn_id->handler_mutex); 1351 err1: 1352 mutex_unlock(&listen_id->handler_mutex); 1353 if (conn_id) 1354 rdma_destroy_id(&conn_id->id); 1355 return ret; 1356 } 1357 1358 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr) 1359 { 1360 if (addr->sa_family == AF_IB) 1361 return ((struct sockaddr_ib *) addr)->sib_sid; 1362 1363 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr))); 1364 } 1365 EXPORT_SYMBOL(rdma_get_service_id); 1366 1367 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr, 1368 struct ib_cm_compare_data *compare) 1369 { 1370 struct cma_hdr *cma_data, *cma_mask; 1371 __be32 ip4_addr; 1372 struct in6_addr ip6_addr; 1373 1374 memset(compare, 0, sizeof *compare); 1375 cma_data = (void *) compare->data; 1376 cma_mask = (void *) compare->mask; 1377 1378 switch (addr->sa_family) { 1379 case AF_INET: 1380 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr; 1381 cma_set_ip_ver(cma_data, 4); 1382 cma_set_ip_ver(cma_mask, 0xF); 1383 if (!cma_any_addr(addr)) { 1384 cma_data->dst_addr.ip4.addr = ip4_addr; 1385 cma_mask->dst_addr.ip4.addr = htonl(~0); 1386 } 1387 break; 1388 case AF_INET6: 1389 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr; 1390 cma_set_ip_ver(cma_data, 6); 1391 cma_set_ip_ver(cma_mask, 0xF); 1392 if (!cma_any_addr(addr)) { 1393 cma_data->dst_addr.ip6 = ip6_addr; 1394 memset(&cma_mask->dst_addr.ip6, 0xFF, 1395 sizeof cma_mask->dst_addr.ip6); 1396 } 1397 break; 1398 default: 1399 break; 1400 } 1401 } 1402 1403 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event) 1404 { 1405 struct rdma_id_private *id_priv = iw_id->context; 1406 struct rdma_cm_event event; 1407 int ret = 0; 1408 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr; 1409 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr; 1410 1411 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT)) 1412 return 0; 1413 1414 memset(&event, 0, sizeof event); 1415 switch (iw_event->event) { 1416 case IW_CM_EVENT_CLOSE: 1417 event.event = RDMA_CM_EVENT_DISCONNECTED; 1418 break; 1419 case IW_CM_EVENT_CONNECT_REPLY: 1420 memcpy(cma_src_addr(id_priv), laddr, 1421 rdma_addr_size(laddr)); 1422 memcpy(cma_dst_addr(id_priv), raddr, 1423 rdma_addr_size(raddr)); 1424 switch (iw_event->status) { 1425 case 0: 1426 event.event = RDMA_CM_EVENT_ESTABLISHED; 1427 event.param.conn.initiator_depth = iw_event->ird; 1428 event.param.conn.responder_resources = iw_event->ord; 1429 break; 1430 case -ECONNRESET: 1431 case -ECONNREFUSED: 1432 event.event = RDMA_CM_EVENT_REJECTED; 1433 break; 1434 case -ETIMEDOUT: 1435 event.event = RDMA_CM_EVENT_UNREACHABLE; 1436 break; 1437 default: 1438 event.event = RDMA_CM_EVENT_CONNECT_ERROR; 1439 break; 1440 } 1441 break; 1442 case IW_CM_EVENT_ESTABLISHED: 1443 event.event = RDMA_CM_EVENT_ESTABLISHED; 1444 event.param.conn.initiator_depth = iw_event->ird; 1445 event.param.conn.responder_resources = iw_event->ord; 1446 break; 1447 default: 1448 BUG_ON(1); 1449 } 1450 1451 event.status = iw_event->status; 1452 event.param.conn.private_data = iw_event->private_data; 1453 event.param.conn.private_data_len = iw_event->private_data_len; 1454 ret = id_priv->id.event_handler(&id_priv->id, &event); 1455 if (ret) { 1456 /* Destroy the CM ID by returning a non-zero value. */ 1457 id_priv->cm_id.iw = NULL; 1458 cma_exch(id_priv, RDMA_CM_DESTROYING); 1459 mutex_unlock(&id_priv->handler_mutex); 1460 rdma_destroy_id(&id_priv->id); 1461 return ret; 1462 } 1463 1464 mutex_unlock(&id_priv->handler_mutex); 1465 return ret; 1466 } 1467 1468 static int iw_conn_req_handler(struct iw_cm_id *cm_id, 1469 struct iw_cm_event *iw_event) 1470 { 1471 struct rdma_cm_id *new_cm_id; 1472 struct rdma_id_private *listen_id, *conn_id; 1473 struct rdma_cm_event event; 1474 int ret; 1475 struct ib_device_attr attr; 1476 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr; 1477 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr; 1478 1479 listen_id = cm_id->context; 1480 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN)) 1481 return -ECONNABORTED; 1482 1483 /* Create a new RDMA id for the new IW CM ID */ 1484 new_cm_id = rdma_create_id(listen_id->id.event_handler, 1485 listen_id->id.context, 1486 RDMA_PS_TCP, IB_QPT_RC); 1487 if (IS_ERR(new_cm_id)) { 1488 ret = -ENOMEM; 1489 goto out; 1490 } 1491 conn_id = container_of(new_cm_id, struct rdma_id_private, id); 1492 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING); 1493 conn_id->state = RDMA_CM_CONNECT; 1494 1495 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL); 1496 if (ret) { 1497 mutex_unlock(&conn_id->handler_mutex); 1498 rdma_destroy_id(new_cm_id); 1499 goto out; 1500 } 1501 1502 ret = cma_acquire_dev(conn_id, listen_id); 1503 if (ret) { 1504 mutex_unlock(&conn_id->handler_mutex); 1505 rdma_destroy_id(new_cm_id); 1506 goto out; 1507 } 1508 1509 conn_id->cm_id.iw = cm_id; 1510 cm_id->context = conn_id; 1511 cm_id->cm_handler = cma_iw_handler; 1512 1513 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr)); 1514 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr)); 1515 1516 ret = ib_query_device(conn_id->id.device, &attr); 1517 if (ret) { 1518 mutex_unlock(&conn_id->handler_mutex); 1519 rdma_destroy_id(new_cm_id); 1520 goto out; 1521 } 1522 1523 memset(&event, 0, sizeof event); 1524 event.event = RDMA_CM_EVENT_CONNECT_REQUEST; 1525 event.param.conn.private_data = iw_event->private_data; 1526 event.param.conn.private_data_len = iw_event->private_data_len; 1527 event.param.conn.initiator_depth = iw_event->ird; 1528 event.param.conn.responder_resources = iw_event->ord; 1529 1530 /* 1531 * Protect against the user destroying conn_id from another thread 1532 * until we're done accessing it. 1533 */ 1534 atomic_inc(&conn_id->refcount); 1535 ret = conn_id->id.event_handler(&conn_id->id, &event); 1536 if (ret) { 1537 /* User wants to destroy the CM ID */ 1538 conn_id->cm_id.iw = NULL; 1539 cma_exch(conn_id, RDMA_CM_DESTROYING); 1540 mutex_unlock(&conn_id->handler_mutex); 1541 cma_deref_id(conn_id); 1542 rdma_destroy_id(&conn_id->id); 1543 goto out; 1544 } 1545 1546 mutex_unlock(&conn_id->handler_mutex); 1547 cma_deref_id(conn_id); 1548 1549 out: 1550 mutex_unlock(&listen_id->handler_mutex); 1551 return ret; 1552 } 1553 1554 static int cma_ib_listen(struct rdma_id_private *id_priv) 1555 { 1556 struct ib_cm_compare_data compare_data; 1557 struct sockaddr *addr; 1558 struct ib_cm_id *id; 1559 __be64 svc_id; 1560 int ret; 1561 1562 id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv); 1563 if (IS_ERR(id)) 1564 return PTR_ERR(id); 1565 1566 id_priv->cm_id.ib = id; 1567 1568 addr = cma_src_addr(id_priv); 1569 svc_id = rdma_get_service_id(&id_priv->id, addr); 1570 if (cma_any_addr(addr) && !id_priv->afonly) 1571 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL); 1572 else { 1573 cma_set_compare_data(id_priv->id.ps, addr, &compare_data); 1574 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data); 1575 } 1576 1577 if (ret) { 1578 ib_destroy_cm_id(id_priv->cm_id.ib); 1579 id_priv->cm_id.ib = NULL; 1580 } 1581 1582 return ret; 1583 } 1584 1585 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog) 1586 { 1587 int ret; 1588 struct iw_cm_id *id; 1589 1590 id = iw_create_cm_id(id_priv->id.device, 1591 iw_conn_req_handler, 1592 id_priv); 1593 if (IS_ERR(id)) 1594 return PTR_ERR(id); 1595 1596 id_priv->cm_id.iw = id; 1597 1598 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv), 1599 rdma_addr_size(cma_src_addr(id_priv))); 1600 1601 ret = iw_cm_listen(id_priv->cm_id.iw, backlog); 1602 1603 if (ret) { 1604 iw_destroy_cm_id(id_priv->cm_id.iw); 1605 id_priv->cm_id.iw = NULL; 1606 } 1607 1608 return ret; 1609 } 1610 1611 static int cma_listen_handler(struct rdma_cm_id *id, 1612 struct rdma_cm_event *event) 1613 { 1614 struct rdma_id_private *id_priv = id->context; 1615 1616 id->context = id_priv->id.context; 1617 id->event_handler = id_priv->id.event_handler; 1618 return id_priv->id.event_handler(id, event); 1619 } 1620 1621 static void cma_listen_on_dev(struct rdma_id_private *id_priv, 1622 struct cma_device *cma_dev) 1623 { 1624 struct rdma_id_private *dev_id_priv; 1625 struct rdma_cm_id *id; 1626 int ret; 1627 1628 if (cma_family(id_priv) == AF_IB && 1629 rdma_node_get_transport(cma_dev->device->node_type) != RDMA_TRANSPORT_IB) 1630 return; 1631 1632 id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps, 1633 id_priv->id.qp_type); 1634 if (IS_ERR(id)) 1635 return; 1636 1637 dev_id_priv = container_of(id, struct rdma_id_private, id); 1638 1639 dev_id_priv->state = RDMA_CM_ADDR_BOUND; 1640 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv), 1641 rdma_addr_size(cma_src_addr(id_priv))); 1642 1643 cma_attach_to_dev(dev_id_priv, cma_dev); 1644 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list); 1645 atomic_inc(&id_priv->refcount); 1646 dev_id_priv->internal_id = 1; 1647 dev_id_priv->afonly = id_priv->afonly; 1648 1649 ret = rdma_listen(id, id_priv->backlog); 1650 if (ret) 1651 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, " 1652 "listening on device %s\n", ret, cma_dev->device->name); 1653 } 1654 1655 static void cma_listen_on_all(struct rdma_id_private *id_priv) 1656 { 1657 struct cma_device *cma_dev; 1658 1659 mutex_lock(&lock); 1660 list_add_tail(&id_priv->list, &listen_any_list); 1661 list_for_each_entry(cma_dev, &dev_list, list) 1662 cma_listen_on_dev(id_priv, cma_dev); 1663 mutex_unlock(&lock); 1664 } 1665 1666 void rdma_set_service_type(struct rdma_cm_id *id, int tos) 1667 { 1668 struct rdma_id_private *id_priv; 1669 1670 id_priv = container_of(id, struct rdma_id_private, id); 1671 id_priv->tos = (u8) tos; 1672 } 1673 EXPORT_SYMBOL(rdma_set_service_type); 1674 1675 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec, 1676 void *context) 1677 { 1678 struct cma_work *work = context; 1679 struct rdma_route *route; 1680 1681 route = &work->id->id.route; 1682 1683 if (!status) { 1684 route->num_paths = 1; 1685 *route->path_rec = *path_rec; 1686 } else { 1687 work->old_state = RDMA_CM_ROUTE_QUERY; 1688 work->new_state = RDMA_CM_ADDR_RESOLVED; 1689 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR; 1690 work->event.status = status; 1691 } 1692 1693 queue_work(cma_wq, &work->work); 1694 } 1695 1696 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms, 1697 struct cma_work *work) 1698 { 1699 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 1700 struct ib_sa_path_rec path_rec; 1701 ib_sa_comp_mask comp_mask; 1702 struct sockaddr_in6 *sin6; 1703 struct sockaddr_ib *sib; 1704 1705 memset(&path_rec, 0, sizeof path_rec); 1706 rdma_addr_get_sgid(dev_addr, &path_rec.sgid); 1707 rdma_addr_get_dgid(dev_addr, &path_rec.dgid); 1708 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr)); 1709 path_rec.numb_path = 1; 1710 path_rec.reversible = 1; 1711 path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv)); 1712 1713 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID | 1714 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH | 1715 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID; 1716 1717 switch (cma_family(id_priv)) { 1718 case AF_INET: 1719 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos); 1720 comp_mask |= IB_SA_PATH_REC_QOS_CLASS; 1721 break; 1722 case AF_INET6: 1723 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv); 1724 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20); 1725 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS; 1726 break; 1727 case AF_IB: 1728 sib = (struct sockaddr_ib *) cma_src_addr(id_priv); 1729 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20); 1730 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS; 1731 break; 1732 } 1733 1734 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device, 1735 id_priv->id.port_num, &path_rec, 1736 comp_mask, timeout_ms, 1737 GFP_KERNEL, cma_query_handler, 1738 work, &id_priv->query); 1739 1740 return (id_priv->query_id < 0) ? id_priv->query_id : 0; 1741 } 1742 1743 static void cma_work_handler(struct work_struct *_work) 1744 { 1745 struct cma_work *work = container_of(_work, struct cma_work, work); 1746 struct rdma_id_private *id_priv = work->id; 1747 int destroy = 0; 1748 1749 mutex_lock(&id_priv->handler_mutex); 1750 if (!cma_comp_exch(id_priv, work->old_state, work->new_state)) 1751 goto out; 1752 1753 if (id_priv->id.event_handler(&id_priv->id, &work->event)) { 1754 cma_exch(id_priv, RDMA_CM_DESTROYING); 1755 destroy = 1; 1756 } 1757 out: 1758 mutex_unlock(&id_priv->handler_mutex); 1759 cma_deref_id(id_priv); 1760 if (destroy) 1761 rdma_destroy_id(&id_priv->id); 1762 kfree(work); 1763 } 1764 1765 static void cma_ndev_work_handler(struct work_struct *_work) 1766 { 1767 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work); 1768 struct rdma_id_private *id_priv = work->id; 1769 int destroy = 0; 1770 1771 mutex_lock(&id_priv->handler_mutex); 1772 if (id_priv->state == RDMA_CM_DESTROYING || 1773 id_priv->state == RDMA_CM_DEVICE_REMOVAL) 1774 goto out; 1775 1776 if (id_priv->id.event_handler(&id_priv->id, &work->event)) { 1777 cma_exch(id_priv, RDMA_CM_DESTROYING); 1778 destroy = 1; 1779 } 1780 1781 out: 1782 mutex_unlock(&id_priv->handler_mutex); 1783 cma_deref_id(id_priv); 1784 if (destroy) 1785 rdma_destroy_id(&id_priv->id); 1786 kfree(work); 1787 } 1788 1789 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms) 1790 { 1791 struct rdma_route *route = &id_priv->id.route; 1792 struct cma_work *work; 1793 int ret; 1794 1795 work = kzalloc(sizeof *work, GFP_KERNEL); 1796 if (!work) 1797 return -ENOMEM; 1798 1799 work->id = id_priv; 1800 INIT_WORK(&work->work, cma_work_handler); 1801 work->old_state = RDMA_CM_ROUTE_QUERY; 1802 work->new_state = RDMA_CM_ROUTE_RESOLVED; 1803 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 1804 1805 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL); 1806 if (!route->path_rec) { 1807 ret = -ENOMEM; 1808 goto err1; 1809 } 1810 1811 ret = cma_query_ib_route(id_priv, timeout_ms, work); 1812 if (ret) 1813 goto err2; 1814 1815 return 0; 1816 err2: 1817 kfree(route->path_rec); 1818 route->path_rec = NULL; 1819 err1: 1820 kfree(work); 1821 return ret; 1822 } 1823 1824 int rdma_set_ib_paths(struct rdma_cm_id *id, 1825 struct ib_sa_path_rec *path_rec, int num_paths) 1826 { 1827 struct rdma_id_private *id_priv; 1828 int ret; 1829 1830 id_priv = container_of(id, struct rdma_id_private, id); 1831 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, 1832 RDMA_CM_ROUTE_RESOLVED)) 1833 return -EINVAL; 1834 1835 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths, 1836 GFP_KERNEL); 1837 if (!id->route.path_rec) { 1838 ret = -ENOMEM; 1839 goto err; 1840 } 1841 1842 id->route.num_paths = num_paths; 1843 return 0; 1844 err: 1845 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED); 1846 return ret; 1847 } 1848 EXPORT_SYMBOL(rdma_set_ib_paths); 1849 1850 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms) 1851 { 1852 struct cma_work *work; 1853 1854 work = kzalloc(sizeof *work, GFP_KERNEL); 1855 if (!work) 1856 return -ENOMEM; 1857 1858 work->id = id_priv; 1859 INIT_WORK(&work->work, cma_work_handler); 1860 work->old_state = RDMA_CM_ROUTE_QUERY; 1861 work->new_state = RDMA_CM_ROUTE_RESOLVED; 1862 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 1863 queue_work(cma_wq, &work->work); 1864 return 0; 1865 } 1866 1867 static int iboe_tos_to_sl(struct net_device *ndev, int tos) 1868 { 1869 int prio; 1870 struct net_device *dev; 1871 1872 prio = rt_tos2priority(tos); 1873 dev = ndev->priv_flags & IFF_802_1Q_VLAN ? 1874 vlan_dev_real_dev(ndev) : ndev; 1875 1876 if (dev->num_tc) 1877 return netdev_get_prio_tc_map(dev, prio); 1878 1879 #if IS_ENABLED(CONFIG_VLAN_8021Q) 1880 if (ndev->priv_flags & IFF_802_1Q_VLAN) 1881 return (vlan_dev_get_egress_qos_mask(ndev, prio) & 1882 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT; 1883 #endif 1884 return 0; 1885 } 1886 1887 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv) 1888 { 1889 struct rdma_route *route = &id_priv->id.route; 1890 struct rdma_addr *addr = &route->addr; 1891 struct cma_work *work; 1892 int ret; 1893 struct net_device *ndev = NULL; 1894 1895 1896 work = kzalloc(sizeof *work, GFP_KERNEL); 1897 if (!work) 1898 return -ENOMEM; 1899 1900 work->id = id_priv; 1901 INIT_WORK(&work->work, cma_work_handler); 1902 1903 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL); 1904 if (!route->path_rec) { 1905 ret = -ENOMEM; 1906 goto err1; 1907 } 1908 1909 route->num_paths = 1; 1910 1911 if (addr->dev_addr.bound_dev_if) 1912 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if); 1913 if (!ndev) { 1914 ret = -ENODEV; 1915 goto err2; 1916 } 1917 1918 route->path_rec->vlan_id = rdma_vlan_dev_vlan_id(ndev); 1919 memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN); 1920 memcpy(route->path_rec->smac, ndev->dev_addr, ndev->addr_len); 1921 1922 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr, 1923 &route->path_rec->sgid); 1924 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr, 1925 &route->path_rec->dgid); 1926 1927 route->path_rec->hop_limit = 1; 1928 route->path_rec->reversible = 1; 1929 route->path_rec->pkey = cpu_to_be16(0xffff); 1930 route->path_rec->mtu_selector = IB_SA_EQ; 1931 route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos); 1932 route->path_rec->mtu = iboe_get_mtu(ndev->mtu); 1933 route->path_rec->rate_selector = IB_SA_EQ; 1934 route->path_rec->rate = iboe_get_rate(ndev); 1935 dev_put(ndev); 1936 route->path_rec->packet_life_time_selector = IB_SA_EQ; 1937 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME; 1938 if (!route->path_rec->mtu) { 1939 ret = -EINVAL; 1940 goto err2; 1941 } 1942 1943 work->old_state = RDMA_CM_ROUTE_QUERY; 1944 work->new_state = RDMA_CM_ROUTE_RESOLVED; 1945 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 1946 work->event.status = 0; 1947 1948 queue_work(cma_wq, &work->work); 1949 1950 return 0; 1951 1952 err2: 1953 kfree(route->path_rec); 1954 route->path_rec = NULL; 1955 err1: 1956 kfree(work); 1957 return ret; 1958 } 1959 1960 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms) 1961 { 1962 struct rdma_id_private *id_priv; 1963 int ret; 1964 1965 id_priv = container_of(id, struct rdma_id_private, id); 1966 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY)) 1967 return -EINVAL; 1968 1969 atomic_inc(&id_priv->refcount); 1970 switch (rdma_node_get_transport(id->device->node_type)) { 1971 case RDMA_TRANSPORT_IB: 1972 switch (rdma_port_get_link_layer(id->device, id->port_num)) { 1973 case IB_LINK_LAYER_INFINIBAND: 1974 ret = cma_resolve_ib_route(id_priv, timeout_ms); 1975 break; 1976 case IB_LINK_LAYER_ETHERNET: 1977 ret = cma_resolve_iboe_route(id_priv); 1978 break; 1979 default: 1980 ret = -ENOSYS; 1981 } 1982 break; 1983 case RDMA_TRANSPORT_IWARP: 1984 ret = cma_resolve_iw_route(id_priv, timeout_ms); 1985 break; 1986 default: 1987 ret = -ENOSYS; 1988 break; 1989 } 1990 if (ret) 1991 goto err; 1992 1993 return 0; 1994 err: 1995 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED); 1996 cma_deref_id(id_priv); 1997 return ret; 1998 } 1999 EXPORT_SYMBOL(rdma_resolve_route); 2000 2001 static void cma_set_loopback(struct sockaddr *addr) 2002 { 2003 switch (addr->sa_family) { 2004 case AF_INET: 2005 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK); 2006 break; 2007 case AF_INET6: 2008 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr, 2009 0, 0, 0, htonl(1)); 2010 break; 2011 default: 2012 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr, 2013 0, 0, 0, htonl(1)); 2014 break; 2015 } 2016 } 2017 2018 static int cma_bind_loopback(struct rdma_id_private *id_priv) 2019 { 2020 struct cma_device *cma_dev, *cur_dev; 2021 struct ib_port_attr port_attr; 2022 union ib_gid gid; 2023 u16 pkey; 2024 int ret; 2025 u8 p; 2026 2027 cma_dev = NULL; 2028 mutex_lock(&lock); 2029 list_for_each_entry(cur_dev, &dev_list, list) { 2030 if (cma_family(id_priv) == AF_IB && 2031 rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB) 2032 continue; 2033 2034 if (!cma_dev) 2035 cma_dev = cur_dev; 2036 2037 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) { 2038 if (!ib_query_port(cur_dev->device, p, &port_attr) && 2039 port_attr.state == IB_PORT_ACTIVE) { 2040 cma_dev = cur_dev; 2041 goto port_found; 2042 } 2043 } 2044 } 2045 2046 if (!cma_dev) { 2047 ret = -ENODEV; 2048 goto out; 2049 } 2050 2051 p = 1; 2052 2053 port_found: 2054 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid); 2055 if (ret) 2056 goto out; 2057 2058 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey); 2059 if (ret) 2060 goto out; 2061 2062 id_priv->id.route.addr.dev_addr.dev_type = 2063 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ? 2064 ARPHRD_INFINIBAND : ARPHRD_ETHER; 2065 2066 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid); 2067 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey); 2068 id_priv->id.port_num = p; 2069 cma_attach_to_dev(id_priv, cma_dev); 2070 cma_set_loopback(cma_src_addr(id_priv)); 2071 out: 2072 mutex_unlock(&lock); 2073 return ret; 2074 } 2075 2076 static void addr_handler(int status, struct sockaddr *src_addr, 2077 struct rdma_dev_addr *dev_addr, void *context) 2078 { 2079 struct rdma_id_private *id_priv = context; 2080 struct rdma_cm_event event; 2081 2082 memset(&event, 0, sizeof event); 2083 mutex_lock(&id_priv->handler_mutex); 2084 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, 2085 RDMA_CM_ADDR_RESOLVED)) 2086 goto out; 2087 2088 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr)); 2089 if (!status && !id_priv->cma_dev) 2090 status = cma_acquire_dev(id_priv, NULL); 2091 2092 if (status) { 2093 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, 2094 RDMA_CM_ADDR_BOUND)) 2095 goto out; 2096 event.event = RDMA_CM_EVENT_ADDR_ERROR; 2097 event.status = status; 2098 } else 2099 event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 2100 2101 if (id_priv->id.event_handler(&id_priv->id, &event)) { 2102 cma_exch(id_priv, RDMA_CM_DESTROYING); 2103 mutex_unlock(&id_priv->handler_mutex); 2104 cma_deref_id(id_priv); 2105 rdma_destroy_id(&id_priv->id); 2106 return; 2107 } 2108 out: 2109 mutex_unlock(&id_priv->handler_mutex); 2110 cma_deref_id(id_priv); 2111 } 2112 2113 static int cma_resolve_loopback(struct rdma_id_private *id_priv) 2114 { 2115 struct cma_work *work; 2116 union ib_gid gid; 2117 int ret; 2118 2119 work = kzalloc(sizeof *work, GFP_KERNEL); 2120 if (!work) 2121 return -ENOMEM; 2122 2123 if (!id_priv->cma_dev) { 2124 ret = cma_bind_loopback(id_priv); 2125 if (ret) 2126 goto err; 2127 } 2128 2129 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid); 2130 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid); 2131 2132 work->id = id_priv; 2133 INIT_WORK(&work->work, cma_work_handler); 2134 work->old_state = RDMA_CM_ADDR_QUERY; 2135 work->new_state = RDMA_CM_ADDR_RESOLVED; 2136 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 2137 queue_work(cma_wq, &work->work); 2138 return 0; 2139 err: 2140 kfree(work); 2141 return ret; 2142 } 2143 2144 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv) 2145 { 2146 struct cma_work *work; 2147 int ret; 2148 2149 work = kzalloc(sizeof *work, GFP_KERNEL); 2150 if (!work) 2151 return -ENOMEM; 2152 2153 if (!id_priv->cma_dev) { 2154 ret = cma_resolve_ib_dev(id_priv); 2155 if (ret) 2156 goto err; 2157 } 2158 2159 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *) 2160 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr)); 2161 2162 work->id = id_priv; 2163 INIT_WORK(&work->work, cma_work_handler); 2164 work->old_state = RDMA_CM_ADDR_QUERY; 2165 work->new_state = RDMA_CM_ADDR_RESOLVED; 2166 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 2167 queue_work(cma_wq, &work->work); 2168 return 0; 2169 err: 2170 kfree(work); 2171 return ret; 2172 } 2173 2174 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr, 2175 struct sockaddr *dst_addr) 2176 { 2177 if (!src_addr || !src_addr->sa_family) { 2178 src_addr = (struct sockaddr *) &id->route.addr.src_addr; 2179 src_addr->sa_family = dst_addr->sa_family; 2180 if (dst_addr->sa_family == AF_INET6) { 2181 ((struct sockaddr_in6 *) src_addr)->sin6_scope_id = 2182 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id; 2183 } else if (dst_addr->sa_family == AF_IB) { 2184 ((struct sockaddr_ib *) src_addr)->sib_pkey = 2185 ((struct sockaddr_ib *) dst_addr)->sib_pkey; 2186 } 2187 } 2188 return rdma_bind_addr(id, src_addr); 2189 } 2190 2191 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr, 2192 struct sockaddr *dst_addr, int timeout_ms) 2193 { 2194 struct rdma_id_private *id_priv; 2195 int ret; 2196 2197 id_priv = container_of(id, struct rdma_id_private, id); 2198 if (id_priv->state == RDMA_CM_IDLE) { 2199 ret = cma_bind_addr(id, src_addr, dst_addr); 2200 if (ret) 2201 return ret; 2202 } 2203 2204 if (cma_family(id_priv) != dst_addr->sa_family) 2205 return -EINVAL; 2206 2207 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) 2208 return -EINVAL; 2209 2210 atomic_inc(&id_priv->refcount); 2211 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr)); 2212 if (cma_any_addr(dst_addr)) { 2213 ret = cma_resolve_loopback(id_priv); 2214 } else { 2215 if (dst_addr->sa_family == AF_IB) { 2216 ret = cma_resolve_ib_addr(id_priv); 2217 } else { 2218 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv), 2219 dst_addr, &id->route.addr.dev_addr, 2220 timeout_ms, addr_handler, id_priv); 2221 } 2222 } 2223 if (ret) 2224 goto err; 2225 2226 return 0; 2227 err: 2228 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND); 2229 cma_deref_id(id_priv); 2230 return ret; 2231 } 2232 EXPORT_SYMBOL(rdma_resolve_addr); 2233 2234 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse) 2235 { 2236 struct rdma_id_private *id_priv; 2237 unsigned long flags; 2238 int ret; 2239 2240 id_priv = container_of(id, struct rdma_id_private, id); 2241 spin_lock_irqsave(&id_priv->lock, flags); 2242 if (reuse || id_priv->state == RDMA_CM_IDLE) { 2243 id_priv->reuseaddr = reuse; 2244 ret = 0; 2245 } else { 2246 ret = -EINVAL; 2247 } 2248 spin_unlock_irqrestore(&id_priv->lock, flags); 2249 return ret; 2250 } 2251 EXPORT_SYMBOL(rdma_set_reuseaddr); 2252 2253 int rdma_set_afonly(struct rdma_cm_id *id, int afonly) 2254 { 2255 struct rdma_id_private *id_priv; 2256 unsigned long flags; 2257 int ret; 2258 2259 id_priv = container_of(id, struct rdma_id_private, id); 2260 spin_lock_irqsave(&id_priv->lock, flags); 2261 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) { 2262 id_priv->options |= (1 << CMA_OPTION_AFONLY); 2263 id_priv->afonly = afonly; 2264 ret = 0; 2265 } else { 2266 ret = -EINVAL; 2267 } 2268 spin_unlock_irqrestore(&id_priv->lock, flags); 2269 return ret; 2270 } 2271 EXPORT_SYMBOL(rdma_set_afonly); 2272 2273 static void cma_bind_port(struct rdma_bind_list *bind_list, 2274 struct rdma_id_private *id_priv) 2275 { 2276 struct sockaddr *addr; 2277 struct sockaddr_ib *sib; 2278 u64 sid, mask; 2279 __be16 port; 2280 2281 addr = cma_src_addr(id_priv); 2282 port = htons(bind_list->port); 2283 2284 switch (addr->sa_family) { 2285 case AF_INET: 2286 ((struct sockaddr_in *) addr)->sin_port = port; 2287 break; 2288 case AF_INET6: 2289 ((struct sockaddr_in6 *) addr)->sin6_port = port; 2290 break; 2291 case AF_IB: 2292 sib = (struct sockaddr_ib *) addr; 2293 sid = be64_to_cpu(sib->sib_sid); 2294 mask = be64_to_cpu(sib->sib_sid_mask); 2295 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port)); 2296 sib->sib_sid_mask = cpu_to_be64(~0ULL); 2297 break; 2298 } 2299 id_priv->bind_list = bind_list; 2300 hlist_add_head(&id_priv->node, &bind_list->owners); 2301 } 2302 2303 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv, 2304 unsigned short snum) 2305 { 2306 struct rdma_bind_list *bind_list; 2307 int ret; 2308 2309 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL); 2310 if (!bind_list) 2311 return -ENOMEM; 2312 2313 ret = idr_alloc(ps, bind_list, snum, snum + 1, GFP_KERNEL); 2314 if (ret < 0) 2315 goto err; 2316 2317 bind_list->ps = ps; 2318 bind_list->port = (unsigned short)ret; 2319 cma_bind_port(bind_list, id_priv); 2320 return 0; 2321 err: 2322 kfree(bind_list); 2323 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret; 2324 } 2325 2326 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv) 2327 { 2328 static unsigned int last_used_port; 2329 int low, high, remaining; 2330 unsigned int rover; 2331 2332 inet_get_local_port_range(&init_net, &low, &high); 2333 remaining = (high - low) + 1; 2334 rover = prandom_u32() % remaining + low; 2335 retry: 2336 if (last_used_port != rover && 2337 !idr_find(ps, (unsigned short) rover)) { 2338 int ret = cma_alloc_port(ps, id_priv, rover); 2339 /* 2340 * Remember previously used port number in order to avoid 2341 * re-using same port immediately after it is closed. 2342 */ 2343 if (!ret) 2344 last_used_port = rover; 2345 if (ret != -EADDRNOTAVAIL) 2346 return ret; 2347 } 2348 if (--remaining) { 2349 rover++; 2350 if ((rover < low) || (rover > high)) 2351 rover = low; 2352 goto retry; 2353 } 2354 return -EADDRNOTAVAIL; 2355 } 2356 2357 /* 2358 * Check that the requested port is available. This is called when trying to 2359 * bind to a specific port, or when trying to listen on a bound port. In 2360 * the latter case, the provided id_priv may already be on the bind_list, but 2361 * we still need to check that it's okay to start listening. 2362 */ 2363 static int cma_check_port(struct rdma_bind_list *bind_list, 2364 struct rdma_id_private *id_priv, uint8_t reuseaddr) 2365 { 2366 struct rdma_id_private *cur_id; 2367 struct sockaddr *addr, *cur_addr; 2368 2369 addr = cma_src_addr(id_priv); 2370 hlist_for_each_entry(cur_id, &bind_list->owners, node) { 2371 if (id_priv == cur_id) 2372 continue; 2373 2374 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr && 2375 cur_id->reuseaddr) 2376 continue; 2377 2378 cur_addr = cma_src_addr(cur_id); 2379 if (id_priv->afonly && cur_id->afonly && 2380 (addr->sa_family != cur_addr->sa_family)) 2381 continue; 2382 2383 if (cma_any_addr(addr) || cma_any_addr(cur_addr)) 2384 return -EADDRNOTAVAIL; 2385 2386 if (!cma_addr_cmp(addr, cur_addr)) 2387 return -EADDRINUSE; 2388 } 2389 return 0; 2390 } 2391 2392 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv) 2393 { 2394 struct rdma_bind_list *bind_list; 2395 unsigned short snum; 2396 int ret; 2397 2398 snum = ntohs(cma_port(cma_src_addr(id_priv))); 2399 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE)) 2400 return -EACCES; 2401 2402 bind_list = idr_find(ps, snum); 2403 if (!bind_list) { 2404 ret = cma_alloc_port(ps, id_priv, snum); 2405 } else { 2406 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr); 2407 if (!ret) 2408 cma_bind_port(bind_list, id_priv); 2409 } 2410 return ret; 2411 } 2412 2413 static int cma_bind_listen(struct rdma_id_private *id_priv) 2414 { 2415 struct rdma_bind_list *bind_list = id_priv->bind_list; 2416 int ret = 0; 2417 2418 mutex_lock(&lock); 2419 if (bind_list->owners.first->next) 2420 ret = cma_check_port(bind_list, id_priv, 0); 2421 mutex_unlock(&lock); 2422 return ret; 2423 } 2424 2425 static struct idr *cma_select_inet_ps(struct rdma_id_private *id_priv) 2426 { 2427 switch (id_priv->id.ps) { 2428 case RDMA_PS_TCP: 2429 return &tcp_ps; 2430 case RDMA_PS_UDP: 2431 return &udp_ps; 2432 case RDMA_PS_IPOIB: 2433 return &ipoib_ps; 2434 case RDMA_PS_IB: 2435 return &ib_ps; 2436 default: 2437 return NULL; 2438 } 2439 } 2440 2441 static struct idr *cma_select_ib_ps(struct rdma_id_private *id_priv) 2442 { 2443 struct idr *ps = NULL; 2444 struct sockaddr_ib *sib; 2445 u64 sid_ps, mask, sid; 2446 2447 sib = (struct sockaddr_ib *) cma_src_addr(id_priv); 2448 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK; 2449 sid = be64_to_cpu(sib->sib_sid) & mask; 2450 2451 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) { 2452 sid_ps = RDMA_IB_IP_PS_IB; 2453 ps = &ib_ps; 2454 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) && 2455 (sid == (RDMA_IB_IP_PS_TCP & mask))) { 2456 sid_ps = RDMA_IB_IP_PS_TCP; 2457 ps = &tcp_ps; 2458 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) && 2459 (sid == (RDMA_IB_IP_PS_UDP & mask))) { 2460 sid_ps = RDMA_IB_IP_PS_UDP; 2461 ps = &udp_ps; 2462 } 2463 2464 if (ps) { 2465 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib))); 2466 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK | 2467 be64_to_cpu(sib->sib_sid_mask)); 2468 } 2469 return ps; 2470 } 2471 2472 static int cma_get_port(struct rdma_id_private *id_priv) 2473 { 2474 struct idr *ps; 2475 int ret; 2476 2477 if (cma_family(id_priv) != AF_IB) 2478 ps = cma_select_inet_ps(id_priv); 2479 else 2480 ps = cma_select_ib_ps(id_priv); 2481 if (!ps) 2482 return -EPROTONOSUPPORT; 2483 2484 mutex_lock(&lock); 2485 if (cma_any_port(cma_src_addr(id_priv))) 2486 ret = cma_alloc_any_port(ps, id_priv); 2487 else 2488 ret = cma_use_port(ps, id_priv); 2489 mutex_unlock(&lock); 2490 2491 return ret; 2492 } 2493 2494 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr, 2495 struct sockaddr *addr) 2496 { 2497 #if IS_ENABLED(CONFIG_IPV6) 2498 struct sockaddr_in6 *sin6; 2499 2500 if (addr->sa_family != AF_INET6) 2501 return 0; 2502 2503 sin6 = (struct sockaddr_in6 *) addr; 2504 2505 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL)) 2506 return 0; 2507 2508 if (!sin6->sin6_scope_id) 2509 return -EINVAL; 2510 2511 dev_addr->bound_dev_if = sin6->sin6_scope_id; 2512 #endif 2513 return 0; 2514 } 2515 2516 int rdma_listen(struct rdma_cm_id *id, int backlog) 2517 { 2518 struct rdma_id_private *id_priv; 2519 int ret; 2520 2521 id_priv = container_of(id, struct rdma_id_private, id); 2522 if (id_priv->state == RDMA_CM_IDLE) { 2523 id->route.addr.src_addr.ss_family = AF_INET; 2524 ret = rdma_bind_addr(id, cma_src_addr(id_priv)); 2525 if (ret) 2526 return ret; 2527 } 2528 2529 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) 2530 return -EINVAL; 2531 2532 if (id_priv->reuseaddr) { 2533 ret = cma_bind_listen(id_priv); 2534 if (ret) 2535 goto err; 2536 } 2537 2538 id_priv->backlog = backlog; 2539 if (id->device) { 2540 switch (rdma_node_get_transport(id->device->node_type)) { 2541 case RDMA_TRANSPORT_IB: 2542 ret = cma_ib_listen(id_priv); 2543 if (ret) 2544 goto err; 2545 break; 2546 case RDMA_TRANSPORT_IWARP: 2547 ret = cma_iw_listen(id_priv, backlog); 2548 if (ret) 2549 goto err; 2550 break; 2551 default: 2552 ret = -ENOSYS; 2553 goto err; 2554 } 2555 } else 2556 cma_listen_on_all(id_priv); 2557 2558 return 0; 2559 err: 2560 id_priv->backlog = 0; 2561 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND); 2562 return ret; 2563 } 2564 EXPORT_SYMBOL(rdma_listen); 2565 2566 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr) 2567 { 2568 struct rdma_id_private *id_priv; 2569 int ret; 2570 2571 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 && 2572 addr->sa_family != AF_IB) 2573 return -EAFNOSUPPORT; 2574 2575 id_priv = container_of(id, struct rdma_id_private, id); 2576 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND)) 2577 return -EINVAL; 2578 2579 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr); 2580 if (ret) 2581 goto err1; 2582 2583 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr)); 2584 if (!cma_any_addr(addr)) { 2585 ret = cma_translate_addr(addr, &id->route.addr.dev_addr); 2586 if (ret) 2587 goto err1; 2588 2589 ret = cma_acquire_dev(id_priv, NULL); 2590 if (ret) 2591 goto err1; 2592 } 2593 2594 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) { 2595 if (addr->sa_family == AF_INET) 2596 id_priv->afonly = 1; 2597 #if IS_ENABLED(CONFIG_IPV6) 2598 else if (addr->sa_family == AF_INET6) 2599 id_priv->afonly = init_net.ipv6.sysctl.bindv6only; 2600 #endif 2601 } 2602 ret = cma_get_port(id_priv); 2603 if (ret) 2604 goto err2; 2605 2606 return 0; 2607 err2: 2608 if (id_priv->cma_dev) 2609 cma_release_dev(id_priv); 2610 err1: 2611 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE); 2612 return ret; 2613 } 2614 EXPORT_SYMBOL(rdma_bind_addr); 2615 2616 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv) 2617 { 2618 struct cma_hdr *cma_hdr; 2619 2620 cma_hdr = hdr; 2621 cma_hdr->cma_version = CMA_VERSION; 2622 if (cma_family(id_priv) == AF_INET) { 2623 struct sockaddr_in *src4, *dst4; 2624 2625 src4 = (struct sockaddr_in *) cma_src_addr(id_priv); 2626 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv); 2627 2628 cma_set_ip_ver(cma_hdr, 4); 2629 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr; 2630 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr; 2631 cma_hdr->port = src4->sin_port; 2632 } else if (cma_family(id_priv) == AF_INET6) { 2633 struct sockaddr_in6 *src6, *dst6; 2634 2635 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv); 2636 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv); 2637 2638 cma_set_ip_ver(cma_hdr, 6); 2639 cma_hdr->src_addr.ip6 = src6->sin6_addr; 2640 cma_hdr->dst_addr.ip6 = dst6->sin6_addr; 2641 cma_hdr->port = src6->sin6_port; 2642 } 2643 return 0; 2644 } 2645 2646 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id, 2647 struct ib_cm_event *ib_event) 2648 { 2649 struct rdma_id_private *id_priv = cm_id->context; 2650 struct rdma_cm_event event; 2651 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd; 2652 int ret = 0; 2653 2654 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT)) 2655 return 0; 2656 2657 memset(&event, 0, sizeof event); 2658 switch (ib_event->event) { 2659 case IB_CM_SIDR_REQ_ERROR: 2660 event.event = RDMA_CM_EVENT_UNREACHABLE; 2661 event.status = -ETIMEDOUT; 2662 break; 2663 case IB_CM_SIDR_REP_RECEIVED: 2664 event.param.ud.private_data = ib_event->private_data; 2665 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE; 2666 if (rep->status != IB_SIDR_SUCCESS) { 2667 event.event = RDMA_CM_EVENT_UNREACHABLE; 2668 event.status = ib_event->param.sidr_rep_rcvd.status; 2669 break; 2670 } 2671 ret = cma_set_qkey(id_priv, rep->qkey); 2672 if (ret) { 2673 event.event = RDMA_CM_EVENT_ADDR_ERROR; 2674 event.status = ret; 2675 break; 2676 } 2677 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num, 2678 id_priv->id.route.path_rec, 2679 &event.param.ud.ah_attr); 2680 event.param.ud.qp_num = rep->qpn; 2681 event.param.ud.qkey = rep->qkey; 2682 event.event = RDMA_CM_EVENT_ESTABLISHED; 2683 event.status = 0; 2684 break; 2685 default: 2686 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n", 2687 ib_event->event); 2688 goto out; 2689 } 2690 2691 ret = id_priv->id.event_handler(&id_priv->id, &event); 2692 if (ret) { 2693 /* Destroy the CM ID by returning a non-zero value. */ 2694 id_priv->cm_id.ib = NULL; 2695 cma_exch(id_priv, RDMA_CM_DESTROYING); 2696 mutex_unlock(&id_priv->handler_mutex); 2697 rdma_destroy_id(&id_priv->id); 2698 return ret; 2699 } 2700 out: 2701 mutex_unlock(&id_priv->handler_mutex); 2702 return ret; 2703 } 2704 2705 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv, 2706 struct rdma_conn_param *conn_param) 2707 { 2708 struct ib_cm_sidr_req_param req; 2709 struct ib_cm_id *id; 2710 void *private_data; 2711 int offset, ret; 2712 2713 memset(&req, 0, sizeof req); 2714 offset = cma_user_data_offset(id_priv); 2715 req.private_data_len = offset + conn_param->private_data_len; 2716 if (req.private_data_len < conn_param->private_data_len) 2717 return -EINVAL; 2718 2719 if (req.private_data_len) { 2720 private_data = kzalloc(req.private_data_len, GFP_ATOMIC); 2721 if (!private_data) 2722 return -ENOMEM; 2723 } else { 2724 private_data = NULL; 2725 } 2726 2727 if (conn_param->private_data && conn_param->private_data_len) 2728 memcpy(private_data + offset, conn_param->private_data, 2729 conn_param->private_data_len); 2730 2731 if (private_data) { 2732 ret = cma_format_hdr(private_data, id_priv); 2733 if (ret) 2734 goto out; 2735 req.private_data = private_data; 2736 } 2737 2738 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler, 2739 id_priv); 2740 if (IS_ERR(id)) { 2741 ret = PTR_ERR(id); 2742 goto out; 2743 } 2744 id_priv->cm_id.ib = id; 2745 2746 req.path = id_priv->id.route.path_rec; 2747 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv)); 2748 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8); 2749 req.max_cm_retries = CMA_MAX_CM_RETRIES; 2750 2751 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req); 2752 if (ret) { 2753 ib_destroy_cm_id(id_priv->cm_id.ib); 2754 id_priv->cm_id.ib = NULL; 2755 } 2756 out: 2757 kfree(private_data); 2758 return ret; 2759 } 2760 2761 static int cma_connect_ib(struct rdma_id_private *id_priv, 2762 struct rdma_conn_param *conn_param) 2763 { 2764 struct ib_cm_req_param req; 2765 struct rdma_route *route; 2766 void *private_data; 2767 struct ib_cm_id *id; 2768 int offset, ret; 2769 2770 memset(&req, 0, sizeof req); 2771 offset = cma_user_data_offset(id_priv); 2772 req.private_data_len = offset + conn_param->private_data_len; 2773 if (req.private_data_len < conn_param->private_data_len) 2774 return -EINVAL; 2775 2776 if (req.private_data_len) { 2777 private_data = kzalloc(req.private_data_len, GFP_ATOMIC); 2778 if (!private_data) 2779 return -ENOMEM; 2780 } else { 2781 private_data = NULL; 2782 } 2783 2784 if (conn_param->private_data && conn_param->private_data_len) 2785 memcpy(private_data + offset, conn_param->private_data, 2786 conn_param->private_data_len); 2787 2788 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv); 2789 if (IS_ERR(id)) { 2790 ret = PTR_ERR(id); 2791 goto out; 2792 } 2793 id_priv->cm_id.ib = id; 2794 2795 route = &id_priv->id.route; 2796 if (private_data) { 2797 ret = cma_format_hdr(private_data, id_priv); 2798 if (ret) 2799 goto out; 2800 req.private_data = private_data; 2801 } 2802 2803 req.primary_path = &route->path_rec[0]; 2804 if (route->num_paths == 2) 2805 req.alternate_path = &route->path_rec[1]; 2806 2807 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv)); 2808 req.qp_num = id_priv->qp_num; 2809 req.qp_type = id_priv->id.qp_type; 2810 req.starting_psn = id_priv->seq_num; 2811 req.responder_resources = conn_param->responder_resources; 2812 req.initiator_depth = conn_param->initiator_depth; 2813 req.flow_control = conn_param->flow_control; 2814 req.retry_count = min_t(u8, 7, conn_param->retry_count); 2815 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count); 2816 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT; 2817 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT; 2818 req.max_cm_retries = CMA_MAX_CM_RETRIES; 2819 req.srq = id_priv->srq ? 1 : 0; 2820 2821 ret = ib_send_cm_req(id_priv->cm_id.ib, &req); 2822 out: 2823 if (ret && !IS_ERR(id)) { 2824 ib_destroy_cm_id(id); 2825 id_priv->cm_id.ib = NULL; 2826 } 2827 2828 kfree(private_data); 2829 return ret; 2830 } 2831 2832 static int cma_connect_iw(struct rdma_id_private *id_priv, 2833 struct rdma_conn_param *conn_param) 2834 { 2835 struct iw_cm_id *cm_id; 2836 int ret; 2837 struct iw_cm_conn_param iw_param; 2838 2839 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv); 2840 if (IS_ERR(cm_id)) 2841 return PTR_ERR(cm_id); 2842 2843 id_priv->cm_id.iw = cm_id; 2844 2845 memcpy(&cm_id->local_addr, cma_src_addr(id_priv), 2846 rdma_addr_size(cma_src_addr(id_priv))); 2847 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv), 2848 rdma_addr_size(cma_dst_addr(id_priv))); 2849 2850 ret = cma_modify_qp_rtr(id_priv, conn_param); 2851 if (ret) 2852 goto out; 2853 2854 if (conn_param) { 2855 iw_param.ord = conn_param->initiator_depth; 2856 iw_param.ird = conn_param->responder_resources; 2857 iw_param.private_data = conn_param->private_data; 2858 iw_param.private_data_len = conn_param->private_data_len; 2859 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num; 2860 } else { 2861 memset(&iw_param, 0, sizeof iw_param); 2862 iw_param.qpn = id_priv->qp_num; 2863 } 2864 ret = iw_cm_connect(cm_id, &iw_param); 2865 out: 2866 if (ret) { 2867 iw_destroy_cm_id(cm_id); 2868 id_priv->cm_id.iw = NULL; 2869 } 2870 return ret; 2871 } 2872 2873 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param) 2874 { 2875 struct rdma_id_private *id_priv; 2876 int ret; 2877 2878 id_priv = container_of(id, struct rdma_id_private, id); 2879 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT)) 2880 return -EINVAL; 2881 2882 if (!id->qp) { 2883 id_priv->qp_num = conn_param->qp_num; 2884 id_priv->srq = conn_param->srq; 2885 } 2886 2887 switch (rdma_node_get_transport(id->device->node_type)) { 2888 case RDMA_TRANSPORT_IB: 2889 if (id->qp_type == IB_QPT_UD) 2890 ret = cma_resolve_ib_udp(id_priv, conn_param); 2891 else 2892 ret = cma_connect_ib(id_priv, conn_param); 2893 break; 2894 case RDMA_TRANSPORT_IWARP: 2895 ret = cma_connect_iw(id_priv, conn_param); 2896 break; 2897 default: 2898 ret = -ENOSYS; 2899 break; 2900 } 2901 if (ret) 2902 goto err; 2903 2904 return 0; 2905 err: 2906 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED); 2907 return ret; 2908 } 2909 EXPORT_SYMBOL(rdma_connect); 2910 2911 static int cma_accept_ib(struct rdma_id_private *id_priv, 2912 struct rdma_conn_param *conn_param) 2913 { 2914 struct ib_cm_rep_param rep; 2915 int ret; 2916 2917 ret = cma_modify_qp_rtr(id_priv, conn_param); 2918 if (ret) 2919 goto out; 2920 2921 ret = cma_modify_qp_rts(id_priv, conn_param); 2922 if (ret) 2923 goto out; 2924 2925 memset(&rep, 0, sizeof rep); 2926 rep.qp_num = id_priv->qp_num; 2927 rep.starting_psn = id_priv->seq_num; 2928 rep.private_data = conn_param->private_data; 2929 rep.private_data_len = conn_param->private_data_len; 2930 rep.responder_resources = conn_param->responder_resources; 2931 rep.initiator_depth = conn_param->initiator_depth; 2932 rep.failover_accepted = 0; 2933 rep.flow_control = conn_param->flow_control; 2934 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count); 2935 rep.srq = id_priv->srq ? 1 : 0; 2936 2937 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep); 2938 out: 2939 return ret; 2940 } 2941 2942 static int cma_accept_iw(struct rdma_id_private *id_priv, 2943 struct rdma_conn_param *conn_param) 2944 { 2945 struct iw_cm_conn_param iw_param; 2946 int ret; 2947 2948 ret = cma_modify_qp_rtr(id_priv, conn_param); 2949 if (ret) 2950 return ret; 2951 2952 iw_param.ord = conn_param->initiator_depth; 2953 iw_param.ird = conn_param->responder_resources; 2954 iw_param.private_data = conn_param->private_data; 2955 iw_param.private_data_len = conn_param->private_data_len; 2956 if (id_priv->id.qp) { 2957 iw_param.qpn = id_priv->qp_num; 2958 } else 2959 iw_param.qpn = conn_param->qp_num; 2960 2961 return iw_cm_accept(id_priv->cm_id.iw, &iw_param); 2962 } 2963 2964 static int cma_send_sidr_rep(struct rdma_id_private *id_priv, 2965 enum ib_cm_sidr_status status, u32 qkey, 2966 const void *private_data, int private_data_len) 2967 { 2968 struct ib_cm_sidr_rep_param rep; 2969 int ret; 2970 2971 memset(&rep, 0, sizeof rep); 2972 rep.status = status; 2973 if (status == IB_SIDR_SUCCESS) { 2974 ret = cma_set_qkey(id_priv, qkey); 2975 if (ret) 2976 return ret; 2977 rep.qp_num = id_priv->qp_num; 2978 rep.qkey = id_priv->qkey; 2979 } 2980 rep.private_data = private_data; 2981 rep.private_data_len = private_data_len; 2982 2983 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep); 2984 } 2985 2986 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param) 2987 { 2988 struct rdma_id_private *id_priv; 2989 int ret; 2990 2991 id_priv = container_of(id, struct rdma_id_private, id); 2992 2993 id_priv->owner = task_pid_nr(current); 2994 2995 if (!cma_comp(id_priv, RDMA_CM_CONNECT)) 2996 return -EINVAL; 2997 2998 if (!id->qp && conn_param) { 2999 id_priv->qp_num = conn_param->qp_num; 3000 id_priv->srq = conn_param->srq; 3001 } 3002 3003 switch (rdma_node_get_transport(id->device->node_type)) { 3004 case RDMA_TRANSPORT_IB: 3005 if (id->qp_type == IB_QPT_UD) { 3006 if (conn_param) 3007 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS, 3008 conn_param->qkey, 3009 conn_param->private_data, 3010 conn_param->private_data_len); 3011 else 3012 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS, 3013 0, NULL, 0); 3014 } else { 3015 if (conn_param) 3016 ret = cma_accept_ib(id_priv, conn_param); 3017 else 3018 ret = cma_rep_recv(id_priv); 3019 } 3020 break; 3021 case RDMA_TRANSPORT_IWARP: 3022 ret = cma_accept_iw(id_priv, conn_param); 3023 break; 3024 default: 3025 ret = -ENOSYS; 3026 break; 3027 } 3028 3029 if (ret) 3030 goto reject; 3031 3032 return 0; 3033 reject: 3034 cma_modify_qp_err(id_priv); 3035 rdma_reject(id, NULL, 0); 3036 return ret; 3037 } 3038 EXPORT_SYMBOL(rdma_accept); 3039 3040 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event) 3041 { 3042 struct rdma_id_private *id_priv; 3043 int ret; 3044 3045 id_priv = container_of(id, struct rdma_id_private, id); 3046 if (!id_priv->cm_id.ib) 3047 return -EINVAL; 3048 3049 switch (id->device->node_type) { 3050 case RDMA_NODE_IB_CA: 3051 ret = ib_cm_notify(id_priv->cm_id.ib, event); 3052 break; 3053 default: 3054 ret = 0; 3055 break; 3056 } 3057 return ret; 3058 } 3059 EXPORT_SYMBOL(rdma_notify); 3060 3061 int rdma_reject(struct rdma_cm_id *id, const void *private_data, 3062 u8 private_data_len) 3063 { 3064 struct rdma_id_private *id_priv; 3065 int ret; 3066 3067 id_priv = container_of(id, struct rdma_id_private, id); 3068 if (!id_priv->cm_id.ib) 3069 return -EINVAL; 3070 3071 switch (rdma_node_get_transport(id->device->node_type)) { 3072 case RDMA_TRANSPORT_IB: 3073 if (id->qp_type == IB_QPT_UD) 3074 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0, 3075 private_data, private_data_len); 3076 else 3077 ret = ib_send_cm_rej(id_priv->cm_id.ib, 3078 IB_CM_REJ_CONSUMER_DEFINED, NULL, 3079 0, private_data, private_data_len); 3080 break; 3081 case RDMA_TRANSPORT_IWARP: 3082 ret = iw_cm_reject(id_priv->cm_id.iw, 3083 private_data, private_data_len); 3084 break; 3085 default: 3086 ret = -ENOSYS; 3087 break; 3088 } 3089 return ret; 3090 } 3091 EXPORT_SYMBOL(rdma_reject); 3092 3093 int rdma_disconnect(struct rdma_cm_id *id) 3094 { 3095 struct rdma_id_private *id_priv; 3096 int ret; 3097 3098 id_priv = container_of(id, struct rdma_id_private, id); 3099 if (!id_priv->cm_id.ib) 3100 return -EINVAL; 3101 3102 switch (rdma_node_get_transport(id->device->node_type)) { 3103 case RDMA_TRANSPORT_IB: 3104 ret = cma_modify_qp_err(id_priv); 3105 if (ret) 3106 goto out; 3107 /* Initiate or respond to a disconnect. */ 3108 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) 3109 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0); 3110 break; 3111 case RDMA_TRANSPORT_IWARP: 3112 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0); 3113 break; 3114 default: 3115 ret = -EINVAL; 3116 break; 3117 } 3118 out: 3119 return ret; 3120 } 3121 EXPORT_SYMBOL(rdma_disconnect); 3122 3123 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast) 3124 { 3125 struct rdma_id_private *id_priv; 3126 struct cma_multicast *mc = multicast->context; 3127 struct rdma_cm_event event; 3128 int ret; 3129 3130 id_priv = mc->id_priv; 3131 if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) && 3132 cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED)) 3133 return 0; 3134 3135 if (!status) 3136 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey)); 3137 mutex_lock(&id_priv->qp_mutex); 3138 if (!status && id_priv->id.qp) 3139 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid, 3140 be16_to_cpu(multicast->rec.mlid)); 3141 mutex_unlock(&id_priv->qp_mutex); 3142 3143 memset(&event, 0, sizeof event); 3144 event.status = status; 3145 event.param.ud.private_data = mc->context; 3146 if (!status) { 3147 event.event = RDMA_CM_EVENT_MULTICAST_JOIN; 3148 ib_init_ah_from_mcmember(id_priv->id.device, 3149 id_priv->id.port_num, &multicast->rec, 3150 &event.param.ud.ah_attr); 3151 event.param.ud.qp_num = 0xFFFFFF; 3152 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey); 3153 } else 3154 event.event = RDMA_CM_EVENT_MULTICAST_ERROR; 3155 3156 ret = id_priv->id.event_handler(&id_priv->id, &event); 3157 if (ret) { 3158 cma_exch(id_priv, RDMA_CM_DESTROYING); 3159 mutex_unlock(&id_priv->handler_mutex); 3160 rdma_destroy_id(&id_priv->id); 3161 return 0; 3162 } 3163 3164 mutex_unlock(&id_priv->handler_mutex); 3165 return 0; 3166 } 3167 3168 static void cma_set_mgid(struct rdma_id_private *id_priv, 3169 struct sockaddr *addr, union ib_gid *mgid) 3170 { 3171 unsigned char mc_map[MAX_ADDR_LEN]; 3172 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 3173 struct sockaddr_in *sin = (struct sockaddr_in *) addr; 3174 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr; 3175 3176 if (cma_any_addr(addr)) { 3177 memset(mgid, 0, sizeof *mgid); 3178 } else if ((addr->sa_family == AF_INET6) && 3179 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) == 3180 0xFF10A01B)) { 3181 /* IPv6 address is an SA assigned MGID. */ 3182 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid); 3183 } else if (addr->sa_family == AF_IB) { 3184 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid); 3185 } else if ((addr->sa_family == AF_INET6)) { 3186 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map); 3187 if (id_priv->id.ps == RDMA_PS_UDP) 3188 mc_map[7] = 0x01; /* Use RDMA CM signature */ 3189 *mgid = *(union ib_gid *) (mc_map + 4); 3190 } else { 3191 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map); 3192 if (id_priv->id.ps == RDMA_PS_UDP) 3193 mc_map[7] = 0x01; /* Use RDMA CM signature */ 3194 *mgid = *(union ib_gid *) (mc_map + 4); 3195 } 3196 } 3197 3198 static int cma_join_ib_multicast(struct rdma_id_private *id_priv, 3199 struct cma_multicast *mc) 3200 { 3201 struct ib_sa_mcmember_rec rec; 3202 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 3203 ib_sa_comp_mask comp_mask; 3204 int ret; 3205 3206 ib_addr_get_mgid(dev_addr, &rec.mgid); 3207 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num, 3208 &rec.mgid, &rec); 3209 if (ret) 3210 return ret; 3211 3212 ret = cma_set_qkey(id_priv, 0); 3213 if (ret) 3214 return ret; 3215 3216 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid); 3217 rec.qkey = cpu_to_be32(id_priv->qkey); 3218 rdma_addr_get_sgid(dev_addr, &rec.port_gid); 3219 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr)); 3220 rec.join_state = 1; 3221 3222 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID | 3223 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE | 3224 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL | 3225 IB_SA_MCMEMBER_REC_FLOW_LABEL | 3226 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS; 3227 3228 if (id_priv->id.ps == RDMA_PS_IPOIB) 3229 comp_mask |= IB_SA_MCMEMBER_REC_RATE | 3230 IB_SA_MCMEMBER_REC_RATE_SELECTOR | 3231 IB_SA_MCMEMBER_REC_MTU_SELECTOR | 3232 IB_SA_MCMEMBER_REC_MTU | 3233 IB_SA_MCMEMBER_REC_HOP_LIMIT; 3234 3235 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device, 3236 id_priv->id.port_num, &rec, 3237 comp_mask, GFP_KERNEL, 3238 cma_ib_mc_handler, mc); 3239 return PTR_ERR_OR_ZERO(mc->multicast.ib); 3240 } 3241 3242 static void iboe_mcast_work_handler(struct work_struct *work) 3243 { 3244 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work); 3245 struct cma_multicast *mc = mw->mc; 3246 struct ib_sa_multicast *m = mc->multicast.ib; 3247 3248 mc->multicast.ib->context = mc; 3249 cma_ib_mc_handler(0, m); 3250 kref_put(&mc->mcref, release_mc); 3251 kfree(mw); 3252 } 3253 3254 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid) 3255 { 3256 struct sockaddr_in *sin = (struct sockaddr_in *)addr; 3257 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr; 3258 3259 if (cma_any_addr(addr)) { 3260 memset(mgid, 0, sizeof *mgid); 3261 } else if (addr->sa_family == AF_INET6) { 3262 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid); 3263 } else { 3264 mgid->raw[0] = 0xff; 3265 mgid->raw[1] = 0x0e; 3266 mgid->raw[2] = 0; 3267 mgid->raw[3] = 0; 3268 mgid->raw[4] = 0; 3269 mgid->raw[5] = 0; 3270 mgid->raw[6] = 0; 3271 mgid->raw[7] = 0; 3272 mgid->raw[8] = 0; 3273 mgid->raw[9] = 0; 3274 mgid->raw[10] = 0xff; 3275 mgid->raw[11] = 0xff; 3276 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr; 3277 } 3278 } 3279 3280 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv, 3281 struct cma_multicast *mc) 3282 { 3283 struct iboe_mcast_work *work; 3284 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 3285 int err; 3286 struct sockaddr *addr = (struct sockaddr *)&mc->addr; 3287 struct net_device *ndev = NULL; 3288 3289 if (cma_zero_addr((struct sockaddr *)&mc->addr)) 3290 return -EINVAL; 3291 3292 work = kzalloc(sizeof *work, GFP_KERNEL); 3293 if (!work) 3294 return -ENOMEM; 3295 3296 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL); 3297 if (!mc->multicast.ib) { 3298 err = -ENOMEM; 3299 goto out1; 3300 } 3301 3302 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid); 3303 3304 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff); 3305 if (id_priv->id.ps == RDMA_PS_UDP) 3306 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY); 3307 3308 if (dev_addr->bound_dev_if) 3309 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if); 3310 if (!ndev) { 3311 err = -ENODEV; 3312 goto out2; 3313 } 3314 mc->multicast.ib->rec.rate = iboe_get_rate(ndev); 3315 mc->multicast.ib->rec.hop_limit = 1; 3316 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu); 3317 dev_put(ndev); 3318 if (!mc->multicast.ib->rec.mtu) { 3319 err = -EINVAL; 3320 goto out2; 3321 } 3322 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr, 3323 &mc->multicast.ib->rec.port_gid); 3324 work->id = id_priv; 3325 work->mc = mc; 3326 INIT_WORK(&work->work, iboe_mcast_work_handler); 3327 kref_get(&mc->mcref); 3328 queue_work(cma_wq, &work->work); 3329 3330 return 0; 3331 3332 out2: 3333 kfree(mc->multicast.ib); 3334 out1: 3335 kfree(work); 3336 return err; 3337 } 3338 3339 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr, 3340 void *context) 3341 { 3342 struct rdma_id_private *id_priv; 3343 struct cma_multicast *mc; 3344 int ret; 3345 3346 id_priv = container_of(id, struct rdma_id_private, id); 3347 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) && 3348 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED)) 3349 return -EINVAL; 3350 3351 mc = kmalloc(sizeof *mc, GFP_KERNEL); 3352 if (!mc) 3353 return -ENOMEM; 3354 3355 memcpy(&mc->addr, addr, rdma_addr_size(addr)); 3356 mc->context = context; 3357 mc->id_priv = id_priv; 3358 3359 spin_lock(&id_priv->lock); 3360 list_add(&mc->list, &id_priv->mc_list); 3361 spin_unlock(&id_priv->lock); 3362 3363 switch (rdma_node_get_transport(id->device->node_type)) { 3364 case RDMA_TRANSPORT_IB: 3365 switch (rdma_port_get_link_layer(id->device, id->port_num)) { 3366 case IB_LINK_LAYER_INFINIBAND: 3367 ret = cma_join_ib_multicast(id_priv, mc); 3368 break; 3369 case IB_LINK_LAYER_ETHERNET: 3370 kref_init(&mc->mcref); 3371 ret = cma_iboe_join_multicast(id_priv, mc); 3372 break; 3373 default: 3374 ret = -EINVAL; 3375 } 3376 break; 3377 default: 3378 ret = -ENOSYS; 3379 break; 3380 } 3381 3382 if (ret) { 3383 spin_lock_irq(&id_priv->lock); 3384 list_del(&mc->list); 3385 spin_unlock_irq(&id_priv->lock); 3386 kfree(mc); 3387 } 3388 return ret; 3389 } 3390 EXPORT_SYMBOL(rdma_join_multicast); 3391 3392 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr) 3393 { 3394 struct rdma_id_private *id_priv; 3395 struct cma_multicast *mc; 3396 3397 id_priv = container_of(id, struct rdma_id_private, id); 3398 spin_lock_irq(&id_priv->lock); 3399 list_for_each_entry(mc, &id_priv->mc_list, list) { 3400 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) { 3401 list_del(&mc->list); 3402 spin_unlock_irq(&id_priv->lock); 3403 3404 if (id->qp) 3405 ib_detach_mcast(id->qp, 3406 &mc->multicast.ib->rec.mgid, 3407 be16_to_cpu(mc->multicast.ib->rec.mlid)); 3408 if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) { 3409 switch (rdma_port_get_link_layer(id->device, id->port_num)) { 3410 case IB_LINK_LAYER_INFINIBAND: 3411 ib_sa_free_multicast(mc->multicast.ib); 3412 kfree(mc); 3413 break; 3414 case IB_LINK_LAYER_ETHERNET: 3415 kref_put(&mc->mcref, release_mc); 3416 break; 3417 default: 3418 break; 3419 } 3420 } 3421 return; 3422 } 3423 } 3424 spin_unlock_irq(&id_priv->lock); 3425 } 3426 EXPORT_SYMBOL(rdma_leave_multicast); 3427 3428 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv) 3429 { 3430 struct rdma_dev_addr *dev_addr; 3431 struct cma_ndev_work *work; 3432 3433 dev_addr = &id_priv->id.route.addr.dev_addr; 3434 3435 if ((dev_addr->bound_dev_if == ndev->ifindex) && 3436 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) { 3437 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n", 3438 ndev->name, &id_priv->id); 3439 work = kzalloc(sizeof *work, GFP_KERNEL); 3440 if (!work) 3441 return -ENOMEM; 3442 3443 INIT_WORK(&work->work, cma_ndev_work_handler); 3444 work->id = id_priv; 3445 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE; 3446 atomic_inc(&id_priv->refcount); 3447 queue_work(cma_wq, &work->work); 3448 } 3449 3450 return 0; 3451 } 3452 3453 static int cma_netdev_callback(struct notifier_block *self, unsigned long event, 3454 void *ptr) 3455 { 3456 struct net_device *ndev = netdev_notifier_info_to_dev(ptr); 3457 struct cma_device *cma_dev; 3458 struct rdma_id_private *id_priv; 3459 int ret = NOTIFY_DONE; 3460 3461 if (dev_net(ndev) != &init_net) 3462 return NOTIFY_DONE; 3463 3464 if (event != NETDEV_BONDING_FAILOVER) 3465 return NOTIFY_DONE; 3466 3467 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING)) 3468 return NOTIFY_DONE; 3469 3470 mutex_lock(&lock); 3471 list_for_each_entry(cma_dev, &dev_list, list) 3472 list_for_each_entry(id_priv, &cma_dev->id_list, list) { 3473 ret = cma_netdev_change(ndev, id_priv); 3474 if (ret) 3475 goto out; 3476 } 3477 3478 out: 3479 mutex_unlock(&lock); 3480 return ret; 3481 } 3482 3483 static struct notifier_block cma_nb = { 3484 .notifier_call = cma_netdev_callback 3485 }; 3486 3487 static void cma_add_one(struct ib_device *device) 3488 { 3489 struct cma_device *cma_dev; 3490 struct rdma_id_private *id_priv; 3491 3492 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL); 3493 if (!cma_dev) 3494 return; 3495 3496 cma_dev->device = device; 3497 3498 init_completion(&cma_dev->comp); 3499 atomic_set(&cma_dev->refcount, 1); 3500 INIT_LIST_HEAD(&cma_dev->id_list); 3501 ib_set_client_data(device, &cma_client, cma_dev); 3502 3503 mutex_lock(&lock); 3504 list_add_tail(&cma_dev->list, &dev_list); 3505 list_for_each_entry(id_priv, &listen_any_list, list) 3506 cma_listen_on_dev(id_priv, cma_dev); 3507 mutex_unlock(&lock); 3508 } 3509 3510 static int cma_remove_id_dev(struct rdma_id_private *id_priv) 3511 { 3512 struct rdma_cm_event event; 3513 enum rdma_cm_state state; 3514 int ret = 0; 3515 3516 /* Record that we want to remove the device */ 3517 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL); 3518 if (state == RDMA_CM_DESTROYING) 3519 return 0; 3520 3521 cma_cancel_operation(id_priv, state); 3522 mutex_lock(&id_priv->handler_mutex); 3523 3524 /* Check for destruction from another callback. */ 3525 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL)) 3526 goto out; 3527 3528 memset(&event, 0, sizeof event); 3529 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL; 3530 ret = id_priv->id.event_handler(&id_priv->id, &event); 3531 out: 3532 mutex_unlock(&id_priv->handler_mutex); 3533 return ret; 3534 } 3535 3536 static void cma_process_remove(struct cma_device *cma_dev) 3537 { 3538 struct rdma_id_private *id_priv; 3539 int ret; 3540 3541 mutex_lock(&lock); 3542 while (!list_empty(&cma_dev->id_list)) { 3543 id_priv = list_entry(cma_dev->id_list.next, 3544 struct rdma_id_private, list); 3545 3546 list_del(&id_priv->listen_list); 3547 list_del_init(&id_priv->list); 3548 atomic_inc(&id_priv->refcount); 3549 mutex_unlock(&lock); 3550 3551 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv); 3552 cma_deref_id(id_priv); 3553 if (ret) 3554 rdma_destroy_id(&id_priv->id); 3555 3556 mutex_lock(&lock); 3557 } 3558 mutex_unlock(&lock); 3559 3560 cma_deref_dev(cma_dev); 3561 wait_for_completion(&cma_dev->comp); 3562 } 3563 3564 static void cma_remove_one(struct ib_device *device) 3565 { 3566 struct cma_device *cma_dev; 3567 3568 cma_dev = ib_get_client_data(device, &cma_client); 3569 if (!cma_dev) 3570 return; 3571 3572 mutex_lock(&lock); 3573 list_del(&cma_dev->list); 3574 mutex_unlock(&lock); 3575 3576 cma_process_remove(cma_dev); 3577 kfree(cma_dev); 3578 } 3579 3580 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb) 3581 { 3582 struct nlmsghdr *nlh; 3583 struct rdma_cm_id_stats *id_stats; 3584 struct rdma_id_private *id_priv; 3585 struct rdma_cm_id *id = NULL; 3586 struct cma_device *cma_dev; 3587 int i_dev = 0, i_id = 0; 3588 3589 /* 3590 * We export all of the IDs as a sequence of messages. Each 3591 * ID gets its own netlink message. 3592 */ 3593 mutex_lock(&lock); 3594 3595 list_for_each_entry(cma_dev, &dev_list, list) { 3596 if (i_dev < cb->args[0]) { 3597 i_dev++; 3598 continue; 3599 } 3600 3601 i_id = 0; 3602 list_for_each_entry(id_priv, &cma_dev->id_list, list) { 3603 if (i_id < cb->args[1]) { 3604 i_id++; 3605 continue; 3606 } 3607 3608 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq, 3609 sizeof *id_stats, RDMA_NL_RDMA_CM, 3610 RDMA_NL_RDMA_CM_ID_STATS); 3611 if (!id_stats) 3612 goto out; 3613 3614 memset(id_stats, 0, sizeof *id_stats); 3615 id = &id_priv->id; 3616 id_stats->node_type = id->route.addr.dev_addr.dev_type; 3617 id_stats->port_num = id->port_num; 3618 id_stats->bound_dev_if = 3619 id->route.addr.dev_addr.bound_dev_if; 3620 3621 if (ibnl_put_attr(skb, nlh, 3622 rdma_addr_size(cma_src_addr(id_priv)), 3623 cma_src_addr(id_priv), 3624 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) 3625 goto out; 3626 if (ibnl_put_attr(skb, nlh, 3627 rdma_addr_size(cma_src_addr(id_priv)), 3628 cma_dst_addr(id_priv), 3629 RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) 3630 goto out; 3631 3632 id_stats->pid = id_priv->owner; 3633 id_stats->port_space = id->ps; 3634 id_stats->cm_state = id_priv->state; 3635 id_stats->qp_num = id_priv->qp_num; 3636 id_stats->qp_type = id->qp_type; 3637 3638 i_id++; 3639 } 3640 3641 cb->args[1] = 0; 3642 i_dev++; 3643 } 3644 3645 out: 3646 mutex_unlock(&lock); 3647 cb->args[0] = i_dev; 3648 cb->args[1] = i_id; 3649 3650 return skb->len; 3651 } 3652 3653 static const struct ibnl_client_cbs cma_cb_table[] = { 3654 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats, 3655 .module = THIS_MODULE }, 3656 }; 3657 3658 static int __init cma_init(void) 3659 { 3660 int ret; 3661 3662 cma_wq = create_singlethread_workqueue("rdma_cm"); 3663 if (!cma_wq) 3664 return -ENOMEM; 3665 3666 ib_sa_register_client(&sa_client); 3667 rdma_addr_register_client(&addr_client); 3668 register_netdevice_notifier(&cma_nb); 3669 3670 ret = ib_register_client(&cma_client); 3671 if (ret) 3672 goto err; 3673 3674 if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table)) 3675 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n"); 3676 3677 return 0; 3678 3679 err: 3680 unregister_netdevice_notifier(&cma_nb); 3681 rdma_addr_unregister_client(&addr_client); 3682 ib_sa_unregister_client(&sa_client); 3683 destroy_workqueue(cma_wq); 3684 return ret; 3685 } 3686 3687 static void __exit cma_cleanup(void) 3688 { 3689 ibnl_remove_client(RDMA_NL_RDMA_CM); 3690 ib_unregister_client(&cma_client); 3691 unregister_netdevice_notifier(&cma_nb); 3692 rdma_addr_unregister_client(&addr_client); 3693 ib_sa_unregister_client(&sa_client); 3694 destroy_workqueue(cma_wq); 3695 idr_destroy(&tcp_ps); 3696 idr_destroy(&udp_ps); 3697 idr_destroy(&ipoib_ps); 3698 idr_destroy(&ib_ps); 3699 } 3700 3701 module_init(cma_init); 3702 module_exit(cma_cleanup); 3703