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/igmp.h> 42 #include <linux/idr.h> 43 #include <linux/inetdevice.h> 44 #include <linux/slab.h> 45 #include <linux/module.h> 46 #include <net/route.h> 47 48 #include <net/net_namespace.h> 49 #include <net/netns/generic.h> 50 #include <net/tcp.h> 51 #include <net/ipv6.h> 52 #include <net/ip_fib.h> 53 #include <net/ip6_route.h> 54 55 #include <rdma/rdma_cm.h> 56 #include <rdma/rdma_cm_ib.h> 57 #include <rdma/rdma_netlink.h> 58 #include <rdma/ib.h> 59 #include <rdma/ib_cache.h> 60 #include <rdma/ib_cm.h> 61 #include <rdma/ib_sa.h> 62 #include <rdma/iw_cm.h> 63 64 #include "core_priv.h" 65 66 MODULE_AUTHOR("Sean Hefty"); 67 MODULE_DESCRIPTION("Generic RDMA CM Agent"); 68 MODULE_LICENSE("Dual BSD/GPL"); 69 70 #define CMA_CM_RESPONSE_TIMEOUT 20 71 #define CMA_MAX_CM_RETRIES 15 72 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24) 73 #define CMA_IBOE_PACKET_LIFETIME 18 74 75 static const char * const cma_events[] = { 76 [RDMA_CM_EVENT_ADDR_RESOLVED] = "address resolved", 77 [RDMA_CM_EVENT_ADDR_ERROR] = "address error", 78 [RDMA_CM_EVENT_ROUTE_RESOLVED] = "route resolved ", 79 [RDMA_CM_EVENT_ROUTE_ERROR] = "route error", 80 [RDMA_CM_EVENT_CONNECT_REQUEST] = "connect request", 81 [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response", 82 [RDMA_CM_EVENT_CONNECT_ERROR] = "connect error", 83 [RDMA_CM_EVENT_UNREACHABLE] = "unreachable", 84 [RDMA_CM_EVENT_REJECTED] = "rejected", 85 [RDMA_CM_EVENT_ESTABLISHED] = "established", 86 [RDMA_CM_EVENT_DISCONNECTED] = "disconnected", 87 [RDMA_CM_EVENT_DEVICE_REMOVAL] = "device removal", 88 [RDMA_CM_EVENT_MULTICAST_JOIN] = "multicast join", 89 [RDMA_CM_EVENT_MULTICAST_ERROR] = "multicast error", 90 [RDMA_CM_EVENT_ADDR_CHANGE] = "address change", 91 [RDMA_CM_EVENT_TIMEWAIT_EXIT] = "timewait exit", 92 }; 93 94 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event) 95 { 96 size_t index = event; 97 98 return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ? 99 cma_events[index] : "unrecognized event"; 100 } 101 EXPORT_SYMBOL(rdma_event_msg); 102 103 static void cma_add_one(struct ib_device *device); 104 static void cma_remove_one(struct ib_device *device, void *client_data); 105 106 static struct ib_client cma_client = { 107 .name = "cma", 108 .add = cma_add_one, 109 .remove = cma_remove_one 110 }; 111 112 static struct ib_sa_client sa_client; 113 static struct rdma_addr_client addr_client; 114 static LIST_HEAD(dev_list); 115 static LIST_HEAD(listen_any_list); 116 static DEFINE_MUTEX(lock); 117 static struct workqueue_struct *cma_wq; 118 static int cma_pernet_id; 119 120 struct cma_pernet { 121 struct idr tcp_ps; 122 struct idr udp_ps; 123 struct idr ipoib_ps; 124 struct idr ib_ps; 125 }; 126 127 static struct cma_pernet *cma_pernet(struct net *net) 128 { 129 return net_generic(net, cma_pernet_id); 130 } 131 132 static struct idr *cma_pernet_idr(struct net *net, enum rdma_port_space ps) 133 { 134 struct cma_pernet *pernet = cma_pernet(net); 135 136 switch (ps) { 137 case RDMA_PS_TCP: 138 return &pernet->tcp_ps; 139 case RDMA_PS_UDP: 140 return &pernet->udp_ps; 141 case RDMA_PS_IPOIB: 142 return &pernet->ipoib_ps; 143 case RDMA_PS_IB: 144 return &pernet->ib_ps; 145 default: 146 return NULL; 147 } 148 } 149 150 struct cma_device { 151 struct list_head list; 152 struct ib_device *device; 153 struct completion comp; 154 atomic_t refcount; 155 struct list_head id_list; 156 enum ib_gid_type *default_gid_type; 157 }; 158 159 struct rdma_bind_list { 160 enum rdma_port_space ps; 161 struct hlist_head owners; 162 unsigned short port; 163 }; 164 165 static int cma_ps_alloc(struct net *net, enum rdma_port_space ps, 166 struct rdma_bind_list *bind_list, int snum) 167 { 168 struct idr *idr = cma_pernet_idr(net, ps); 169 170 return idr_alloc(idr, bind_list, snum, snum + 1, GFP_KERNEL); 171 } 172 173 static struct rdma_bind_list *cma_ps_find(struct net *net, 174 enum rdma_port_space ps, int snum) 175 { 176 struct idr *idr = cma_pernet_idr(net, ps); 177 178 return idr_find(idr, snum); 179 } 180 181 static void cma_ps_remove(struct net *net, enum rdma_port_space ps, int snum) 182 { 183 struct idr *idr = cma_pernet_idr(net, ps); 184 185 idr_remove(idr, snum); 186 } 187 188 enum { 189 CMA_OPTION_AFONLY, 190 }; 191 192 void cma_ref_dev(struct cma_device *cma_dev) 193 { 194 atomic_inc(&cma_dev->refcount); 195 } 196 197 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter, 198 void *cookie) 199 { 200 struct cma_device *cma_dev; 201 struct cma_device *found_cma_dev = NULL; 202 203 mutex_lock(&lock); 204 205 list_for_each_entry(cma_dev, &dev_list, list) 206 if (filter(cma_dev->device, cookie)) { 207 found_cma_dev = cma_dev; 208 break; 209 } 210 211 if (found_cma_dev) 212 cma_ref_dev(found_cma_dev); 213 mutex_unlock(&lock); 214 return found_cma_dev; 215 } 216 217 int cma_get_default_gid_type(struct cma_device *cma_dev, 218 unsigned int port) 219 { 220 if (port < rdma_start_port(cma_dev->device) || 221 port > rdma_end_port(cma_dev->device)) 222 return -EINVAL; 223 224 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)]; 225 } 226 227 int cma_set_default_gid_type(struct cma_device *cma_dev, 228 unsigned int port, 229 enum ib_gid_type default_gid_type) 230 { 231 unsigned long supported_gids; 232 233 if (port < rdma_start_port(cma_dev->device) || 234 port > rdma_end_port(cma_dev->device)) 235 return -EINVAL; 236 237 supported_gids = roce_gid_type_mask_support(cma_dev->device, port); 238 239 if (!(supported_gids & 1 << default_gid_type)) 240 return -EINVAL; 241 242 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] = 243 default_gid_type; 244 245 return 0; 246 } 247 248 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev) 249 { 250 return cma_dev->device; 251 } 252 253 /* 254 * Device removal can occur at anytime, so we need extra handling to 255 * serialize notifying the user of device removal with other callbacks. 256 * We do this by disabling removal notification while a callback is in process, 257 * and reporting it after the callback completes. 258 */ 259 struct rdma_id_private { 260 struct rdma_cm_id id; 261 262 struct rdma_bind_list *bind_list; 263 struct hlist_node node; 264 struct list_head list; /* listen_any_list or cma_device.list */ 265 struct list_head listen_list; /* per device listens */ 266 struct cma_device *cma_dev; 267 struct list_head mc_list; 268 269 int internal_id; 270 enum rdma_cm_state state; 271 spinlock_t lock; 272 struct mutex qp_mutex; 273 274 struct completion comp; 275 atomic_t refcount; 276 struct mutex handler_mutex; 277 278 int backlog; 279 int timeout_ms; 280 struct ib_sa_query *query; 281 int query_id; 282 union { 283 struct ib_cm_id *ib; 284 struct iw_cm_id *iw; 285 } cm_id; 286 287 u32 seq_num; 288 u32 qkey; 289 u32 qp_num; 290 pid_t owner; 291 u32 options; 292 u8 srq; 293 u8 tos; 294 u8 reuseaddr; 295 u8 afonly; 296 enum ib_gid_type gid_type; 297 }; 298 299 struct cma_multicast { 300 struct rdma_id_private *id_priv; 301 union { 302 struct ib_sa_multicast *ib; 303 } multicast; 304 struct list_head list; 305 void *context; 306 struct sockaddr_storage addr; 307 struct kref mcref; 308 bool igmp_joined; 309 }; 310 311 struct cma_work { 312 struct work_struct work; 313 struct rdma_id_private *id; 314 enum rdma_cm_state old_state; 315 enum rdma_cm_state new_state; 316 struct rdma_cm_event event; 317 }; 318 319 struct cma_ndev_work { 320 struct work_struct work; 321 struct rdma_id_private *id; 322 struct rdma_cm_event event; 323 }; 324 325 struct iboe_mcast_work { 326 struct work_struct work; 327 struct rdma_id_private *id; 328 struct cma_multicast *mc; 329 }; 330 331 union cma_ip_addr { 332 struct in6_addr ip6; 333 struct { 334 __be32 pad[3]; 335 __be32 addr; 336 } ip4; 337 }; 338 339 struct cma_hdr { 340 u8 cma_version; 341 u8 ip_version; /* IP version: 7:4 */ 342 __be16 port; 343 union cma_ip_addr src_addr; 344 union cma_ip_addr dst_addr; 345 }; 346 347 #define CMA_VERSION 0x00 348 349 struct cma_req_info { 350 struct ib_device *device; 351 int port; 352 union ib_gid local_gid; 353 __be64 service_id; 354 u16 pkey; 355 bool has_gid:1; 356 }; 357 358 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp) 359 { 360 unsigned long flags; 361 int ret; 362 363 spin_lock_irqsave(&id_priv->lock, flags); 364 ret = (id_priv->state == comp); 365 spin_unlock_irqrestore(&id_priv->lock, flags); 366 return ret; 367 } 368 369 static int cma_comp_exch(struct rdma_id_private *id_priv, 370 enum rdma_cm_state comp, enum rdma_cm_state exch) 371 { 372 unsigned long flags; 373 int ret; 374 375 spin_lock_irqsave(&id_priv->lock, flags); 376 if ((ret = (id_priv->state == comp))) 377 id_priv->state = exch; 378 spin_unlock_irqrestore(&id_priv->lock, flags); 379 return ret; 380 } 381 382 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv, 383 enum rdma_cm_state exch) 384 { 385 unsigned long flags; 386 enum rdma_cm_state old; 387 388 spin_lock_irqsave(&id_priv->lock, flags); 389 old = id_priv->state; 390 id_priv->state = exch; 391 spin_unlock_irqrestore(&id_priv->lock, flags); 392 return old; 393 } 394 395 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr) 396 { 397 return hdr->ip_version >> 4; 398 } 399 400 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver) 401 { 402 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF); 403 } 404 405 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join) 406 { 407 struct in_device *in_dev = NULL; 408 409 if (ndev) { 410 rtnl_lock(); 411 in_dev = __in_dev_get_rtnl(ndev); 412 if (in_dev) { 413 if (join) 414 ip_mc_inc_group(in_dev, 415 *(__be32 *)(mgid->raw + 12)); 416 else 417 ip_mc_dec_group(in_dev, 418 *(__be32 *)(mgid->raw + 12)); 419 } 420 rtnl_unlock(); 421 } 422 return (in_dev) ? 0 : -ENODEV; 423 } 424 425 static void _cma_attach_to_dev(struct rdma_id_private *id_priv, 426 struct cma_device *cma_dev) 427 { 428 cma_ref_dev(cma_dev); 429 id_priv->cma_dev = cma_dev; 430 id_priv->gid_type = 0; 431 id_priv->id.device = cma_dev->device; 432 id_priv->id.route.addr.dev_addr.transport = 433 rdma_node_get_transport(cma_dev->device->node_type); 434 list_add_tail(&id_priv->list, &cma_dev->id_list); 435 } 436 437 static void cma_attach_to_dev(struct rdma_id_private *id_priv, 438 struct cma_device *cma_dev) 439 { 440 _cma_attach_to_dev(id_priv, cma_dev); 441 id_priv->gid_type = 442 cma_dev->default_gid_type[id_priv->id.port_num - 443 rdma_start_port(cma_dev->device)]; 444 } 445 446 void cma_deref_dev(struct cma_device *cma_dev) 447 { 448 if (atomic_dec_and_test(&cma_dev->refcount)) 449 complete(&cma_dev->comp); 450 } 451 452 static inline void release_mc(struct kref *kref) 453 { 454 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref); 455 456 kfree(mc->multicast.ib); 457 kfree(mc); 458 } 459 460 static void cma_release_dev(struct rdma_id_private *id_priv) 461 { 462 mutex_lock(&lock); 463 list_del(&id_priv->list); 464 cma_deref_dev(id_priv->cma_dev); 465 id_priv->cma_dev = NULL; 466 mutex_unlock(&lock); 467 } 468 469 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv) 470 { 471 return (struct sockaddr *) &id_priv->id.route.addr.src_addr; 472 } 473 474 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv) 475 { 476 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr; 477 } 478 479 static inline unsigned short cma_family(struct rdma_id_private *id_priv) 480 { 481 return id_priv->id.route.addr.src_addr.ss_family; 482 } 483 484 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey) 485 { 486 struct ib_sa_mcmember_rec rec; 487 int ret = 0; 488 489 if (id_priv->qkey) { 490 if (qkey && id_priv->qkey != qkey) 491 return -EINVAL; 492 return 0; 493 } 494 495 if (qkey) { 496 id_priv->qkey = qkey; 497 return 0; 498 } 499 500 switch (id_priv->id.ps) { 501 case RDMA_PS_UDP: 502 case RDMA_PS_IB: 503 id_priv->qkey = RDMA_UDP_QKEY; 504 break; 505 case RDMA_PS_IPOIB: 506 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid); 507 ret = ib_sa_get_mcmember_rec(id_priv->id.device, 508 id_priv->id.port_num, &rec.mgid, 509 &rec); 510 if (!ret) 511 id_priv->qkey = be32_to_cpu(rec.qkey); 512 break; 513 default: 514 break; 515 } 516 return ret; 517 } 518 519 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr) 520 { 521 dev_addr->dev_type = ARPHRD_INFINIBAND; 522 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr); 523 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey)); 524 } 525 526 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr) 527 { 528 int ret; 529 530 if (addr->sa_family != AF_IB) { 531 ret = rdma_translate_ip(addr, dev_addr, NULL); 532 } else { 533 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr); 534 ret = 0; 535 } 536 537 return ret; 538 } 539 540 static inline int cma_validate_port(struct ib_device *device, u8 port, 541 enum ib_gid_type gid_type, 542 union ib_gid *gid, int dev_type, 543 int bound_if_index) 544 { 545 int ret = -ENODEV; 546 struct net_device *ndev = NULL; 547 548 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port)) 549 return ret; 550 551 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port)) 552 return ret; 553 554 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) { 555 ndev = dev_get_by_index(&init_net, bound_if_index); 556 if (ndev && ndev->flags & IFF_LOOPBACK) { 557 pr_info("detected loopback device\n"); 558 dev_put(ndev); 559 560 if (!device->get_netdev) 561 return -EOPNOTSUPP; 562 563 ndev = device->get_netdev(device, port); 564 if (!ndev) 565 return -ENODEV; 566 } 567 } else { 568 gid_type = IB_GID_TYPE_IB; 569 } 570 571 ret = ib_find_cached_gid_by_port(device, gid, gid_type, port, 572 ndev, NULL); 573 574 if (ndev) 575 dev_put(ndev); 576 577 return ret; 578 } 579 580 static int cma_acquire_dev(struct rdma_id_private *id_priv, 581 struct rdma_id_private *listen_id_priv) 582 { 583 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 584 struct cma_device *cma_dev; 585 union ib_gid gid, iboe_gid, *gidp; 586 int ret = -ENODEV; 587 u8 port; 588 589 if (dev_addr->dev_type != ARPHRD_INFINIBAND && 590 id_priv->id.ps == RDMA_PS_IPOIB) 591 return -EINVAL; 592 593 mutex_lock(&lock); 594 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr, 595 &iboe_gid); 596 597 memcpy(&gid, dev_addr->src_dev_addr + 598 rdma_addr_gid_offset(dev_addr), sizeof gid); 599 600 if (listen_id_priv) { 601 cma_dev = listen_id_priv->cma_dev; 602 port = listen_id_priv->id.port_num; 603 gidp = rdma_protocol_roce(cma_dev->device, port) ? 604 &iboe_gid : &gid; 605 606 ret = cma_validate_port(cma_dev->device, port, 607 rdma_protocol_ib(cma_dev->device, port) ? 608 IB_GID_TYPE_IB : 609 listen_id_priv->gid_type, gidp, 610 dev_addr->dev_type, 611 dev_addr->bound_dev_if); 612 if (!ret) { 613 id_priv->id.port_num = port; 614 goto out; 615 } 616 } 617 618 list_for_each_entry(cma_dev, &dev_list, list) { 619 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) { 620 if (listen_id_priv && 621 listen_id_priv->cma_dev == cma_dev && 622 listen_id_priv->id.port_num == port) 623 continue; 624 625 gidp = rdma_protocol_roce(cma_dev->device, port) ? 626 &iboe_gid : &gid; 627 628 ret = cma_validate_port(cma_dev->device, port, 629 rdma_protocol_ib(cma_dev->device, port) ? 630 IB_GID_TYPE_IB : 631 cma_dev->default_gid_type[port - 1], 632 gidp, dev_addr->dev_type, 633 dev_addr->bound_dev_if); 634 if (!ret) { 635 id_priv->id.port_num = port; 636 goto out; 637 } 638 } 639 } 640 641 out: 642 if (!ret) 643 cma_attach_to_dev(id_priv, cma_dev); 644 645 mutex_unlock(&lock); 646 return ret; 647 } 648 649 /* 650 * Select the source IB device and address to reach the destination IB address. 651 */ 652 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv) 653 { 654 struct cma_device *cma_dev, *cur_dev; 655 struct sockaddr_ib *addr; 656 union ib_gid gid, sgid, *dgid; 657 u16 pkey, index; 658 u8 p; 659 int i; 660 661 cma_dev = NULL; 662 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv); 663 dgid = (union ib_gid *) &addr->sib_addr; 664 pkey = ntohs(addr->sib_pkey); 665 666 list_for_each_entry(cur_dev, &dev_list, list) { 667 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) { 668 if (!rdma_cap_af_ib(cur_dev->device, p)) 669 continue; 670 671 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index)) 672 continue; 673 674 for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i, 675 &gid, NULL); 676 i++) { 677 if (!memcmp(&gid, dgid, sizeof(gid))) { 678 cma_dev = cur_dev; 679 sgid = gid; 680 id_priv->id.port_num = p; 681 goto found; 682 } 683 684 if (!cma_dev && (gid.global.subnet_prefix == 685 dgid->global.subnet_prefix)) { 686 cma_dev = cur_dev; 687 sgid = gid; 688 id_priv->id.port_num = p; 689 } 690 } 691 } 692 } 693 694 if (!cma_dev) 695 return -ENODEV; 696 697 found: 698 cma_attach_to_dev(id_priv, cma_dev); 699 addr = (struct sockaddr_ib *) cma_src_addr(id_priv); 700 memcpy(&addr->sib_addr, &sgid, sizeof sgid); 701 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr); 702 return 0; 703 } 704 705 static void cma_deref_id(struct rdma_id_private *id_priv) 706 { 707 if (atomic_dec_and_test(&id_priv->refcount)) 708 complete(&id_priv->comp); 709 } 710 711 static int cma_disable_callback(struct rdma_id_private *id_priv, 712 enum rdma_cm_state state) 713 { 714 mutex_lock(&id_priv->handler_mutex); 715 if (id_priv->state != state) { 716 mutex_unlock(&id_priv->handler_mutex); 717 return -EINVAL; 718 } 719 return 0; 720 } 721 722 struct rdma_cm_id *rdma_create_id(struct net *net, 723 rdma_cm_event_handler event_handler, 724 void *context, enum rdma_port_space ps, 725 enum ib_qp_type qp_type) 726 { 727 struct rdma_id_private *id_priv; 728 729 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL); 730 if (!id_priv) 731 return ERR_PTR(-ENOMEM); 732 733 id_priv->owner = task_pid_nr(current); 734 id_priv->state = RDMA_CM_IDLE; 735 id_priv->id.context = context; 736 id_priv->id.event_handler = event_handler; 737 id_priv->id.ps = ps; 738 id_priv->id.qp_type = qp_type; 739 spin_lock_init(&id_priv->lock); 740 mutex_init(&id_priv->qp_mutex); 741 init_completion(&id_priv->comp); 742 atomic_set(&id_priv->refcount, 1); 743 mutex_init(&id_priv->handler_mutex); 744 INIT_LIST_HEAD(&id_priv->listen_list); 745 INIT_LIST_HEAD(&id_priv->mc_list); 746 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num); 747 id_priv->id.route.addr.dev_addr.net = get_net(net); 748 749 return &id_priv->id; 750 } 751 EXPORT_SYMBOL(rdma_create_id); 752 753 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp) 754 { 755 struct ib_qp_attr qp_attr; 756 int qp_attr_mask, ret; 757 758 qp_attr.qp_state = IB_QPS_INIT; 759 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 760 if (ret) 761 return ret; 762 763 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 764 if (ret) 765 return ret; 766 767 qp_attr.qp_state = IB_QPS_RTR; 768 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE); 769 if (ret) 770 return ret; 771 772 qp_attr.qp_state = IB_QPS_RTS; 773 qp_attr.sq_psn = 0; 774 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN); 775 776 return ret; 777 } 778 779 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp) 780 { 781 struct ib_qp_attr qp_attr; 782 int qp_attr_mask, ret; 783 784 qp_attr.qp_state = IB_QPS_INIT; 785 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 786 if (ret) 787 return ret; 788 789 return ib_modify_qp(qp, &qp_attr, qp_attr_mask); 790 } 791 792 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd, 793 struct ib_qp_init_attr *qp_init_attr) 794 { 795 struct rdma_id_private *id_priv; 796 struct ib_qp *qp; 797 int ret; 798 799 id_priv = container_of(id, struct rdma_id_private, id); 800 if (id->device != pd->device) 801 return -EINVAL; 802 803 qp_init_attr->port_num = id->port_num; 804 qp = ib_create_qp(pd, qp_init_attr); 805 if (IS_ERR(qp)) 806 return PTR_ERR(qp); 807 808 if (id->qp_type == IB_QPT_UD) 809 ret = cma_init_ud_qp(id_priv, qp); 810 else 811 ret = cma_init_conn_qp(id_priv, qp); 812 if (ret) 813 goto err; 814 815 id->qp = qp; 816 id_priv->qp_num = qp->qp_num; 817 id_priv->srq = (qp->srq != NULL); 818 return 0; 819 err: 820 ib_destroy_qp(qp); 821 return ret; 822 } 823 EXPORT_SYMBOL(rdma_create_qp); 824 825 void rdma_destroy_qp(struct rdma_cm_id *id) 826 { 827 struct rdma_id_private *id_priv; 828 829 id_priv = container_of(id, struct rdma_id_private, id); 830 mutex_lock(&id_priv->qp_mutex); 831 ib_destroy_qp(id_priv->id.qp); 832 id_priv->id.qp = NULL; 833 mutex_unlock(&id_priv->qp_mutex); 834 } 835 EXPORT_SYMBOL(rdma_destroy_qp); 836 837 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv, 838 struct rdma_conn_param *conn_param) 839 { 840 struct ib_qp_attr qp_attr; 841 int qp_attr_mask, ret; 842 union ib_gid sgid; 843 844 mutex_lock(&id_priv->qp_mutex); 845 if (!id_priv->id.qp) { 846 ret = 0; 847 goto out; 848 } 849 850 /* Need to update QP attributes from default values. */ 851 qp_attr.qp_state = IB_QPS_INIT; 852 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 853 if (ret) 854 goto out; 855 856 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 857 if (ret) 858 goto out; 859 860 qp_attr.qp_state = IB_QPS_RTR; 861 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 862 if (ret) 863 goto out; 864 865 ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num, 866 qp_attr.ah_attr.grh.sgid_index, &sgid, NULL); 867 if (ret) 868 goto out; 869 870 BUG_ON(id_priv->cma_dev->device != id_priv->id.device); 871 872 if (conn_param) 873 qp_attr.max_dest_rd_atomic = conn_param->responder_resources; 874 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 875 out: 876 mutex_unlock(&id_priv->qp_mutex); 877 return ret; 878 } 879 880 static int cma_modify_qp_rts(struct rdma_id_private *id_priv, 881 struct rdma_conn_param *conn_param) 882 { 883 struct ib_qp_attr qp_attr; 884 int qp_attr_mask, ret; 885 886 mutex_lock(&id_priv->qp_mutex); 887 if (!id_priv->id.qp) { 888 ret = 0; 889 goto out; 890 } 891 892 qp_attr.qp_state = IB_QPS_RTS; 893 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 894 if (ret) 895 goto out; 896 897 if (conn_param) 898 qp_attr.max_rd_atomic = conn_param->initiator_depth; 899 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 900 out: 901 mutex_unlock(&id_priv->qp_mutex); 902 return ret; 903 } 904 905 static int cma_modify_qp_err(struct rdma_id_private *id_priv) 906 { 907 struct ib_qp_attr qp_attr; 908 int ret; 909 910 mutex_lock(&id_priv->qp_mutex); 911 if (!id_priv->id.qp) { 912 ret = 0; 913 goto out; 914 } 915 916 qp_attr.qp_state = IB_QPS_ERR; 917 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE); 918 out: 919 mutex_unlock(&id_priv->qp_mutex); 920 return ret; 921 } 922 923 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv, 924 struct ib_qp_attr *qp_attr, int *qp_attr_mask) 925 { 926 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 927 int ret; 928 u16 pkey; 929 930 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num)) 931 pkey = 0xffff; 932 else 933 pkey = ib_addr_get_pkey(dev_addr); 934 935 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num, 936 pkey, &qp_attr->pkey_index); 937 if (ret) 938 return ret; 939 940 qp_attr->port_num = id_priv->id.port_num; 941 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT; 942 943 if (id_priv->id.qp_type == IB_QPT_UD) { 944 ret = cma_set_qkey(id_priv, 0); 945 if (ret) 946 return ret; 947 948 qp_attr->qkey = id_priv->qkey; 949 *qp_attr_mask |= IB_QP_QKEY; 950 } else { 951 qp_attr->qp_access_flags = 0; 952 *qp_attr_mask |= IB_QP_ACCESS_FLAGS; 953 } 954 return 0; 955 } 956 957 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr, 958 int *qp_attr_mask) 959 { 960 struct rdma_id_private *id_priv; 961 int ret = 0; 962 963 id_priv = container_of(id, struct rdma_id_private, id); 964 if (rdma_cap_ib_cm(id->device, id->port_num)) { 965 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD)) 966 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask); 967 else 968 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr, 969 qp_attr_mask); 970 971 if (qp_attr->qp_state == IB_QPS_RTR) 972 qp_attr->rq_psn = id_priv->seq_num; 973 } else if (rdma_cap_iw_cm(id->device, id->port_num)) { 974 if (!id_priv->cm_id.iw) { 975 qp_attr->qp_access_flags = 0; 976 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS; 977 } else 978 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr, 979 qp_attr_mask); 980 } else 981 ret = -ENOSYS; 982 983 return ret; 984 } 985 EXPORT_SYMBOL(rdma_init_qp_attr); 986 987 static inline int cma_zero_addr(struct sockaddr *addr) 988 { 989 switch (addr->sa_family) { 990 case AF_INET: 991 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr); 992 case AF_INET6: 993 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr); 994 case AF_IB: 995 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr); 996 default: 997 return 0; 998 } 999 } 1000 1001 static inline int cma_loopback_addr(struct sockaddr *addr) 1002 { 1003 switch (addr->sa_family) { 1004 case AF_INET: 1005 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr); 1006 case AF_INET6: 1007 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr); 1008 case AF_IB: 1009 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr); 1010 default: 1011 return 0; 1012 } 1013 } 1014 1015 static inline int cma_any_addr(struct sockaddr *addr) 1016 { 1017 return cma_zero_addr(addr) || cma_loopback_addr(addr); 1018 } 1019 1020 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst) 1021 { 1022 if (src->sa_family != dst->sa_family) 1023 return -1; 1024 1025 switch (src->sa_family) { 1026 case AF_INET: 1027 return ((struct sockaddr_in *) src)->sin_addr.s_addr != 1028 ((struct sockaddr_in *) dst)->sin_addr.s_addr; 1029 case AF_INET6: 1030 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr, 1031 &((struct sockaddr_in6 *) dst)->sin6_addr); 1032 default: 1033 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr, 1034 &((struct sockaddr_ib *) dst)->sib_addr); 1035 } 1036 } 1037 1038 static __be16 cma_port(struct sockaddr *addr) 1039 { 1040 struct sockaddr_ib *sib; 1041 1042 switch (addr->sa_family) { 1043 case AF_INET: 1044 return ((struct sockaddr_in *) addr)->sin_port; 1045 case AF_INET6: 1046 return ((struct sockaddr_in6 *) addr)->sin6_port; 1047 case AF_IB: 1048 sib = (struct sockaddr_ib *) addr; 1049 return htons((u16) (be64_to_cpu(sib->sib_sid) & 1050 be64_to_cpu(sib->sib_sid_mask))); 1051 default: 1052 return 0; 1053 } 1054 } 1055 1056 static inline int cma_any_port(struct sockaddr *addr) 1057 { 1058 return !cma_port(addr); 1059 } 1060 1061 static void cma_save_ib_info(struct sockaddr *src_addr, 1062 struct sockaddr *dst_addr, 1063 struct rdma_cm_id *listen_id, 1064 struct ib_sa_path_rec *path) 1065 { 1066 struct sockaddr_ib *listen_ib, *ib; 1067 1068 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr; 1069 if (src_addr) { 1070 ib = (struct sockaddr_ib *)src_addr; 1071 ib->sib_family = AF_IB; 1072 if (path) { 1073 ib->sib_pkey = path->pkey; 1074 ib->sib_flowinfo = path->flow_label; 1075 memcpy(&ib->sib_addr, &path->sgid, 16); 1076 ib->sib_sid = path->service_id; 1077 ib->sib_scope_id = 0; 1078 } else { 1079 ib->sib_pkey = listen_ib->sib_pkey; 1080 ib->sib_flowinfo = listen_ib->sib_flowinfo; 1081 ib->sib_addr = listen_ib->sib_addr; 1082 ib->sib_sid = listen_ib->sib_sid; 1083 ib->sib_scope_id = listen_ib->sib_scope_id; 1084 } 1085 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL); 1086 } 1087 if (dst_addr) { 1088 ib = (struct sockaddr_ib *)dst_addr; 1089 ib->sib_family = AF_IB; 1090 if (path) { 1091 ib->sib_pkey = path->pkey; 1092 ib->sib_flowinfo = path->flow_label; 1093 memcpy(&ib->sib_addr, &path->dgid, 16); 1094 } 1095 } 1096 } 1097 1098 static void cma_save_ip4_info(struct sockaddr *src_addr, 1099 struct sockaddr *dst_addr, 1100 struct cma_hdr *hdr, 1101 __be16 local_port) 1102 { 1103 struct sockaddr_in *ip4; 1104 1105 if (src_addr) { 1106 ip4 = (struct sockaddr_in *)src_addr; 1107 ip4->sin_family = AF_INET; 1108 ip4->sin_addr.s_addr = hdr->dst_addr.ip4.addr; 1109 ip4->sin_port = local_port; 1110 } 1111 1112 if (dst_addr) { 1113 ip4 = (struct sockaddr_in *)dst_addr; 1114 ip4->sin_family = AF_INET; 1115 ip4->sin_addr.s_addr = hdr->src_addr.ip4.addr; 1116 ip4->sin_port = hdr->port; 1117 } 1118 } 1119 1120 static void cma_save_ip6_info(struct sockaddr *src_addr, 1121 struct sockaddr *dst_addr, 1122 struct cma_hdr *hdr, 1123 __be16 local_port) 1124 { 1125 struct sockaddr_in6 *ip6; 1126 1127 if (src_addr) { 1128 ip6 = (struct sockaddr_in6 *)src_addr; 1129 ip6->sin6_family = AF_INET6; 1130 ip6->sin6_addr = hdr->dst_addr.ip6; 1131 ip6->sin6_port = local_port; 1132 } 1133 1134 if (dst_addr) { 1135 ip6 = (struct sockaddr_in6 *)dst_addr; 1136 ip6->sin6_family = AF_INET6; 1137 ip6->sin6_addr = hdr->src_addr.ip6; 1138 ip6->sin6_port = hdr->port; 1139 } 1140 } 1141 1142 static u16 cma_port_from_service_id(__be64 service_id) 1143 { 1144 return (u16)be64_to_cpu(service_id); 1145 } 1146 1147 static int cma_save_ip_info(struct sockaddr *src_addr, 1148 struct sockaddr *dst_addr, 1149 struct ib_cm_event *ib_event, 1150 __be64 service_id) 1151 { 1152 struct cma_hdr *hdr; 1153 __be16 port; 1154 1155 hdr = ib_event->private_data; 1156 if (hdr->cma_version != CMA_VERSION) 1157 return -EINVAL; 1158 1159 port = htons(cma_port_from_service_id(service_id)); 1160 1161 switch (cma_get_ip_ver(hdr)) { 1162 case 4: 1163 cma_save_ip4_info(src_addr, dst_addr, hdr, port); 1164 break; 1165 case 6: 1166 cma_save_ip6_info(src_addr, dst_addr, hdr, port); 1167 break; 1168 default: 1169 return -EAFNOSUPPORT; 1170 } 1171 1172 return 0; 1173 } 1174 1175 static int cma_save_net_info(struct sockaddr *src_addr, 1176 struct sockaddr *dst_addr, 1177 struct rdma_cm_id *listen_id, 1178 struct ib_cm_event *ib_event, 1179 sa_family_t sa_family, __be64 service_id) 1180 { 1181 if (sa_family == AF_IB) { 1182 if (ib_event->event == IB_CM_REQ_RECEIVED) 1183 cma_save_ib_info(src_addr, dst_addr, listen_id, 1184 ib_event->param.req_rcvd.primary_path); 1185 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) 1186 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL); 1187 return 0; 1188 } 1189 1190 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id); 1191 } 1192 1193 static int cma_save_req_info(const struct ib_cm_event *ib_event, 1194 struct cma_req_info *req) 1195 { 1196 const struct ib_cm_req_event_param *req_param = 1197 &ib_event->param.req_rcvd; 1198 const struct ib_cm_sidr_req_event_param *sidr_param = 1199 &ib_event->param.sidr_req_rcvd; 1200 1201 switch (ib_event->event) { 1202 case IB_CM_REQ_RECEIVED: 1203 req->device = req_param->listen_id->device; 1204 req->port = req_param->port; 1205 memcpy(&req->local_gid, &req_param->primary_path->sgid, 1206 sizeof(req->local_gid)); 1207 req->has_gid = true; 1208 req->service_id = req_param->primary_path->service_id; 1209 req->pkey = be16_to_cpu(req_param->primary_path->pkey); 1210 if (req->pkey != req_param->bth_pkey) 1211 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n" 1212 "RDMA CMA: in the future this may cause the request to be dropped\n", 1213 req_param->bth_pkey, req->pkey); 1214 break; 1215 case IB_CM_SIDR_REQ_RECEIVED: 1216 req->device = sidr_param->listen_id->device; 1217 req->port = sidr_param->port; 1218 req->has_gid = false; 1219 req->service_id = sidr_param->service_id; 1220 req->pkey = sidr_param->pkey; 1221 if (req->pkey != sidr_param->bth_pkey) 1222 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n" 1223 "RDMA CMA: in the future this may cause the request to be dropped\n", 1224 sidr_param->bth_pkey, req->pkey); 1225 break; 1226 default: 1227 return -EINVAL; 1228 } 1229 1230 return 0; 1231 } 1232 1233 static bool validate_ipv4_net_dev(struct net_device *net_dev, 1234 const struct sockaddr_in *dst_addr, 1235 const struct sockaddr_in *src_addr) 1236 { 1237 __be32 daddr = dst_addr->sin_addr.s_addr, 1238 saddr = src_addr->sin_addr.s_addr; 1239 struct fib_result res; 1240 struct flowi4 fl4; 1241 int err; 1242 bool ret; 1243 1244 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) || 1245 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) || 1246 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) || 1247 ipv4_is_loopback(saddr)) 1248 return false; 1249 1250 memset(&fl4, 0, sizeof(fl4)); 1251 fl4.flowi4_iif = net_dev->ifindex; 1252 fl4.daddr = daddr; 1253 fl4.saddr = saddr; 1254 1255 rcu_read_lock(); 1256 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0); 1257 ret = err == 0 && FIB_RES_DEV(res) == net_dev; 1258 rcu_read_unlock(); 1259 1260 return ret; 1261 } 1262 1263 static bool validate_ipv6_net_dev(struct net_device *net_dev, 1264 const struct sockaddr_in6 *dst_addr, 1265 const struct sockaddr_in6 *src_addr) 1266 { 1267 #if IS_ENABLED(CONFIG_IPV6) 1268 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) & 1269 IPV6_ADDR_LINKLOCAL; 1270 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr, 1271 &src_addr->sin6_addr, net_dev->ifindex, 1272 strict); 1273 bool ret; 1274 1275 if (!rt) 1276 return false; 1277 1278 ret = rt->rt6i_idev->dev == net_dev; 1279 ip6_rt_put(rt); 1280 1281 return ret; 1282 #else 1283 return false; 1284 #endif 1285 } 1286 1287 static bool validate_net_dev(struct net_device *net_dev, 1288 const struct sockaddr *daddr, 1289 const struct sockaddr *saddr) 1290 { 1291 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr; 1292 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr; 1293 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr; 1294 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr; 1295 1296 switch (daddr->sa_family) { 1297 case AF_INET: 1298 return saddr->sa_family == AF_INET && 1299 validate_ipv4_net_dev(net_dev, daddr4, saddr4); 1300 1301 case AF_INET6: 1302 return saddr->sa_family == AF_INET6 && 1303 validate_ipv6_net_dev(net_dev, daddr6, saddr6); 1304 1305 default: 1306 return false; 1307 } 1308 } 1309 1310 static struct net_device *cma_get_net_dev(struct ib_cm_event *ib_event, 1311 const struct cma_req_info *req) 1312 { 1313 struct sockaddr_storage listen_addr_storage, src_addr_storage; 1314 struct sockaddr *listen_addr = (struct sockaddr *)&listen_addr_storage, 1315 *src_addr = (struct sockaddr *)&src_addr_storage; 1316 struct net_device *net_dev; 1317 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL; 1318 int err; 1319 1320 err = cma_save_ip_info(listen_addr, src_addr, ib_event, 1321 req->service_id); 1322 if (err) 1323 return ERR_PTR(err); 1324 1325 net_dev = ib_get_net_dev_by_params(req->device, req->port, req->pkey, 1326 gid, listen_addr); 1327 if (!net_dev) 1328 return ERR_PTR(-ENODEV); 1329 1330 if (!validate_net_dev(net_dev, listen_addr, src_addr)) { 1331 dev_put(net_dev); 1332 return ERR_PTR(-EHOSTUNREACH); 1333 } 1334 1335 return net_dev; 1336 } 1337 1338 static enum rdma_port_space rdma_ps_from_service_id(__be64 service_id) 1339 { 1340 return (be64_to_cpu(service_id) >> 16) & 0xffff; 1341 } 1342 1343 static bool cma_match_private_data(struct rdma_id_private *id_priv, 1344 const struct cma_hdr *hdr) 1345 { 1346 struct sockaddr *addr = cma_src_addr(id_priv); 1347 __be32 ip4_addr; 1348 struct in6_addr ip6_addr; 1349 1350 if (cma_any_addr(addr) && !id_priv->afonly) 1351 return true; 1352 1353 switch (addr->sa_family) { 1354 case AF_INET: 1355 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr; 1356 if (cma_get_ip_ver(hdr) != 4) 1357 return false; 1358 if (!cma_any_addr(addr) && 1359 hdr->dst_addr.ip4.addr != ip4_addr) 1360 return false; 1361 break; 1362 case AF_INET6: 1363 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr; 1364 if (cma_get_ip_ver(hdr) != 6) 1365 return false; 1366 if (!cma_any_addr(addr) && 1367 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr))) 1368 return false; 1369 break; 1370 case AF_IB: 1371 return true; 1372 default: 1373 return false; 1374 } 1375 1376 return true; 1377 } 1378 1379 static bool cma_protocol_roce_dev_port(struct ib_device *device, int port_num) 1380 { 1381 enum rdma_link_layer ll = rdma_port_get_link_layer(device, port_num); 1382 enum rdma_transport_type transport = 1383 rdma_node_get_transport(device->node_type); 1384 1385 return ll == IB_LINK_LAYER_ETHERNET && transport == RDMA_TRANSPORT_IB; 1386 } 1387 1388 static bool cma_protocol_roce(const struct rdma_cm_id *id) 1389 { 1390 struct ib_device *device = id->device; 1391 const int port_num = id->port_num ?: rdma_start_port(device); 1392 1393 return cma_protocol_roce_dev_port(device, port_num); 1394 } 1395 1396 static bool cma_match_net_dev(const struct rdma_cm_id *id, 1397 const struct net_device *net_dev, 1398 u8 port_num) 1399 { 1400 const struct rdma_addr *addr = &id->route.addr; 1401 1402 if (!net_dev) 1403 /* This request is an AF_IB request or a RoCE request */ 1404 return (!id->port_num || id->port_num == port_num) && 1405 (addr->src_addr.ss_family == AF_IB || 1406 cma_protocol_roce_dev_port(id->device, port_num)); 1407 1408 return !addr->dev_addr.bound_dev_if || 1409 (net_eq(dev_net(net_dev), addr->dev_addr.net) && 1410 addr->dev_addr.bound_dev_if == net_dev->ifindex); 1411 } 1412 1413 static struct rdma_id_private *cma_find_listener( 1414 const struct rdma_bind_list *bind_list, 1415 const struct ib_cm_id *cm_id, 1416 const struct ib_cm_event *ib_event, 1417 const struct cma_req_info *req, 1418 const struct net_device *net_dev) 1419 { 1420 struct rdma_id_private *id_priv, *id_priv_dev; 1421 1422 if (!bind_list) 1423 return ERR_PTR(-EINVAL); 1424 1425 hlist_for_each_entry(id_priv, &bind_list->owners, node) { 1426 if (cma_match_private_data(id_priv, ib_event->private_data)) { 1427 if (id_priv->id.device == cm_id->device && 1428 cma_match_net_dev(&id_priv->id, net_dev, req->port)) 1429 return id_priv; 1430 list_for_each_entry(id_priv_dev, 1431 &id_priv->listen_list, 1432 listen_list) { 1433 if (id_priv_dev->id.device == cm_id->device && 1434 cma_match_net_dev(&id_priv_dev->id, net_dev, req->port)) 1435 return id_priv_dev; 1436 } 1437 } 1438 } 1439 1440 return ERR_PTR(-EINVAL); 1441 } 1442 1443 static struct rdma_id_private *cma_id_from_event(struct ib_cm_id *cm_id, 1444 struct ib_cm_event *ib_event, 1445 struct net_device **net_dev) 1446 { 1447 struct cma_req_info req; 1448 struct rdma_bind_list *bind_list; 1449 struct rdma_id_private *id_priv; 1450 int err; 1451 1452 err = cma_save_req_info(ib_event, &req); 1453 if (err) 1454 return ERR_PTR(err); 1455 1456 *net_dev = cma_get_net_dev(ib_event, &req); 1457 if (IS_ERR(*net_dev)) { 1458 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) { 1459 /* Assuming the protocol is AF_IB */ 1460 *net_dev = NULL; 1461 } else if (cma_protocol_roce_dev_port(req.device, req.port)) { 1462 /* TODO find the net dev matching the request parameters 1463 * through the RoCE GID table */ 1464 *net_dev = NULL; 1465 } else { 1466 return ERR_CAST(*net_dev); 1467 } 1468 } 1469 1470 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net, 1471 rdma_ps_from_service_id(req.service_id), 1472 cma_port_from_service_id(req.service_id)); 1473 id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev); 1474 if (IS_ERR(id_priv) && *net_dev) { 1475 dev_put(*net_dev); 1476 *net_dev = NULL; 1477 } 1478 1479 return id_priv; 1480 } 1481 1482 static inline int cma_user_data_offset(struct rdma_id_private *id_priv) 1483 { 1484 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr); 1485 } 1486 1487 static void cma_cancel_route(struct rdma_id_private *id_priv) 1488 { 1489 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) { 1490 if (id_priv->query) 1491 ib_sa_cancel_query(id_priv->query_id, id_priv->query); 1492 } 1493 } 1494 1495 static void cma_cancel_listens(struct rdma_id_private *id_priv) 1496 { 1497 struct rdma_id_private *dev_id_priv; 1498 1499 /* 1500 * Remove from listen_any_list to prevent added devices from spawning 1501 * additional listen requests. 1502 */ 1503 mutex_lock(&lock); 1504 list_del(&id_priv->list); 1505 1506 while (!list_empty(&id_priv->listen_list)) { 1507 dev_id_priv = list_entry(id_priv->listen_list.next, 1508 struct rdma_id_private, listen_list); 1509 /* sync with device removal to avoid duplicate destruction */ 1510 list_del_init(&dev_id_priv->list); 1511 list_del(&dev_id_priv->listen_list); 1512 mutex_unlock(&lock); 1513 1514 rdma_destroy_id(&dev_id_priv->id); 1515 mutex_lock(&lock); 1516 } 1517 mutex_unlock(&lock); 1518 } 1519 1520 static void cma_cancel_operation(struct rdma_id_private *id_priv, 1521 enum rdma_cm_state state) 1522 { 1523 switch (state) { 1524 case RDMA_CM_ADDR_QUERY: 1525 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr); 1526 break; 1527 case RDMA_CM_ROUTE_QUERY: 1528 cma_cancel_route(id_priv); 1529 break; 1530 case RDMA_CM_LISTEN: 1531 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev) 1532 cma_cancel_listens(id_priv); 1533 break; 1534 default: 1535 break; 1536 } 1537 } 1538 1539 static void cma_release_port(struct rdma_id_private *id_priv) 1540 { 1541 struct rdma_bind_list *bind_list = id_priv->bind_list; 1542 struct net *net = id_priv->id.route.addr.dev_addr.net; 1543 1544 if (!bind_list) 1545 return; 1546 1547 mutex_lock(&lock); 1548 hlist_del(&id_priv->node); 1549 if (hlist_empty(&bind_list->owners)) { 1550 cma_ps_remove(net, bind_list->ps, bind_list->port); 1551 kfree(bind_list); 1552 } 1553 mutex_unlock(&lock); 1554 } 1555 1556 static void cma_leave_mc_groups(struct rdma_id_private *id_priv) 1557 { 1558 struct cma_multicast *mc; 1559 1560 while (!list_empty(&id_priv->mc_list)) { 1561 mc = container_of(id_priv->mc_list.next, 1562 struct cma_multicast, list); 1563 list_del(&mc->list); 1564 if (rdma_cap_ib_mcast(id_priv->cma_dev->device, 1565 id_priv->id.port_num)) { 1566 ib_sa_free_multicast(mc->multicast.ib); 1567 kfree(mc); 1568 } else { 1569 if (mc->igmp_joined) { 1570 struct rdma_dev_addr *dev_addr = 1571 &id_priv->id.route.addr.dev_addr; 1572 struct net_device *ndev = NULL; 1573 1574 if (dev_addr->bound_dev_if) 1575 ndev = dev_get_by_index(&init_net, 1576 dev_addr->bound_dev_if); 1577 if (ndev) { 1578 cma_igmp_send(ndev, 1579 &mc->multicast.ib->rec.mgid, 1580 false); 1581 dev_put(ndev); 1582 } 1583 } 1584 kref_put(&mc->mcref, release_mc); 1585 } 1586 } 1587 } 1588 1589 void rdma_destroy_id(struct rdma_cm_id *id) 1590 { 1591 struct rdma_id_private *id_priv; 1592 enum rdma_cm_state state; 1593 1594 id_priv = container_of(id, struct rdma_id_private, id); 1595 state = cma_exch(id_priv, RDMA_CM_DESTROYING); 1596 cma_cancel_operation(id_priv, state); 1597 1598 /* 1599 * Wait for any active callback to finish. New callbacks will find 1600 * the id_priv state set to destroying and abort. 1601 */ 1602 mutex_lock(&id_priv->handler_mutex); 1603 mutex_unlock(&id_priv->handler_mutex); 1604 1605 if (id_priv->cma_dev) { 1606 if (rdma_cap_ib_cm(id_priv->id.device, 1)) { 1607 if (id_priv->cm_id.ib) 1608 ib_destroy_cm_id(id_priv->cm_id.ib); 1609 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) { 1610 if (id_priv->cm_id.iw) 1611 iw_destroy_cm_id(id_priv->cm_id.iw); 1612 } 1613 cma_leave_mc_groups(id_priv); 1614 cma_release_dev(id_priv); 1615 } 1616 1617 cma_release_port(id_priv); 1618 cma_deref_id(id_priv); 1619 wait_for_completion(&id_priv->comp); 1620 1621 if (id_priv->internal_id) 1622 cma_deref_id(id_priv->id.context); 1623 1624 kfree(id_priv->id.route.path_rec); 1625 put_net(id_priv->id.route.addr.dev_addr.net); 1626 kfree(id_priv); 1627 } 1628 EXPORT_SYMBOL(rdma_destroy_id); 1629 1630 static int cma_rep_recv(struct rdma_id_private *id_priv) 1631 { 1632 int ret; 1633 1634 ret = cma_modify_qp_rtr(id_priv, NULL); 1635 if (ret) 1636 goto reject; 1637 1638 ret = cma_modify_qp_rts(id_priv, NULL); 1639 if (ret) 1640 goto reject; 1641 1642 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0); 1643 if (ret) 1644 goto reject; 1645 1646 return 0; 1647 reject: 1648 cma_modify_qp_err(id_priv); 1649 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED, 1650 NULL, 0, NULL, 0); 1651 return ret; 1652 } 1653 1654 static void cma_set_rep_event_data(struct rdma_cm_event *event, 1655 struct ib_cm_rep_event_param *rep_data, 1656 void *private_data) 1657 { 1658 event->param.conn.private_data = private_data; 1659 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE; 1660 event->param.conn.responder_resources = rep_data->responder_resources; 1661 event->param.conn.initiator_depth = rep_data->initiator_depth; 1662 event->param.conn.flow_control = rep_data->flow_control; 1663 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count; 1664 event->param.conn.srq = rep_data->srq; 1665 event->param.conn.qp_num = rep_data->remote_qpn; 1666 } 1667 1668 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event) 1669 { 1670 struct rdma_id_private *id_priv = cm_id->context; 1671 struct rdma_cm_event event; 1672 int ret = 0; 1673 1674 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT && 1675 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) || 1676 (ib_event->event == IB_CM_TIMEWAIT_EXIT && 1677 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT))) 1678 return 0; 1679 1680 memset(&event, 0, sizeof event); 1681 switch (ib_event->event) { 1682 case IB_CM_REQ_ERROR: 1683 case IB_CM_REP_ERROR: 1684 event.event = RDMA_CM_EVENT_UNREACHABLE; 1685 event.status = -ETIMEDOUT; 1686 break; 1687 case IB_CM_REP_RECEIVED: 1688 if (id_priv->id.qp) { 1689 event.status = cma_rep_recv(id_priv); 1690 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR : 1691 RDMA_CM_EVENT_ESTABLISHED; 1692 } else { 1693 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE; 1694 } 1695 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd, 1696 ib_event->private_data); 1697 break; 1698 case IB_CM_RTU_RECEIVED: 1699 case IB_CM_USER_ESTABLISHED: 1700 event.event = RDMA_CM_EVENT_ESTABLISHED; 1701 break; 1702 case IB_CM_DREQ_ERROR: 1703 event.status = -ETIMEDOUT; /* fall through */ 1704 case IB_CM_DREQ_RECEIVED: 1705 case IB_CM_DREP_RECEIVED: 1706 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT, 1707 RDMA_CM_DISCONNECT)) 1708 goto out; 1709 event.event = RDMA_CM_EVENT_DISCONNECTED; 1710 break; 1711 case IB_CM_TIMEWAIT_EXIT: 1712 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT; 1713 break; 1714 case IB_CM_MRA_RECEIVED: 1715 /* ignore event */ 1716 goto out; 1717 case IB_CM_REJ_RECEIVED: 1718 cma_modify_qp_err(id_priv); 1719 event.status = ib_event->param.rej_rcvd.reason; 1720 event.event = RDMA_CM_EVENT_REJECTED; 1721 event.param.conn.private_data = ib_event->private_data; 1722 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE; 1723 break; 1724 default: 1725 pr_err("RDMA CMA: unexpected IB CM event: %d\n", 1726 ib_event->event); 1727 goto out; 1728 } 1729 1730 ret = id_priv->id.event_handler(&id_priv->id, &event); 1731 if (ret) { 1732 /* Destroy the CM ID by returning a non-zero value. */ 1733 id_priv->cm_id.ib = NULL; 1734 cma_exch(id_priv, RDMA_CM_DESTROYING); 1735 mutex_unlock(&id_priv->handler_mutex); 1736 rdma_destroy_id(&id_priv->id); 1737 return ret; 1738 } 1739 out: 1740 mutex_unlock(&id_priv->handler_mutex); 1741 return ret; 1742 } 1743 1744 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id, 1745 struct ib_cm_event *ib_event, 1746 struct net_device *net_dev) 1747 { 1748 struct rdma_id_private *id_priv; 1749 struct rdma_cm_id *id; 1750 struct rdma_route *rt; 1751 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family; 1752 const __be64 service_id = 1753 ib_event->param.req_rcvd.primary_path->service_id; 1754 int ret; 1755 1756 id = rdma_create_id(listen_id->route.addr.dev_addr.net, 1757 listen_id->event_handler, listen_id->context, 1758 listen_id->ps, ib_event->param.req_rcvd.qp_type); 1759 if (IS_ERR(id)) 1760 return NULL; 1761 1762 id_priv = container_of(id, struct rdma_id_private, id); 1763 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr, 1764 (struct sockaddr *)&id->route.addr.dst_addr, 1765 listen_id, ib_event, ss_family, service_id)) 1766 goto err; 1767 1768 rt = &id->route; 1769 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1; 1770 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths, 1771 GFP_KERNEL); 1772 if (!rt->path_rec) 1773 goto err; 1774 1775 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path; 1776 if (rt->num_paths == 2) 1777 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path; 1778 1779 if (net_dev) { 1780 ret = rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL); 1781 if (ret) 1782 goto err; 1783 } else { 1784 if (!cma_protocol_roce(listen_id) && 1785 cma_any_addr(cma_src_addr(id_priv))) { 1786 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND; 1787 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid); 1788 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey)); 1789 } else if (!cma_any_addr(cma_src_addr(id_priv))) { 1790 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr); 1791 if (ret) 1792 goto err; 1793 } 1794 } 1795 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid); 1796 1797 id_priv->state = RDMA_CM_CONNECT; 1798 return id_priv; 1799 1800 err: 1801 rdma_destroy_id(id); 1802 return NULL; 1803 } 1804 1805 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id, 1806 struct ib_cm_event *ib_event, 1807 struct net_device *net_dev) 1808 { 1809 struct rdma_id_private *id_priv; 1810 struct rdma_cm_id *id; 1811 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family; 1812 struct net *net = listen_id->route.addr.dev_addr.net; 1813 int ret; 1814 1815 id = rdma_create_id(net, listen_id->event_handler, listen_id->context, 1816 listen_id->ps, IB_QPT_UD); 1817 if (IS_ERR(id)) 1818 return NULL; 1819 1820 id_priv = container_of(id, struct rdma_id_private, id); 1821 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr, 1822 (struct sockaddr *)&id->route.addr.dst_addr, 1823 listen_id, ib_event, ss_family, 1824 ib_event->param.sidr_req_rcvd.service_id)) 1825 goto err; 1826 1827 if (net_dev) { 1828 ret = rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL); 1829 if (ret) 1830 goto err; 1831 } else { 1832 if (!cma_any_addr(cma_src_addr(id_priv))) { 1833 ret = cma_translate_addr(cma_src_addr(id_priv), 1834 &id->route.addr.dev_addr); 1835 if (ret) 1836 goto err; 1837 } 1838 } 1839 1840 id_priv->state = RDMA_CM_CONNECT; 1841 return id_priv; 1842 err: 1843 rdma_destroy_id(id); 1844 return NULL; 1845 } 1846 1847 static void cma_set_req_event_data(struct rdma_cm_event *event, 1848 struct ib_cm_req_event_param *req_data, 1849 void *private_data, int offset) 1850 { 1851 event->param.conn.private_data = private_data + offset; 1852 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset; 1853 event->param.conn.responder_resources = req_data->responder_resources; 1854 event->param.conn.initiator_depth = req_data->initiator_depth; 1855 event->param.conn.flow_control = req_data->flow_control; 1856 event->param.conn.retry_count = req_data->retry_count; 1857 event->param.conn.rnr_retry_count = req_data->rnr_retry_count; 1858 event->param.conn.srq = req_data->srq; 1859 event->param.conn.qp_num = req_data->remote_qpn; 1860 } 1861 1862 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event) 1863 { 1864 return (((ib_event->event == IB_CM_REQ_RECEIVED) && 1865 (ib_event->param.req_rcvd.qp_type == id->qp_type)) || 1866 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) && 1867 (id->qp_type == IB_QPT_UD)) || 1868 (!id->qp_type)); 1869 } 1870 1871 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event) 1872 { 1873 struct rdma_id_private *listen_id, *conn_id; 1874 struct rdma_cm_event event; 1875 struct net_device *net_dev; 1876 int offset, ret; 1877 1878 listen_id = cma_id_from_event(cm_id, ib_event, &net_dev); 1879 if (IS_ERR(listen_id)) 1880 return PTR_ERR(listen_id); 1881 1882 if (!cma_check_req_qp_type(&listen_id->id, ib_event)) { 1883 ret = -EINVAL; 1884 goto net_dev_put; 1885 } 1886 1887 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN)) { 1888 ret = -ECONNABORTED; 1889 goto net_dev_put; 1890 } 1891 1892 memset(&event, 0, sizeof event); 1893 offset = cma_user_data_offset(listen_id); 1894 event.event = RDMA_CM_EVENT_CONNECT_REQUEST; 1895 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) { 1896 conn_id = cma_new_udp_id(&listen_id->id, ib_event, net_dev); 1897 event.param.ud.private_data = ib_event->private_data + offset; 1898 event.param.ud.private_data_len = 1899 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset; 1900 } else { 1901 conn_id = cma_new_conn_id(&listen_id->id, ib_event, net_dev); 1902 cma_set_req_event_data(&event, &ib_event->param.req_rcvd, 1903 ib_event->private_data, offset); 1904 } 1905 if (!conn_id) { 1906 ret = -ENOMEM; 1907 goto err1; 1908 } 1909 1910 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING); 1911 ret = cma_acquire_dev(conn_id, listen_id); 1912 if (ret) 1913 goto err2; 1914 1915 conn_id->cm_id.ib = cm_id; 1916 cm_id->context = conn_id; 1917 cm_id->cm_handler = cma_ib_handler; 1918 1919 /* 1920 * Protect against the user destroying conn_id from another thread 1921 * until we're done accessing it. 1922 */ 1923 atomic_inc(&conn_id->refcount); 1924 ret = conn_id->id.event_handler(&conn_id->id, &event); 1925 if (ret) 1926 goto err3; 1927 /* 1928 * Acquire mutex to prevent user executing rdma_destroy_id() 1929 * while we're accessing the cm_id. 1930 */ 1931 mutex_lock(&lock); 1932 if (cma_comp(conn_id, RDMA_CM_CONNECT) && 1933 (conn_id->id.qp_type != IB_QPT_UD)) 1934 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0); 1935 mutex_unlock(&lock); 1936 mutex_unlock(&conn_id->handler_mutex); 1937 mutex_unlock(&listen_id->handler_mutex); 1938 cma_deref_id(conn_id); 1939 if (net_dev) 1940 dev_put(net_dev); 1941 return 0; 1942 1943 err3: 1944 cma_deref_id(conn_id); 1945 /* Destroy the CM ID by returning a non-zero value. */ 1946 conn_id->cm_id.ib = NULL; 1947 err2: 1948 cma_exch(conn_id, RDMA_CM_DESTROYING); 1949 mutex_unlock(&conn_id->handler_mutex); 1950 err1: 1951 mutex_unlock(&listen_id->handler_mutex); 1952 if (conn_id) 1953 rdma_destroy_id(&conn_id->id); 1954 1955 net_dev_put: 1956 if (net_dev) 1957 dev_put(net_dev); 1958 1959 return ret; 1960 } 1961 1962 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr) 1963 { 1964 if (addr->sa_family == AF_IB) 1965 return ((struct sockaddr_ib *) addr)->sib_sid; 1966 1967 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr))); 1968 } 1969 EXPORT_SYMBOL(rdma_get_service_id); 1970 1971 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event) 1972 { 1973 struct rdma_id_private *id_priv = iw_id->context; 1974 struct rdma_cm_event event; 1975 int ret = 0; 1976 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr; 1977 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr; 1978 1979 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT)) 1980 return 0; 1981 1982 memset(&event, 0, sizeof event); 1983 switch (iw_event->event) { 1984 case IW_CM_EVENT_CLOSE: 1985 event.event = RDMA_CM_EVENT_DISCONNECTED; 1986 break; 1987 case IW_CM_EVENT_CONNECT_REPLY: 1988 memcpy(cma_src_addr(id_priv), laddr, 1989 rdma_addr_size(laddr)); 1990 memcpy(cma_dst_addr(id_priv), raddr, 1991 rdma_addr_size(raddr)); 1992 switch (iw_event->status) { 1993 case 0: 1994 event.event = RDMA_CM_EVENT_ESTABLISHED; 1995 event.param.conn.initiator_depth = iw_event->ird; 1996 event.param.conn.responder_resources = iw_event->ord; 1997 break; 1998 case -ECONNRESET: 1999 case -ECONNREFUSED: 2000 event.event = RDMA_CM_EVENT_REJECTED; 2001 break; 2002 case -ETIMEDOUT: 2003 event.event = RDMA_CM_EVENT_UNREACHABLE; 2004 break; 2005 default: 2006 event.event = RDMA_CM_EVENT_CONNECT_ERROR; 2007 break; 2008 } 2009 break; 2010 case IW_CM_EVENT_ESTABLISHED: 2011 event.event = RDMA_CM_EVENT_ESTABLISHED; 2012 event.param.conn.initiator_depth = iw_event->ird; 2013 event.param.conn.responder_resources = iw_event->ord; 2014 break; 2015 default: 2016 BUG_ON(1); 2017 } 2018 2019 event.status = iw_event->status; 2020 event.param.conn.private_data = iw_event->private_data; 2021 event.param.conn.private_data_len = iw_event->private_data_len; 2022 ret = id_priv->id.event_handler(&id_priv->id, &event); 2023 if (ret) { 2024 /* Destroy the CM ID by returning a non-zero value. */ 2025 id_priv->cm_id.iw = NULL; 2026 cma_exch(id_priv, RDMA_CM_DESTROYING); 2027 mutex_unlock(&id_priv->handler_mutex); 2028 rdma_destroy_id(&id_priv->id); 2029 return ret; 2030 } 2031 2032 mutex_unlock(&id_priv->handler_mutex); 2033 return ret; 2034 } 2035 2036 static int iw_conn_req_handler(struct iw_cm_id *cm_id, 2037 struct iw_cm_event *iw_event) 2038 { 2039 struct rdma_cm_id *new_cm_id; 2040 struct rdma_id_private *listen_id, *conn_id; 2041 struct rdma_cm_event event; 2042 int ret; 2043 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr; 2044 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr; 2045 2046 listen_id = cm_id->context; 2047 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN)) 2048 return -ECONNABORTED; 2049 2050 /* Create a new RDMA id for the new IW CM ID */ 2051 new_cm_id = rdma_create_id(listen_id->id.route.addr.dev_addr.net, 2052 listen_id->id.event_handler, 2053 listen_id->id.context, 2054 RDMA_PS_TCP, IB_QPT_RC); 2055 if (IS_ERR(new_cm_id)) { 2056 ret = -ENOMEM; 2057 goto out; 2058 } 2059 conn_id = container_of(new_cm_id, struct rdma_id_private, id); 2060 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING); 2061 conn_id->state = RDMA_CM_CONNECT; 2062 2063 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL); 2064 if (ret) { 2065 mutex_unlock(&conn_id->handler_mutex); 2066 rdma_destroy_id(new_cm_id); 2067 goto out; 2068 } 2069 2070 ret = cma_acquire_dev(conn_id, listen_id); 2071 if (ret) { 2072 mutex_unlock(&conn_id->handler_mutex); 2073 rdma_destroy_id(new_cm_id); 2074 goto out; 2075 } 2076 2077 conn_id->cm_id.iw = cm_id; 2078 cm_id->context = conn_id; 2079 cm_id->cm_handler = cma_iw_handler; 2080 2081 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr)); 2082 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr)); 2083 2084 memset(&event, 0, sizeof event); 2085 event.event = RDMA_CM_EVENT_CONNECT_REQUEST; 2086 event.param.conn.private_data = iw_event->private_data; 2087 event.param.conn.private_data_len = iw_event->private_data_len; 2088 event.param.conn.initiator_depth = iw_event->ird; 2089 event.param.conn.responder_resources = iw_event->ord; 2090 2091 /* 2092 * Protect against the user destroying conn_id from another thread 2093 * until we're done accessing it. 2094 */ 2095 atomic_inc(&conn_id->refcount); 2096 ret = conn_id->id.event_handler(&conn_id->id, &event); 2097 if (ret) { 2098 /* User wants to destroy the CM ID */ 2099 conn_id->cm_id.iw = NULL; 2100 cma_exch(conn_id, RDMA_CM_DESTROYING); 2101 mutex_unlock(&conn_id->handler_mutex); 2102 cma_deref_id(conn_id); 2103 rdma_destroy_id(&conn_id->id); 2104 goto out; 2105 } 2106 2107 mutex_unlock(&conn_id->handler_mutex); 2108 cma_deref_id(conn_id); 2109 2110 out: 2111 mutex_unlock(&listen_id->handler_mutex); 2112 return ret; 2113 } 2114 2115 static int cma_ib_listen(struct rdma_id_private *id_priv) 2116 { 2117 struct sockaddr *addr; 2118 struct ib_cm_id *id; 2119 __be64 svc_id; 2120 2121 addr = cma_src_addr(id_priv); 2122 svc_id = rdma_get_service_id(&id_priv->id, addr); 2123 id = ib_cm_insert_listen(id_priv->id.device, cma_req_handler, svc_id); 2124 if (IS_ERR(id)) 2125 return PTR_ERR(id); 2126 id_priv->cm_id.ib = id; 2127 2128 return 0; 2129 } 2130 2131 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog) 2132 { 2133 int ret; 2134 struct iw_cm_id *id; 2135 2136 id = iw_create_cm_id(id_priv->id.device, 2137 iw_conn_req_handler, 2138 id_priv); 2139 if (IS_ERR(id)) 2140 return PTR_ERR(id); 2141 2142 id->tos = id_priv->tos; 2143 id_priv->cm_id.iw = id; 2144 2145 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv), 2146 rdma_addr_size(cma_src_addr(id_priv))); 2147 2148 ret = iw_cm_listen(id_priv->cm_id.iw, backlog); 2149 2150 if (ret) { 2151 iw_destroy_cm_id(id_priv->cm_id.iw); 2152 id_priv->cm_id.iw = NULL; 2153 } 2154 2155 return ret; 2156 } 2157 2158 static int cma_listen_handler(struct rdma_cm_id *id, 2159 struct rdma_cm_event *event) 2160 { 2161 struct rdma_id_private *id_priv = id->context; 2162 2163 id->context = id_priv->id.context; 2164 id->event_handler = id_priv->id.event_handler; 2165 return id_priv->id.event_handler(id, event); 2166 } 2167 2168 static void cma_listen_on_dev(struct rdma_id_private *id_priv, 2169 struct cma_device *cma_dev) 2170 { 2171 struct rdma_id_private *dev_id_priv; 2172 struct rdma_cm_id *id; 2173 struct net *net = id_priv->id.route.addr.dev_addr.net; 2174 int ret; 2175 2176 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1)) 2177 return; 2178 2179 id = rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps, 2180 id_priv->id.qp_type); 2181 if (IS_ERR(id)) 2182 return; 2183 2184 dev_id_priv = container_of(id, struct rdma_id_private, id); 2185 2186 dev_id_priv->state = RDMA_CM_ADDR_BOUND; 2187 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv), 2188 rdma_addr_size(cma_src_addr(id_priv))); 2189 2190 _cma_attach_to_dev(dev_id_priv, cma_dev); 2191 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list); 2192 atomic_inc(&id_priv->refcount); 2193 dev_id_priv->internal_id = 1; 2194 dev_id_priv->afonly = id_priv->afonly; 2195 2196 ret = rdma_listen(id, id_priv->backlog); 2197 if (ret) 2198 pr_warn("RDMA CMA: cma_listen_on_dev, error %d, listening on device %s\n", 2199 ret, cma_dev->device->name); 2200 } 2201 2202 static void cma_listen_on_all(struct rdma_id_private *id_priv) 2203 { 2204 struct cma_device *cma_dev; 2205 2206 mutex_lock(&lock); 2207 list_add_tail(&id_priv->list, &listen_any_list); 2208 list_for_each_entry(cma_dev, &dev_list, list) 2209 cma_listen_on_dev(id_priv, cma_dev); 2210 mutex_unlock(&lock); 2211 } 2212 2213 void rdma_set_service_type(struct rdma_cm_id *id, int tos) 2214 { 2215 struct rdma_id_private *id_priv; 2216 2217 id_priv = container_of(id, struct rdma_id_private, id); 2218 id_priv->tos = (u8) tos; 2219 } 2220 EXPORT_SYMBOL(rdma_set_service_type); 2221 2222 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec, 2223 void *context) 2224 { 2225 struct cma_work *work = context; 2226 struct rdma_route *route; 2227 2228 route = &work->id->id.route; 2229 2230 if (!status) { 2231 route->num_paths = 1; 2232 *route->path_rec = *path_rec; 2233 } else { 2234 work->old_state = RDMA_CM_ROUTE_QUERY; 2235 work->new_state = RDMA_CM_ADDR_RESOLVED; 2236 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR; 2237 work->event.status = status; 2238 } 2239 2240 queue_work(cma_wq, &work->work); 2241 } 2242 2243 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms, 2244 struct cma_work *work) 2245 { 2246 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 2247 struct ib_sa_path_rec path_rec; 2248 ib_sa_comp_mask comp_mask; 2249 struct sockaddr_in6 *sin6; 2250 struct sockaddr_ib *sib; 2251 2252 memset(&path_rec, 0, sizeof path_rec); 2253 rdma_addr_get_sgid(dev_addr, &path_rec.sgid); 2254 rdma_addr_get_dgid(dev_addr, &path_rec.dgid); 2255 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr)); 2256 path_rec.numb_path = 1; 2257 path_rec.reversible = 1; 2258 path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv)); 2259 2260 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID | 2261 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH | 2262 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID; 2263 2264 switch (cma_family(id_priv)) { 2265 case AF_INET: 2266 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos); 2267 comp_mask |= IB_SA_PATH_REC_QOS_CLASS; 2268 break; 2269 case AF_INET6: 2270 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv); 2271 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20); 2272 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS; 2273 break; 2274 case AF_IB: 2275 sib = (struct sockaddr_ib *) cma_src_addr(id_priv); 2276 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20); 2277 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS; 2278 break; 2279 } 2280 2281 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device, 2282 id_priv->id.port_num, &path_rec, 2283 comp_mask, timeout_ms, 2284 GFP_KERNEL, cma_query_handler, 2285 work, &id_priv->query); 2286 2287 return (id_priv->query_id < 0) ? id_priv->query_id : 0; 2288 } 2289 2290 static void cma_work_handler(struct work_struct *_work) 2291 { 2292 struct cma_work *work = container_of(_work, struct cma_work, work); 2293 struct rdma_id_private *id_priv = work->id; 2294 int destroy = 0; 2295 2296 mutex_lock(&id_priv->handler_mutex); 2297 if (!cma_comp_exch(id_priv, work->old_state, work->new_state)) 2298 goto out; 2299 2300 if (id_priv->id.event_handler(&id_priv->id, &work->event)) { 2301 cma_exch(id_priv, RDMA_CM_DESTROYING); 2302 destroy = 1; 2303 } 2304 out: 2305 mutex_unlock(&id_priv->handler_mutex); 2306 cma_deref_id(id_priv); 2307 if (destroy) 2308 rdma_destroy_id(&id_priv->id); 2309 kfree(work); 2310 } 2311 2312 static void cma_ndev_work_handler(struct work_struct *_work) 2313 { 2314 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work); 2315 struct rdma_id_private *id_priv = work->id; 2316 int destroy = 0; 2317 2318 mutex_lock(&id_priv->handler_mutex); 2319 if (id_priv->state == RDMA_CM_DESTROYING || 2320 id_priv->state == RDMA_CM_DEVICE_REMOVAL) 2321 goto out; 2322 2323 if (id_priv->id.event_handler(&id_priv->id, &work->event)) { 2324 cma_exch(id_priv, RDMA_CM_DESTROYING); 2325 destroy = 1; 2326 } 2327 2328 out: 2329 mutex_unlock(&id_priv->handler_mutex); 2330 cma_deref_id(id_priv); 2331 if (destroy) 2332 rdma_destroy_id(&id_priv->id); 2333 kfree(work); 2334 } 2335 2336 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms) 2337 { 2338 struct rdma_route *route = &id_priv->id.route; 2339 struct cma_work *work; 2340 int ret; 2341 2342 work = kzalloc(sizeof *work, GFP_KERNEL); 2343 if (!work) 2344 return -ENOMEM; 2345 2346 work->id = id_priv; 2347 INIT_WORK(&work->work, cma_work_handler); 2348 work->old_state = RDMA_CM_ROUTE_QUERY; 2349 work->new_state = RDMA_CM_ROUTE_RESOLVED; 2350 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 2351 2352 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL); 2353 if (!route->path_rec) { 2354 ret = -ENOMEM; 2355 goto err1; 2356 } 2357 2358 ret = cma_query_ib_route(id_priv, timeout_ms, work); 2359 if (ret) 2360 goto err2; 2361 2362 return 0; 2363 err2: 2364 kfree(route->path_rec); 2365 route->path_rec = NULL; 2366 err1: 2367 kfree(work); 2368 return ret; 2369 } 2370 2371 int rdma_set_ib_paths(struct rdma_cm_id *id, 2372 struct ib_sa_path_rec *path_rec, int num_paths) 2373 { 2374 struct rdma_id_private *id_priv; 2375 int ret; 2376 2377 id_priv = container_of(id, struct rdma_id_private, id); 2378 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, 2379 RDMA_CM_ROUTE_RESOLVED)) 2380 return -EINVAL; 2381 2382 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths, 2383 GFP_KERNEL); 2384 if (!id->route.path_rec) { 2385 ret = -ENOMEM; 2386 goto err; 2387 } 2388 2389 id->route.num_paths = num_paths; 2390 return 0; 2391 err: 2392 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED); 2393 return ret; 2394 } 2395 EXPORT_SYMBOL(rdma_set_ib_paths); 2396 2397 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms) 2398 { 2399 struct cma_work *work; 2400 2401 work = kzalloc(sizeof *work, GFP_KERNEL); 2402 if (!work) 2403 return -ENOMEM; 2404 2405 work->id = id_priv; 2406 INIT_WORK(&work->work, cma_work_handler); 2407 work->old_state = RDMA_CM_ROUTE_QUERY; 2408 work->new_state = RDMA_CM_ROUTE_RESOLVED; 2409 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 2410 queue_work(cma_wq, &work->work); 2411 return 0; 2412 } 2413 2414 static int iboe_tos_to_sl(struct net_device *ndev, int tos) 2415 { 2416 int prio; 2417 struct net_device *dev; 2418 2419 prio = rt_tos2priority(tos); 2420 dev = ndev->priv_flags & IFF_802_1Q_VLAN ? 2421 vlan_dev_real_dev(ndev) : ndev; 2422 2423 if (dev->num_tc) 2424 return netdev_get_prio_tc_map(dev, prio); 2425 2426 #if IS_ENABLED(CONFIG_VLAN_8021Q) 2427 if (ndev->priv_flags & IFF_802_1Q_VLAN) 2428 return (vlan_dev_get_egress_qos_mask(ndev, prio) & 2429 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT; 2430 #endif 2431 return 0; 2432 } 2433 2434 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv) 2435 { 2436 struct rdma_route *route = &id_priv->id.route; 2437 struct rdma_addr *addr = &route->addr; 2438 struct cma_work *work; 2439 int ret; 2440 struct net_device *ndev = NULL; 2441 2442 2443 work = kzalloc(sizeof *work, GFP_KERNEL); 2444 if (!work) 2445 return -ENOMEM; 2446 2447 work->id = id_priv; 2448 INIT_WORK(&work->work, cma_work_handler); 2449 2450 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL); 2451 if (!route->path_rec) { 2452 ret = -ENOMEM; 2453 goto err1; 2454 } 2455 2456 route->num_paths = 1; 2457 2458 if (addr->dev_addr.bound_dev_if) { 2459 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if); 2460 if (!ndev) 2461 return -ENODEV; 2462 2463 if (ndev->flags & IFF_LOOPBACK) { 2464 dev_put(ndev); 2465 if (!id_priv->id.device->get_netdev) 2466 return -EOPNOTSUPP; 2467 2468 ndev = id_priv->id.device->get_netdev(id_priv->id.device, 2469 id_priv->id.port_num); 2470 if (!ndev) 2471 return -ENODEV; 2472 } 2473 2474 route->path_rec->net = &init_net; 2475 route->path_rec->ifindex = ndev->ifindex; 2476 route->path_rec->gid_type = id_priv->gid_type; 2477 } 2478 if (!ndev) { 2479 ret = -ENODEV; 2480 goto err2; 2481 } 2482 2483 memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN); 2484 2485 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr, 2486 &route->path_rec->sgid); 2487 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr, 2488 &route->path_rec->dgid); 2489 2490 /* Use the hint from IP Stack to select GID Type */ 2491 if (route->path_rec->gid_type < ib_network_to_gid_type(addr->dev_addr.network)) 2492 route->path_rec->gid_type = ib_network_to_gid_type(addr->dev_addr.network); 2493 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB) 2494 /* TODO: get the hoplimit from the inet/inet6 device */ 2495 route->path_rec->hop_limit = addr->dev_addr.hoplimit; 2496 else 2497 route->path_rec->hop_limit = 1; 2498 route->path_rec->reversible = 1; 2499 route->path_rec->pkey = cpu_to_be16(0xffff); 2500 route->path_rec->mtu_selector = IB_SA_EQ; 2501 route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos); 2502 route->path_rec->mtu = iboe_get_mtu(ndev->mtu); 2503 route->path_rec->rate_selector = IB_SA_EQ; 2504 route->path_rec->rate = iboe_get_rate(ndev); 2505 dev_put(ndev); 2506 route->path_rec->packet_life_time_selector = IB_SA_EQ; 2507 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME; 2508 if (!route->path_rec->mtu) { 2509 ret = -EINVAL; 2510 goto err2; 2511 } 2512 2513 work->old_state = RDMA_CM_ROUTE_QUERY; 2514 work->new_state = RDMA_CM_ROUTE_RESOLVED; 2515 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 2516 work->event.status = 0; 2517 2518 queue_work(cma_wq, &work->work); 2519 2520 return 0; 2521 2522 err2: 2523 kfree(route->path_rec); 2524 route->path_rec = NULL; 2525 err1: 2526 kfree(work); 2527 return ret; 2528 } 2529 2530 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms) 2531 { 2532 struct rdma_id_private *id_priv; 2533 int ret; 2534 2535 id_priv = container_of(id, struct rdma_id_private, id); 2536 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY)) 2537 return -EINVAL; 2538 2539 atomic_inc(&id_priv->refcount); 2540 if (rdma_cap_ib_sa(id->device, id->port_num)) 2541 ret = cma_resolve_ib_route(id_priv, timeout_ms); 2542 else if (rdma_protocol_roce(id->device, id->port_num)) 2543 ret = cma_resolve_iboe_route(id_priv); 2544 else if (rdma_protocol_iwarp(id->device, id->port_num)) 2545 ret = cma_resolve_iw_route(id_priv, timeout_ms); 2546 else 2547 ret = -ENOSYS; 2548 2549 if (ret) 2550 goto err; 2551 2552 return 0; 2553 err: 2554 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED); 2555 cma_deref_id(id_priv); 2556 return ret; 2557 } 2558 EXPORT_SYMBOL(rdma_resolve_route); 2559 2560 static void cma_set_loopback(struct sockaddr *addr) 2561 { 2562 switch (addr->sa_family) { 2563 case AF_INET: 2564 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK); 2565 break; 2566 case AF_INET6: 2567 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr, 2568 0, 0, 0, htonl(1)); 2569 break; 2570 default: 2571 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr, 2572 0, 0, 0, htonl(1)); 2573 break; 2574 } 2575 } 2576 2577 static int cma_bind_loopback(struct rdma_id_private *id_priv) 2578 { 2579 struct cma_device *cma_dev, *cur_dev; 2580 struct ib_port_attr port_attr; 2581 union ib_gid gid; 2582 u16 pkey; 2583 int ret; 2584 u8 p; 2585 2586 cma_dev = NULL; 2587 mutex_lock(&lock); 2588 list_for_each_entry(cur_dev, &dev_list, list) { 2589 if (cma_family(id_priv) == AF_IB && 2590 !rdma_cap_ib_cm(cur_dev->device, 1)) 2591 continue; 2592 2593 if (!cma_dev) 2594 cma_dev = cur_dev; 2595 2596 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) { 2597 if (!ib_query_port(cur_dev->device, p, &port_attr) && 2598 port_attr.state == IB_PORT_ACTIVE) { 2599 cma_dev = cur_dev; 2600 goto port_found; 2601 } 2602 } 2603 } 2604 2605 if (!cma_dev) { 2606 ret = -ENODEV; 2607 goto out; 2608 } 2609 2610 p = 1; 2611 2612 port_found: 2613 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid, NULL); 2614 if (ret) 2615 goto out; 2616 2617 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey); 2618 if (ret) 2619 goto out; 2620 2621 id_priv->id.route.addr.dev_addr.dev_type = 2622 (rdma_protocol_ib(cma_dev->device, p)) ? 2623 ARPHRD_INFINIBAND : ARPHRD_ETHER; 2624 2625 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid); 2626 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey); 2627 id_priv->id.port_num = p; 2628 cma_attach_to_dev(id_priv, cma_dev); 2629 cma_set_loopback(cma_src_addr(id_priv)); 2630 out: 2631 mutex_unlock(&lock); 2632 return ret; 2633 } 2634 2635 static void addr_handler(int status, struct sockaddr *src_addr, 2636 struct rdma_dev_addr *dev_addr, void *context) 2637 { 2638 struct rdma_id_private *id_priv = context; 2639 struct rdma_cm_event event; 2640 2641 memset(&event, 0, sizeof event); 2642 mutex_lock(&id_priv->handler_mutex); 2643 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, 2644 RDMA_CM_ADDR_RESOLVED)) 2645 goto out; 2646 2647 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr)); 2648 if (!status && !id_priv->cma_dev) 2649 status = cma_acquire_dev(id_priv, NULL); 2650 2651 if (status) { 2652 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, 2653 RDMA_CM_ADDR_BOUND)) 2654 goto out; 2655 event.event = RDMA_CM_EVENT_ADDR_ERROR; 2656 event.status = status; 2657 } else 2658 event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 2659 2660 if (id_priv->id.event_handler(&id_priv->id, &event)) { 2661 cma_exch(id_priv, RDMA_CM_DESTROYING); 2662 mutex_unlock(&id_priv->handler_mutex); 2663 cma_deref_id(id_priv); 2664 rdma_destroy_id(&id_priv->id); 2665 return; 2666 } 2667 out: 2668 mutex_unlock(&id_priv->handler_mutex); 2669 cma_deref_id(id_priv); 2670 } 2671 2672 static int cma_resolve_loopback(struct rdma_id_private *id_priv) 2673 { 2674 struct cma_work *work; 2675 union ib_gid gid; 2676 int ret; 2677 2678 work = kzalloc(sizeof *work, GFP_KERNEL); 2679 if (!work) 2680 return -ENOMEM; 2681 2682 if (!id_priv->cma_dev) { 2683 ret = cma_bind_loopback(id_priv); 2684 if (ret) 2685 goto err; 2686 } 2687 2688 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid); 2689 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid); 2690 2691 work->id = id_priv; 2692 INIT_WORK(&work->work, cma_work_handler); 2693 work->old_state = RDMA_CM_ADDR_QUERY; 2694 work->new_state = RDMA_CM_ADDR_RESOLVED; 2695 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 2696 queue_work(cma_wq, &work->work); 2697 return 0; 2698 err: 2699 kfree(work); 2700 return ret; 2701 } 2702 2703 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv) 2704 { 2705 struct cma_work *work; 2706 int ret; 2707 2708 work = kzalloc(sizeof *work, GFP_KERNEL); 2709 if (!work) 2710 return -ENOMEM; 2711 2712 if (!id_priv->cma_dev) { 2713 ret = cma_resolve_ib_dev(id_priv); 2714 if (ret) 2715 goto err; 2716 } 2717 2718 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *) 2719 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr)); 2720 2721 work->id = id_priv; 2722 INIT_WORK(&work->work, cma_work_handler); 2723 work->old_state = RDMA_CM_ADDR_QUERY; 2724 work->new_state = RDMA_CM_ADDR_RESOLVED; 2725 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 2726 queue_work(cma_wq, &work->work); 2727 return 0; 2728 err: 2729 kfree(work); 2730 return ret; 2731 } 2732 2733 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr, 2734 struct sockaddr *dst_addr) 2735 { 2736 if (!src_addr || !src_addr->sa_family) { 2737 src_addr = (struct sockaddr *) &id->route.addr.src_addr; 2738 src_addr->sa_family = dst_addr->sa_family; 2739 if (dst_addr->sa_family == AF_INET6) { 2740 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr; 2741 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr; 2742 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id; 2743 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL) 2744 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id; 2745 } else if (dst_addr->sa_family == AF_IB) { 2746 ((struct sockaddr_ib *) src_addr)->sib_pkey = 2747 ((struct sockaddr_ib *) dst_addr)->sib_pkey; 2748 } 2749 } 2750 return rdma_bind_addr(id, src_addr); 2751 } 2752 2753 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr, 2754 struct sockaddr *dst_addr, int timeout_ms) 2755 { 2756 struct rdma_id_private *id_priv; 2757 int ret; 2758 2759 id_priv = container_of(id, struct rdma_id_private, id); 2760 if (id_priv->state == RDMA_CM_IDLE) { 2761 ret = cma_bind_addr(id, src_addr, dst_addr); 2762 if (ret) 2763 return ret; 2764 } 2765 2766 if (cma_family(id_priv) != dst_addr->sa_family) 2767 return -EINVAL; 2768 2769 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) 2770 return -EINVAL; 2771 2772 atomic_inc(&id_priv->refcount); 2773 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr)); 2774 if (cma_any_addr(dst_addr)) { 2775 ret = cma_resolve_loopback(id_priv); 2776 } else { 2777 if (dst_addr->sa_family == AF_IB) { 2778 ret = cma_resolve_ib_addr(id_priv); 2779 } else { 2780 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv), 2781 dst_addr, &id->route.addr.dev_addr, 2782 timeout_ms, addr_handler, id_priv); 2783 } 2784 } 2785 if (ret) 2786 goto err; 2787 2788 return 0; 2789 err: 2790 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND); 2791 cma_deref_id(id_priv); 2792 return ret; 2793 } 2794 EXPORT_SYMBOL(rdma_resolve_addr); 2795 2796 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse) 2797 { 2798 struct rdma_id_private *id_priv; 2799 unsigned long flags; 2800 int ret; 2801 2802 id_priv = container_of(id, struct rdma_id_private, id); 2803 spin_lock_irqsave(&id_priv->lock, flags); 2804 if (reuse || id_priv->state == RDMA_CM_IDLE) { 2805 id_priv->reuseaddr = reuse; 2806 ret = 0; 2807 } else { 2808 ret = -EINVAL; 2809 } 2810 spin_unlock_irqrestore(&id_priv->lock, flags); 2811 return ret; 2812 } 2813 EXPORT_SYMBOL(rdma_set_reuseaddr); 2814 2815 int rdma_set_afonly(struct rdma_cm_id *id, int afonly) 2816 { 2817 struct rdma_id_private *id_priv; 2818 unsigned long flags; 2819 int ret; 2820 2821 id_priv = container_of(id, struct rdma_id_private, id); 2822 spin_lock_irqsave(&id_priv->lock, flags); 2823 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) { 2824 id_priv->options |= (1 << CMA_OPTION_AFONLY); 2825 id_priv->afonly = afonly; 2826 ret = 0; 2827 } else { 2828 ret = -EINVAL; 2829 } 2830 spin_unlock_irqrestore(&id_priv->lock, flags); 2831 return ret; 2832 } 2833 EXPORT_SYMBOL(rdma_set_afonly); 2834 2835 static void cma_bind_port(struct rdma_bind_list *bind_list, 2836 struct rdma_id_private *id_priv) 2837 { 2838 struct sockaddr *addr; 2839 struct sockaddr_ib *sib; 2840 u64 sid, mask; 2841 __be16 port; 2842 2843 addr = cma_src_addr(id_priv); 2844 port = htons(bind_list->port); 2845 2846 switch (addr->sa_family) { 2847 case AF_INET: 2848 ((struct sockaddr_in *) addr)->sin_port = port; 2849 break; 2850 case AF_INET6: 2851 ((struct sockaddr_in6 *) addr)->sin6_port = port; 2852 break; 2853 case AF_IB: 2854 sib = (struct sockaddr_ib *) addr; 2855 sid = be64_to_cpu(sib->sib_sid); 2856 mask = be64_to_cpu(sib->sib_sid_mask); 2857 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port)); 2858 sib->sib_sid_mask = cpu_to_be64(~0ULL); 2859 break; 2860 } 2861 id_priv->bind_list = bind_list; 2862 hlist_add_head(&id_priv->node, &bind_list->owners); 2863 } 2864 2865 static int cma_alloc_port(enum rdma_port_space ps, 2866 struct rdma_id_private *id_priv, unsigned short snum) 2867 { 2868 struct rdma_bind_list *bind_list; 2869 int ret; 2870 2871 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL); 2872 if (!bind_list) 2873 return -ENOMEM; 2874 2875 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list, 2876 snum); 2877 if (ret < 0) 2878 goto err; 2879 2880 bind_list->ps = ps; 2881 bind_list->port = (unsigned short)ret; 2882 cma_bind_port(bind_list, id_priv); 2883 return 0; 2884 err: 2885 kfree(bind_list); 2886 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret; 2887 } 2888 2889 static int cma_alloc_any_port(enum rdma_port_space ps, 2890 struct rdma_id_private *id_priv) 2891 { 2892 static unsigned int last_used_port; 2893 int low, high, remaining; 2894 unsigned int rover; 2895 struct net *net = id_priv->id.route.addr.dev_addr.net; 2896 2897 inet_get_local_port_range(net, &low, &high); 2898 remaining = (high - low) + 1; 2899 rover = prandom_u32() % remaining + low; 2900 retry: 2901 if (last_used_port != rover && 2902 !cma_ps_find(net, ps, (unsigned short)rover)) { 2903 int ret = cma_alloc_port(ps, id_priv, rover); 2904 /* 2905 * Remember previously used port number in order to avoid 2906 * re-using same port immediately after it is closed. 2907 */ 2908 if (!ret) 2909 last_used_port = rover; 2910 if (ret != -EADDRNOTAVAIL) 2911 return ret; 2912 } 2913 if (--remaining) { 2914 rover++; 2915 if ((rover < low) || (rover > high)) 2916 rover = low; 2917 goto retry; 2918 } 2919 return -EADDRNOTAVAIL; 2920 } 2921 2922 /* 2923 * Check that the requested port is available. This is called when trying to 2924 * bind to a specific port, or when trying to listen on a bound port. In 2925 * the latter case, the provided id_priv may already be on the bind_list, but 2926 * we still need to check that it's okay to start listening. 2927 */ 2928 static int cma_check_port(struct rdma_bind_list *bind_list, 2929 struct rdma_id_private *id_priv, uint8_t reuseaddr) 2930 { 2931 struct rdma_id_private *cur_id; 2932 struct sockaddr *addr, *cur_addr; 2933 2934 addr = cma_src_addr(id_priv); 2935 hlist_for_each_entry(cur_id, &bind_list->owners, node) { 2936 if (id_priv == cur_id) 2937 continue; 2938 2939 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr && 2940 cur_id->reuseaddr) 2941 continue; 2942 2943 cur_addr = cma_src_addr(cur_id); 2944 if (id_priv->afonly && cur_id->afonly && 2945 (addr->sa_family != cur_addr->sa_family)) 2946 continue; 2947 2948 if (cma_any_addr(addr) || cma_any_addr(cur_addr)) 2949 return -EADDRNOTAVAIL; 2950 2951 if (!cma_addr_cmp(addr, cur_addr)) 2952 return -EADDRINUSE; 2953 } 2954 return 0; 2955 } 2956 2957 static int cma_use_port(enum rdma_port_space ps, 2958 struct rdma_id_private *id_priv) 2959 { 2960 struct rdma_bind_list *bind_list; 2961 unsigned short snum; 2962 int ret; 2963 2964 snum = ntohs(cma_port(cma_src_addr(id_priv))); 2965 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE)) 2966 return -EACCES; 2967 2968 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum); 2969 if (!bind_list) { 2970 ret = cma_alloc_port(ps, id_priv, snum); 2971 } else { 2972 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr); 2973 if (!ret) 2974 cma_bind_port(bind_list, id_priv); 2975 } 2976 return ret; 2977 } 2978 2979 static int cma_bind_listen(struct rdma_id_private *id_priv) 2980 { 2981 struct rdma_bind_list *bind_list = id_priv->bind_list; 2982 int ret = 0; 2983 2984 mutex_lock(&lock); 2985 if (bind_list->owners.first->next) 2986 ret = cma_check_port(bind_list, id_priv, 0); 2987 mutex_unlock(&lock); 2988 return ret; 2989 } 2990 2991 static enum rdma_port_space cma_select_inet_ps( 2992 struct rdma_id_private *id_priv) 2993 { 2994 switch (id_priv->id.ps) { 2995 case RDMA_PS_TCP: 2996 case RDMA_PS_UDP: 2997 case RDMA_PS_IPOIB: 2998 case RDMA_PS_IB: 2999 return id_priv->id.ps; 3000 default: 3001 3002 return 0; 3003 } 3004 } 3005 3006 static enum rdma_port_space cma_select_ib_ps(struct rdma_id_private *id_priv) 3007 { 3008 enum rdma_port_space ps = 0; 3009 struct sockaddr_ib *sib; 3010 u64 sid_ps, mask, sid; 3011 3012 sib = (struct sockaddr_ib *) cma_src_addr(id_priv); 3013 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK; 3014 sid = be64_to_cpu(sib->sib_sid) & mask; 3015 3016 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) { 3017 sid_ps = RDMA_IB_IP_PS_IB; 3018 ps = RDMA_PS_IB; 3019 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) && 3020 (sid == (RDMA_IB_IP_PS_TCP & mask))) { 3021 sid_ps = RDMA_IB_IP_PS_TCP; 3022 ps = RDMA_PS_TCP; 3023 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) && 3024 (sid == (RDMA_IB_IP_PS_UDP & mask))) { 3025 sid_ps = RDMA_IB_IP_PS_UDP; 3026 ps = RDMA_PS_UDP; 3027 } 3028 3029 if (ps) { 3030 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib))); 3031 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK | 3032 be64_to_cpu(sib->sib_sid_mask)); 3033 } 3034 return ps; 3035 } 3036 3037 static int cma_get_port(struct rdma_id_private *id_priv) 3038 { 3039 enum rdma_port_space ps; 3040 int ret; 3041 3042 if (cma_family(id_priv) != AF_IB) 3043 ps = cma_select_inet_ps(id_priv); 3044 else 3045 ps = cma_select_ib_ps(id_priv); 3046 if (!ps) 3047 return -EPROTONOSUPPORT; 3048 3049 mutex_lock(&lock); 3050 if (cma_any_port(cma_src_addr(id_priv))) 3051 ret = cma_alloc_any_port(ps, id_priv); 3052 else 3053 ret = cma_use_port(ps, id_priv); 3054 mutex_unlock(&lock); 3055 3056 return ret; 3057 } 3058 3059 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr, 3060 struct sockaddr *addr) 3061 { 3062 #if IS_ENABLED(CONFIG_IPV6) 3063 struct sockaddr_in6 *sin6; 3064 3065 if (addr->sa_family != AF_INET6) 3066 return 0; 3067 3068 sin6 = (struct sockaddr_in6 *) addr; 3069 3070 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL)) 3071 return 0; 3072 3073 if (!sin6->sin6_scope_id) 3074 return -EINVAL; 3075 3076 dev_addr->bound_dev_if = sin6->sin6_scope_id; 3077 #endif 3078 return 0; 3079 } 3080 3081 int rdma_listen(struct rdma_cm_id *id, int backlog) 3082 { 3083 struct rdma_id_private *id_priv; 3084 int ret; 3085 3086 id_priv = container_of(id, struct rdma_id_private, id); 3087 if (id_priv->state == RDMA_CM_IDLE) { 3088 id->route.addr.src_addr.ss_family = AF_INET; 3089 ret = rdma_bind_addr(id, cma_src_addr(id_priv)); 3090 if (ret) 3091 return ret; 3092 } 3093 3094 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) 3095 return -EINVAL; 3096 3097 if (id_priv->reuseaddr) { 3098 ret = cma_bind_listen(id_priv); 3099 if (ret) 3100 goto err; 3101 } 3102 3103 id_priv->backlog = backlog; 3104 if (id->device) { 3105 if (rdma_cap_ib_cm(id->device, 1)) { 3106 ret = cma_ib_listen(id_priv); 3107 if (ret) 3108 goto err; 3109 } else if (rdma_cap_iw_cm(id->device, 1)) { 3110 ret = cma_iw_listen(id_priv, backlog); 3111 if (ret) 3112 goto err; 3113 } else { 3114 ret = -ENOSYS; 3115 goto err; 3116 } 3117 } else 3118 cma_listen_on_all(id_priv); 3119 3120 return 0; 3121 err: 3122 id_priv->backlog = 0; 3123 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND); 3124 return ret; 3125 } 3126 EXPORT_SYMBOL(rdma_listen); 3127 3128 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr) 3129 { 3130 struct rdma_id_private *id_priv; 3131 int ret; 3132 3133 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 && 3134 addr->sa_family != AF_IB) 3135 return -EAFNOSUPPORT; 3136 3137 id_priv = container_of(id, struct rdma_id_private, id); 3138 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND)) 3139 return -EINVAL; 3140 3141 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr); 3142 if (ret) 3143 goto err1; 3144 3145 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr)); 3146 if (!cma_any_addr(addr)) { 3147 ret = cma_translate_addr(addr, &id->route.addr.dev_addr); 3148 if (ret) 3149 goto err1; 3150 3151 ret = cma_acquire_dev(id_priv, NULL); 3152 if (ret) 3153 goto err1; 3154 } 3155 3156 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) { 3157 if (addr->sa_family == AF_INET) 3158 id_priv->afonly = 1; 3159 #if IS_ENABLED(CONFIG_IPV6) 3160 else if (addr->sa_family == AF_INET6) { 3161 struct net *net = id_priv->id.route.addr.dev_addr.net; 3162 3163 id_priv->afonly = net->ipv6.sysctl.bindv6only; 3164 } 3165 #endif 3166 } 3167 ret = cma_get_port(id_priv); 3168 if (ret) 3169 goto err2; 3170 3171 return 0; 3172 err2: 3173 if (id_priv->cma_dev) 3174 cma_release_dev(id_priv); 3175 err1: 3176 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE); 3177 return ret; 3178 } 3179 EXPORT_SYMBOL(rdma_bind_addr); 3180 3181 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv) 3182 { 3183 struct cma_hdr *cma_hdr; 3184 3185 cma_hdr = hdr; 3186 cma_hdr->cma_version = CMA_VERSION; 3187 if (cma_family(id_priv) == AF_INET) { 3188 struct sockaddr_in *src4, *dst4; 3189 3190 src4 = (struct sockaddr_in *) cma_src_addr(id_priv); 3191 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv); 3192 3193 cma_set_ip_ver(cma_hdr, 4); 3194 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr; 3195 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr; 3196 cma_hdr->port = src4->sin_port; 3197 } else if (cma_family(id_priv) == AF_INET6) { 3198 struct sockaddr_in6 *src6, *dst6; 3199 3200 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv); 3201 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv); 3202 3203 cma_set_ip_ver(cma_hdr, 6); 3204 cma_hdr->src_addr.ip6 = src6->sin6_addr; 3205 cma_hdr->dst_addr.ip6 = dst6->sin6_addr; 3206 cma_hdr->port = src6->sin6_port; 3207 } 3208 return 0; 3209 } 3210 3211 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id, 3212 struct ib_cm_event *ib_event) 3213 { 3214 struct rdma_id_private *id_priv = cm_id->context; 3215 struct rdma_cm_event event; 3216 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd; 3217 int ret = 0; 3218 3219 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT)) 3220 return 0; 3221 3222 memset(&event, 0, sizeof event); 3223 switch (ib_event->event) { 3224 case IB_CM_SIDR_REQ_ERROR: 3225 event.event = RDMA_CM_EVENT_UNREACHABLE; 3226 event.status = -ETIMEDOUT; 3227 break; 3228 case IB_CM_SIDR_REP_RECEIVED: 3229 event.param.ud.private_data = ib_event->private_data; 3230 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE; 3231 if (rep->status != IB_SIDR_SUCCESS) { 3232 event.event = RDMA_CM_EVENT_UNREACHABLE; 3233 event.status = ib_event->param.sidr_rep_rcvd.status; 3234 break; 3235 } 3236 ret = cma_set_qkey(id_priv, rep->qkey); 3237 if (ret) { 3238 event.event = RDMA_CM_EVENT_ADDR_ERROR; 3239 event.status = ret; 3240 break; 3241 } 3242 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num, 3243 id_priv->id.route.path_rec, 3244 &event.param.ud.ah_attr); 3245 event.param.ud.qp_num = rep->qpn; 3246 event.param.ud.qkey = rep->qkey; 3247 event.event = RDMA_CM_EVENT_ESTABLISHED; 3248 event.status = 0; 3249 break; 3250 default: 3251 pr_err("RDMA CMA: unexpected IB CM event: %d\n", 3252 ib_event->event); 3253 goto out; 3254 } 3255 3256 ret = id_priv->id.event_handler(&id_priv->id, &event); 3257 if (ret) { 3258 /* Destroy the CM ID by returning a non-zero value. */ 3259 id_priv->cm_id.ib = NULL; 3260 cma_exch(id_priv, RDMA_CM_DESTROYING); 3261 mutex_unlock(&id_priv->handler_mutex); 3262 rdma_destroy_id(&id_priv->id); 3263 return ret; 3264 } 3265 out: 3266 mutex_unlock(&id_priv->handler_mutex); 3267 return ret; 3268 } 3269 3270 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv, 3271 struct rdma_conn_param *conn_param) 3272 { 3273 struct ib_cm_sidr_req_param req; 3274 struct ib_cm_id *id; 3275 void *private_data; 3276 int offset, ret; 3277 3278 memset(&req, 0, sizeof req); 3279 offset = cma_user_data_offset(id_priv); 3280 req.private_data_len = offset + conn_param->private_data_len; 3281 if (req.private_data_len < conn_param->private_data_len) 3282 return -EINVAL; 3283 3284 if (req.private_data_len) { 3285 private_data = kzalloc(req.private_data_len, GFP_ATOMIC); 3286 if (!private_data) 3287 return -ENOMEM; 3288 } else { 3289 private_data = NULL; 3290 } 3291 3292 if (conn_param->private_data && conn_param->private_data_len) 3293 memcpy(private_data + offset, conn_param->private_data, 3294 conn_param->private_data_len); 3295 3296 if (private_data) { 3297 ret = cma_format_hdr(private_data, id_priv); 3298 if (ret) 3299 goto out; 3300 req.private_data = private_data; 3301 } 3302 3303 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler, 3304 id_priv); 3305 if (IS_ERR(id)) { 3306 ret = PTR_ERR(id); 3307 goto out; 3308 } 3309 id_priv->cm_id.ib = id; 3310 3311 req.path = id_priv->id.route.path_rec; 3312 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv)); 3313 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8); 3314 req.max_cm_retries = CMA_MAX_CM_RETRIES; 3315 3316 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req); 3317 if (ret) { 3318 ib_destroy_cm_id(id_priv->cm_id.ib); 3319 id_priv->cm_id.ib = NULL; 3320 } 3321 out: 3322 kfree(private_data); 3323 return ret; 3324 } 3325 3326 static int cma_connect_ib(struct rdma_id_private *id_priv, 3327 struct rdma_conn_param *conn_param) 3328 { 3329 struct ib_cm_req_param req; 3330 struct rdma_route *route; 3331 void *private_data; 3332 struct ib_cm_id *id; 3333 int offset, ret; 3334 3335 memset(&req, 0, sizeof req); 3336 offset = cma_user_data_offset(id_priv); 3337 req.private_data_len = offset + conn_param->private_data_len; 3338 if (req.private_data_len < conn_param->private_data_len) 3339 return -EINVAL; 3340 3341 if (req.private_data_len) { 3342 private_data = kzalloc(req.private_data_len, GFP_ATOMIC); 3343 if (!private_data) 3344 return -ENOMEM; 3345 } else { 3346 private_data = NULL; 3347 } 3348 3349 if (conn_param->private_data && conn_param->private_data_len) 3350 memcpy(private_data + offset, conn_param->private_data, 3351 conn_param->private_data_len); 3352 3353 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv); 3354 if (IS_ERR(id)) { 3355 ret = PTR_ERR(id); 3356 goto out; 3357 } 3358 id_priv->cm_id.ib = id; 3359 3360 route = &id_priv->id.route; 3361 if (private_data) { 3362 ret = cma_format_hdr(private_data, id_priv); 3363 if (ret) 3364 goto out; 3365 req.private_data = private_data; 3366 } 3367 3368 req.primary_path = &route->path_rec[0]; 3369 if (route->num_paths == 2) 3370 req.alternate_path = &route->path_rec[1]; 3371 3372 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv)); 3373 req.qp_num = id_priv->qp_num; 3374 req.qp_type = id_priv->id.qp_type; 3375 req.starting_psn = id_priv->seq_num; 3376 req.responder_resources = conn_param->responder_resources; 3377 req.initiator_depth = conn_param->initiator_depth; 3378 req.flow_control = conn_param->flow_control; 3379 req.retry_count = min_t(u8, 7, conn_param->retry_count); 3380 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count); 3381 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT; 3382 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT; 3383 req.max_cm_retries = CMA_MAX_CM_RETRIES; 3384 req.srq = id_priv->srq ? 1 : 0; 3385 3386 ret = ib_send_cm_req(id_priv->cm_id.ib, &req); 3387 out: 3388 if (ret && !IS_ERR(id)) { 3389 ib_destroy_cm_id(id); 3390 id_priv->cm_id.ib = NULL; 3391 } 3392 3393 kfree(private_data); 3394 return ret; 3395 } 3396 3397 static int cma_connect_iw(struct rdma_id_private *id_priv, 3398 struct rdma_conn_param *conn_param) 3399 { 3400 struct iw_cm_id *cm_id; 3401 int ret; 3402 struct iw_cm_conn_param iw_param; 3403 3404 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv); 3405 if (IS_ERR(cm_id)) 3406 return PTR_ERR(cm_id); 3407 3408 cm_id->tos = id_priv->tos; 3409 id_priv->cm_id.iw = cm_id; 3410 3411 memcpy(&cm_id->local_addr, cma_src_addr(id_priv), 3412 rdma_addr_size(cma_src_addr(id_priv))); 3413 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv), 3414 rdma_addr_size(cma_dst_addr(id_priv))); 3415 3416 ret = cma_modify_qp_rtr(id_priv, conn_param); 3417 if (ret) 3418 goto out; 3419 3420 if (conn_param) { 3421 iw_param.ord = conn_param->initiator_depth; 3422 iw_param.ird = conn_param->responder_resources; 3423 iw_param.private_data = conn_param->private_data; 3424 iw_param.private_data_len = conn_param->private_data_len; 3425 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num; 3426 } else { 3427 memset(&iw_param, 0, sizeof iw_param); 3428 iw_param.qpn = id_priv->qp_num; 3429 } 3430 ret = iw_cm_connect(cm_id, &iw_param); 3431 out: 3432 if (ret) { 3433 iw_destroy_cm_id(cm_id); 3434 id_priv->cm_id.iw = NULL; 3435 } 3436 return ret; 3437 } 3438 3439 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param) 3440 { 3441 struct rdma_id_private *id_priv; 3442 int ret; 3443 3444 id_priv = container_of(id, struct rdma_id_private, id); 3445 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT)) 3446 return -EINVAL; 3447 3448 if (!id->qp) { 3449 id_priv->qp_num = conn_param->qp_num; 3450 id_priv->srq = conn_param->srq; 3451 } 3452 3453 if (rdma_cap_ib_cm(id->device, id->port_num)) { 3454 if (id->qp_type == IB_QPT_UD) 3455 ret = cma_resolve_ib_udp(id_priv, conn_param); 3456 else 3457 ret = cma_connect_ib(id_priv, conn_param); 3458 } else if (rdma_cap_iw_cm(id->device, id->port_num)) 3459 ret = cma_connect_iw(id_priv, conn_param); 3460 else 3461 ret = -ENOSYS; 3462 if (ret) 3463 goto err; 3464 3465 return 0; 3466 err: 3467 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED); 3468 return ret; 3469 } 3470 EXPORT_SYMBOL(rdma_connect); 3471 3472 static int cma_accept_ib(struct rdma_id_private *id_priv, 3473 struct rdma_conn_param *conn_param) 3474 { 3475 struct ib_cm_rep_param rep; 3476 int ret; 3477 3478 ret = cma_modify_qp_rtr(id_priv, conn_param); 3479 if (ret) 3480 goto out; 3481 3482 ret = cma_modify_qp_rts(id_priv, conn_param); 3483 if (ret) 3484 goto out; 3485 3486 memset(&rep, 0, sizeof rep); 3487 rep.qp_num = id_priv->qp_num; 3488 rep.starting_psn = id_priv->seq_num; 3489 rep.private_data = conn_param->private_data; 3490 rep.private_data_len = conn_param->private_data_len; 3491 rep.responder_resources = conn_param->responder_resources; 3492 rep.initiator_depth = conn_param->initiator_depth; 3493 rep.failover_accepted = 0; 3494 rep.flow_control = conn_param->flow_control; 3495 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count); 3496 rep.srq = id_priv->srq ? 1 : 0; 3497 3498 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep); 3499 out: 3500 return ret; 3501 } 3502 3503 static int cma_accept_iw(struct rdma_id_private *id_priv, 3504 struct rdma_conn_param *conn_param) 3505 { 3506 struct iw_cm_conn_param iw_param; 3507 int ret; 3508 3509 ret = cma_modify_qp_rtr(id_priv, conn_param); 3510 if (ret) 3511 return ret; 3512 3513 iw_param.ord = conn_param->initiator_depth; 3514 iw_param.ird = conn_param->responder_resources; 3515 iw_param.private_data = conn_param->private_data; 3516 iw_param.private_data_len = conn_param->private_data_len; 3517 if (id_priv->id.qp) { 3518 iw_param.qpn = id_priv->qp_num; 3519 } else 3520 iw_param.qpn = conn_param->qp_num; 3521 3522 return iw_cm_accept(id_priv->cm_id.iw, &iw_param); 3523 } 3524 3525 static int cma_send_sidr_rep(struct rdma_id_private *id_priv, 3526 enum ib_cm_sidr_status status, u32 qkey, 3527 const void *private_data, int private_data_len) 3528 { 3529 struct ib_cm_sidr_rep_param rep; 3530 int ret; 3531 3532 memset(&rep, 0, sizeof rep); 3533 rep.status = status; 3534 if (status == IB_SIDR_SUCCESS) { 3535 ret = cma_set_qkey(id_priv, qkey); 3536 if (ret) 3537 return ret; 3538 rep.qp_num = id_priv->qp_num; 3539 rep.qkey = id_priv->qkey; 3540 } 3541 rep.private_data = private_data; 3542 rep.private_data_len = private_data_len; 3543 3544 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep); 3545 } 3546 3547 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param) 3548 { 3549 struct rdma_id_private *id_priv; 3550 int ret; 3551 3552 id_priv = container_of(id, struct rdma_id_private, id); 3553 3554 id_priv->owner = task_pid_nr(current); 3555 3556 if (!cma_comp(id_priv, RDMA_CM_CONNECT)) 3557 return -EINVAL; 3558 3559 if (!id->qp && conn_param) { 3560 id_priv->qp_num = conn_param->qp_num; 3561 id_priv->srq = conn_param->srq; 3562 } 3563 3564 if (rdma_cap_ib_cm(id->device, id->port_num)) { 3565 if (id->qp_type == IB_QPT_UD) { 3566 if (conn_param) 3567 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS, 3568 conn_param->qkey, 3569 conn_param->private_data, 3570 conn_param->private_data_len); 3571 else 3572 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS, 3573 0, NULL, 0); 3574 } else { 3575 if (conn_param) 3576 ret = cma_accept_ib(id_priv, conn_param); 3577 else 3578 ret = cma_rep_recv(id_priv); 3579 } 3580 } else if (rdma_cap_iw_cm(id->device, id->port_num)) 3581 ret = cma_accept_iw(id_priv, conn_param); 3582 else 3583 ret = -ENOSYS; 3584 3585 if (ret) 3586 goto reject; 3587 3588 return 0; 3589 reject: 3590 cma_modify_qp_err(id_priv); 3591 rdma_reject(id, NULL, 0); 3592 return ret; 3593 } 3594 EXPORT_SYMBOL(rdma_accept); 3595 3596 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event) 3597 { 3598 struct rdma_id_private *id_priv; 3599 int ret; 3600 3601 id_priv = container_of(id, struct rdma_id_private, id); 3602 if (!id_priv->cm_id.ib) 3603 return -EINVAL; 3604 3605 switch (id->device->node_type) { 3606 case RDMA_NODE_IB_CA: 3607 ret = ib_cm_notify(id_priv->cm_id.ib, event); 3608 break; 3609 default: 3610 ret = 0; 3611 break; 3612 } 3613 return ret; 3614 } 3615 EXPORT_SYMBOL(rdma_notify); 3616 3617 int rdma_reject(struct rdma_cm_id *id, const void *private_data, 3618 u8 private_data_len) 3619 { 3620 struct rdma_id_private *id_priv; 3621 int ret; 3622 3623 id_priv = container_of(id, struct rdma_id_private, id); 3624 if (!id_priv->cm_id.ib) 3625 return -EINVAL; 3626 3627 if (rdma_cap_ib_cm(id->device, id->port_num)) { 3628 if (id->qp_type == IB_QPT_UD) 3629 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0, 3630 private_data, private_data_len); 3631 else 3632 ret = ib_send_cm_rej(id_priv->cm_id.ib, 3633 IB_CM_REJ_CONSUMER_DEFINED, NULL, 3634 0, private_data, private_data_len); 3635 } else if (rdma_cap_iw_cm(id->device, id->port_num)) { 3636 ret = iw_cm_reject(id_priv->cm_id.iw, 3637 private_data, private_data_len); 3638 } else 3639 ret = -ENOSYS; 3640 3641 return ret; 3642 } 3643 EXPORT_SYMBOL(rdma_reject); 3644 3645 int rdma_disconnect(struct rdma_cm_id *id) 3646 { 3647 struct rdma_id_private *id_priv; 3648 int ret; 3649 3650 id_priv = container_of(id, struct rdma_id_private, id); 3651 if (!id_priv->cm_id.ib) 3652 return -EINVAL; 3653 3654 if (rdma_cap_ib_cm(id->device, id->port_num)) { 3655 ret = cma_modify_qp_err(id_priv); 3656 if (ret) 3657 goto out; 3658 /* Initiate or respond to a disconnect. */ 3659 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) 3660 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0); 3661 } else if (rdma_cap_iw_cm(id->device, id->port_num)) { 3662 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0); 3663 } else 3664 ret = -EINVAL; 3665 3666 out: 3667 return ret; 3668 } 3669 EXPORT_SYMBOL(rdma_disconnect); 3670 3671 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast) 3672 { 3673 struct rdma_id_private *id_priv; 3674 struct cma_multicast *mc = multicast->context; 3675 struct rdma_cm_event event; 3676 int ret; 3677 3678 id_priv = mc->id_priv; 3679 if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) && 3680 cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED)) 3681 return 0; 3682 3683 if (!status) 3684 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey)); 3685 mutex_lock(&id_priv->qp_mutex); 3686 if (!status && id_priv->id.qp) 3687 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid, 3688 be16_to_cpu(multicast->rec.mlid)); 3689 mutex_unlock(&id_priv->qp_mutex); 3690 3691 memset(&event, 0, sizeof event); 3692 event.status = status; 3693 event.param.ud.private_data = mc->context; 3694 if (!status) { 3695 struct rdma_dev_addr *dev_addr = 3696 &id_priv->id.route.addr.dev_addr; 3697 struct net_device *ndev = 3698 dev_get_by_index(&init_net, dev_addr->bound_dev_if); 3699 enum ib_gid_type gid_type = 3700 id_priv->cma_dev->default_gid_type[id_priv->id.port_num - 3701 rdma_start_port(id_priv->cma_dev->device)]; 3702 3703 event.event = RDMA_CM_EVENT_MULTICAST_JOIN; 3704 ib_init_ah_from_mcmember(id_priv->id.device, 3705 id_priv->id.port_num, &multicast->rec, 3706 ndev, gid_type, 3707 &event.param.ud.ah_attr); 3708 event.param.ud.qp_num = 0xFFFFFF; 3709 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey); 3710 if (ndev) 3711 dev_put(ndev); 3712 } else 3713 event.event = RDMA_CM_EVENT_MULTICAST_ERROR; 3714 3715 ret = id_priv->id.event_handler(&id_priv->id, &event); 3716 if (ret) { 3717 cma_exch(id_priv, RDMA_CM_DESTROYING); 3718 mutex_unlock(&id_priv->handler_mutex); 3719 rdma_destroy_id(&id_priv->id); 3720 return 0; 3721 } 3722 3723 mutex_unlock(&id_priv->handler_mutex); 3724 return 0; 3725 } 3726 3727 static void cma_set_mgid(struct rdma_id_private *id_priv, 3728 struct sockaddr *addr, union ib_gid *mgid) 3729 { 3730 unsigned char mc_map[MAX_ADDR_LEN]; 3731 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 3732 struct sockaddr_in *sin = (struct sockaddr_in *) addr; 3733 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr; 3734 3735 if (cma_any_addr(addr)) { 3736 memset(mgid, 0, sizeof *mgid); 3737 } else if ((addr->sa_family == AF_INET6) && 3738 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) == 3739 0xFF10A01B)) { 3740 /* IPv6 address is an SA assigned MGID. */ 3741 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid); 3742 } else if (addr->sa_family == AF_IB) { 3743 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid); 3744 } else if ((addr->sa_family == AF_INET6)) { 3745 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map); 3746 if (id_priv->id.ps == RDMA_PS_UDP) 3747 mc_map[7] = 0x01; /* Use RDMA CM signature */ 3748 *mgid = *(union ib_gid *) (mc_map + 4); 3749 } else { 3750 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map); 3751 if (id_priv->id.ps == RDMA_PS_UDP) 3752 mc_map[7] = 0x01; /* Use RDMA CM signature */ 3753 *mgid = *(union ib_gid *) (mc_map + 4); 3754 } 3755 } 3756 3757 static int cma_join_ib_multicast(struct rdma_id_private *id_priv, 3758 struct cma_multicast *mc) 3759 { 3760 struct ib_sa_mcmember_rec rec; 3761 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 3762 ib_sa_comp_mask comp_mask; 3763 int ret; 3764 3765 ib_addr_get_mgid(dev_addr, &rec.mgid); 3766 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num, 3767 &rec.mgid, &rec); 3768 if (ret) 3769 return ret; 3770 3771 ret = cma_set_qkey(id_priv, 0); 3772 if (ret) 3773 return ret; 3774 3775 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid); 3776 rec.qkey = cpu_to_be32(id_priv->qkey); 3777 rdma_addr_get_sgid(dev_addr, &rec.port_gid); 3778 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr)); 3779 rec.join_state = 1; 3780 3781 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID | 3782 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE | 3783 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL | 3784 IB_SA_MCMEMBER_REC_FLOW_LABEL | 3785 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS; 3786 3787 if (id_priv->id.ps == RDMA_PS_IPOIB) 3788 comp_mask |= IB_SA_MCMEMBER_REC_RATE | 3789 IB_SA_MCMEMBER_REC_RATE_SELECTOR | 3790 IB_SA_MCMEMBER_REC_MTU_SELECTOR | 3791 IB_SA_MCMEMBER_REC_MTU | 3792 IB_SA_MCMEMBER_REC_HOP_LIMIT; 3793 3794 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device, 3795 id_priv->id.port_num, &rec, 3796 comp_mask, GFP_KERNEL, 3797 cma_ib_mc_handler, mc); 3798 return PTR_ERR_OR_ZERO(mc->multicast.ib); 3799 } 3800 3801 static void iboe_mcast_work_handler(struct work_struct *work) 3802 { 3803 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work); 3804 struct cma_multicast *mc = mw->mc; 3805 struct ib_sa_multicast *m = mc->multicast.ib; 3806 3807 mc->multicast.ib->context = mc; 3808 cma_ib_mc_handler(0, m); 3809 kref_put(&mc->mcref, release_mc); 3810 kfree(mw); 3811 } 3812 3813 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid) 3814 { 3815 struct sockaddr_in *sin = (struct sockaddr_in *)addr; 3816 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr; 3817 3818 if (cma_any_addr(addr)) { 3819 memset(mgid, 0, sizeof *mgid); 3820 } else if (addr->sa_family == AF_INET6) { 3821 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid); 3822 } else { 3823 mgid->raw[0] = 0xff; 3824 mgid->raw[1] = 0x0e; 3825 mgid->raw[2] = 0; 3826 mgid->raw[3] = 0; 3827 mgid->raw[4] = 0; 3828 mgid->raw[5] = 0; 3829 mgid->raw[6] = 0; 3830 mgid->raw[7] = 0; 3831 mgid->raw[8] = 0; 3832 mgid->raw[9] = 0; 3833 mgid->raw[10] = 0xff; 3834 mgid->raw[11] = 0xff; 3835 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr; 3836 } 3837 } 3838 3839 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv, 3840 struct cma_multicast *mc) 3841 { 3842 struct iboe_mcast_work *work; 3843 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 3844 int err = 0; 3845 struct sockaddr *addr = (struct sockaddr *)&mc->addr; 3846 struct net_device *ndev = NULL; 3847 enum ib_gid_type gid_type; 3848 3849 if (cma_zero_addr((struct sockaddr *)&mc->addr)) 3850 return -EINVAL; 3851 3852 work = kzalloc(sizeof *work, GFP_KERNEL); 3853 if (!work) 3854 return -ENOMEM; 3855 3856 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL); 3857 if (!mc->multicast.ib) { 3858 err = -ENOMEM; 3859 goto out1; 3860 } 3861 3862 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid); 3863 3864 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff); 3865 if (id_priv->id.ps == RDMA_PS_UDP) 3866 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY); 3867 3868 if (dev_addr->bound_dev_if) 3869 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if); 3870 if (!ndev) { 3871 err = -ENODEV; 3872 goto out2; 3873 } 3874 mc->multicast.ib->rec.rate = iboe_get_rate(ndev); 3875 mc->multicast.ib->rec.hop_limit = 1; 3876 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu); 3877 3878 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num - 3879 rdma_start_port(id_priv->cma_dev->device)]; 3880 if (addr->sa_family == AF_INET) { 3881 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) 3882 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid, 3883 true); 3884 if (!err) { 3885 mc->igmp_joined = true; 3886 mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT; 3887 } 3888 } else { 3889 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) 3890 err = -ENOTSUPP; 3891 } 3892 dev_put(ndev); 3893 if (err || !mc->multicast.ib->rec.mtu) { 3894 if (!err) 3895 err = -EINVAL; 3896 goto out2; 3897 } 3898 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr, 3899 &mc->multicast.ib->rec.port_gid); 3900 work->id = id_priv; 3901 work->mc = mc; 3902 INIT_WORK(&work->work, iboe_mcast_work_handler); 3903 kref_get(&mc->mcref); 3904 queue_work(cma_wq, &work->work); 3905 3906 return 0; 3907 3908 out2: 3909 kfree(mc->multicast.ib); 3910 out1: 3911 kfree(work); 3912 return err; 3913 } 3914 3915 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr, 3916 void *context) 3917 { 3918 struct rdma_id_private *id_priv; 3919 struct cma_multicast *mc; 3920 int ret; 3921 3922 id_priv = container_of(id, struct rdma_id_private, id); 3923 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) && 3924 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED)) 3925 return -EINVAL; 3926 3927 mc = kmalloc(sizeof *mc, GFP_KERNEL); 3928 if (!mc) 3929 return -ENOMEM; 3930 3931 memcpy(&mc->addr, addr, rdma_addr_size(addr)); 3932 mc->context = context; 3933 mc->id_priv = id_priv; 3934 mc->igmp_joined = false; 3935 spin_lock(&id_priv->lock); 3936 list_add(&mc->list, &id_priv->mc_list); 3937 spin_unlock(&id_priv->lock); 3938 3939 if (rdma_protocol_roce(id->device, id->port_num)) { 3940 kref_init(&mc->mcref); 3941 ret = cma_iboe_join_multicast(id_priv, mc); 3942 } else if (rdma_cap_ib_mcast(id->device, id->port_num)) 3943 ret = cma_join_ib_multicast(id_priv, mc); 3944 else 3945 ret = -ENOSYS; 3946 3947 if (ret) { 3948 spin_lock_irq(&id_priv->lock); 3949 list_del(&mc->list); 3950 spin_unlock_irq(&id_priv->lock); 3951 kfree(mc); 3952 } 3953 return ret; 3954 } 3955 EXPORT_SYMBOL(rdma_join_multicast); 3956 3957 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr) 3958 { 3959 struct rdma_id_private *id_priv; 3960 struct cma_multicast *mc; 3961 3962 id_priv = container_of(id, struct rdma_id_private, id); 3963 spin_lock_irq(&id_priv->lock); 3964 list_for_each_entry(mc, &id_priv->mc_list, list) { 3965 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) { 3966 list_del(&mc->list); 3967 spin_unlock_irq(&id_priv->lock); 3968 3969 if (id->qp) 3970 ib_detach_mcast(id->qp, 3971 &mc->multicast.ib->rec.mgid, 3972 be16_to_cpu(mc->multicast.ib->rec.mlid)); 3973 3974 BUG_ON(id_priv->cma_dev->device != id->device); 3975 3976 if (rdma_cap_ib_mcast(id->device, id->port_num)) { 3977 ib_sa_free_multicast(mc->multicast.ib); 3978 kfree(mc); 3979 } else if (rdma_protocol_roce(id->device, id->port_num)) { 3980 if (mc->igmp_joined) { 3981 struct rdma_dev_addr *dev_addr = 3982 &id->route.addr.dev_addr; 3983 struct net_device *ndev = NULL; 3984 3985 if (dev_addr->bound_dev_if) 3986 ndev = dev_get_by_index(&init_net, 3987 dev_addr->bound_dev_if); 3988 if (ndev) { 3989 cma_igmp_send(ndev, 3990 &mc->multicast.ib->rec.mgid, 3991 false); 3992 dev_put(ndev); 3993 } 3994 mc->igmp_joined = false; 3995 } 3996 kref_put(&mc->mcref, release_mc); 3997 } 3998 return; 3999 } 4000 } 4001 spin_unlock_irq(&id_priv->lock); 4002 } 4003 EXPORT_SYMBOL(rdma_leave_multicast); 4004 4005 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv) 4006 { 4007 struct rdma_dev_addr *dev_addr; 4008 struct cma_ndev_work *work; 4009 4010 dev_addr = &id_priv->id.route.addr.dev_addr; 4011 4012 if ((dev_addr->bound_dev_if == ndev->ifindex) && 4013 (net_eq(dev_net(ndev), dev_addr->net)) && 4014 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) { 4015 pr_info("RDMA CM addr change for ndev %s used by id %p\n", 4016 ndev->name, &id_priv->id); 4017 work = kzalloc(sizeof *work, GFP_KERNEL); 4018 if (!work) 4019 return -ENOMEM; 4020 4021 INIT_WORK(&work->work, cma_ndev_work_handler); 4022 work->id = id_priv; 4023 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE; 4024 atomic_inc(&id_priv->refcount); 4025 queue_work(cma_wq, &work->work); 4026 } 4027 4028 return 0; 4029 } 4030 4031 static int cma_netdev_callback(struct notifier_block *self, unsigned long event, 4032 void *ptr) 4033 { 4034 struct net_device *ndev = netdev_notifier_info_to_dev(ptr); 4035 struct cma_device *cma_dev; 4036 struct rdma_id_private *id_priv; 4037 int ret = NOTIFY_DONE; 4038 4039 if (event != NETDEV_BONDING_FAILOVER) 4040 return NOTIFY_DONE; 4041 4042 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING)) 4043 return NOTIFY_DONE; 4044 4045 mutex_lock(&lock); 4046 list_for_each_entry(cma_dev, &dev_list, list) 4047 list_for_each_entry(id_priv, &cma_dev->id_list, list) { 4048 ret = cma_netdev_change(ndev, id_priv); 4049 if (ret) 4050 goto out; 4051 } 4052 4053 out: 4054 mutex_unlock(&lock); 4055 return ret; 4056 } 4057 4058 static struct notifier_block cma_nb = { 4059 .notifier_call = cma_netdev_callback 4060 }; 4061 4062 static void cma_add_one(struct ib_device *device) 4063 { 4064 struct cma_device *cma_dev; 4065 struct rdma_id_private *id_priv; 4066 unsigned int i; 4067 unsigned long supported_gids = 0; 4068 4069 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL); 4070 if (!cma_dev) 4071 return; 4072 4073 cma_dev->device = device; 4074 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt, 4075 sizeof(*cma_dev->default_gid_type), 4076 GFP_KERNEL); 4077 if (!cma_dev->default_gid_type) { 4078 kfree(cma_dev); 4079 return; 4080 } 4081 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) { 4082 supported_gids = roce_gid_type_mask_support(device, i); 4083 WARN_ON(!supported_gids); 4084 cma_dev->default_gid_type[i - rdma_start_port(device)] = 4085 find_first_bit(&supported_gids, BITS_PER_LONG); 4086 } 4087 4088 init_completion(&cma_dev->comp); 4089 atomic_set(&cma_dev->refcount, 1); 4090 INIT_LIST_HEAD(&cma_dev->id_list); 4091 ib_set_client_data(device, &cma_client, cma_dev); 4092 4093 mutex_lock(&lock); 4094 list_add_tail(&cma_dev->list, &dev_list); 4095 list_for_each_entry(id_priv, &listen_any_list, list) 4096 cma_listen_on_dev(id_priv, cma_dev); 4097 mutex_unlock(&lock); 4098 } 4099 4100 static int cma_remove_id_dev(struct rdma_id_private *id_priv) 4101 { 4102 struct rdma_cm_event event; 4103 enum rdma_cm_state state; 4104 int ret = 0; 4105 4106 /* Record that we want to remove the device */ 4107 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL); 4108 if (state == RDMA_CM_DESTROYING) 4109 return 0; 4110 4111 cma_cancel_operation(id_priv, state); 4112 mutex_lock(&id_priv->handler_mutex); 4113 4114 /* Check for destruction from another callback. */ 4115 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL)) 4116 goto out; 4117 4118 memset(&event, 0, sizeof event); 4119 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL; 4120 ret = id_priv->id.event_handler(&id_priv->id, &event); 4121 out: 4122 mutex_unlock(&id_priv->handler_mutex); 4123 return ret; 4124 } 4125 4126 static void cma_process_remove(struct cma_device *cma_dev) 4127 { 4128 struct rdma_id_private *id_priv; 4129 int ret; 4130 4131 mutex_lock(&lock); 4132 while (!list_empty(&cma_dev->id_list)) { 4133 id_priv = list_entry(cma_dev->id_list.next, 4134 struct rdma_id_private, list); 4135 4136 list_del(&id_priv->listen_list); 4137 list_del_init(&id_priv->list); 4138 atomic_inc(&id_priv->refcount); 4139 mutex_unlock(&lock); 4140 4141 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv); 4142 cma_deref_id(id_priv); 4143 if (ret) 4144 rdma_destroy_id(&id_priv->id); 4145 4146 mutex_lock(&lock); 4147 } 4148 mutex_unlock(&lock); 4149 4150 cma_deref_dev(cma_dev); 4151 wait_for_completion(&cma_dev->comp); 4152 } 4153 4154 static void cma_remove_one(struct ib_device *device, void *client_data) 4155 { 4156 struct cma_device *cma_dev = client_data; 4157 4158 if (!cma_dev) 4159 return; 4160 4161 mutex_lock(&lock); 4162 list_del(&cma_dev->list); 4163 mutex_unlock(&lock); 4164 4165 cma_process_remove(cma_dev); 4166 kfree(cma_dev->default_gid_type); 4167 kfree(cma_dev); 4168 } 4169 4170 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb) 4171 { 4172 struct nlmsghdr *nlh; 4173 struct rdma_cm_id_stats *id_stats; 4174 struct rdma_id_private *id_priv; 4175 struct rdma_cm_id *id = NULL; 4176 struct cma_device *cma_dev; 4177 int i_dev = 0, i_id = 0; 4178 4179 /* 4180 * We export all of the IDs as a sequence of messages. Each 4181 * ID gets its own netlink message. 4182 */ 4183 mutex_lock(&lock); 4184 4185 list_for_each_entry(cma_dev, &dev_list, list) { 4186 if (i_dev < cb->args[0]) { 4187 i_dev++; 4188 continue; 4189 } 4190 4191 i_id = 0; 4192 list_for_each_entry(id_priv, &cma_dev->id_list, list) { 4193 if (i_id < cb->args[1]) { 4194 i_id++; 4195 continue; 4196 } 4197 4198 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq, 4199 sizeof *id_stats, RDMA_NL_RDMA_CM, 4200 RDMA_NL_RDMA_CM_ID_STATS, 4201 NLM_F_MULTI); 4202 if (!id_stats) 4203 goto out; 4204 4205 memset(id_stats, 0, sizeof *id_stats); 4206 id = &id_priv->id; 4207 id_stats->node_type = id->route.addr.dev_addr.dev_type; 4208 id_stats->port_num = id->port_num; 4209 id_stats->bound_dev_if = 4210 id->route.addr.dev_addr.bound_dev_if; 4211 4212 if (ibnl_put_attr(skb, nlh, 4213 rdma_addr_size(cma_src_addr(id_priv)), 4214 cma_src_addr(id_priv), 4215 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) 4216 goto out; 4217 if (ibnl_put_attr(skb, nlh, 4218 rdma_addr_size(cma_src_addr(id_priv)), 4219 cma_dst_addr(id_priv), 4220 RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) 4221 goto out; 4222 4223 id_stats->pid = id_priv->owner; 4224 id_stats->port_space = id->ps; 4225 id_stats->cm_state = id_priv->state; 4226 id_stats->qp_num = id_priv->qp_num; 4227 id_stats->qp_type = id->qp_type; 4228 4229 i_id++; 4230 } 4231 4232 cb->args[1] = 0; 4233 i_dev++; 4234 } 4235 4236 out: 4237 mutex_unlock(&lock); 4238 cb->args[0] = i_dev; 4239 cb->args[1] = i_id; 4240 4241 return skb->len; 4242 } 4243 4244 static const struct ibnl_client_cbs cma_cb_table[] = { 4245 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats, 4246 .module = THIS_MODULE }, 4247 }; 4248 4249 static int cma_init_net(struct net *net) 4250 { 4251 struct cma_pernet *pernet = cma_pernet(net); 4252 4253 idr_init(&pernet->tcp_ps); 4254 idr_init(&pernet->udp_ps); 4255 idr_init(&pernet->ipoib_ps); 4256 idr_init(&pernet->ib_ps); 4257 4258 return 0; 4259 } 4260 4261 static void cma_exit_net(struct net *net) 4262 { 4263 struct cma_pernet *pernet = cma_pernet(net); 4264 4265 idr_destroy(&pernet->tcp_ps); 4266 idr_destroy(&pernet->udp_ps); 4267 idr_destroy(&pernet->ipoib_ps); 4268 idr_destroy(&pernet->ib_ps); 4269 } 4270 4271 static struct pernet_operations cma_pernet_operations = { 4272 .init = cma_init_net, 4273 .exit = cma_exit_net, 4274 .id = &cma_pernet_id, 4275 .size = sizeof(struct cma_pernet), 4276 }; 4277 4278 static int __init cma_init(void) 4279 { 4280 int ret; 4281 4282 cma_wq = create_singlethread_workqueue("rdma_cm"); 4283 if (!cma_wq) 4284 return -ENOMEM; 4285 4286 ret = register_pernet_subsys(&cma_pernet_operations); 4287 if (ret) 4288 goto err_wq; 4289 4290 ib_sa_register_client(&sa_client); 4291 rdma_addr_register_client(&addr_client); 4292 register_netdevice_notifier(&cma_nb); 4293 4294 ret = ib_register_client(&cma_client); 4295 if (ret) 4296 goto err; 4297 4298 if (ibnl_add_client(RDMA_NL_RDMA_CM, ARRAY_SIZE(cma_cb_table), 4299 cma_cb_table)) 4300 pr_warn("RDMA CMA: failed to add netlink callback\n"); 4301 cma_configfs_init(); 4302 4303 return 0; 4304 4305 err: 4306 unregister_netdevice_notifier(&cma_nb); 4307 rdma_addr_unregister_client(&addr_client); 4308 ib_sa_unregister_client(&sa_client); 4309 err_wq: 4310 destroy_workqueue(cma_wq); 4311 return ret; 4312 } 4313 4314 static void __exit cma_cleanup(void) 4315 { 4316 cma_configfs_exit(); 4317 ibnl_remove_client(RDMA_NL_RDMA_CM); 4318 ib_unregister_client(&cma_client); 4319 unregister_netdevice_notifier(&cma_nb); 4320 rdma_addr_unregister_client(&addr_client); 4321 ib_sa_unregister_client(&sa_client); 4322 unregister_pernet_subsys(&cma_pernet_operations); 4323 destroy_workqueue(cma_wq); 4324 } 4325 4326 module_init(cma_init); 4327 module_exit(cma_cleanup); 4328