1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB 2 /* 3 * Copyright (c) 2005 Voltaire Inc. All rights reserved. 4 * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved. 5 * Copyright (c) 1999-2019, Mellanox Technologies, Inc. All rights reserved. 6 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved. 7 */ 8 9 #include <linux/completion.h> 10 #include <linux/in.h> 11 #include <linux/in6.h> 12 #include <linux/mutex.h> 13 #include <linux/random.h> 14 #include <linux/igmp.h> 15 #include <linux/xarray.h> 16 #include <linux/inetdevice.h> 17 #include <linux/slab.h> 18 #include <linux/module.h> 19 #include <net/route.h> 20 21 #include <net/net_namespace.h> 22 #include <net/netns/generic.h> 23 #include <net/tcp.h> 24 #include <net/ipv6.h> 25 #include <net/ip_fib.h> 26 #include <net/ip6_route.h> 27 28 #include <rdma/rdma_cm.h> 29 #include <rdma/rdma_cm_ib.h> 30 #include <rdma/rdma_netlink.h> 31 #include <rdma/ib.h> 32 #include <rdma/ib_cache.h> 33 #include <rdma/ib_cm.h> 34 #include <rdma/ib_sa.h> 35 #include <rdma/iw_cm.h> 36 37 #include "core_priv.h" 38 #include "cma_priv.h" 39 #include "cma_trace.h" 40 41 MODULE_AUTHOR("Sean Hefty"); 42 MODULE_DESCRIPTION("Generic RDMA CM Agent"); 43 MODULE_LICENSE("Dual BSD/GPL"); 44 45 #define CMA_CM_RESPONSE_TIMEOUT 20 46 #define CMA_QUERY_CLASSPORT_INFO_TIMEOUT 3000 47 #define CMA_MAX_CM_RETRIES 15 48 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24) 49 #define CMA_IBOE_PACKET_LIFETIME 18 50 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP 51 52 static const char * const cma_events[] = { 53 [RDMA_CM_EVENT_ADDR_RESOLVED] = "address resolved", 54 [RDMA_CM_EVENT_ADDR_ERROR] = "address error", 55 [RDMA_CM_EVENT_ROUTE_RESOLVED] = "route resolved ", 56 [RDMA_CM_EVENT_ROUTE_ERROR] = "route error", 57 [RDMA_CM_EVENT_CONNECT_REQUEST] = "connect request", 58 [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response", 59 [RDMA_CM_EVENT_CONNECT_ERROR] = "connect error", 60 [RDMA_CM_EVENT_UNREACHABLE] = "unreachable", 61 [RDMA_CM_EVENT_REJECTED] = "rejected", 62 [RDMA_CM_EVENT_ESTABLISHED] = "established", 63 [RDMA_CM_EVENT_DISCONNECTED] = "disconnected", 64 [RDMA_CM_EVENT_DEVICE_REMOVAL] = "device removal", 65 [RDMA_CM_EVENT_MULTICAST_JOIN] = "multicast join", 66 [RDMA_CM_EVENT_MULTICAST_ERROR] = "multicast error", 67 [RDMA_CM_EVENT_ADDR_CHANGE] = "address change", 68 [RDMA_CM_EVENT_TIMEWAIT_EXIT] = "timewait exit", 69 }; 70 71 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event) 72 { 73 size_t index = event; 74 75 return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ? 76 cma_events[index] : "unrecognized event"; 77 } 78 EXPORT_SYMBOL(rdma_event_msg); 79 80 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id, 81 int reason) 82 { 83 if (rdma_ib_or_roce(id->device, id->port_num)) 84 return ibcm_reject_msg(reason); 85 86 if (rdma_protocol_iwarp(id->device, id->port_num)) 87 return iwcm_reject_msg(reason); 88 89 WARN_ON_ONCE(1); 90 return "unrecognized transport"; 91 } 92 EXPORT_SYMBOL(rdma_reject_msg); 93 94 bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason) 95 { 96 if (rdma_ib_or_roce(id->device, id->port_num)) 97 return reason == IB_CM_REJ_CONSUMER_DEFINED; 98 99 if (rdma_protocol_iwarp(id->device, id->port_num)) 100 return reason == -ECONNREFUSED; 101 102 WARN_ON_ONCE(1); 103 return false; 104 } 105 EXPORT_SYMBOL(rdma_is_consumer_reject); 106 107 const void *rdma_consumer_reject_data(struct rdma_cm_id *id, 108 struct rdma_cm_event *ev, u8 *data_len) 109 { 110 const void *p; 111 112 if (rdma_is_consumer_reject(id, ev->status)) { 113 *data_len = ev->param.conn.private_data_len; 114 p = ev->param.conn.private_data; 115 } else { 116 *data_len = 0; 117 p = NULL; 118 } 119 return p; 120 } 121 EXPORT_SYMBOL(rdma_consumer_reject_data); 122 123 /** 124 * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id. 125 * @id: Communication Identifier 126 */ 127 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id) 128 { 129 struct rdma_id_private *id_priv; 130 131 id_priv = container_of(id, struct rdma_id_private, id); 132 if (id->device->node_type == RDMA_NODE_RNIC) 133 return id_priv->cm_id.iw; 134 return NULL; 135 } 136 EXPORT_SYMBOL(rdma_iw_cm_id); 137 138 /** 139 * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack. 140 * @res: rdma resource tracking entry pointer 141 */ 142 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res) 143 { 144 struct rdma_id_private *id_priv = 145 container_of(res, struct rdma_id_private, res); 146 147 return &id_priv->id; 148 } 149 EXPORT_SYMBOL(rdma_res_to_id); 150 151 static void cma_add_one(struct ib_device *device); 152 static void cma_remove_one(struct ib_device *device, void *client_data); 153 154 static struct ib_client cma_client = { 155 .name = "cma", 156 .add = cma_add_one, 157 .remove = cma_remove_one 158 }; 159 160 static struct ib_sa_client sa_client; 161 static LIST_HEAD(dev_list); 162 static LIST_HEAD(listen_any_list); 163 static DEFINE_MUTEX(lock); 164 static struct workqueue_struct *cma_wq; 165 static unsigned int cma_pernet_id; 166 167 struct cma_pernet { 168 struct xarray tcp_ps; 169 struct xarray udp_ps; 170 struct xarray ipoib_ps; 171 struct xarray ib_ps; 172 }; 173 174 static struct cma_pernet *cma_pernet(struct net *net) 175 { 176 return net_generic(net, cma_pernet_id); 177 } 178 179 static 180 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps) 181 { 182 struct cma_pernet *pernet = cma_pernet(net); 183 184 switch (ps) { 185 case RDMA_PS_TCP: 186 return &pernet->tcp_ps; 187 case RDMA_PS_UDP: 188 return &pernet->udp_ps; 189 case RDMA_PS_IPOIB: 190 return &pernet->ipoib_ps; 191 case RDMA_PS_IB: 192 return &pernet->ib_ps; 193 default: 194 return NULL; 195 } 196 } 197 198 struct cma_device { 199 struct list_head list; 200 struct ib_device *device; 201 struct completion comp; 202 atomic_t refcount; 203 struct list_head id_list; 204 enum ib_gid_type *default_gid_type; 205 u8 *default_roce_tos; 206 }; 207 208 struct rdma_bind_list { 209 enum rdma_ucm_port_space ps; 210 struct hlist_head owners; 211 unsigned short port; 212 }; 213 214 struct class_port_info_context { 215 struct ib_class_port_info *class_port_info; 216 struct ib_device *device; 217 struct completion done; 218 struct ib_sa_query *sa_query; 219 u8 port_num; 220 }; 221 222 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps, 223 struct rdma_bind_list *bind_list, int snum) 224 { 225 struct xarray *xa = cma_pernet_xa(net, ps); 226 227 return xa_insert(xa, snum, bind_list, GFP_KERNEL); 228 } 229 230 static struct rdma_bind_list *cma_ps_find(struct net *net, 231 enum rdma_ucm_port_space ps, int snum) 232 { 233 struct xarray *xa = cma_pernet_xa(net, ps); 234 235 return xa_load(xa, snum); 236 } 237 238 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps, 239 int snum) 240 { 241 struct xarray *xa = cma_pernet_xa(net, ps); 242 243 xa_erase(xa, snum); 244 } 245 246 enum { 247 CMA_OPTION_AFONLY, 248 }; 249 250 void cma_ref_dev(struct cma_device *cma_dev) 251 { 252 atomic_inc(&cma_dev->refcount); 253 } 254 255 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter, 256 void *cookie) 257 { 258 struct cma_device *cma_dev; 259 struct cma_device *found_cma_dev = NULL; 260 261 mutex_lock(&lock); 262 263 list_for_each_entry(cma_dev, &dev_list, list) 264 if (filter(cma_dev->device, cookie)) { 265 found_cma_dev = cma_dev; 266 break; 267 } 268 269 if (found_cma_dev) 270 cma_ref_dev(found_cma_dev); 271 mutex_unlock(&lock); 272 return found_cma_dev; 273 } 274 275 int cma_get_default_gid_type(struct cma_device *cma_dev, 276 unsigned int port) 277 { 278 if (!rdma_is_port_valid(cma_dev->device, port)) 279 return -EINVAL; 280 281 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)]; 282 } 283 284 int cma_set_default_gid_type(struct cma_device *cma_dev, 285 unsigned int port, 286 enum ib_gid_type default_gid_type) 287 { 288 unsigned long supported_gids; 289 290 if (!rdma_is_port_valid(cma_dev->device, port)) 291 return -EINVAL; 292 293 supported_gids = roce_gid_type_mask_support(cma_dev->device, port); 294 295 if (!(supported_gids & 1 << default_gid_type)) 296 return -EINVAL; 297 298 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] = 299 default_gid_type; 300 301 return 0; 302 } 303 304 int cma_get_default_roce_tos(struct cma_device *cma_dev, unsigned int port) 305 { 306 if (!rdma_is_port_valid(cma_dev->device, port)) 307 return -EINVAL; 308 309 return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)]; 310 } 311 312 int cma_set_default_roce_tos(struct cma_device *cma_dev, unsigned int port, 313 u8 default_roce_tos) 314 { 315 if (!rdma_is_port_valid(cma_dev->device, port)) 316 return -EINVAL; 317 318 cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] = 319 default_roce_tos; 320 321 return 0; 322 } 323 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev) 324 { 325 return cma_dev->device; 326 } 327 328 /* 329 * Device removal can occur at anytime, so we need extra handling to 330 * serialize notifying the user of device removal with other callbacks. 331 * We do this by disabling removal notification while a callback is in process, 332 * and reporting it after the callback completes. 333 */ 334 335 struct cma_multicast { 336 struct rdma_id_private *id_priv; 337 union { 338 struct ib_sa_multicast *ib; 339 } multicast; 340 struct list_head list; 341 void *context; 342 struct sockaddr_storage addr; 343 struct kref mcref; 344 u8 join_state; 345 }; 346 347 struct cma_work { 348 struct work_struct work; 349 struct rdma_id_private *id; 350 enum rdma_cm_state old_state; 351 enum rdma_cm_state new_state; 352 struct rdma_cm_event event; 353 }; 354 355 struct cma_ndev_work { 356 struct work_struct work; 357 struct rdma_id_private *id; 358 struct rdma_cm_event event; 359 }; 360 361 struct iboe_mcast_work { 362 struct work_struct work; 363 struct rdma_id_private *id; 364 struct cma_multicast *mc; 365 }; 366 367 union cma_ip_addr { 368 struct in6_addr ip6; 369 struct { 370 __be32 pad[3]; 371 __be32 addr; 372 } ip4; 373 }; 374 375 struct cma_hdr { 376 u8 cma_version; 377 u8 ip_version; /* IP version: 7:4 */ 378 __be16 port; 379 union cma_ip_addr src_addr; 380 union cma_ip_addr dst_addr; 381 }; 382 383 #define CMA_VERSION 0x00 384 385 struct cma_req_info { 386 struct sockaddr_storage listen_addr_storage; 387 struct sockaddr_storage src_addr_storage; 388 struct ib_device *device; 389 union ib_gid local_gid; 390 __be64 service_id; 391 int port; 392 bool has_gid; 393 u16 pkey; 394 }; 395 396 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp) 397 { 398 unsigned long flags; 399 int ret; 400 401 spin_lock_irqsave(&id_priv->lock, flags); 402 ret = (id_priv->state == comp); 403 spin_unlock_irqrestore(&id_priv->lock, flags); 404 return ret; 405 } 406 407 static int cma_comp_exch(struct rdma_id_private *id_priv, 408 enum rdma_cm_state comp, enum rdma_cm_state exch) 409 { 410 unsigned long flags; 411 int ret; 412 413 spin_lock_irqsave(&id_priv->lock, flags); 414 if ((ret = (id_priv->state == comp))) 415 id_priv->state = exch; 416 spin_unlock_irqrestore(&id_priv->lock, flags); 417 return ret; 418 } 419 420 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv, 421 enum rdma_cm_state exch) 422 { 423 unsigned long flags; 424 enum rdma_cm_state old; 425 426 spin_lock_irqsave(&id_priv->lock, flags); 427 old = id_priv->state; 428 id_priv->state = exch; 429 spin_unlock_irqrestore(&id_priv->lock, flags); 430 return old; 431 } 432 433 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr) 434 { 435 return hdr->ip_version >> 4; 436 } 437 438 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver) 439 { 440 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF); 441 } 442 443 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join) 444 { 445 struct in_device *in_dev = NULL; 446 447 if (ndev) { 448 rtnl_lock(); 449 in_dev = __in_dev_get_rtnl(ndev); 450 if (in_dev) { 451 if (join) 452 ip_mc_inc_group(in_dev, 453 *(__be32 *)(mgid->raw + 12)); 454 else 455 ip_mc_dec_group(in_dev, 456 *(__be32 *)(mgid->raw + 12)); 457 } 458 rtnl_unlock(); 459 } 460 return (in_dev) ? 0 : -ENODEV; 461 } 462 463 static void _cma_attach_to_dev(struct rdma_id_private *id_priv, 464 struct cma_device *cma_dev) 465 { 466 cma_ref_dev(cma_dev); 467 id_priv->cma_dev = cma_dev; 468 id_priv->id.device = cma_dev->device; 469 id_priv->id.route.addr.dev_addr.transport = 470 rdma_node_get_transport(cma_dev->device->node_type); 471 list_add_tail(&id_priv->list, &cma_dev->id_list); 472 if (id_priv->res.kern_name) 473 rdma_restrack_kadd(&id_priv->res); 474 else 475 rdma_restrack_uadd(&id_priv->res); 476 } 477 478 static void cma_attach_to_dev(struct rdma_id_private *id_priv, 479 struct cma_device *cma_dev) 480 { 481 _cma_attach_to_dev(id_priv, cma_dev); 482 id_priv->gid_type = 483 cma_dev->default_gid_type[id_priv->id.port_num - 484 rdma_start_port(cma_dev->device)]; 485 } 486 487 void cma_deref_dev(struct cma_device *cma_dev) 488 { 489 if (atomic_dec_and_test(&cma_dev->refcount)) 490 complete(&cma_dev->comp); 491 } 492 493 static inline void release_mc(struct kref *kref) 494 { 495 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref); 496 497 kfree(mc->multicast.ib); 498 kfree(mc); 499 } 500 501 static void cma_release_dev(struct rdma_id_private *id_priv) 502 { 503 mutex_lock(&lock); 504 list_del(&id_priv->list); 505 cma_deref_dev(id_priv->cma_dev); 506 id_priv->cma_dev = NULL; 507 mutex_unlock(&lock); 508 } 509 510 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv) 511 { 512 return (struct sockaddr *) &id_priv->id.route.addr.src_addr; 513 } 514 515 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv) 516 { 517 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr; 518 } 519 520 static inline unsigned short cma_family(struct rdma_id_private *id_priv) 521 { 522 return id_priv->id.route.addr.src_addr.ss_family; 523 } 524 525 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey) 526 { 527 struct ib_sa_mcmember_rec rec; 528 int ret = 0; 529 530 if (id_priv->qkey) { 531 if (qkey && id_priv->qkey != qkey) 532 return -EINVAL; 533 return 0; 534 } 535 536 if (qkey) { 537 id_priv->qkey = qkey; 538 return 0; 539 } 540 541 switch (id_priv->id.ps) { 542 case RDMA_PS_UDP: 543 case RDMA_PS_IB: 544 id_priv->qkey = RDMA_UDP_QKEY; 545 break; 546 case RDMA_PS_IPOIB: 547 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid); 548 ret = ib_sa_get_mcmember_rec(id_priv->id.device, 549 id_priv->id.port_num, &rec.mgid, 550 &rec); 551 if (!ret) 552 id_priv->qkey = be32_to_cpu(rec.qkey); 553 break; 554 default: 555 break; 556 } 557 return ret; 558 } 559 560 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr) 561 { 562 dev_addr->dev_type = ARPHRD_INFINIBAND; 563 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr); 564 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey)); 565 } 566 567 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr) 568 { 569 int ret; 570 571 if (addr->sa_family != AF_IB) { 572 ret = rdma_translate_ip(addr, dev_addr); 573 } else { 574 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr); 575 ret = 0; 576 } 577 578 return ret; 579 } 580 581 static const struct ib_gid_attr * 582 cma_validate_port(struct ib_device *device, u8 port, 583 enum ib_gid_type gid_type, 584 union ib_gid *gid, 585 struct rdma_id_private *id_priv) 586 { 587 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 588 int bound_if_index = dev_addr->bound_dev_if; 589 const struct ib_gid_attr *sgid_attr; 590 int dev_type = dev_addr->dev_type; 591 struct net_device *ndev = NULL; 592 593 if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net)) 594 return ERR_PTR(-ENODEV); 595 596 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port)) 597 return ERR_PTR(-ENODEV); 598 599 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port)) 600 return ERR_PTR(-ENODEV); 601 602 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) { 603 ndev = dev_get_by_index(dev_addr->net, bound_if_index); 604 if (!ndev) 605 return ERR_PTR(-ENODEV); 606 } else { 607 gid_type = IB_GID_TYPE_IB; 608 } 609 610 sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev); 611 if (ndev) 612 dev_put(ndev); 613 return sgid_attr; 614 } 615 616 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv, 617 const struct ib_gid_attr *sgid_attr) 618 { 619 WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr); 620 id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr; 621 } 622 623 /** 624 * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute 625 * based on source ip address. 626 * @id_priv: cm_id which should be bound to cma device 627 * 628 * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute 629 * based on source IP address. It returns 0 on success or error code otherwise. 630 * It is applicable to active and passive side cm_id. 631 */ 632 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv) 633 { 634 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 635 const struct ib_gid_attr *sgid_attr; 636 union ib_gid gid, iboe_gid, *gidp; 637 struct cma_device *cma_dev; 638 enum ib_gid_type gid_type; 639 int ret = -ENODEV; 640 unsigned int port; 641 642 if (dev_addr->dev_type != ARPHRD_INFINIBAND && 643 id_priv->id.ps == RDMA_PS_IPOIB) 644 return -EINVAL; 645 646 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr, 647 &iboe_gid); 648 649 memcpy(&gid, dev_addr->src_dev_addr + 650 rdma_addr_gid_offset(dev_addr), sizeof(gid)); 651 652 mutex_lock(&lock); 653 list_for_each_entry(cma_dev, &dev_list, list) { 654 rdma_for_each_port (cma_dev->device, port) { 655 gidp = rdma_protocol_roce(cma_dev->device, port) ? 656 &iboe_gid : &gid; 657 gid_type = cma_dev->default_gid_type[port - 1]; 658 sgid_attr = cma_validate_port(cma_dev->device, port, 659 gid_type, gidp, id_priv); 660 if (!IS_ERR(sgid_attr)) { 661 id_priv->id.port_num = port; 662 cma_bind_sgid_attr(id_priv, sgid_attr); 663 cma_attach_to_dev(id_priv, cma_dev); 664 ret = 0; 665 goto out; 666 } 667 } 668 } 669 out: 670 mutex_unlock(&lock); 671 return ret; 672 } 673 674 /** 675 * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute 676 * @id_priv: cm id to bind to cma device 677 * @listen_id_priv: listener cm id to match against 678 * @req: Pointer to req structure containaining incoming 679 * request information 680 * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when 681 * rdma device matches for listen_id and incoming request. It also verifies 682 * that a GID table entry is present for the source address. 683 * Returns 0 on success, or returns error code otherwise. 684 */ 685 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv, 686 const struct rdma_id_private *listen_id_priv, 687 struct cma_req_info *req) 688 { 689 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 690 const struct ib_gid_attr *sgid_attr; 691 enum ib_gid_type gid_type; 692 union ib_gid gid; 693 694 if (dev_addr->dev_type != ARPHRD_INFINIBAND && 695 id_priv->id.ps == RDMA_PS_IPOIB) 696 return -EINVAL; 697 698 if (rdma_protocol_roce(req->device, req->port)) 699 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr, 700 &gid); 701 else 702 memcpy(&gid, dev_addr->src_dev_addr + 703 rdma_addr_gid_offset(dev_addr), sizeof(gid)); 704 705 gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1]; 706 sgid_attr = cma_validate_port(req->device, req->port, 707 gid_type, &gid, id_priv); 708 if (IS_ERR(sgid_attr)) 709 return PTR_ERR(sgid_attr); 710 711 id_priv->id.port_num = req->port; 712 cma_bind_sgid_attr(id_priv, sgid_attr); 713 /* Need to acquire lock to protect against reader 714 * of cma_dev->id_list such as cma_netdev_callback() and 715 * cma_process_remove(). 716 */ 717 mutex_lock(&lock); 718 cma_attach_to_dev(id_priv, listen_id_priv->cma_dev); 719 mutex_unlock(&lock); 720 return 0; 721 } 722 723 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv, 724 const struct rdma_id_private *listen_id_priv) 725 { 726 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 727 const struct ib_gid_attr *sgid_attr; 728 struct cma_device *cma_dev; 729 enum ib_gid_type gid_type; 730 int ret = -ENODEV; 731 union ib_gid gid; 732 u8 port; 733 734 if (dev_addr->dev_type != ARPHRD_INFINIBAND && 735 id_priv->id.ps == RDMA_PS_IPOIB) 736 return -EINVAL; 737 738 memcpy(&gid, dev_addr->src_dev_addr + 739 rdma_addr_gid_offset(dev_addr), sizeof(gid)); 740 741 mutex_lock(&lock); 742 743 cma_dev = listen_id_priv->cma_dev; 744 port = listen_id_priv->id.port_num; 745 gid_type = listen_id_priv->gid_type; 746 sgid_attr = cma_validate_port(cma_dev->device, port, 747 gid_type, &gid, id_priv); 748 if (!IS_ERR(sgid_attr)) { 749 id_priv->id.port_num = port; 750 cma_bind_sgid_attr(id_priv, sgid_attr); 751 ret = 0; 752 goto out; 753 } 754 755 list_for_each_entry(cma_dev, &dev_list, list) { 756 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) { 757 if (listen_id_priv->cma_dev == cma_dev && 758 listen_id_priv->id.port_num == port) 759 continue; 760 761 gid_type = cma_dev->default_gid_type[port - 1]; 762 sgid_attr = cma_validate_port(cma_dev->device, port, 763 gid_type, &gid, id_priv); 764 if (!IS_ERR(sgid_attr)) { 765 id_priv->id.port_num = port; 766 cma_bind_sgid_attr(id_priv, sgid_attr); 767 ret = 0; 768 goto out; 769 } 770 } 771 } 772 773 out: 774 if (!ret) 775 cma_attach_to_dev(id_priv, cma_dev); 776 777 mutex_unlock(&lock); 778 return ret; 779 } 780 781 /* 782 * Select the source IB device and address to reach the destination IB address. 783 */ 784 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv) 785 { 786 struct cma_device *cma_dev, *cur_dev; 787 struct sockaddr_ib *addr; 788 union ib_gid gid, sgid, *dgid; 789 u16 pkey, index; 790 u8 p; 791 enum ib_port_state port_state; 792 int i; 793 794 cma_dev = NULL; 795 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv); 796 dgid = (union ib_gid *) &addr->sib_addr; 797 pkey = ntohs(addr->sib_pkey); 798 799 mutex_lock(&lock); 800 list_for_each_entry(cur_dev, &dev_list, list) { 801 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) { 802 if (!rdma_cap_af_ib(cur_dev->device, p)) 803 continue; 804 805 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index)) 806 continue; 807 808 if (ib_get_cached_port_state(cur_dev->device, p, &port_state)) 809 continue; 810 for (i = 0; !rdma_query_gid(cur_dev->device, 811 p, i, &gid); 812 i++) { 813 if (!memcmp(&gid, dgid, sizeof(gid))) { 814 cma_dev = cur_dev; 815 sgid = gid; 816 id_priv->id.port_num = p; 817 goto found; 818 } 819 820 if (!cma_dev && (gid.global.subnet_prefix == 821 dgid->global.subnet_prefix) && 822 port_state == IB_PORT_ACTIVE) { 823 cma_dev = cur_dev; 824 sgid = gid; 825 id_priv->id.port_num = p; 826 goto found; 827 } 828 } 829 } 830 } 831 mutex_unlock(&lock); 832 return -ENODEV; 833 834 found: 835 cma_attach_to_dev(id_priv, cma_dev); 836 mutex_unlock(&lock); 837 addr = (struct sockaddr_ib *)cma_src_addr(id_priv); 838 memcpy(&addr->sib_addr, &sgid, sizeof(sgid)); 839 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr); 840 return 0; 841 } 842 843 static void cma_deref_id(struct rdma_id_private *id_priv) 844 { 845 if (atomic_dec_and_test(&id_priv->refcount)) 846 complete(&id_priv->comp); 847 } 848 849 struct rdma_cm_id *__rdma_create_id(struct net *net, 850 rdma_cm_event_handler event_handler, 851 void *context, enum rdma_ucm_port_space ps, 852 enum ib_qp_type qp_type, const char *caller) 853 { 854 struct rdma_id_private *id_priv; 855 856 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL); 857 if (!id_priv) 858 return ERR_PTR(-ENOMEM); 859 860 rdma_restrack_set_task(&id_priv->res, caller); 861 id_priv->res.type = RDMA_RESTRACK_CM_ID; 862 id_priv->state = RDMA_CM_IDLE; 863 id_priv->id.context = context; 864 id_priv->id.event_handler = event_handler; 865 id_priv->id.ps = ps; 866 id_priv->id.qp_type = qp_type; 867 id_priv->tos_set = false; 868 id_priv->timeout_set = false; 869 id_priv->gid_type = IB_GID_TYPE_IB; 870 spin_lock_init(&id_priv->lock); 871 mutex_init(&id_priv->qp_mutex); 872 init_completion(&id_priv->comp); 873 atomic_set(&id_priv->refcount, 1); 874 mutex_init(&id_priv->handler_mutex); 875 INIT_LIST_HEAD(&id_priv->listen_list); 876 INIT_LIST_HEAD(&id_priv->mc_list); 877 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num); 878 id_priv->id.route.addr.dev_addr.net = get_net(net); 879 id_priv->seq_num &= 0x00ffffff; 880 881 trace_cm_id_create(id_priv); 882 return &id_priv->id; 883 } 884 EXPORT_SYMBOL(__rdma_create_id); 885 886 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp) 887 { 888 struct ib_qp_attr qp_attr; 889 int qp_attr_mask, ret; 890 891 qp_attr.qp_state = IB_QPS_INIT; 892 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 893 if (ret) 894 return ret; 895 896 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 897 if (ret) 898 return ret; 899 900 qp_attr.qp_state = IB_QPS_RTR; 901 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE); 902 if (ret) 903 return ret; 904 905 qp_attr.qp_state = IB_QPS_RTS; 906 qp_attr.sq_psn = 0; 907 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN); 908 909 return ret; 910 } 911 912 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp) 913 { 914 struct ib_qp_attr qp_attr; 915 int qp_attr_mask, ret; 916 917 qp_attr.qp_state = IB_QPS_INIT; 918 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 919 if (ret) 920 return ret; 921 922 return ib_modify_qp(qp, &qp_attr, qp_attr_mask); 923 } 924 925 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd, 926 struct ib_qp_init_attr *qp_init_attr) 927 { 928 struct rdma_id_private *id_priv; 929 struct ib_qp *qp; 930 int ret; 931 932 id_priv = container_of(id, struct rdma_id_private, id); 933 if (id->device != pd->device) { 934 ret = -EINVAL; 935 goto out_err; 936 } 937 938 qp_init_attr->port_num = id->port_num; 939 qp = ib_create_qp(pd, qp_init_attr); 940 if (IS_ERR(qp)) { 941 ret = PTR_ERR(qp); 942 goto out_err; 943 } 944 945 if (id->qp_type == IB_QPT_UD) 946 ret = cma_init_ud_qp(id_priv, qp); 947 else 948 ret = cma_init_conn_qp(id_priv, qp); 949 if (ret) 950 goto out_destroy; 951 952 id->qp = qp; 953 id_priv->qp_num = qp->qp_num; 954 id_priv->srq = (qp->srq != NULL); 955 trace_cm_qp_create(id_priv, pd, qp_init_attr, 0); 956 return 0; 957 out_destroy: 958 ib_destroy_qp(qp); 959 out_err: 960 trace_cm_qp_create(id_priv, pd, qp_init_attr, ret); 961 return ret; 962 } 963 EXPORT_SYMBOL(rdma_create_qp); 964 965 void rdma_destroy_qp(struct rdma_cm_id *id) 966 { 967 struct rdma_id_private *id_priv; 968 969 id_priv = container_of(id, struct rdma_id_private, id); 970 trace_cm_qp_destroy(id_priv); 971 mutex_lock(&id_priv->qp_mutex); 972 ib_destroy_qp(id_priv->id.qp); 973 id_priv->id.qp = NULL; 974 mutex_unlock(&id_priv->qp_mutex); 975 } 976 EXPORT_SYMBOL(rdma_destroy_qp); 977 978 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv, 979 struct rdma_conn_param *conn_param) 980 { 981 struct ib_qp_attr qp_attr; 982 int qp_attr_mask, ret; 983 984 mutex_lock(&id_priv->qp_mutex); 985 if (!id_priv->id.qp) { 986 ret = 0; 987 goto out; 988 } 989 990 /* Need to update QP attributes from default values. */ 991 qp_attr.qp_state = IB_QPS_INIT; 992 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 993 if (ret) 994 goto out; 995 996 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 997 if (ret) 998 goto out; 999 1000 qp_attr.qp_state = IB_QPS_RTR; 1001 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 1002 if (ret) 1003 goto out; 1004 1005 BUG_ON(id_priv->cma_dev->device != id_priv->id.device); 1006 1007 if (conn_param) 1008 qp_attr.max_dest_rd_atomic = conn_param->responder_resources; 1009 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 1010 out: 1011 mutex_unlock(&id_priv->qp_mutex); 1012 return ret; 1013 } 1014 1015 static int cma_modify_qp_rts(struct rdma_id_private *id_priv, 1016 struct rdma_conn_param *conn_param) 1017 { 1018 struct ib_qp_attr qp_attr; 1019 int qp_attr_mask, ret; 1020 1021 mutex_lock(&id_priv->qp_mutex); 1022 if (!id_priv->id.qp) { 1023 ret = 0; 1024 goto out; 1025 } 1026 1027 qp_attr.qp_state = IB_QPS_RTS; 1028 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 1029 if (ret) 1030 goto out; 1031 1032 if (conn_param) 1033 qp_attr.max_rd_atomic = conn_param->initiator_depth; 1034 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 1035 out: 1036 mutex_unlock(&id_priv->qp_mutex); 1037 return ret; 1038 } 1039 1040 static int cma_modify_qp_err(struct rdma_id_private *id_priv) 1041 { 1042 struct ib_qp_attr qp_attr; 1043 int ret; 1044 1045 mutex_lock(&id_priv->qp_mutex); 1046 if (!id_priv->id.qp) { 1047 ret = 0; 1048 goto out; 1049 } 1050 1051 qp_attr.qp_state = IB_QPS_ERR; 1052 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE); 1053 out: 1054 mutex_unlock(&id_priv->qp_mutex); 1055 return ret; 1056 } 1057 1058 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv, 1059 struct ib_qp_attr *qp_attr, int *qp_attr_mask) 1060 { 1061 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 1062 int ret; 1063 u16 pkey; 1064 1065 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num)) 1066 pkey = 0xffff; 1067 else 1068 pkey = ib_addr_get_pkey(dev_addr); 1069 1070 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num, 1071 pkey, &qp_attr->pkey_index); 1072 if (ret) 1073 return ret; 1074 1075 qp_attr->port_num = id_priv->id.port_num; 1076 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT; 1077 1078 if (id_priv->id.qp_type == IB_QPT_UD) { 1079 ret = cma_set_qkey(id_priv, 0); 1080 if (ret) 1081 return ret; 1082 1083 qp_attr->qkey = id_priv->qkey; 1084 *qp_attr_mask |= IB_QP_QKEY; 1085 } else { 1086 qp_attr->qp_access_flags = 0; 1087 *qp_attr_mask |= IB_QP_ACCESS_FLAGS; 1088 } 1089 return 0; 1090 } 1091 1092 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr, 1093 int *qp_attr_mask) 1094 { 1095 struct rdma_id_private *id_priv; 1096 int ret = 0; 1097 1098 id_priv = container_of(id, struct rdma_id_private, id); 1099 if (rdma_cap_ib_cm(id->device, id->port_num)) { 1100 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD)) 1101 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask); 1102 else 1103 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr, 1104 qp_attr_mask); 1105 1106 if (qp_attr->qp_state == IB_QPS_RTR) 1107 qp_attr->rq_psn = id_priv->seq_num; 1108 } else if (rdma_cap_iw_cm(id->device, id->port_num)) { 1109 if (!id_priv->cm_id.iw) { 1110 qp_attr->qp_access_flags = 0; 1111 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS; 1112 } else 1113 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr, 1114 qp_attr_mask); 1115 qp_attr->port_num = id_priv->id.port_num; 1116 *qp_attr_mask |= IB_QP_PORT; 1117 } else 1118 ret = -ENOSYS; 1119 1120 if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set) 1121 qp_attr->timeout = id_priv->timeout; 1122 1123 return ret; 1124 } 1125 EXPORT_SYMBOL(rdma_init_qp_attr); 1126 1127 static inline bool cma_zero_addr(const struct sockaddr *addr) 1128 { 1129 switch (addr->sa_family) { 1130 case AF_INET: 1131 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr); 1132 case AF_INET6: 1133 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr); 1134 case AF_IB: 1135 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr); 1136 default: 1137 return false; 1138 } 1139 } 1140 1141 static inline bool cma_loopback_addr(const struct sockaddr *addr) 1142 { 1143 switch (addr->sa_family) { 1144 case AF_INET: 1145 return ipv4_is_loopback( 1146 ((struct sockaddr_in *)addr)->sin_addr.s_addr); 1147 case AF_INET6: 1148 return ipv6_addr_loopback( 1149 &((struct sockaddr_in6 *)addr)->sin6_addr); 1150 case AF_IB: 1151 return ib_addr_loopback( 1152 &((struct sockaddr_ib *)addr)->sib_addr); 1153 default: 1154 return false; 1155 } 1156 } 1157 1158 static inline bool cma_any_addr(const struct sockaddr *addr) 1159 { 1160 return cma_zero_addr(addr) || cma_loopback_addr(addr); 1161 } 1162 1163 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst) 1164 { 1165 if (src->sa_family != dst->sa_family) 1166 return -1; 1167 1168 switch (src->sa_family) { 1169 case AF_INET: 1170 return ((struct sockaddr_in *)src)->sin_addr.s_addr != 1171 ((struct sockaddr_in *)dst)->sin_addr.s_addr; 1172 case AF_INET6: { 1173 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src; 1174 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst; 1175 bool link_local; 1176 1177 if (ipv6_addr_cmp(&src_addr6->sin6_addr, 1178 &dst_addr6->sin6_addr)) 1179 return 1; 1180 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) & 1181 IPV6_ADDR_LINKLOCAL; 1182 /* Link local must match their scope_ids */ 1183 return link_local ? (src_addr6->sin6_scope_id != 1184 dst_addr6->sin6_scope_id) : 1185 0; 1186 } 1187 1188 default: 1189 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr, 1190 &((struct sockaddr_ib *) dst)->sib_addr); 1191 } 1192 } 1193 1194 static __be16 cma_port(const struct sockaddr *addr) 1195 { 1196 struct sockaddr_ib *sib; 1197 1198 switch (addr->sa_family) { 1199 case AF_INET: 1200 return ((struct sockaddr_in *) addr)->sin_port; 1201 case AF_INET6: 1202 return ((struct sockaddr_in6 *) addr)->sin6_port; 1203 case AF_IB: 1204 sib = (struct sockaddr_ib *) addr; 1205 return htons((u16) (be64_to_cpu(sib->sib_sid) & 1206 be64_to_cpu(sib->sib_sid_mask))); 1207 default: 1208 return 0; 1209 } 1210 } 1211 1212 static inline int cma_any_port(const struct sockaddr *addr) 1213 { 1214 return !cma_port(addr); 1215 } 1216 1217 static void cma_save_ib_info(struct sockaddr *src_addr, 1218 struct sockaddr *dst_addr, 1219 const struct rdma_cm_id *listen_id, 1220 const struct sa_path_rec *path) 1221 { 1222 struct sockaddr_ib *listen_ib, *ib; 1223 1224 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr; 1225 if (src_addr) { 1226 ib = (struct sockaddr_ib *)src_addr; 1227 ib->sib_family = AF_IB; 1228 if (path) { 1229 ib->sib_pkey = path->pkey; 1230 ib->sib_flowinfo = path->flow_label; 1231 memcpy(&ib->sib_addr, &path->sgid, 16); 1232 ib->sib_sid = path->service_id; 1233 ib->sib_scope_id = 0; 1234 } else { 1235 ib->sib_pkey = listen_ib->sib_pkey; 1236 ib->sib_flowinfo = listen_ib->sib_flowinfo; 1237 ib->sib_addr = listen_ib->sib_addr; 1238 ib->sib_sid = listen_ib->sib_sid; 1239 ib->sib_scope_id = listen_ib->sib_scope_id; 1240 } 1241 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL); 1242 } 1243 if (dst_addr) { 1244 ib = (struct sockaddr_ib *)dst_addr; 1245 ib->sib_family = AF_IB; 1246 if (path) { 1247 ib->sib_pkey = path->pkey; 1248 ib->sib_flowinfo = path->flow_label; 1249 memcpy(&ib->sib_addr, &path->dgid, 16); 1250 } 1251 } 1252 } 1253 1254 static void cma_save_ip4_info(struct sockaddr_in *src_addr, 1255 struct sockaddr_in *dst_addr, 1256 struct cma_hdr *hdr, 1257 __be16 local_port) 1258 { 1259 if (src_addr) { 1260 *src_addr = (struct sockaddr_in) { 1261 .sin_family = AF_INET, 1262 .sin_addr.s_addr = hdr->dst_addr.ip4.addr, 1263 .sin_port = local_port, 1264 }; 1265 } 1266 1267 if (dst_addr) { 1268 *dst_addr = (struct sockaddr_in) { 1269 .sin_family = AF_INET, 1270 .sin_addr.s_addr = hdr->src_addr.ip4.addr, 1271 .sin_port = hdr->port, 1272 }; 1273 } 1274 } 1275 1276 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr, 1277 struct sockaddr_in6 *dst_addr, 1278 struct cma_hdr *hdr, 1279 __be16 local_port) 1280 { 1281 if (src_addr) { 1282 *src_addr = (struct sockaddr_in6) { 1283 .sin6_family = AF_INET6, 1284 .sin6_addr = hdr->dst_addr.ip6, 1285 .sin6_port = local_port, 1286 }; 1287 } 1288 1289 if (dst_addr) { 1290 *dst_addr = (struct sockaddr_in6) { 1291 .sin6_family = AF_INET6, 1292 .sin6_addr = hdr->src_addr.ip6, 1293 .sin6_port = hdr->port, 1294 }; 1295 } 1296 } 1297 1298 static u16 cma_port_from_service_id(__be64 service_id) 1299 { 1300 return (u16)be64_to_cpu(service_id); 1301 } 1302 1303 static int cma_save_ip_info(struct sockaddr *src_addr, 1304 struct sockaddr *dst_addr, 1305 const struct ib_cm_event *ib_event, 1306 __be64 service_id) 1307 { 1308 struct cma_hdr *hdr; 1309 __be16 port; 1310 1311 hdr = ib_event->private_data; 1312 if (hdr->cma_version != CMA_VERSION) 1313 return -EINVAL; 1314 1315 port = htons(cma_port_from_service_id(service_id)); 1316 1317 switch (cma_get_ip_ver(hdr)) { 1318 case 4: 1319 cma_save_ip4_info((struct sockaddr_in *)src_addr, 1320 (struct sockaddr_in *)dst_addr, hdr, port); 1321 break; 1322 case 6: 1323 cma_save_ip6_info((struct sockaddr_in6 *)src_addr, 1324 (struct sockaddr_in6 *)dst_addr, hdr, port); 1325 break; 1326 default: 1327 return -EAFNOSUPPORT; 1328 } 1329 1330 return 0; 1331 } 1332 1333 static int cma_save_net_info(struct sockaddr *src_addr, 1334 struct sockaddr *dst_addr, 1335 const struct rdma_cm_id *listen_id, 1336 const struct ib_cm_event *ib_event, 1337 sa_family_t sa_family, __be64 service_id) 1338 { 1339 if (sa_family == AF_IB) { 1340 if (ib_event->event == IB_CM_REQ_RECEIVED) 1341 cma_save_ib_info(src_addr, dst_addr, listen_id, 1342 ib_event->param.req_rcvd.primary_path); 1343 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) 1344 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL); 1345 return 0; 1346 } 1347 1348 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id); 1349 } 1350 1351 static int cma_save_req_info(const struct ib_cm_event *ib_event, 1352 struct cma_req_info *req) 1353 { 1354 const struct ib_cm_req_event_param *req_param = 1355 &ib_event->param.req_rcvd; 1356 const struct ib_cm_sidr_req_event_param *sidr_param = 1357 &ib_event->param.sidr_req_rcvd; 1358 1359 switch (ib_event->event) { 1360 case IB_CM_REQ_RECEIVED: 1361 req->device = req_param->listen_id->device; 1362 req->port = req_param->port; 1363 memcpy(&req->local_gid, &req_param->primary_path->sgid, 1364 sizeof(req->local_gid)); 1365 req->has_gid = true; 1366 req->service_id = req_param->primary_path->service_id; 1367 req->pkey = be16_to_cpu(req_param->primary_path->pkey); 1368 if (req->pkey != req_param->bth_pkey) 1369 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n" 1370 "RDMA CMA: in the future this may cause the request to be dropped\n", 1371 req_param->bth_pkey, req->pkey); 1372 break; 1373 case IB_CM_SIDR_REQ_RECEIVED: 1374 req->device = sidr_param->listen_id->device; 1375 req->port = sidr_param->port; 1376 req->has_gid = false; 1377 req->service_id = sidr_param->service_id; 1378 req->pkey = sidr_param->pkey; 1379 if (req->pkey != sidr_param->bth_pkey) 1380 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n" 1381 "RDMA CMA: in the future this may cause the request to be dropped\n", 1382 sidr_param->bth_pkey, req->pkey); 1383 break; 1384 default: 1385 return -EINVAL; 1386 } 1387 1388 return 0; 1389 } 1390 1391 static bool validate_ipv4_net_dev(struct net_device *net_dev, 1392 const struct sockaddr_in *dst_addr, 1393 const struct sockaddr_in *src_addr) 1394 { 1395 __be32 daddr = dst_addr->sin_addr.s_addr, 1396 saddr = src_addr->sin_addr.s_addr; 1397 struct fib_result res; 1398 struct flowi4 fl4; 1399 int err; 1400 bool ret; 1401 1402 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) || 1403 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) || 1404 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) || 1405 ipv4_is_loopback(saddr)) 1406 return false; 1407 1408 memset(&fl4, 0, sizeof(fl4)); 1409 fl4.flowi4_iif = net_dev->ifindex; 1410 fl4.daddr = daddr; 1411 fl4.saddr = saddr; 1412 1413 rcu_read_lock(); 1414 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0); 1415 ret = err == 0 && FIB_RES_DEV(res) == net_dev; 1416 rcu_read_unlock(); 1417 1418 return ret; 1419 } 1420 1421 static bool validate_ipv6_net_dev(struct net_device *net_dev, 1422 const struct sockaddr_in6 *dst_addr, 1423 const struct sockaddr_in6 *src_addr) 1424 { 1425 #if IS_ENABLED(CONFIG_IPV6) 1426 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) & 1427 IPV6_ADDR_LINKLOCAL; 1428 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr, 1429 &src_addr->sin6_addr, net_dev->ifindex, 1430 NULL, strict); 1431 bool ret; 1432 1433 if (!rt) 1434 return false; 1435 1436 ret = rt->rt6i_idev->dev == net_dev; 1437 ip6_rt_put(rt); 1438 1439 return ret; 1440 #else 1441 return false; 1442 #endif 1443 } 1444 1445 static bool validate_net_dev(struct net_device *net_dev, 1446 const struct sockaddr *daddr, 1447 const struct sockaddr *saddr) 1448 { 1449 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr; 1450 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr; 1451 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr; 1452 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr; 1453 1454 switch (daddr->sa_family) { 1455 case AF_INET: 1456 return saddr->sa_family == AF_INET && 1457 validate_ipv4_net_dev(net_dev, daddr4, saddr4); 1458 1459 case AF_INET6: 1460 return saddr->sa_family == AF_INET6 && 1461 validate_ipv6_net_dev(net_dev, daddr6, saddr6); 1462 1463 default: 1464 return false; 1465 } 1466 } 1467 1468 static struct net_device * 1469 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event) 1470 { 1471 const struct ib_gid_attr *sgid_attr = NULL; 1472 struct net_device *ndev; 1473 1474 if (ib_event->event == IB_CM_REQ_RECEIVED) 1475 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr; 1476 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) 1477 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr; 1478 1479 if (!sgid_attr) 1480 return NULL; 1481 1482 rcu_read_lock(); 1483 ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr); 1484 if (IS_ERR(ndev)) 1485 ndev = NULL; 1486 else 1487 dev_hold(ndev); 1488 rcu_read_unlock(); 1489 return ndev; 1490 } 1491 1492 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event, 1493 struct cma_req_info *req) 1494 { 1495 struct sockaddr *listen_addr = 1496 (struct sockaddr *)&req->listen_addr_storage; 1497 struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage; 1498 struct net_device *net_dev; 1499 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL; 1500 int err; 1501 1502 err = cma_save_ip_info(listen_addr, src_addr, ib_event, 1503 req->service_id); 1504 if (err) 1505 return ERR_PTR(err); 1506 1507 if (rdma_protocol_roce(req->device, req->port)) 1508 net_dev = roce_get_net_dev_by_cm_event(ib_event); 1509 else 1510 net_dev = ib_get_net_dev_by_params(req->device, req->port, 1511 req->pkey, 1512 gid, listen_addr); 1513 if (!net_dev) 1514 return ERR_PTR(-ENODEV); 1515 1516 return net_dev; 1517 } 1518 1519 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id) 1520 { 1521 return (be64_to_cpu(service_id) >> 16) & 0xffff; 1522 } 1523 1524 static bool cma_match_private_data(struct rdma_id_private *id_priv, 1525 const struct cma_hdr *hdr) 1526 { 1527 struct sockaddr *addr = cma_src_addr(id_priv); 1528 __be32 ip4_addr; 1529 struct in6_addr ip6_addr; 1530 1531 if (cma_any_addr(addr) && !id_priv->afonly) 1532 return true; 1533 1534 switch (addr->sa_family) { 1535 case AF_INET: 1536 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr; 1537 if (cma_get_ip_ver(hdr) != 4) 1538 return false; 1539 if (!cma_any_addr(addr) && 1540 hdr->dst_addr.ip4.addr != ip4_addr) 1541 return false; 1542 break; 1543 case AF_INET6: 1544 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr; 1545 if (cma_get_ip_ver(hdr) != 6) 1546 return false; 1547 if (!cma_any_addr(addr) && 1548 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr))) 1549 return false; 1550 break; 1551 case AF_IB: 1552 return true; 1553 default: 1554 return false; 1555 } 1556 1557 return true; 1558 } 1559 1560 static bool cma_protocol_roce(const struct rdma_cm_id *id) 1561 { 1562 struct ib_device *device = id->device; 1563 const int port_num = id->port_num ?: rdma_start_port(device); 1564 1565 return rdma_protocol_roce(device, port_num); 1566 } 1567 1568 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req) 1569 { 1570 const struct sockaddr *daddr = 1571 (const struct sockaddr *)&req->listen_addr_storage; 1572 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr; 1573 1574 /* Returns true if the req is for IPv6 link local */ 1575 return (daddr->sa_family == AF_INET6 && 1576 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)); 1577 } 1578 1579 static bool cma_match_net_dev(const struct rdma_cm_id *id, 1580 const struct net_device *net_dev, 1581 const struct cma_req_info *req) 1582 { 1583 const struct rdma_addr *addr = &id->route.addr; 1584 1585 if (!net_dev) 1586 /* This request is an AF_IB request */ 1587 return (!id->port_num || id->port_num == req->port) && 1588 (addr->src_addr.ss_family == AF_IB); 1589 1590 /* 1591 * If the request is not for IPv6 link local, allow matching 1592 * request to any netdevice of the one or multiport rdma device. 1593 */ 1594 if (!cma_is_req_ipv6_ll(req)) 1595 return true; 1596 /* 1597 * Net namespaces must match, and if the listner is listening 1598 * on a specific netdevice than netdevice must match as well. 1599 */ 1600 if (net_eq(dev_net(net_dev), addr->dev_addr.net) && 1601 (!!addr->dev_addr.bound_dev_if == 1602 (addr->dev_addr.bound_dev_if == net_dev->ifindex))) 1603 return true; 1604 else 1605 return false; 1606 } 1607 1608 static struct rdma_id_private *cma_find_listener( 1609 const struct rdma_bind_list *bind_list, 1610 const struct ib_cm_id *cm_id, 1611 const struct ib_cm_event *ib_event, 1612 const struct cma_req_info *req, 1613 const struct net_device *net_dev) 1614 { 1615 struct rdma_id_private *id_priv, *id_priv_dev; 1616 1617 if (!bind_list) 1618 return ERR_PTR(-EINVAL); 1619 1620 hlist_for_each_entry(id_priv, &bind_list->owners, node) { 1621 if (cma_match_private_data(id_priv, ib_event->private_data)) { 1622 if (id_priv->id.device == cm_id->device && 1623 cma_match_net_dev(&id_priv->id, net_dev, req)) 1624 return id_priv; 1625 list_for_each_entry(id_priv_dev, 1626 &id_priv->listen_list, 1627 listen_list) { 1628 if (id_priv_dev->id.device == cm_id->device && 1629 cma_match_net_dev(&id_priv_dev->id, 1630 net_dev, req)) 1631 return id_priv_dev; 1632 } 1633 } 1634 } 1635 1636 return ERR_PTR(-EINVAL); 1637 } 1638 1639 static struct rdma_id_private * 1640 cma_ib_id_from_event(struct ib_cm_id *cm_id, 1641 const struct ib_cm_event *ib_event, 1642 struct cma_req_info *req, 1643 struct net_device **net_dev) 1644 { 1645 struct rdma_bind_list *bind_list; 1646 struct rdma_id_private *id_priv; 1647 int err; 1648 1649 err = cma_save_req_info(ib_event, req); 1650 if (err) 1651 return ERR_PTR(err); 1652 1653 *net_dev = cma_get_net_dev(ib_event, req); 1654 if (IS_ERR(*net_dev)) { 1655 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) { 1656 /* Assuming the protocol is AF_IB */ 1657 *net_dev = NULL; 1658 } else { 1659 return ERR_CAST(*net_dev); 1660 } 1661 } 1662 1663 /* 1664 * Net namespace might be getting deleted while route lookup, 1665 * cm_id lookup is in progress. Therefore, perform netdevice 1666 * validation, cm_id lookup under rcu lock. 1667 * RCU lock along with netdevice state check, synchronizes with 1668 * netdevice migrating to different net namespace and also avoids 1669 * case where net namespace doesn't get deleted while lookup is in 1670 * progress. 1671 * If the device state is not IFF_UP, its properties such as ifindex 1672 * and nd_net cannot be trusted to remain valid without rcu lock. 1673 * net/core/dev.c change_net_namespace() ensures to synchronize with 1674 * ongoing operations on net device after device is closed using 1675 * synchronize_net(). 1676 */ 1677 rcu_read_lock(); 1678 if (*net_dev) { 1679 /* 1680 * If netdevice is down, it is likely that it is administratively 1681 * down or it might be migrating to different namespace. 1682 * In that case avoid further processing, as the net namespace 1683 * or ifindex may change. 1684 */ 1685 if (((*net_dev)->flags & IFF_UP) == 0) { 1686 id_priv = ERR_PTR(-EHOSTUNREACH); 1687 goto err; 1688 } 1689 1690 if (!validate_net_dev(*net_dev, 1691 (struct sockaddr *)&req->listen_addr_storage, 1692 (struct sockaddr *)&req->src_addr_storage)) { 1693 id_priv = ERR_PTR(-EHOSTUNREACH); 1694 goto err; 1695 } 1696 } 1697 1698 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net, 1699 rdma_ps_from_service_id(req->service_id), 1700 cma_port_from_service_id(req->service_id)); 1701 id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev); 1702 err: 1703 rcu_read_unlock(); 1704 if (IS_ERR(id_priv) && *net_dev) { 1705 dev_put(*net_dev); 1706 *net_dev = NULL; 1707 } 1708 return id_priv; 1709 } 1710 1711 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv) 1712 { 1713 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr); 1714 } 1715 1716 static void cma_cancel_route(struct rdma_id_private *id_priv) 1717 { 1718 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) { 1719 if (id_priv->query) 1720 ib_sa_cancel_query(id_priv->query_id, id_priv->query); 1721 } 1722 } 1723 1724 static void cma_cancel_listens(struct rdma_id_private *id_priv) 1725 { 1726 struct rdma_id_private *dev_id_priv; 1727 1728 /* 1729 * Remove from listen_any_list to prevent added devices from spawning 1730 * additional listen requests. 1731 */ 1732 mutex_lock(&lock); 1733 list_del(&id_priv->list); 1734 1735 while (!list_empty(&id_priv->listen_list)) { 1736 dev_id_priv = list_entry(id_priv->listen_list.next, 1737 struct rdma_id_private, listen_list); 1738 /* sync with device removal to avoid duplicate destruction */ 1739 list_del_init(&dev_id_priv->list); 1740 list_del(&dev_id_priv->listen_list); 1741 mutex_unlock(&lock); 1742 1743 rdma_destroy_id(&dev_id_priv->id); 1744 mutex_lock(&lock); 1745 } 1746 mutex_unlock(&lock); 1747 } 1748 1749 static void cma_cancel_operation(struct rdma_id_private *id_priv, 1750 enum rdma_cm_state state) 1751 { 1752 switch (state) { 1753 case RDMA_CM_ADDR_QUERY: 1754 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr); 1755 break; 1756 case RDMA_CM_ROUTE_QUERY: 1757 cma_cancel_route(id_priv); 1758 break; 1759 case RDMA_CM_LISTEN: 1760 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev) 1761 cma_cancel_listens(id_priv); 1762 break; 1763 default: 1764 break; 1765 } 1766 } 1767 1768 static void cma_release_port(struct rdma_id_private *id_priv) 1769 { 1770 struct rdma_bind_list *bind_list = id_priv->bind_list; 1771 struct net *net = id_priv->id.route.addr.dev_addr.net; 1772 1773 if (!bind_list) 1774 return; 1775 1776 mutex_lock(&lock); 1777 hlist_del(&id_priv->node); 1778 if (hlist_empty(&bind_list->owners)) { 1779 cma_ps_remove(net, bind_list->ps, bind_list->port); 1780 kfree(bind_list); 1781 } 1782 mutex_unlock(&lock); 1783 } 1784 1785 static void cma_leave_roce_mc_group(struct rdma_id_private *id_priv, 1786 struct cma_multicast *mc) 1787 { 1788 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 1789 struct net_device *ndev = NULL; 1790 1791 if (dev_addr->bound_dev_if) 1792 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if); 1793 if (ndev) { 1794 cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid, false); 1795 dev_put(ndev); 1796 } 1797 kref_put(&mc->mcref, release_mc); 1798 } 1799 1800 static void cma_leave_mc_groups(struct rdma_id_private *id_priv) 1801 { 1802 struct cma_multicast *mc; 1803 1804 while (!list_empty(&id_priv->mc_list)) { 1805 mc = container_of(id_priv->mc_list.next, 1806 struct cma_multicast, list); 1807 list_del(&mc->list); 1808 if (rdma_cap_ib_mcast(id_priv->cma_dev->device, 1809 id_priv->id.port_num)) { 1810 ib_sa_free_multicast(mc->multicast.ib); 1811 kfree(mc); 1812 } else { 1813 cma_leave_roce_mc_group(id_priv, mc); 1814 } 1815 } 1816 } 1817 1818 void rdma_destroy_id(struct rdma_cm_id *id) 1819 { 1820 struct rdma_id_private *id_priv; 1821 enum rdma_cm_state state; 1822 1823 id_priv = container_of(id, struct rdma_id_private, id); 1824 trace_cm_id_destroy(id_priv); 1825 state = cma_exch(id_priv, RDMA_CM_DESTROYING); 1826 cma_cancel_operation(id_priv, state); 1827 1828 /* 1829 * Wait for any active callback to finish. New callbacks will find 1830 * the id_priv state set to destroying and abort. 1831 */ 1832 mutex_lock(&id_priv->handler_mutex); 1833 mutex_unlock(&id_priv->handler_mutex); 1834 1835 rdma_restrack_del(&id_priv->res); 1836 if (id_priv->cma_dev) { 1837 if (rdma_cap_ib_cm(id_priv->id.device, 1)) { 1838 if (id_priv->cm_id.ib) 1839 ib_destroy_cm_id(id_priv->cm_id.ib); 1840 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) { 1841 if (id_priv->cm_id.iw) 1842 iw_destroy_cm_id(id_priv->cm_id.iw); 1843 } 1844 cma_leave_mc_groups(id_priv); 1845 cma_release_dev(id_priv); 1846 } 1847 1848 cma_release_port(id_priv); 1849 cma_deref_id(id_priv); 1850 wait_for_completion(&id_priv->comp); 1851 1852 if (id_priv->internal_id) 1853 cma_deref_id(id_priv->id.context); 1854 1855 kfree(id_priv->id.route.path_rec); 1856 1857 if (id_priv->id.route.addr.dev_addr.sgid_attr) 1858 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr); 1859 1860 put_net(id_priv->id.route.addr.dev_addr.net); 1861 kfree(id_priv); 1862 } 1863 EXPORT_SYMBOL(rdma_destroy_id); 1864 1865 static int cma_rep_recv(struct rdma_id_private *id_priv) 1866 { 1867 int ret; 1868 1869 ret = cma_modify_qp_rtr(id_priv, NULL); 1870 if (ret) 1871 goto reject; 1872 1873 ret = cma_modify_qp_rts(id_priv, NULL); 1874 if (ret) 1875 goto reject; 1876 1877 trace_cm_send_rtu(id_priv); 1878 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0); 1879 if (ret) 1880 goto reject; 1881 1882 return 0; 1883 reject: 1884 pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret); 1885 cma_modify_qp_err(id_priv); 1886 trace_cm_send_rej(id_priv); 1887 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED, 1888 NULL, 0, NULL, 0); 1889 return ret; 1890 } 1891 1892 static void cma_set_rep_event_data(struct rdma_cm_event *event, 1893 const struct ib_cm_rep_event_param *rep_data, 1894 void *private_data) 1895 { 1896 event->param.conn.private_data = private_data; 1897 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE; 1898 event->param.conn.responder_resources = rep_data->responder_resources; 1899 event->param.conn.initiator_depth = rep_data->initiator_depth; 1900 event->param.conn.flow_control = rep_data->flow_control; 1901 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count; 1902 event->param.conn.srq = rep_data->srq; 1903 event->param.conn.qp_num = rep_data->remote_qpn; 1904 } 1905 1906 static int cma_cm_event_handler(struct rdma_id_private *id_priv, 1907 struct rdma_cm_event *event) 1908 { 1909 int ret; 1910 1911 trace_cm_event_handler(id_priv, event); 1912 ret = id_priv->id.event_handler(&id_priv->id, event); 1913 trace_cm_event_done(id_priv, event, ret); 1914 return ret; 1915 } 1916 1917 static int cma_ib_handler(struct ib_cm_id *cm_id, 1918 const struct ib_cm_event *ib_event) 1919 { 1920 struct rdma_id_private *id_priv = cm_id->context; 1921 struct rdma_cm_event event = {}; 1922 int ret = 0; 1923 1924 mutex_lock(&id_priv->handler_mutex); 1925 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT && 1926 id_priv->state != RDMA_CM_CONNECT) || 1927 (ib_event->event == IB_CM_TIMEWAIT_EXIT && 1928 id_priv->state != RDMA_CM_DISCONNECT)) 1929 goto out; 1930 1931 switch (ib_event->event) { 1932 case IB_CM_REQ_ERROR: 1933 case IB_CM_REP_ERROR: 1934 event.event = RDMA_CM_EVENT_UNREACHABLE; 1935 event.status = -ETIMEDOUT; 1936 break; 1937 case IB_CM_REP_RECEIVED: 1938 if (cma_comp(id_priv, RDMA_CM_CONNECT) && 1939 (id_priv->id.qp_type != IB_QPT_UD)) { 1940 trace_cm_send_mra(id_priv); 1941 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0); 1942 } 1943 if (id_priv->id.qp) { 1944 event.status = cma_rep_recv(id_priv); 1945 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR : 1946 RDMA_CM_EVENT_ESTABLISHED; 1947 } else { 1948 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE; 1949 } 1950 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd, 1951 ib_event->private_data); 1952 break; 1953 case IB_CM_RTU_RECEIVED: 1954 case IB_CM_USER_ESTABLISHED: 1955 event.event = RDMA_CM_EVENT_ESTABLISHED; 1956 break; 1957 case IB_CM_DREQ_ERROR: 1958 event.status = -ETIMEDOUT; /* fall through */ 1959 case IB_CM_DREQ_RECEIVED: 1960 case IB_CM_DREP_RECEIVED: 1961 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT, 1962 RDMA_CM_DISCONNECT)) 1963 goto out; 1964 event.event = RDMA_CM_EVENT_DISCONNECTED; 1965 break; 1966 case IB_CM_TIMEWAIT_EXIT: 1967 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT; 1968 break; 1969 case IB_CM_MRA_RECEIVED: 1970 /* ignore event */ 1971 goto out; 1972 case IB_CM_REJ_RECEIVED: 1973 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id, 1974 ib_event->param.rej_rcvd.reason)); 1975 cma_modify_qp_err(id_priv); 1976 event.status = ib_event->param.rej_rcvd.reason; 1977 event.event = RDMA_CM_EVENT_REJECTED; 1978 event.param.conn.private_data = ib_event->private_data; 1979 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE; 1980 break; 1981 default: 1982 pr_err("RDMA CMA: unexpected IB CM event: %d\n", 1983 ib_event->event); 1984 goto out; 1985 } 1986 1987 ret = cma_cm_event_handler(id_priv, &event); 1988 if (ret) { 1989 /* Destroy the CM ID by returning a non-zero value. */ 1990 id_priv->cm_id.ib = NULL; 1991 cma_exch(id_priv, RDMA_CM_DESTROYING); 1992 mutex_unlock(&id_priv->handler_mutex); 1993 rdma_destroy_id(&id_priv->id); 1994 return ret; 1995 } 1996 out: 1997 mutex_unlock(&id_priv->handler_mutex); 1998 return ret; 1999 } 2000 2001 static struct rdma_id_private * 2002 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id, 2003 const struct ib_cm_event *ib_event, 2004 struct net_device *net_dev) 2005 { 2006 struct rdma_id_private *listen_id_priv; 2007 struct rdma_id_private *id_priv; 2008 struct rdma_cm_id *id; 2009 struct rdma_route *rt; 2010 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family; 2011 struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path; 2012 const __be64 service_id = 2013 ib_event->param.req_rcvd.primary_path->service_id; 2014 int ret; 2015 2016 listen_id_priv = container_of(listen_id, struct rdma_id_private, id); 2017 id = __rdma_create_id(listen_id->route.addr.dev_addr.net, 2018 listen_id->event_handler, listen_id->context, 2019 listen_id->ps, ib_event->param.req_rcvd.qp_type, 2020 listen_id_priv->res.kern_name); 2021 if (IS_ERR(id)) 2022 return NULL; 2023 2024 id_priv = container_of(id, struct rdma_id_private, id); 2025 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr, 2026 (struct sockaddr *)&id->route.addr.dst_addr, 2027 listen_id, ib_event, ss_family, service_id)) 2028 goto err; 2029 2030 rt = &id->route; 2031 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1; 2032 rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec), 2033 GFP_KERNEL); 2034 if (!rt->path_rec) 2035 goto err; 2036 2037 rt->path_rec[0] = *path; 2038 if (rt->num_paths == 2) 2039 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path; 2040 2041 if (net_dev) { 2042 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev); 2043 } else { 2044 if (!cma_protocol_roce(listen_id) && 2045 cma_any_addr(cma_src_addr(id_priv))) { 2046 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND; 2047 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid); 2048 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey)); 2049 } else if (!cma_any_addr(cma_src_addr(id_priv))) { 2050 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr); 2051 if (ret) 2052 goto err; 2053 } 2054 } 2055 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid); 2056 2057 id_priv->state = RDMA_CM_CONNECT; 2058 return id_priv; 2059 2060 err: 2061 rdma_destroy_id(id); 2062 return NULL; 2063 } 2064 2065 static struct rdma_id_private * 2066 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id, 2067 const struct ib_cm_event *ib_event, 2068 struct net_device *net_dev) 2069 { 2070 const struct rdma_id_private *listen_id_priv; 2071 struct rdma_id_private *id_priv; 2072 struct rdma_cm_id *id; 2073 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family; 2074 struct net *net = listen_id->route.addr.dev_addr.net; 2075 int ret; 2076 2077 listen_id_priv = container_of(listen_id, struct rdma_id_private, id); 2078 id = __rdma_create_id(net, listen_id->event_handler, listen_id->context, 2079 listen_id->ps, IB_QPT_UD, 2080 listen_id_priv->res.kern_name); 2081 if (IS_ERR(id)) 2082 return NULL; 2083 2084 id_priv = container_of(id, struct rdma_id_private, id); 2085 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr, 2086 (struct sockaddr *)&id->route.addr.dst_addr, 2087 listen_id, ib_event, ss_family, 2088 ib_event->param.sidr_req_rcvd.service_id)) 2089 goto err; 2090 2091 if (net_dev) { 2092 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev); 2093 } else { 2094 if (!cma_any_addr(cma_src_addr(id_priv))) { 2095 ret = cma_translate_addr(cma_src_addr(id_priv), 2096 &id->route.addr.dev_addr); 2097 if (ret) 2098 goto err; 2099 } 2100 } 2101 2102 id_priv->state = RDMA_CM_CONNECT; 2103 return id_priv; 2104 err: 2105 rdma_destroy_id(id); 2106 return NULL; 2107 } 2108 2109 static void cma_set_req_event_data(struct rdma_cm_event *event, 2110 const struct ib_cm_req_event_param *req_data, 2111 void *private_data, int offset) 2112 { 2113 event->param.conn.private_data = private_data + offset; 2114 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset; 2115 event->param.conn.responder_resources = req_data->responder_resources; 2116 event->param.conn.initiator_depth = req_data->initiator_depth; 2117 event->param.conn.flow_control = req_data->flow_control; 2118 event->param.conn.retry_count = req_data->retry_count; 2119 event->param.conn.rnr_retry_count = req_data->rnr_retry_count; 2120 event->param.conn.srq = req_data->srq; 2121 event->param.conn.qp_num = req_data->remote_qpn; 2122 } 2123 2124 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id, 2125 const struct ib_cm_event *ib_event) 2126 { 2127 return (((ib_event->event == IB_CM_REQ_RECEIVED) && 2128 (ib_event->param.req_rcvd.qp_type == id->qp_type)) || 2129 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) && 2130 (id->qp_type == IB_QPT_UD)) || 2131 (!id->qp_type)); 2132 } 2133 2134 static int cma_ib_req_handler(struct ib_cm_id *cm_id, 2135 const struct ib_cm_event *ib_event) 2136 { 2137 struct rdma_id_private *listen_id, *conn_id = NULL; 2138 struct rdma_cm_event event = {}; 2139 struct cma_req_info req = {}; 2140 struct net_device *net_dev; 2141 u8 offset; 2142 int ret; 2143 2144 listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev); 2145 if (IS_ERR(listen_id)) 2146 return PTR_ERR(listen_id); 2147 2148 trace_cm_req_handler(listen_id, ib_event->event); 2149 if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) { 2150 ret = -EINVAL; 2151 goto net_dev_put; 2152 } 2153 2154 mutex_lock(&listen_id->handler_mutex); 2155 if (listen_id->state != RDMA_CM_LISTEN) { 2156 ret = -ECONNABORTED; 2157 goto err1; 2158 } 2159 2160 offset = cma_user_data_offset(listen_id); 2161 event.event = RDMA_CM_EVENT_CONNECT_REQUEST; 2162 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) { 2163 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev); 2164 event.param.ud.private_data = ib_event->private_data + offset; 2165 event.param.ud.private_data_len = 2166 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset; 2167 } else { 2168 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev); 2169 cma_set_req_event_data(&event, &ib_event->param.req_rcvd, 2170 ib_event->private_data, offset); 2171 } 2172 if (!conn_id) { 2173 ret = -ENOMEM; 2174 goto err1; 2175 } 2176 2177 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING); 2178 ret = cma_ib_acquire_dev(conn_id, listen_id, &req); 2179 if (ret) 2180 goto err2; 2181 2182 conn_id->cm_id.ib = cm_id; 2183 cm_id->context = conn_id; 2184 cm_id->cm_handler = cma_ib_handler; 2185 2186 /* 2187 * Protect against the user destroying conn_id from another thread 2188 * until we're done accessing it. 2189 */ 2190 atomic_inc(&conn_id->refcount); 2191 ret = cma_cm_event_handler(conn_id, &event); 2192 if (ret) 2193 goto err3; 2194 /* 2195 * Acquire mutex to prevent user executing rdma_destroy_id() 2196 * while we're accessing the cm_id. 2197 */ 2198 mutex_lock(&lock); 2199 if (cma_comp(conn_id, RDMA_CM_CONNECT) && 2200 (conn_id->id.qp_type != IB_QPT_UD)) { 2201 trace_cm_send_mra(cm_id->context); 2202 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0); 2203 } 2204 mutex_unlock(&lock); 2205 mutex_unlock(&conn_id->handler_mutex); 2206 mutex_unlock(&listen_id->handler_mutex); 2207 cma_deref_id(conn_id); 2208 if (net_dev) 2209 dev_put(net_dev); 2210 return 0; 2211 2212 err3: 2213 cma_deref_id(conn_id); 2214 /* Destroy the CM ID by returning a non-zero value. */ 2215 conn_id->cm_id.ib = NULL; 2216 err2: 2217 cma_exch(conn_id, RDMA_CM_DESTROYING); 2218 mutex_unlock(&conn_id->handler_mutex); 2219 err1: 2220 mutex_unlock(&listen_id->handler_mutex); 2221 if (conn_id) 2222 rdma_destroy_id(&conn_id->id); 2223 2224 net_dev_put: 2225 if (net_dev) 2226 dev_put(net_dev); 2227 2228 return ret; 2229 } 2230 2231 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr) 2232 { 2233 if (addr->sa_family == AF_IB) 2234 return ((struct sockaddr_ib *) addr)->sib_sid; 2235 2236 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr))); 2237 } 2238 EXPORT_SYMBOL(rdma_get_service_id); 2239 2240 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid, 2241 union ib_gid *dgid) 2242 { 2243 struct rdma_addr *addr = &cm_id->route.addr; 2244 2245 if (!cm_id->device) { 2246 if (sgid) 2247 memset(sgid, 0, sizeof(*sgid)); 2248 if (dgid) 2249 memset(dgid, 0, sizeof(*dgid)); 2250 return; 2251 } 2252 2253 if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) { 2254 if (sgid) 2255 rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid); 2256 if (dgid) 2257 rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid); 2258 } else { 2259 if (sgid) 2260 rdma_addr_get_sgid(&addr->dev_addr, sgid); 2261 if (dgid) 2262 rdma_addr_get_dgid(&addr->dev_addr, dgid); 2263 } 2264 } 2265 EXPORT_SYMBOL(rdma_read_gids); 2266 2267 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event) 2268 { 2269 struct rdma_id_private *id_priv = iw_id->context; 2270 struct rdma_cm_event event = {}; 2271 int ret = 0; 2272 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr; 2273 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr; 2274 2275 mutex_lock(&id_priv->handler_mutex); 2276 if (id_priv->state != RDMA_CM_CONNECT) 2277 goto out; 2278 2279 switch (iw_event->event) { 2280 case IW_CM_EVENT_CLOSE: 2281 event.event = RDMA_CM_EVENT_DISCONNECTED; 2282 break; 2283 case IW_CM_EVENT_CONNECT_REPLY: 2284 memcpy(cma_src_addr(id_priv), laddr, 2285 rdma_addr_size(laddr)); 2286 memcpy(cma_dst_addr(id_priv), raddr, 2287 rdma_addr_size(raddr)); 2288 switch (iw_event->status) { 2289 case 0: 2290 event.event = RDMA_CM_EVENT_ESTABLISHED; 2291 event.param.conn.initiator_depth = iw_event->ird; 2292 event.param.conn.responder_resources = iw_event->ord; 2293 break; 2294 case -ECONNRESET: 2295 case -ECONNREFUSED: 2296 event.event = RDMA_CM_EVENT_REJECTED; 2297 break; 2298 case -ETIMEDOUT: 2299 event.event = RDMA_CM_EVENT_UNREACHABLE; 2300 break; 2301 default: 2302 event.event = RDMA_CM_EVENT_CONNECT_ERROR; 2303 break; 2304 } 2305 break; 2306 case IW_CM_EVENT_ESTABLISHED: 2307 event.event = RDMA_CM_EVENT_ESTABLISHED; 2308 event.param.conn.initiator_depth = iw_event->ird; 2309 event.param.conn.responder_resources = iw_event->ord; 2310 break; 2311 default: 2312 goto out; 2313 } 2314 2315 event.status = iw_event->status; 2316 event.param.conn.private_data = iw_event->private_data; 2317 event.param.conn.private_data_len = iw_event->private_data_len; 2318 ret = cma_cm_event_handler(id_priv, &event); 2319 if (ret) { 2320 /* Destroy the CM ID by returning a non-zero value. */ 2321 id_priv->cm_id.iw = NULL; 2322 cma_exch(id_priv, RDMA_CM_DESTROYING); 2323 mutex_unlock(&id_priv->handler_mutex); 2324 rdma_destroy_id(&id_priv->id); 2325 return ret; 2326 } 2327 2328 out: 2329 mutex_unlock(&id_priv->handler_mutex); 2330 return ret; 2331 } 2332 2333 static int iw_conn_req_handler(struct iw_cm_id *cm_id, 2334 struct iw_cm_event *iw_event) 2335 { 2336 struct rdma_cm_id *new_cm_id; 2337 struct rdma_id_private *listen_id, *conn_id; 2338 struct rdma_cm_event event = {}; 2339 int ret = -ECONNABORTED; 2340 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr; 2341 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr; 2342 2343 event.event = RDMA_CM_EVENT_CONNECT_REQUEST; 2344 event.param.conn.private_data = iw_event->private_data; 2345 event.param.conn.private_data_len = iw_event->private_data_len; 2346 event.param.conn.initiator_depth = iw_event->ird; 2347 event.param.conn.responder_resources = iw_event->ord; 2348 2349 listen_id = cm_id->context; 2350 2351 mutex_lock(&listen_id->handler_mutex); 2352 if (listen_id->state != RDMA_CM_LISTEN) 2353 goto out; 2354 2355 /* Create a new RDMA id for the new IW CM ID */ 2356 new_cm_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net, 2357 listen_id->id.event_handler, 2358 listen_id->id.context, 2359 RDMA_PS_TCP, IB_QPT_RC, 2360 listen_id->res.kern_name); 2361 if (IS_ERR(new_cm_id)) { 2362 ret = -ENOMEM; 2363 goto out; 2364 } 2365 conn_id = container_of(new_cm_id, struct rdma_id_private, id); 2366 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING); 2367 conn_id->state = RDMA_CM_CONNECT; 2368 2369 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr); 2370 if (ret) { 2371 mutex_unlock(&conn_id->handler_mutex); 2372 rdma_destroy_id(new_cm_id); 2373 goto out; 2374 } 2375 2376 ret = cma_iw_acquire_dev(conn_id, listen_id); 2377 if (ret) { 2378 mutex_unlock(&conn_id->handler_mutex); 2379 rdma_destroy_id(new_cm_id); 2380 goto out; 2381 } 2382 2383 conn_id->cm_id.iw = cm_id; 2384 cm_id->context = conn_id; 2385 cm_id->cm_handler = cma_iw_handler; 2386 2387 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr)); 2388 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr)); 2389 2390 /* 2391 * Protect against the user destroying conn_id from another thread 2392 * until we're done accessing it. 2393 */ 2394 atomic_inc(&conn_id->refcount); 2395 ret = cma_cm_event_handler(conn_id, &event); 2396 if (ret) { 2397 /* User wants to destroy the CM ID */ 2398 conn_id->cm_id.iw = NULL; 2399 cma_exch(conn_id, RDMA_CM_DESTROYING); 2400 mutex_unlock(&conn_id->handler_mutex); 2401 mutex_unlock(&listen_id->handler_mutex); 2402 cma_deref_id(conn_id); 2403 rdma_destroy_id(&conn_id->id); 2404 return ret; 2405 } 2406 2407 mutex_unlock(&conn_id->handler_mutex); 2408 cma_deref_id(conn_id); 2409 2410 out: 2411 mutex_unlock(&listen_id->handler_mutex); 2412 return ret; 2413 } 2414 2415 static int cma_ib_listen(struct rdma_id_private *id_priv) 2416 { 2417 struct sockaddr *addr; 2418 struct ib_cm_id *id; 2419 __be64 svc_id; 2420 2421 addr = cma_src_addr(id_priv); 2422 svc_id = rdma_get_service_id(&id_priv->id, addr); 2423 id = ib_cm_insert_listen(id_priv->id.device, 2424 cma_ib_req_handler, svc_id); 2425 if (IS_ERR(id)) 2426 return PTR_ERR(id); 2427 id_priv->cm_id.ib = id; 2428 2429 return 0; 2430 } 2431 2432 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog) 2433 { 2434 int ret; 2435 struct iw_cm_id *id; 2436 2437 id = iw_create_cm_id(id_priv->id.device, 2438 iw_conn_req_handler, 2439 id_priv); 2440 if (IS_ERR(id)) 2441 return PTR_ERR(id); 2442 2443 id->tos = id_priv->tos; 2444 id->tos_set = id_priv->tos_set; 2445 id_priv->cm_id.iw = id; 2446 2447 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv), 2448 rdma_addr_size(cma_src_addr(id_priv))); 2449 2450 ret = iw_cm_listen(id_priv->cm_id.iw, backlog); 2451 2452 if (ret) { 2453 iw_destroy_cm_id(id_priv->cm_id.iw); 2454 id_priv->cm_id.iw = NULL; 2455 } 2456 2457 return ret; 2458 } 2459 2460 static int cma_listen_handler(struct rdma_cm_id *id, 2461 struct rdma_cm_event *event) 2462 { 2463 struct rdma_id_private *id_priv = id->context; 2464 2465 id->context = id_priv->id.context; 2466 id->event_handler = id_priv->id.event_handler; 2467 trace_cm_event_handler(id_priv, event); 2468 return id_priv->id.event_handler(id, event); 2469 } 2470 2471 static void cma_listen_on_dev(struct rdma_id_private *id_priv, 2472 struct cma_device *cma_dev) 2473 { 2474 struct rdma_id_private *dev_id_priv; 2475 struct rdma_cm_id *id; 2476 struct net *net = id_priv->id.route.addr.dev_addr.net; 2477 int ret; 2478 2479 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1)) 2480 return; 2481 2482 id = __rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps, 2483 id_priv->id.qp_type, id_priv->res.kern_name); 2484 if (IS_ERR(id)) 2485 return; 2486 2487 dev_id_priv = container_of(id, struct rdma_id_private, id); 2488 2489 dev_id_priv->state = RDMA_CM_ADDR_BOUND; 2490 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv), 2491 rdma_addr_size(cma_src_addr(id_priv))); 2492 2493 _cma_attach_to_dev(dev_id_priv, cma_dev); 2494 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list); 2495 atomic_inc(&id_priv->refcount); 2496 dev_id_priv->internal_id = 1; 2497 dev_id_priv->afonly = id_priv->afonly; 2498 dev_id_priv->tos_set = id_priv->tos_set; 2499 dev_id_priv->tos = id_priv->tos; 2500 2501 ret = rdma_listen(id, id_priv->backlog); 2502 if (ret) 2503 dev_warn(&cma_dev->device->dev, 2504 "RDMA CMA: cma_listen_on_dev, error %d\n", ret); 2505 } 2506 2507 static void cma_listen_on_all(struct rdma_id_private *id_priv) 2508 { 2509 struct cma_device *cma_dev; 2510 2511 mutex_lock(&lock); 2512 list_add_tail(&id_priv->list, &listen_any_list); 2513 list_for_each_entry(cma_dev, &dev_list, list) 2514 cma_listen_on_dev(id_priv, cma_dev); 2515 mutex_unlock(&lock); 2516 } 2517 2518 void rdma_set_service_type(struct rdma_cm_id *id, int tos) 2519 { 2520 struct rdma_id_private *id_priv; 2521 2522 id_priv = container_of(id, struct rdma_id_private, id); 2523 id_priv->tos = (u8) tos; 2524 id_priv->tos_set = true; 2525 } 2526 EXPORT_SYMBOL(rdma_set_service_type); 2527 2528 /** 2529 * rdma_set_ack_timeout() - Set the ack timeout of QP associated 2530 * with a connection identifier. 2531 * @id: Communication identifier to associated with service type. 2532 * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec. 2533 * 2534 * This function should be called before rdma_connect() on active side, 2535 * and on passive side before rdma_accept(). It is applicable to primary 2536 * path only. The timeout will affect the local side of the QP, it is not 2537 * negotiated with remote side and zero disables the timer. In case it is 2538 * set before rdma_resolve_route, the value will also be used to determine 2539 * PacketLifeTime for RoCE. 2540 * 2541 * Return: 0 for success 2542 */ 2543 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout) 2544 { 2545 struct rdma_id_private *id_priv; 2546 2547 if (id->qp_type != IB_QPT_RC) 2548 return -EINVAL; 2549 2550 id_priv = container_of(id, struct rdma_id_private, id); 2551 id_priv->timeout = timeout; 2552 id_priv->timeout_set = true; 2553 2554 return 0; 2555 } 2556 EXPORT_SYMBOL(rdma_set_ack_timeout); 2557 2558 static void cma_query_handler(int status, struct sa_path_rec *path_rec, 2559 void *context) 2560 { 2561 struct cma_work *work = context; 2562 struct rdma_route *route; 2563 2564 route = &work->id->id.route; 2565 2566 if (!status) { 2567 route->num_paths = 1; 2568 *route->path_rec = *path_rec; 2569 } else { 2570 work->old_state = RDMA_CM_ROUTE_QUERY; 2571 work->new_state = RDMA_CM_ADDR_RESOLVED; 2572 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR; 2573 work->event.status = status; 2574 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n", 2575 status); 2576 } 2577 2578 queue_work(cma_wq, &work->work); 2579 } 2580 2581 static int cma_query_ib_route(struct rdma_id_private *id_priv, 2582 unsigned long timeout_ms, struct cma_work *work) 2583 { 2584 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 2585 struct sa_path_rec path_rec; 2586 ib_sa_comp_mask comp_mask; 2587 struct sockaddr_in6 *sin6; 2588 struct sockaddr_ib *sib; 2589 2590 memset(&path_rec, 0, sizeof path_rec); 2591 2592 if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num)) 2593 path_rec.rec_type = SA_PATH_REC_TYPE_OPA; 2594 else 2595 path_rec.rec_type = SA_PATH_REC_TYPE_IB; 2596 rdma_addr_get_sgid(dev_addr, &path_rec.sgid); 2597 rdma_addr_get_dgid(dev_addr, &path_rec.dgid); 2598 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr)); 2599 path_rec.numb_path = 1; 2600 path_rec.reversible = 1; 2601 path_rec.service_id = rdma_get_service_id(&id_priv->id, 2602 cma_dst_addr(id_priv)); 2603 2604 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID | 2605 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH | 2606 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID; 2607 2608 switch (cma_family(id_priv)) { 2609 case AF_INET: 2610 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos); 2611 comp_mask |= IB_SA_PATH_REC_QOS_CLASS; 2612 break; 2613 case AF_INET6: 2614 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv); 2615 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20); 2616 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS; 2617 break; 2618 case AF_IB: 2619 sib = (struct sockaddr_ib *) cma_src_addr(id_priv); 2620 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20); 2621 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS; 2622 break; 2623 } 2624 2625 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device, 2626 id_priv->id.port_num, &path_rec, 2627 comp_mask, timeout_ms, 2628 GFP_KERNEL, cma_query_handler, 2629 work, &id_priv->query); 2630 2631 return (id_priv->query_id < 0) ? id_priv->query_id : 0; 2632 } 2633 2634 static void cma_work_handler(struct work_struct *_work) 2635 { 2636 struct cma_work *work = container_of(_work, struct cma_work, work); 2637 struct rdma_id_private *id_priv = work->id; 2638 int destroy = 0; 2639 2640 mutex_lock(&id_priv->handler_mutex); 2641 if (!cma_comp_exch(id_priv, work->old_state, work->new_state)) 2642 goto out; 2643 2644 if (cma_cm_event_handler(id_priv, &work->event)) { 2645 cma_exch(id_priv, RDMA_CM_DESTROYING); 2646 destroy = 1; 2647 } 2648 out: 2649 mutex_unlock(&id_priv->handler_mutex); 2650 cma_deref_id(id_priv); 2651 if (destroy) 2652 rdma_destroy_id(&id_priv->id); 2653 kfree(work); 2654 } 2655 2656 static void cma_ndev_work_handler(struct work_struct *_work) 2657 { 2658 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work); 2659 struct rdma_id_private *id_priv = work->id; 2660 int destroy = 0; 2661 2662 mutex_lock(&id_priv->handler_mutex); 2663 if (id_priv->state == RDMA_CM_DESTROYING || 2664 id_priv->state == RDMA_CM_DEVICE_REMOVAL) 2665 goto out; 2666 2667 if (cma_cm_event_handler(id_priv, &work->event)) { 2668 cma_exch(id_priv, RDMA_CM_DESTROYING); 2669 destroy = 1; 2670 } 2671 2672 out: 2673 mutex_unlock(&id_priv->handler_mutex); 2674 cma_deref_id(id_priv); 2675 if (destroy) 2676 rdma_destroy_id(&id_priv->id); 2677 kfree(work); 2678 } 2679 2680 static void cma_init_resolve_route_work(struct cma_work *work, 2681 struct rdma_id_private *id_priv) 2682 { 2683 work->id = id_priv; 2684 INIT_WORK(&work->work, cma_work_handler); 2685 work->old_state = RDMA_CM_ROUTE_QUERY; 2686 work->new_state = RDMA_CM_ROUTE_RESOLVED; 2687 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 2688 } 2689 2690 static void cma_init_resolve_addr_work(struct cma_work *work, 2691 struct rdma_id_private *id_priv) 2692 { 2693 work->id = id_priv; 2694 INIT_WORK(&work->work, cma_work_handler); 2695 work->old_state = RDMA_CM_ADDR_QUERY; 2696 work->new_state = RDMA_CM_ADDR_RESOLVED; 2697 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 2698 } 2699 2700 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, 2701 unsigned long timeout_ms) 2702 { 2703 struct rdma_route *route = &id_priv->id.route; 2704 struct cma_work *work; 2705 int ret; 2706 2707 work = kzalloc(sizeof *work, GFP_KERNEL); 2708 if (!work) 2709 return -ENOMEM; 2710 2711 cma_init_resolve_route_work(work, id_priv); 2712 2713 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL); 2714 if (!route->path_rec) { 2715 ret = -ENOMEM; 2716 goto err1; 2717 } 2718 2719 ret = cma_query_ib_route(id_priv, timeout_ms, work); 2720 if (ret) 2721 goto err2; 2722 2723 return 0; 2724 err2: 2725 kfree(route->path_rec); 2726 route->path_rec = NULL; 2727 err1: 2728 kfree(work); 2729 return ret; 2730 } 2731 2732 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type, 2733 unsigned long supported_gids, 2734 enum ib_gid_type default_gid) 2735 { 2736 if ((network_type == RDMA_NETWORK_IPV4 || 2737 network_type == RDMA_NETWORK_IPV6) && 2738 test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids)) 2739 return IB_GID_TYPE_ROCE_UDP_ENCAP; 2740 2741 return default_gid; 2742 } 2743 2744 /* 2745 * cma_iboe_set_path_rec_l2_fields() is helper function which sets 2746 * path record type based on GID type. 2747 * It also sets up other L2 fields which includes destination mac address 2748 * netdev ifindex, of the path record. 2749 * It returns the netdev of the bound interface for this path record entry. 2750 */ 2751 static struct net_device * 2752 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv) 2753 { 2754 struct rdma_route *route = &id_priv->id.route; 2755 enum ib_gid_type gid_type = IB_GID_TYPE_ROCE; 2756 struct rdma_addr *addr = &route->addr; 2757 unsigned long supported_gids; 2758 struct net_device *ndev; 2759 2760 if (!addr->dev_addr.bound_dev_if) 2761 return NULL; 2762 2763 ndev = dev_get_by_index(addr->dev_addr.net, 2764 addr->dev_addr.bound_dev_if); 2765 if (!ndev) 2766 return NULL; 2767 2768 supported_gids = roce_gid_type_mask_support(id_priv->id.device, 2769 id_priv->id.port_num); 2770 gid_type = cma_route_gid_type(addr->dev_addr.network, 2771 supported_gids, 2772 id_priv->gid_type); 2773 /* Use the hint from IP Stack to select GID Type */ 2774 if (gid_type < ib_network_to_gid_type(addr->dev_addr.network)) 2775 gid_type = ib_network_to_gid_type(addr->dev_addr.network); 2776 route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type); 2777 2778 route->path_rec->roce.route_resolved = true; 2779 sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr); 2780 return ndev; 2781 } 2782 2783 int rdma_set_ib_path(struct rdma_cm_id *id, 2784 struct sa_path_rec *path_rec) 2785 { 2786 struct rdma_id_private *id_priv; 2787 struct net_device *ndev; 2788 int ret; 2789 2790 id_priv = container_of(id, struct rdma_id_private, id); 2791 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, 2792 RDMA_CM_ROUTE_RESOLVED)) 2793 return -EINVAL; 2794 2795 id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec), 2796 GFP_KERNEL); 2797 if (!id->route.path_rec) { 2798 ret = -ENOMEM; 2799 goto err; 2800 } 2801 2802 if (rdma_protocol_roce(id->device, id->port_num)) { 2803 ndev = cma_iboe_set_path_rec_l2_fields(id_priv); 2804 if (!ndev) { 2805 ret = -ENODEV; 2806 goto err_free; 2807 } 2808 dev_put(ndev); 2809 } 2810 2811 id->route.num_paths = 1; 2812 return 0; 2813 2814 err_free: 2815 kfree(id->route.path_rec); 2816 id->route.path_rec = NULL; 2817 err: 2818 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED); 2819 return ret; 2820 } 2821 EXPORT_SYMBOL(rdma_set_ib_path); 2822 2823 static int cma_resolve_iw_route(struct rdma_id_private *id_priv) 2824 { 2825 struct cma_work *work; 2826 2827 work = kzalloc(sizeof *work, GFP_KERNEL); 2828 if (!work) 2829 return -ENOMEM; 2830 2831 cma_init_resolve_route_work(work, id_priv); 2832 queue_work(cma_wq, &work->work); 2833 return 0; 2834 } 2835 2836 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio) 2837 { 2838 struct net_device *dev; 2839 2840 dev = vlan_dev_real_dev(vlan_ndev); 2841 if (dev->num_tc) 2842 return netdev_get_prio_tc_map(dev, prio); 2843 2844 return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) & 2845 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT; 2846 } 2847 2848 struct iboe_prio_tc_map { 2849 int input_prio; 2850 int output_tc; 2851 bool found; 2852 }; 2853 2854 static int get_lower_vlan_dev_tc(struct net_device *dev, void *data) 2855 { 2856 struct iboe_prio_tc_map *map = data; 2857 2858 if (is_vlan_dev(dev)) 2859 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio); 2860 else if (dev->num_tc) 2861 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio); 2862 else 2863 map->output_tc = 0; 2864 /* We are interested only in first level VLAN device, so always 2865 * return 1 to stop iterating over next level devices. 2866 */ 2867 map->found = true; 2868 return 1; 2869 } 2870 2871 static int iboe_tos_to_sl(struct net_device *ndev, int tos) 2872 { 2873 struct iboe_prio_tc_map prio_tc_map = {}; 2874 int prio = rt_tos2priority(tos); 2875 2876 /* If VLAN device, get it directly from the VLAN netdev */ 2877 if (is_vlan_dev(ndev)) 2878 return get_vlan_ndev_tc(ndev, prio); 2879 2880 prio_tc_map.input_prio = prio; 2881 rcu_read_lock(); 2882 netdev_walk_all_lower_dev_rcu(ndev, 2883 get_lower_vlan_dev_tc, 2884 &prio_tc_map); 2885 rcu_read_unlock(); 2886 /* If map is found from lower device, use it; Otherwise 2887 * continue with the current netdevice to get priority to tc map. 2888 */ 2889 if (prio_tc_map.found) 2890 return prio_tc_map.output_tc; 2891 else if (ndev->num_tc) 2892 return netdev_get_prio_tc_map(ndev, prio); 2893 else 2894 return 0; 2895 } 2896 2897 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv) 2898 { 2899 struct rdma_route *route = &id_priv->id.route; 2900 struct rdma_addr *addr = &route->addr; 2901 struct cma_work *work; 2902 int ret; 2903 struct net_device *ndev; 2904 2905 u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num - 2906 rdma_start_port(id_priv->cma_dev->device)]; 2907 u8 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos; 2908 2909 2910 work = kzalloc(sizeof *work, GFP_KERNEL); 2911 if (!work) 2912 return -ENOMEM; 2913 2914 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL); 2915 if (!route->path_rec) { 2916 ret = -ENOMEM; 2917 goto err1; 2918 } 2919 2920 route->num_paths = 1; 2921 2922 ndev = cma_iboe_set_path_rec_l2_fields(id_priv); 2923 if (!ndev) { 2924 ret = -ENODEV; 2925 goto err2; 2926 } 2927 2928 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr, 2929 &route->path_rec->sgid); 2930 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr, 2931 &route->path_rec->dgid); 2932 2933 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB) 2934 /* TODO: get the hoplimit from the inet/inet6 device */ 2935 route->path_rec->hop_limit = addr->dev_addr.hoplimit; 2936 else 2937 route->path_rec->hop_limit = 1; 2938 route->path_rec->reversible = 1; 2939 route->path_rec->pkey = cpu_to_be16(0xffff); 2940 route->path_rec->mtu_selector = IB_SA_EQ; 2941 route->path_rec->sl = iboe_tos_to_sl(ndev, tos); 2942 route->path_rec->traffic_class = tos; 2943 route->path_rec->mtu = iboe_get_mtu(ndev->mtu); 2944 route->path_rec->rate_selector = IB_SA_EQ; 2945 route->path_rec->rate = iboe_get_rate(ndev); 2946 dev_put(ndev); 2947 route->path_rec->packet_life_time_selector = IB_SA_EQ; 2948 /* In case ACK timeout is set, use this value to calculate 2949 * PacketLifeTime. As per IBTA 12.7.34, 2950 * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay). 2951 * Assuming a negligible local ACK delay, we can use 2952 * PacketLifeTime = local ACK timeout/2 2953 * as a reasonable approximation for RoCE networks. 2954 */ 2955 route->path_rec->packet_life_time = id_priv->timeout_set ? 2956 id_priv->timeout - 1 : CMA_IBOE_PACKET_LIFETIME; 2957 2958 if (!route->path_rec->mtu) { 2959 ret = -EINVAL; 2960 goto err2; 2961 } 2962 2963 cma_init_resolve_route_work(work, id_priv); 2964 queue_work(cma_wq, &work->work); 2965 2966 return 0; 2967 2968 err2: 2969 kfree(route->path_rec); 2970 route->path_rec = NULL; 2971 err1: 2972 kfree(work); 2973 return ret; 2974 } 2975 2976 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms) 2977 { 2978 struct rdma_id_private *id_priv; 2979 int ret; 2980 2981 id_priv = container_of(id, struct rdma_id_private, id); 2982 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY)) 2983 return -EINVAL; 2984 2985 atomic_inc(&id_priv->refcount); 2986 if (rdma_cap_ib_sa(id->device, id->port_num)) 2987 ret = cma_resolve_ib_route(id_priv, timeout_ms); 2988 else if (rdma_protocol_roce(id->device, id->port_num)) 2989 ret = cma_resolve_iboe_route(id_priv); 2990 else if (rdma_protocol_iwarp(id->device, id->port_num)) 2991 ret = cma_resolve_iw_route(id_priv); 2992 else 2993 ret = -ENOSYS; 2994 2995 if (ret) 2996 goto err; 2997 2998 return 0; 2999 err: 3000 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED); 3001 cma_deref_id(id_priv); 3002 return ret; 3003 } 3004 EXPORT_SYMBOL(rdma_resolve_route); 3005 3006 static void cma_set_loopback(struct sockaddr *addr) 3007 { 3008 switch (addr->sa_family) { 3009 case AF_INET: 3010 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK); 3011 break; 3012 case AF_INET6: 3013 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr, 3014 0, 0, 0, htonl(1)); 3015 break; 3016 default: 3017 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr, 3018 0, 0, 0, htonl(1)); 3019 break; 3020 } 3021 } 3022 3023 static int cma_bind_loopback(struct rdma_id_private *id_priv) 3024 { 3025 struct cma_device *cma_dev, *cur_dev; 3026 union ib_gid gid; 3027 enum ib_port_state port_state; 3028 u16 pkey; 3029 int ret; 3030 u8 p; 3031 3032 cma_dev = NULL; 3033 mutex_lock(&lock); 3034 list_for_each_entry(cur_dev, &dev_list, list) { 3035 if (cma_family(id_priv) == AF_IB && 3036 !rdma_cap_ib_cm(cur_dev->device, 1)) 3037 continue; 3038 3039 if (!cma_dev) 3040 cma_dev = cur_dev; 3041 3042 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) { 3043 if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) && 3044 port_state == IB_PORT_ACTIVE) { 3045 cma_dev = cur_dev; 3046 goto port_found; 3047 } 3048 } 3049 } 3050 3051 if (!cma_dev) { 3052 ret = -ENODEV; 3053 goto out; 3054 } 3055 3056 p = 1; 3057 3058 port_found: 3059 ret = rdma_query_gid(cma_dev->device, p, 0, &gid); 3060 if (ret) 3061 goto out; 3062 3063 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey); 3064 if (ret) 3065 goto out; 3066 3067 id_priv->id.route.addr.dev_addr.dev_type = 3068 (rdma_protocol_ib(cma_dev->device, p)) ? 3069 ARPHRD_INFINIBAND : ARPHRD_ETHER; 3070 3071 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid); 3072 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey); 3073 id_priv->id.port_num = p; 3074 cma_attach_to_dev(id_priv, cma_dev); 3075 cma_set_loopback(cma_src_addr(id_priv)); 3076 out: 3077 mutex_unlock(&lock); 3078 return ret; 3079 } 3080 3081 static void addr_handler(int status, struct sockaddr *src_addr, 3082 struct rdma_dev_addr *dev_addr, void *context) 3083 { 3084 struct rdma_id_private *id_priv = context; 3085 struct rdma_cm_event event = {}; 3086 struct sockaddr *addr; 3087 struct sockaddr_storage old_addr; 3088 3089 mutex_lock(&id_priv->handler_mutex); 3090 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, 3091 RDMA_CM_ADDR_RESOLVED)) 3092 goto out; 3093 3094 /* 3095 * Store the previous src address, so that if we fail to acquire 3096 * matching rdma device, old address can be restored back, which helps 3097 * to cancel the cma listen operation correctly. 3098 */ 3099 addr = cma_src_addr(id_priv); 3100 memcpy(&old_addr, addr, rdma_addr_size(addr)); 3101 memcpy(addr, src_addr, rdma_addr_size(src_addr)); 3102 if (!status && !id_priv->cma_dev) { 3103 status = cma_acquire_dev_by_src_ip(id_priv); 3104 if (status) 3105 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n", 3106 status); 3107 } else if (status) { 3108 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status); 3109 } 3110 3111 if (status) { 3112 memcpy(addr, &old_addr, 3113 rdma_addr_size((struct sockaddr *)&old_addr)); 3114 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, 3115 RDMA_CM_ADDR_BOUND)) 3116 goto out; 3117 event.event = RDMA_CM_EVENT_ADDR_ERROR; 3118 event.status = status; 3119 } else 3120 event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 3121 3122 if (cma_cm_event_handler(id_priv, &event)) { 3123 cma_exch(id_priv, RDMA_CM_DESTROYING); 3124 mutex_unlock(&id_priv->handler_mutex); 3125 rdma_destroy_id(&id_priv->id); 3126 return; 3127 } 3128 out: 3129 mutex_unlock(&id_priv->handler_mutex); 3130 } 3131 3132 static int cma_resolve_loopback(struct rdma_id_private *id_priv) 3133 { 3134 struct cma_work *work; 3135 union ib_gid gid; 3136 int ret; 3137 3138 work = kzalloc(sizeof *work, GFP_KERNEL); 3139 if (!work) 3140 return -ENOMEM; 3141 3142 if (!id_priv->cma_dev) { 3143 ret = cma_bind_loopback(id_priv); 3144 if (ret) 3145 goto err; 3146 } 3147 3148 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid); 3149 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid); 3150 3151 atomic_inc(&id_priv->refcount); 3152 cma_init_resolve_addr_work(work, id_priv); 3153 queue_work(cma_wq, &work->work); 3154 return 0; 3155 err: 3156 kfree(work); 3157 return ret; 3158 } 3159 3160 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv) 3161 { 3162 struct cma_work *work; 3163 int ret; 3164 3165 work = kzalloc(sizeof *work, GFP_KERNEL); 3166 if (!work) 3167 return -ENOMEM; 3168 3169 if (!id_priv->cma_dev) { 3170 ret = cma_resolve_ib_dev(id_priv); 3171 if (ret) 3172 goto err; 3173 } 3174 3175 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *) 3176 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr)); 3177 3178 atomic_inc(&id_priv->refcount); 3179 cma_init_resolve_addr_work(work, id_priv); 3180 queue_work(cma_wq, &work->work); 3181 return 0; 3182 err: 3183 kfree(work); 3184 return ret; 3185 } 3186 3187 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr, 3188 const struct sockaddr *dst_addr) 3189 { 3190 if (!src_addr || !src_addr->sa_family) { 3191 src_addr = (struct sockaddr *) &id->route.addr.src_addr; 3192 src_addr->sa_family = dst_addr->sa_family; 3193 if (IS_ENABLED(CONFIG_IPV6) && 3194 dst_addr->sa_family == AF_INET6) { 3195 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr; 3196 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr; 3197 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id; 3198 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL) 3199 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id; 3200 } else if (dst_addr->sa_family == AF_IB) { 3201 ((struct sockaddr_ib *) src_addr)->sib_pkey = 3202 ((struct sockaddr_ib *) dst_addr)->sib_pkey; 3203 } 3204 } 3205 return rdma_bind_addr(id, src_addr); 3206 } 3207 3208 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr, 3209 const struct sockaddr *dst_addr, unsigned long timeout_ms) 3210 { 3211 struct rdma_id_private *id_priv; 3212 int ret; 3213 3214 id_priv = container_of(id, struct rdma_id_private, id); 3215 if (id_priv->state == RDMA_CM_IDLE) { 3216 ret = cma_bind_addr(id, src_addr, dst_addr); 3217 if (ret) 3218 return ret; 3219 } 3220 3221 if (cma_family(id_priv) != dst_addr->sa_family) 3222 return -EINVAL; 3223 3224 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) 3225 return -EINVAL; 3226 3227 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr)); 3228 if (cma_any_addr(dst_addr)) { 3229 ret = cma_resolve_loopback(id_priv); 3230 } else { 3231 if (dst_addr->sa_family == AF_IB) { 3232 ret = cma_resolve_ib_addr(id_priv); 3233 } else { 3234 ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr, 3235 &id->route.addr.dev_addr, 3236 timeout_ms, addr_handler, 3237 false, id_priv); 3238 } 3239 } 3240 if (ret) 3241 goto err; 3242 3243 return 0; 3244 err: 3245 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND); 3246 return ret; 3247 } 3248 EXPORT_SYMBOL(rdma_resolve_addr); 3249 3250 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse) 3251 { 3252 struct rdma_id_private *id_priv; 3253 unsigned long flags; 3254 int ret; 3255 3256 id_priv = container_of(id, struct rdma_id_private, id); 3257 spin_lock_irqsave(&id_priv->lock, flags); 3258 if (reuse || id_priv->state == RDMA_CM_IDLE) { 3259 id_priv->reuseaddr = reuse; 3260 ret = 0; 3261 } else { 3262 ret = -EINVAL; 3263 } 3264 spin_unlock_irqrestore(&id_priv->lock, flags); 3265 return ret; 3266 } 3267 EXPORT_SYMBOL(rdma_set_reuseaddr); 3268 3269 int rdma_set_afonly(struct rdma_cm_id *id, int afonly) 3270 { 3271 struct rdma_id_private *id_priv; 3272 unsigned long flags; 3273 int ret; 3274 3275 id_priv = container_of(id, struct rdma_id_private, id); 3276 spin_lock_irqsave(&id_priv->lock, flags); 3277 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) { 3278 id_priv->options |= (1 << CMA_OPTION_AFONLY); 3279 id_priv->afonly = afonly; 3280 ret = 0; 3281 } else { 3282 ret = -EINVAL; 3283 } 3284 spin_unlock_irqrestore(&id_priv->lock, flags); 3285 return ret; 3286 } 3287 EXPORT_SYMBOL(rdma_set_afonly); 3288 3289 static void cma_bind_port(struct rdma_bind_list *bind_list, 3290 struct rdma_id_private *id_priv) 3291 { 3292 struct sockaddr *addr; 3293 struct sockaddr_ib *sib; 3294 u64 sid, mask; 3295 __be16 port; 3296 3297 addr = cma_src_addr(id_priv); 3298 port = htons(bind_list->port); 3299 3300 switch (addr->sa_family) { 3301 case AF_INET: 3302 ((struct sockaddr_in *) addr)->sin_port = port; 3303 break; 3304 case AF_INET6: 3305 ((struct sockaddr_in6 *) addr)->sin6_port = port; 3306 break; 3307 case AF_IB: 3308 sib = (struct sockaddr_ib *) addr; 3309 sid = be64_to_cpu(sib->sib_sid); 3310 mask = be64_to_cpu(sib->sib_sid_mask); 3311 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port)); 3312 sib->sib_sid_mask = cpu_to_be64(~0ULL); 3313 break; 3314 } 3315 id_priv->bind_list = bind_list; 3316 hlist_add_head(&id_priv->node, &bind_list->owners); 3317 } 3318 3319 static int cma_alloc_port(enum rdma_ucm_port_space ps, 3320 struct rdma_id_private *id_priv, unsigned short snum) 3321 { 3322 struct rdma_bind_list *bind_list; 3323 int ret; 3324 3325 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL); 3326 if (!bind_list) 3327 return -ENOMEM; 3328 3329 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list, 3330 snum); 3331 if (ret < 0) 3332 goto err; 3333 3334 bind_list->ps = ps; 3335 bind_list->port = snum; 3336 cma_bind_port(bind_list, id_priv); 3337 return 0; 3338 err: 3339 kfree(bind_list); 3340 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret; 3341 } 3342 3343 static int cma_port_is_unique(struct rdma_bind_list *bind_list, 3344 struct rdma_id_private *id_priv) 3345 { 3346 struct rdma_id_private *cur_id; 3347 struct sockaddr *daddr = cma_dst_addr(id_priv); 3348 struct sockaddr *saddr = cma_src_addr(id_priv); 3349 __be16 dport = cma_port(daddr); 3350 3351 hlist_for_each_entry(cur_id, &bind_list->owners, node) { 3352 struct sockaddr *cur_daddr = cma_dst_addr(cur_id); 3353 struct sockaddr *cur_saddr = cma_src_addr(cur_id); 3354 __be16 cur_dport = cma_port(cur_daddr); 3355 3356 if (id_priv == cur_id) 3357 continue; 3358 3359 /* different dest port -> unique */ 3360 if (!cma_any_port(daddr) && 3361 !cma_any_port(cur_daddr) && 3362 (dport != cur_dport)) 3363 continue; 3364 3365 /* different src address -> unique */ 3366 if (!cma_any_addr(saddr) && 3367 !cma_any_addr(cur_saddr) && 3368 cma_addr_cmp(saddr, cur_saddr)) 3369 continue; 3370 3371 /* different dst address -> unique */ 3372 if (!cma_any_addr(daddr) && 3373 !cma_any_addr(cur_daddr) && 3374 cma_addr_cmp(daddr, cur_daddr)) 3375 continue; 3376 3377 return -EADDRNOTAVAIL; 3378 } 3379 return 0; 3380 } 3381 3382 static int cma_alloc_any_port(enum rdma_ucm_port_space ps, 3383 struct rdma_id_private *id_priv) 3384 { 3385 static unsigned int last_used_port; 3386 int low, high, remaining; 3387 unsigned int rover; 3388 struct net *net = id_priv->id.route.addr.dev_addr.net; 3389 3390 inet_get_local_port_range(net, &low, &high); 3391 remaining = (high - low) + 1; 3392 rover = prandom_u32() % remaining + low; 3393 retry: 3394 if (last_used_port != rover) { 3395 struct rdma_bind_list *bind_list; 3396 int ret; 3397 3398 bind_list = cma_ps_find(net, ps, (unsigned short)rover); 3399 3400 if (!bind_list) { 3401 ret = cma_alloc_port(ps, id_priv, rover); 3402 } else { 3403 ret = cma_port_is_unique(bind_list, id_priv); 3404 if (!ret) 3405 cma_bind_port(bind_list, id_priv); 3406 } 3407 /* 3408 * Remember previously used port number in order to avoid 3409 * re-using same port immediately after it is closed. 3410 */ 3411 if (!ret) 3412 last_used_port = rover; 3413 if (ret != -EADDRNOTAVAIL) 3414 return ret; 3415 } 3416 if (--remaining) { 3417 rover++; 3418 if ((rover < low) || (rover > high)) 3419 rover = low; 3420 goto retry; 3421 } 3422 return -EADDRNOTAVAIL; 3423 } 3424 3425 /* 3426 * Check that the requested port is available. This is called when trying to 3427 * bind to a specific port, or when trying to listen on a bound port. In 3428 * the latter case, the provided id_priv may already be on the bind_list, but 3429 * we still need to check that it's okay to start listening. 3430 */ 3431 static int cma_check_port(struct rdma_bind_list *bind_list, 3432 struct rdma_id_private *id_priv, uint8_t reuseaddr) 3433 { 3434 struct rdma_id_private *cur_id; 3435 struct sockaddr *addr, *cur_addr; 3436 3437 addr = cma_src_addr(id_priv); 3438 hlist_for_each_entry(cur_id, &bind_list->owners, node) { 3439 if (id_priv == cur_id) 3440 continue; 3441 3442 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr && 3443 cur_id->reuseaddr) 3444 continue; 3445 3446 cur_addr = cma_src_addr(cur_id); 3447 if (id_priv->afonly && cur_id->afonly && 3448 (addr->sa_family != cur_addr->sa_family)) 3449 continue; 3450 3451 if (cma_any_addr(addr) || cma_any_addr(cur_addr)) 3452 return -EADDRNOTAVAIL; 3453 3454 if (!cma_addr_cmp(addr, cur_addr)) 3455 return -EADDRINUSE; 3456 } 3457 return 0; 3458 } 3459 3460 static int cma_use_port(enum rdma_ucm_port_space ps, 3461 struct rdma_id_private *id_priv) 3462 { 3463 struct rdma_bind_list *bind_list; 3464 unsigned short snum; 3465 int ret; 3466 3467 snum = ntohs(cma_port(cma_src_addr(id_priv))); 3468 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE)) 3469 return -EACCES; 3470 3471 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum); 3472 if (!bind_list) { 3473 ret = cma_alloc_port(ps, id_priv, snum); 3474 } else { 3475 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr); 3476 if (!ret) 3477 cma_bind_port(bind_list, id_priv); 3478 } 3479 return ret; 3480 } 3481 3482 static int cma_bind_listen(struct rdma_id_private *id_priv) 3483 { 3484 struct rdma_bind_list *bind_list = id_priv->bind_list; 3485 int ret = 0; 3486 3487 mutex_lock(&lock); 3488 if (bind_list->owners.first->next) 3489 ret = cma_check_port(bind_list, id_priv, 0); 3490 mutex_unlock(&lock); 3491 return ret; 3492 } 3493 3494 static enum rdma_ucm_port_space 3495 cma_select_inet_ps(struct rdma_id_private *id_priv) 3496 { 3497 switch (id_priv->id.ps) { 3498 case RDMA_PS_TCP: 3499 case RDMA_PS_UDP: 3500 case RDMA_PS_IPOIB: 3501 case RDMA_PS_IB: 3502 return id_priv->id.ps; 3503 default: 3504 3505 return 0; 3506 } 3507 } 3508 3509 static enum rdma_ucm_port_space 3510 cma_select_ib_ps(struct rdma_id_private *id_priv) 3511 { 3512 enum rdma_ucm_port_space ps = 0; 3513 struct sockaddr_ib *sib; 3514 u64 sid_ps, mask, sid; 3515 3516 sib = (struct sockaddr_ib *) cma_src_addr(id_priv); 3517 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK; 3518 sid = be64_to_cpu(sib->sib_sid) & mask; 3519 3520 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) { 3521 sid_ps = RDMA_IB_IP_PS_IB; 3522 ps = RDMA_PS_IB; 3523 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) && 3524 (sid == (RDMA_IB_IP_PS_TCP & mask))) { 3525 sid_ps = RDMA_IB_IP_PS_TCP; 3526 ps = RDMA_PS_TCP; 3527 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) && 3528 (sid == (RDMA_IB_IP_PS_UDP & mask))) { 3529 sid_ps = RDMA_IB_IP_PS_UDP; 3530 ps = RDMA_PS_UDP; 3531 } 3532 3533 if (ps) { 3534 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib))); 3535 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK | 3536 be64_to_cpu(sib->sib_sid_mask)); 3537 } 3538 return ps; 3539 } 3540 3541 static int cma_get_port(struct rdma_id_private *id_priv) 3542 { 3543 enum rdma_ucm_port_space ps; 3544 int ret; 3545 3546 if (cma_family(id_priv) != AF_IB) 3547 ps = cma_select_inet_ps(id_priv); 3548 else 3549 ps = cma_select_ib_ps(id_priv); 3550 if (!ps) 3551 return -EPROTONOSUPPORT; 3552 3553 mutex_lock(&lock); 3554 if (cma_any_port(cma_src_addr(id_priv))) 3555 ret = cma_alloc_any_port(ps, id_priv); 3556 else 3557 ret = cma_use_port(ps, id_priv); 3558 mutex_unlock(&lock); 3559 3560 return ret; 3561 } 3562 3563 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr, 3564 struct sockaddr *addr) 3565 { 3566 #if IS_ENABLED(CONFIG_IPV6) 3567 struct sockaddr_in6 *sin6; 3568 3569 if (addr->sa_family != AF_INET6) 3570 return 0; 3571 3572 sin6 = (struct sockaddr_in6 *) addr; 3573 3574 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL)) 3575 return 0; 3576 3577 if (!sin6->sin6_scope_id) 3578 return -EINVAL; 3579 3580 dev_addr->bound_dev_if = sin6->sin6_scope_id; 3581 #endif 3582 return 0; 3583 } 3584 3585 int rdma_listen(struct rdma_cm_id *id, int backlog) 3586 { 3587 struct rdma_id_private *id_priv; 3588 int ret; 3589 3590 id_priv = container_of(id, struct rdma_id_private, id); 3591 if (id_priv->state == RDMA_CM_IDLE) { 3592 id->route.addr.src_addr.ss_family = AF_INET; 3593 ret = rdma_bind_addr(id, cma_src_addr(id_priv)); 3594 if (ret) 3595 return ret; 3596 } 3597 3598 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) 3599 return -EINVAL; 3600 3601 if (id_priv->reuseaddr) { 3602 ret = cma_bind_listen(id_priv); 3603 if (ret) 3604 goto err; 3605 } 3606 3607 id_priv->backlog = backlog; 3608 if (id->device) { 3609 if (rdma_cap_ib_cm(id->device, 1)) { 3610 ret = cma_ib_listen(id_priv); 3611 if (ret) 3612 goto err; 3613 } else if (rdma_cap_iw_cm(id->device, 1)) { 3614 ret = cma_iw_listen(id_priv, backlog); 3615 if (ret) 3616 goto err; 3617 } else { 3618 ret = -ENOSYS; 3619 goto err; 3620 } 3621 } else 3622 cma_listen_on_all(id_priv); 3623 3624 return 0; 3625 err: 3626 id_priv->backlog = 0; 3627 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND); 3628 return ret; 3629 } 3630 EXPORT_SYMBOL(rdma_listen); 3631 3632 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr) 3633 { 3634 struct rdma_id_private *id_priv; 3635 int ret; 3636 struct sockaddr *daddr; 3637 3638 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 && 3639 addr->sa_family != AF_IB) 3640 return -EAFNOSUPPORT; 3641 3642 id_priv = container_of(id, struct rdma_id_private, id); 3643 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND)) 3644 return -EINVAL; 3645 3646 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr); 3647 if (ret) 3648 goto err1; 3649 3650 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr)); 3651 if (!cma_any_addr(addr)) { 3652 ret = cma_translate_addr(addr, &id->route.addr.dev_addr); 3653 if (ret) 3654 goto err1; 3655 3656 ret = cma_acquire_dev_by_src_ip(id_priv); 3657 if (ret) 3658 goto err1; 3659 } 3660 3661 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) { 3662 if (addr->sa_family == AF_INET) 3663 id_priv->afonly = 1; 3664 #if IS_ENABLED(CONFIG_IPV6) 3665 else if (addr->sa_family == AF_INET6) { 3666 struct net *net = id_priv->id.route.addr.dev_addr.net; 3667 3668 id_priv->afonly = net->ipv6.sysctl.bindv6only; 3669 } 3670 #endif 3671 } 3672 daddr = cma_dst_addr(id_priv); 3673 daddr->sa_family = addr->sa_family; 3674 3675 ret = cma_get_port(id_priv); 3676 if (ret) 3677 goto err2; 3678 3679 return 0; 3680 err2: 3681 rdma_restrack_del(&id_priv->res); 3682 if (id_priv->cma_dev) 3683 cma_release_dev(id_priv); 3684 err1: 3685 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE); 3686 return ret; 3687 } 3688 EXPORT_SYMBOL(rdma_bind_addr); 3689 3690 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv) 3691 { 3692 struct cma_hdr *cma_hdr; 3693 3694 cma_hdr = hdr; 3695 cma_hdr->cma_version = CMA_VERSION; 3696 if (cma_family(id_priv) == AF_INET) { 3697 struct sockaddr_in *src4, *dst4; 3698 3699 src4 = (struct sockaddr_in *) cma_src_addr(id_priv); 3700 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv); 3701 3702 cma_set_ip_ver(cma_hdr, 4); 3703 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr; 3704 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr; 3705 cma_hdr->port = src4->sin_port; 3706 } else if (cma_family(id_priv) == AF_INET6) { 3707 struct sockaddr_in6 *src6, *dst6; 3708 3709 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv); 3710 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv); 3711 3712 cma_set_ip_ver(cma_hdr, 6); 3713 cma_hdr->src_addr.ip6 = src6->sin6_addr; 3714 cma_hdr->dst_addr.ip6 = dst6->sin6_addr; 3715 cma_hdr->port = src6->sin6_port; 3716 } 3717 return 0; 3718 } 3719 3720 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id, 3721 const struct ib_cm_event *ib_event) 3722 { 3723 struct rdma_id_private *id_priv = cm_id->context; 3724 struct rdma_cm_event event = {}; 3725 const struct ib_cm_sidr_rep_event_param *rep = 3726 &ib_event->param.sidr_rep_rcvd; 3727 int ret = 0; 3728 3729 mutex_lock(&id_priv->handler_mutex); 3730 if (id_priv->state != RDMA_CM_CONNECT) 3731 goto out; 3732 3733 switch (ib_event->event) { 3734 case IB_CM_SIDR_REQ_ERROR: 3735 event.event = RDMA_CM_EVENT_UNREACHABLE; 3736 event.status = -ETIMEDOUT; 3737 break; 3738 case IB_CM_SIDR_REP_RECEIVED: 3739 event.param.ud.private_data = ib_event->private_data; 3740 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE; 3741 if (rep->status != IB_SIDR_SUCCESS) { 3742 event.event = RDMA_CM_EVENT_UNREACHABLE; 3743 event.status = ib_event->param.sidr_rep_rcvd.status; 3744 pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n", 3745 event.status); 3746 break; 3747 } 3748 ret = cma_set_qkey(id_priv, rep->qkey); 3749 if (ret) { 3750 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret); 3751 event.event = RDMA_CM_EVENT_ADDR_ERROR; 3752 event.status = ret; 3753 break; 3754 } 3755 ib_init_ah_attr_from_path(id_priv->id.device, 3756 id_priv->id.port_num, 3757 id_priv->id.route.path_rec, 3758 &event.param.ud.ah_attr, 3759 rep->sgid_attr); 3760 event.param.ud.qp_num = rep->qpn; 3761 event.param.ud.qkey = rep->qkey; 3762 event.event = RDMA_CM_EVENT_ESTABLISHED; 3763 event.status = 0; 3764 break; 3765 default: 3766 pr_err("RDMA CMA: unexpected IB CM event: %d\n", 3767 ib_event->event); 3768 goto out; 3769 } 3770 3771 ret = cma_cm_event_handler(id_priv, &event); 3772 3773 rdma_destroy_ah_attr(&event.param.ud.ah_attr); 3774 if (ret) { 3775 /* Destroy the CM ID by returning a non-zero value. */ 3776 id_priv->cm_id.ib = NULL; 3777 cma_exch(id_priv, RDMA_CM_DESTROYING); 3778 mutex_unlock(&id_priv->handler_mutex); 3779 rdma_destroy_id(&id_priv->id); 3780 return ret; 3781 } 3782 out: 3783 mutex_unlock(&id_priv->handler_mutex); 3784 return ret; 3785 } 3786 3787 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv, 3788 struct rdma_conn_param *conn_param) 3789 { 3790 struct ib_cm_sidr_req_param req; 3791 struct ib_cm_id *id; 3792 void *private_data; 3793 u8 offset; 3794 int ret; 3795 3796 memset(&req, 0, sizeof req); 3797 offset = cma_user_data_offset(id_priv); 3798 req.private_data_len = offset + conn_param->private_data_len; 3799 if (req.private_data_len < conn_param->private_data_len) 3800 return -EINVAL; 3801 3802 if (req.private_data_len) { 3803 private_data = kzalloc(req.private_data_len, GFP_ATOMIC); 3804 if (!private_data) 3805 return -ENOMEM; 3806 } else { 3807 private_data = NULL; 3808 } 3809 3810 if (conn_param->private_data && conn_param->private_data_len) 3811 memcpy(private_data + offset, conn_param->private_data, 3812 conn_param->private_data_len); 3813 3814 if (private_data) { 3815 ret = cma_format_hdr(private_data, id_priv); 3816 if (ret) 3817 goto out; 3818 req.private_data = private_data; 3819 } 3820 3821 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler, 3822 id_priv); 3823 if (IS_ERR(id)) { 3824 ret = PTR_ERR(id); 3825 goto out; 3826 } 3827 id_priv->cm_id.ib = id; 3828 3829 req.path = id_priv->id.route.path_rec; 3830 req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr; 3831 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv)); 3832 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8); 3833 req.max_cm_retries = CMA_MAX_CM_RETRIES; 3834 3835 trace_cm_send_sidr_req(id_priv); 3836 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req); 3837 if (ret) { 3838 ib_destroy_cm_id(id_priv->cm_id.ib); 3839 id_priv->cm_id.ib = NULL; 3840 } 3841 out: 3842 kfree(private_data); 3843 return ret; 3844 } 3845 3846 static int cma_connect_ib(struct rdma_id_private *id_priv, 3847 struct rdma_conn_param *conn_param) 3848 { 3849 struct ib_cm_req_param req; 3850 struct rdma_route *route; 3851 void *private_data; 3852 struct ib_cm_id *id; 3853 u8 offset; 3854 int ret; 3855 3856 memset(&req, 0, sizeof req); 3857 offset = cma_user_data_offset(id_priv); 3858 req.private_data_len = offset + conn_param->private_data_len; 3859 if (req.private_data_len < conn_param->private_data_len) 3860 return -EINVAL; 3861 3862 if (req.private_data_len) { 3863 private_data = kzalloc(req.private_data_len, GFP_ATOMIC); 3864 if (!private_data) 3865 return -ENOMEM; 3866 } else { 3867 private_data = NULL; 3868 } 3869 3870 if (conn_param->private_data && conn_param->private_data_len) 3871 memcpy(private_data + offset, conn_param->private_data, 3872 conn_param->private_data_len); 3873 3874 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv); 3875 if (IS_ERR(id)) { 3876 ret = PTR_ERR(id); 3877 goto out; 3878 } 3879 id_priv->cm_id.ib = id; 3880 3881 route = &id_priv->id.route; 3882 if (private_data) { 3883 ret = cma_format_hdr(private_data, id_priv); 3884 if (ret) 3885 goto out; 3886 req.private_data = private_data; 3887 } 3888 3889 req.primary_path = &route->path_rec[0]; 3890 if (route->num_paths == 2) 3891 req.alternate_path = &route->path_rec[1]; 3892 3893 req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr; 3894 /* Alternate path SGID attribute currently unsupported */ 3895 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv)); 3896 req.qp_num = id_priv->qp_num; 3897 req.qp_type = id_priv->id.qp_type; 3898 req.starting_psn = id_priv->seq_num; 3899 req.responder_resources = conn_param->responder_resources; 3900 req.initiator_depth = conn_param->initiator_depth; 3901 req.flow_control = conn_param->flow_control; 3902 req.retry_count = min_t(u8, 7, conn_param->retry_count); 3903 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count); 3904 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT; 3905 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT; 3906 req.max_cm_retries = CMA_MAX_CM_RETRIES; 3907 req.srq = id_priv->srq ? 1 : 0; 3908 3909 trace_cm_send_req(id_priv); 3910 ret = ib_send_cm_req(id_priv->cm_id.ib, &req); 3911 out: 3912 if (ret && !IS_ERR(id)) { 3913 ib_destroy_cm_id(id); 3914 id_priv->cm_id.ib = NULL; 3915 } 3916 3917 kfree(private_data); 3918 return ret; 3919 } 3920 3921 static int cma_connect_iw(struct rdma_id_private *id_priv, 3922 struct rdma_conn_param *conn_param) 3923 { 3924 struct iw_cm_id *cm_id; 3925 int ret; 3926 struct iw_cm_conn_param iw_param; 3927 3928 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv); 3929 if (IS_ERR(cm_id)) 3930 return PTR_ERR(cm_id); 3931 3932 cm_id->tos = id_priv->tos; 3933 cm_id->tos_set = id_priv->tos_set; 3934 id_priv->cm_id.iw = cm_id; 3935 3936 memcpy(&cm_id->local_addr, cma_src_addr(id_priv), 3937 rdma_addr_size(cma_src_addr(id_priv))); 3938 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv), 3939 rdma_addr_size(cma_dst_addr(id_priv))); 3940 3941 ret = cma_modify_qp_rtr(id_priv, conn_param); 3942 if (ret) 3943 goto out; 3944 3945 if (conn_param) { 3946 iw_param.ord = conn_param->initiator_depth; 3947 iw_param.ird = conn_param->responder_resources; 3948 iw_param.private_data = conn_param->private_data; 3949 iw_param.private_data_len = conn_param->private_data_len; 3950 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num; 3951 } else { 3952 memset(&iw_param, 0, sizeof iw_param); 3953 iw_param.qpn = id_priv->qp_num; 3954 } 3955 ret = iw_cm_connect(cm_id, &iw_param); 3956 out: 3957 if (ret) { 3958 iw_destroy_cm_id(cm_id); 3959 id_priv->cm_id.iw = NULL; 3960 } 3961 return ret; 3962 } 3963 3964 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param) 3965 { 3966 struct rdma_id_private *id_priv; 3967 int ret; 3968 3969 id_priv = container_of(id, struct rdma_id_private, id); 3970 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT)) 3971 return -EINVAL; 3972 3973 if (!id->qp) { 3974 id_priv->qp_num = conn_param->qp_num; 3975 id_priv->srq = conn_param->srq; 3976 } 3977 3978 if (rdma_cap_ib_cm(id->device, id->port_num)) { 3979 if (id->qp_type == IB_QPT_UD) 3980 ret = cma_resolve_ib_udp(id_priv, conn_param); 3981 else 3982 ret = cma_connect_ib(id_priv, conn_param); 3983 } else if (rdma_cap_iw_cm(id->device, id->port_num)) 3984 ret = cma_connect_iw(id_priv, conn_param); 3985 else 3986 ret = -ENOSYS; 3987 if (ret) 3988 goto err; 3989 3990 return 0; 3991 err: 3992 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED); 3993 return ret; 3994 } 3995 EXPORT_SYMBOL(rdma_connect); 3996 3997 static int cma_accept_ib(struct rdma_id_private *id_priv, 3998 struct rdma_conn_param *conn_param) 3999 { 4000 struct ib_cm_rep_param rep; 4001 int ret; 4002 4003 ret = cma_modify_qp_rtr(id_priv, conn_param); 4004 if (ret) 4005 goto out; 4006 4007 ret = cma_modify_qp_rts(id_priv, conn_param); 4008 if (ret) 4009 goto out; 4010 4011 memset(&rep, 0, sizeof rep); 4012 rep.qp_num = id_priv->qp_num; 4013 rep.starting_psn = id_priv->seq_num; 4014 rep.private_data = conn_param->private_data; 4015 rep.private_data_len = conn_param->private_data_len; 4016 rep.responder_resources = conn_param->responder_resources; 4017 rep.initiator_depth = conn_param->initiator_depth; 4018 rep.failover_accepted = 0; 4019 rep.flow_control = conn_param->flow_control; 4020 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count); 4021 rep.srq = id_priv->srq ? 1 : 0; 4022 4023 trace_cm_send_rep(id_priv); 4024 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep); 4025 out: 4026 return ret; 4027 } 4028 4029 static int cma_accept_iw(struct rdma_id_private *id_priv, 4030 struct rdma_conn_param *conn_param) 4031 { 4032 struct iw_cm_conn_param iw_param; 4033 int ret; 4034 4035 if (!conn_param) 4036 return -EINVAL; 4037 4038 ret = cma_modify_qp_rtr(id_priv, conn_param); 4039 if (ret) 4040 return ret; 4041 4042 iw_param.ord = conn_param->initiator_depth; 4043 iw_param.ird = conn_param->responder_resources; 4044 iw_param.private_data = conn_param->private_data; 4045 iw_param.private_data_len = conn_param->private_data_len; 4046 if (id_priv->id.qp) { 4047 iw_param.qpn = id_priv->qp_num; 4048 } else 4049 iw_param.qpn = conn_param->qp_num; 4050 4051 return iw_cm_accept(id_priv->cm_id.iw, &iw_param); 4052 } 4053 4054 static int cma_send_sidr_rep(struct rdma_id_private *id_priv, 4055 enum ib_cm_sidr_status status, u32 qkey, 4056 const void *private_data, int private_data_len) 4057 { 4058 struct ib_cm_sidr_rep_param rep; 4059 int ret; 4060 4061 memset(&rep, 0, sizeof rep); 4062 rep.status = status; 4063 if (status == IB_SIDR_SUCCESS) { 4064 ret = cma_set_qkey(id_priv, qkey); 4065 if (ret) 4066 return ret; 4067 rep.qp_num = id_priv->qp_num; 4068 rep.qkey = id_priv->qkey; 4069 } 4070 rep.private_data = private_data; 4071 rep.private_data_len = private_data_len; 4072 4073 trace_cm_send_sidr_rep(id_priv); 4074 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep); 4075 } 4076 4077 int __rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param, 4078 const char *caller) 4079 { 4080 struct rdma_id_private *id_priv; 4081 int ret; 4082 4083 id_priv = container_of(id, struct rdma_id_private, id); 4084 4085 rdma_restrack_set_task(&id_priv->res, caller); 4086 4087 if (!cma_comp(id_priv, RDMA_CM_CONNECT)) 4088 return -EINVAL; 4089 4090 if (!id->qp && conn_param) { 4091 id_priv->qp_num = conn_param->qp_num; 4092 id_priv->srq = conn_param->srq; 4093 } 4094 4095 if (rdma_cap_ib_cm(id->device, id->port_num)) { 4096 if (id->qp_type == IB_QPT_UD) { 4097 if (conn_param) 4098 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS, 4099 conn_param->qkey, 4100 conn_param->private_data, 4101 conn_param->private_data_len); 4102 else 4103 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS, 4104 0, NULL, 0); 4105 } else { 4106 if (conn_param) 4107 ret = cma_accept_ib(id_priv, conn_param); 4108 else 4109 ret = cma_rep_recv(id_priv); 4110 } 4111 } else if (rdma_cap_iw_cm(id->device, id->port_num)) 4112 ret = cma_accept_iw(id_priv, conn_param); 4113 else 4114 ret = -ENOSYS; 4115 4116 if (ret) 4117 goto reject; 4118 4119 return 0; 4120 reject: 4121 cma_modify_qp_err(id_priv); 4122 rdma_reject(id, NULL, 0); 4123 return ret; 4124 } 4125 EXPORT_SYMBOL(__rdma_accept); 4126 4127 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event) 4128 { 4129 struct rdma_id_private *id_priv; 4130 int ret; 4131 4132 id_priv = container_of(id, struct rdma_id_private, id); 4133 if (!id_priv->cm_id.ib) 4134 return -EINVAL; 4135 4136 switch (id->device->node_type) { 4137 case RDMA_NODE_IB_CA: 4138 ret = ib_cm_notify(id_priv->cm_id.ib, event); 4139 break; 4140 default: 4141 ret = 0; 4142 break; 4143 } 4144 return ret; 4145 } 4146 EXPORT_SYMBOL(rdma_notify); 4147 4148 int rdma_reject(struct rdma_cm_id *id, const void *private_data, 4149 u8 private_data_len) 4150 { 4151 struct rdma_id_private *id_priv; 4152 int ret; 4153 4154 id_priv = container_of(id, struct rdma_id_private, id); 4155 if (!id_priv->cm_id.ib) 4156 return -EINVAL; 4157 4158 if (rdma_cap_ib_cm(id->device, id->port_num)) { 4159 if (id->qp_type == IB_QPT_UD) { 4160 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0, 4161 private_data, private_data_len); 4162 } else { 4163 trace_cm_send_rej(id_priv); 4164 ret = ib_send_cm_rej(id_priv->cm_id.ib, 4165 IB_CM_REJ_CONSUMER_DEFINED, NULL, 4166 0, private_data, private_data_len); 4167 } 4168 } else if (rdma_cap_iw_cm(id->device, id->port_num)) { 4169 ret = iw_cm_reject(id_priv->cm_id.iw, 4170 private_data, private_data_len); 4171 } else 4172 ret = -ENOSYS; 4173 4174 return ret; 4175 } 4176 EXPORT_SYMBOL(rdma_reject); 4177 4178 int rdma_disconnect(struct rdma_cm_id *id) 4179 { 4180 struct rdma_id_private *id_priv; 4181 int ret; 4182 4183 id_priv = container_of(id, struct rdma_id_private, id); 4184 if (!id_priv->cm_id.ib) 4185 return -EINVAL; 4186 4187 if (rdma_cap_ib_cm(id->device, id->port_num)) { 4188 ret = cma_modify_qp_err(id_priv); 4189 if (ret) 4190 goto out; 4191 /* Initiate or respond to a disconnect. */ 4192 trace_cm_disconnect(id_priv); 4193 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) { 4194 if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0)) 4195 trace_cm_sent_drep(id_priv); 4196 } else { 4197 trace_cm_sent_dreq(id_priv); 4198 } 4199 } else if (rdma_cap_iw_cm(id->device, id->port_num)) { 4200 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0); 4201 } else 4202 ret = -EINVAL; 4203 4204 out: 4205 return ret; 4206 } 4207 EXPORT_SYMBOL(rdma_disconnect); 4208 4209 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast) 4210 { 4211 struct rdma_id_private *id_priv; 4212 struct cma_multicast *mc = multicast->context; 4213 struct rdma_cm_event event = {}; 4214 int ret = 0; 4215 4216 id_priv = mc->id_priv; 4217 mutex_lock(&id_priv->handler_mutex); 4218 if (id_priv->state != RDMA_CM_ADDR_BOUND && 4219 id_priv->state != RDMA_CM_ADDR_RESOLVED) 4220 goto out; 4221 4222 if (!status) 4223 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey)); 4224 else 4225 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n", 4226 status); 4227 mutex_lock(&id_priv->qp_mutex); 4228 if (!status && id_priv->id.qp) { 4229 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid, 4230 be16_to_cpu(multicast->rec.mlid)); 4231 if (status) 4232 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to attach QP. status %d\n", 4233 status); 4234 } 4235 mutex_unlock(&id_priv->qp_mutex); 4236 4237 event.status = status; 4238 event.param.ud.private_data = mc->context; 4239 if (!status) { 4240 struct rdma_dev_addr *dev_addr = 4241 &id_priv->id.route.addr.dev_addr; 4242 struct net_device *ndev = 4243 dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if); 4244 enum ib_gid_type gid_type = 4245 id_priv->cma_dev->default_gid_type[id_priv->id.port_num - 4246 rdma_start_port(id_priv->cma_dev->device)]; 4247 4248 event.event = RDMA_CM_EVENT_MULTICAST_JOIN; 4249 ret = ib_init_ah_from_mcmember(id_priv->id.device, 4250 id_priv->id.port_num, 4251 &multicast->rec, 4252 ndev, gid_type, 4253 &event.param.ud.ah_attr); 4254 if (ret) 4255 event.event = RDMA_CM_EVENT_MULTICAST_ERROR; 4256 4257 event.param.ud.qp_num = 0xFFFFFF; 4258 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey); 4259 if (ndev) 4260 dev_put(ndev); 4261 } else 4262 event.event = RDMA_CM_EVENT_MULTICAST_ERROR; 4263 4264 ret = cma_cm_event_handler(id_priv, &event); 4265 4266 rdma_destroy_ah_attr(&event.param.ud.ah_attr); 4267 if (ret) { 4268 cma_exch(id_priv, RDMA_CM_DESTROYING); 4269 mutex_unlock(&id_priv->handler_mutex); 4270 rdma_destroy_id(&id_priv->id); 4271 return 0; 4272 } 4273 4274 out: 4275 mutex_unlock(&id_priv->handler_mutex); 4276 return 0; 4277 } 4278 4279 static void cma_set_mgid(struct rdma_id_private *id_priv, 4280 struct sockaddr *addr, union ib_gid *mgid) 4281 { 4282 unsigned char mc_map[MAX_ADDR_LEN]; 4283 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 4284 struct sockaddr_in *sin = (struct sockaddr_in *) addr; 4285 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr; 4286 4287 if (cma_any_addr(addr)) { 4288 memset(mgid, 0, sizeof *mgid); 4289 } else if ((addr->sa_family == AF_INET6) && 4290 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) == 4291 0xFF10A01B)) { 4292 /* IPv6 address is an SA assigned MGID. */ 4293 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid); 4294 } else if (addr->sa_family == AF_IB) { 4295 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid); 4296 } else if (addr->sa_family == AF_INET6) { 4297 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map); 4298 if (id_priv->id.ps == RDMA_PS_UDP) 4299 mc_map[7] = 0x01; /* Use RDMA CM signature */ 4300 *mgid = *(union ib_gid *) (mc_map + 4); 4301 } else { 4302 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map); 4303 if (id_priv->id.ps == RDMA_PS_UDP) 4304 mc_map[7] = 0x01; /* Use RDMA CM signature */ 4305 *mgid = *(union ib_gid *) (mc_map + 4); 4306 } 4307 } 4308 4309 static int cma_join_ib_multicast(struct rdma_id_private *id_priv, 4310 struct cma_multicast *mc) 4311 { 4312 struct ib_sa_mcmember_rec rec; 4313 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 4314 ib_sa_comp_mask comp_mask; 4315 int ret; 4316 4317 ib_addr_get_mgid(dev_addr, &rec.mgid); 4318 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num, 4319 &rec.mgid, &rec); 4320 if (ret) 4321 return ret; 4322 4323 ret = cma_set_qkey(id_priv, 0); 4324 if (ret) 4325 return ret; 4326 4327 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid); 4328 rec.qkey = cpu_to_be32(id_priv->qkey); 4329 rdma_addr_get_sgid(dev_addr, &rec.port_gid); 4330 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr)); 4331 rec.join_state = mc->join_state; 4332 4333 if ((rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) && 4334 (!ib_sa_sendonly_fullmem_support(&sa_client, 4335 id_priv->id.device, 4336 id_priv->id.port_num))) { 4337 dev_warn( 4338 &id_priv->id.device->dev, 4339 "RDMA CM: port %u Unable to multicast join: SM doesn't support Send Only Full Member option\n", 4340 id_priv->id.port_num); 4341 return -EOPNOTSUPP; 4342 } 4343 4344 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID | 4345 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE | 4346 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL | 4347 IB_SA_MCMEMBER_REC_FLOW_LABEL | 4348 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS; 4349 4350 if (id_priv->id.ps == RDMA_PS_IPOIB) 4351 comp_mask |= IB_SA_MCMEMBER_REC_RATE | 4352 IB_SA_MCMEMBER_REC_RATE_SELECTOR | 4353 IB_SA_MCMEMBER_REC_MTU_SELECTOR | 4354 IB_SA_MCMEMBER_REC_MTU | 4355 IB_SA_MCMEMBER_REC_HOP_LIMIT; 4356 4357 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device, 4358 id_priv->id.port_num, &rec, 4359 comp_mask, GFP_KERNEL, 4360 cma_ib_mc_handler, mc); 4361 return PTR_ERR_OR_ZERO(mc->multicast.ib); 4362 } 4363 4364 static void iboe_mcast_work_handler(struct work_struct *work) 4365 { 4366 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work); 4367 struct cma_multicast *mc = mw->mc; 4368 struct ib_sa_multicast *m = mc->multicast.ib; 4369 4370 mc->multicast.ib->context = mc; 4371 cma_ib_mc_handler(0, m); 4372 kref_put(&mc->mcref, release_mc); 4373 kfree(mw); 4374 } 4375 4376 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid, 4377 enum ib_gid_type gid_type) 4378 { 4379 struct sockaddr_in *sin = (struct sockaddr_in *)addr; 4380 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr; 4381 4382 if (cma_any_addr(addr)) { 4383 memset(mgid, 0, sizeof *mgid); 4384 } else if (addr->sa_family == AF_INET6) { 4385 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid); 4386 } else { 4387 mgid->raw[0] = 4388 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff; 4389 mgid->raw[1] = 4390 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e; 4391 mgid->raw[2] = 0; 4392 mgid->raw[3] = 0; 4393 mgid->raw[4] = 0; 4394 mgid->raw[5] = 0; 4395 mgid->raw[6] = 0; 4396 mgid->raw[7] = 0; 4397 mgid->raw[8] = 0; 4398 mgid->raw[9] = 0; 4399 mgid->raw[10] = 0xff; 4400 mgid->raw[11] = 0xff; 4401 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr; 4402 } 4403 } 4404 4405 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv, 4406 struct cma_multicast *mc) 4407 { 4408 struct iboe_mcast_work *work; 4409 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 4410 int err = 0; 4411 struct sockaddr *addr = (struct sockaddr *)&mc->addr; 4412 struct net_device *ndev = NULL; 4413 enum ib_gid_type gid_type; 4414 bool send_only; 4415 4416 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN); 4417 4418 if (cma_zero_addr((struct sockaddr *)&mc->addr)) 4419 return -EINVAL; 4420 4421 work = kzalloc(sizeof *work, GFP_KERNEL); 4422 if (!work) 4423 return -ENOMEM; 4424 4425 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL); 4426 if (!mc->multicast.ib) { 4427 err = -ENOMEM; 4428 goto out1; 4429 } 4430 4431 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num - 4432 rdma_start_port(id_priv->cma_dev->device)]; 4433 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid, gid_type); 4434 4435 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff); 4436 if (id_priv->id.ps == RDMA_PS_UDP) 4437 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY); 4438 4439 if (dev_addr->bound_dev_if) 4440 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if); 4441 if (!ndev) { 4442 err = -ENODEV; 4443 goto out2; 4444 } 4445 mc->multicast.ib->rec.rate = iboe_get_rate(ndev); 4446 mc->multicast.ib->rec.hop_limit = 1; 4447 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu); 4448 4449 if (addr->sa_family == AF_INET) { 4450 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) { 4451 mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT; 4452 if (!send_only) { 4453 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid, 4454 true); 4455 } 4456 } 4457 } else { 4458 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) 4459 err = -ENOTSUPP; 4460 } 4461 dev_put(ndev); 4462 if (err || !mc->multicast.ib->rec.mtu) { 4463 if (!err) 4464 err = -EINVAL; 4465 goto out2; 4466 } 4467 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr, 4468 &mc->multicast.ib->rec.port_gid); 4469 work->id = id_priv; 4470 work->mc = mc; 4471 INIT_WORK(&work->work, iboe_mcast_work_handler); 4472 kref_get(&mc->mcref); 4473 queue_work(cma_wq, &work->work); 4474 4475 return 0; 4476 4477 out2: 4478 kfree(mc->multicast.ib); 4479 out1: 4480 kfree(work); 4481 return err; 4482 } 4483 4484 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr, 4485 u8 join_state, void *context) 4486 { 4487 struct rdma_id_private *id_priv; 4488 struct cma_multicast *mc; 4489 int ret; 4490 4491 if (!id->device) 4492 return -EINVAL; 4493 4494 id_priv = container_of(id, struct rdma_id_private, id); 4495 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) && 4496 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED)) 4497 return -EINVAL; 4498 4499 mc = kmalloc(sizeof *mc, GFP_KERNEL); 4500 if (!mc) 4501 return -ENOMEM; 4502 4503 memcpy(&mc->addr, addr, rdma_addr_size(addr)); 4504 mc->context = context; 4505 mc->id_priv = id_priv; 4506 mc->join_state = join_state; 4507 4508 if (rdma_protocol_roce(id->device, id->port_num)) { 4509 kref_init(&mc->mcref); 4510 ret = cma_iboe_join_multicast(id_priv, mc); 4511 if (ret) 4512 goto out_err; 4513 } else if (rdma_cap_ib_mcast(id->device, id->port_num)) { 4514 ret = cma_join_ib_multicast(id_priv, mc); 4515 if (ret) 4516 goto out_err; 4517 } else { 4518 ret = -ENOSYS; 4519 goto out_err; 4520 } 4521 4522 spin_lock(&id_priv->lock); 4523 list_add(&mc->list, &id_priv->mc_list); 4524 spin_unlock(&id_priv->lock); 4525 4526 return 0; 4527 out_err: 4528 kfree(mc); 4529 return ret; 4530 } 4531 EXPORT_SYMBOL(rdma_join_multicast); 4532 4533 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr) 4534 { 4535 struct rdma_id_private *id_priv; 4536 struct cma_multicast *mc; 4537 4538 id_priv = container_of(id, struct rdma_id_private, id); 4539 spin_lock_irq(&id_priv->lock); 4540 list_for_each_entry(mc, &id_priv->mc_list, list) { 4541 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) { 4542 list_del(&mc->list); 4543 spin_unlock_irq(&id_priv->lock); 4544 4545 if (id->qp) 4546 ib_detach_mcast(id->qp, 4547 &mc->multicast.ib->rec.mgid, 4548 be16_to_cpu(mc->multicast.ib->rec.mlid)); 4549 4550 BUG_ON(id_priv->cma_dev->device != id->device); 4551 4552 if (rdma_cap_ib_mcast(id->device, id->port_num)) { 4553 ib_sa_free_multicast(mc->multicast.ib); 4554 kfree(mc); 4555 } else if (rdma_protocol_roce(id->device, id->port_num)) { 4556 cma_leave_roce_mc_group(id_priv, mc); 4557 } 4558 return; 4559 } 4560 } 4561 spin_unlock_irq(&id_priv->lock); 4562 } 4563 EXPORT_SYMBOL(rdma_leave_multicast); 4564 4565 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv) 4566 { 4567 struct rdma_dev_addr *dev_addr; 4568 struct cma_ndev_work *work; 4569 4570 dev_addr = &id_priv->id.route.addr.dev_addr; 4571 4572 if ((dev_addr->bound_dev_if == ndev->ifindex) && 4573 (net_eq(dev_net(ndev), dev_addr->net)) && 4574 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) { 4575 pr_info("RDMA CM addr change for ndev %s used by id %p\n", 4576 ndev->name, &id_priv->id); 4577 work = kzalloc(sizeof *work, GFP_KERNEL); 4578 if (!work) 4579 return -ENOMEM; 4580 4581 INIT_WORK(&work->work, cma_ndev_work_handler); 4582 work->id = id_priv; 4583 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE; 4584 atomic_inc(&id_priv->refcount); 4585 queue_work(cma_wq, &work->work); 4586 } 4587 4588 return 0; 4589 } 4590 4591 static int cma_netdev_callback(struct notifier_block *self, unsigned long event, 4592 void *ptr) 4593 { 4594 struct net_device *ndev = netdev_notifier_info_to_dev(ptr); 4595 struct cma_device *cma_dev; 4596 struct rdma_id_private *id_priv; 4597 int ret = NOTIFY_DONE; 4598 4599 if (event != NETDEV_BONDING_FAILOVER) 4600 return NOTIFY_DONE; 4601 4602 if (!netif_is_bond_master(ndev)) 4603 return NOTIFY_DONE; 4604 4605 mutex_lock(&lock); 4606 list_for_each_entry(cma_dev, &dev_list, list) 4607 list_for_each_entry(id_priv, &cma_dev->id_list, list) { 4608 ret = cma_netdev_change(ndev, id_priv); 4609 if (ret) 4610 goto out; 4611 } 4612 4613 out: 4614 mutex_unlock(&lock); 4615 return ret; 4616 } 4617 4618 static struct notifier_block cma_nb = { 4619 .notifier_call = cma_netdev_callback 4620 }; 4621 4622 static void cma_add_one(struct ib_device *device) 4623 { 4624 struct cma_device *cma_dev; 4625 struct rdma_id_private *id_priv; 4626 unsigned int i; 4627 unsigned long supported_gids = 0; 4628 4629 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL); 4630 if (!cma_dev) 4631 return; 4632 4633 cma_dev->device = device; 4634 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt, 4635 sizeof(*cma_dev->default_gid_type), 4636 GFP_KERNEL); 4637 if (!cma_dev->default_gid_type) 4638 goto free_cma_dev; 4639 4640 cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt, 4641 sizeof(*cma_dev->default_roce_tos), 4642 GFP_KERNEL); 4643 if (!cma_dev->default_roce_tos) 4644 goto free_gid_type; 4645 4646 rdma_for_each_port (device, i) { 4647 supported_gids = roce_gid_type_mask_support(device, i); 4648 WARN_ON(!supported_gids); 4649 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE)) 4650 cma_dev->default_gid_type[i - rdma_start_port(device)] = 4651 CMA_PREFERRED_ROCE_GID_TYPE; 4652 else 4653 cma_dev->default_gid_type[i - rdma_start_port(device)] = 4654 find_first_bit(&supported_gids, BITS_PER_LONG); 4655 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0; 4656 } 4657 4658 init_completion(&cma_dev->comp); 4659 atomic_set(&cma_dev->refcount, 1); 4660 INIT_LIST_HEAD(&cma_dev->id_list); 4661 ib_set_client_data(device, &cma_client, cma_dev); 4662 4663 mutex_lock(&lock); 4664 list_add_tail(&cma_dev->list, &dev_list); 4665 list_for_each_entry(id_priv, &listen_any_list, list) 4666 cma_listen_on_dev(id_priv, cma_dev); 4667 mutex_unlock(&lock); 4668 4669 trace_cm_add_one(device); 4670 return; 4671 4672 free_gid_type: 4673 kfree(cma_dev->default_gid_type); 4674 4675 free_cma_dev: 4676 kfree(cma_dev); 4677 4678 return; 4679 } 4680 4681 static int cma_remove_id_dev(struct rdma_id_private *id_priv) 4682 { 4683 struct rdma_cm_event event = {}; 4684 enum rdma_cm_state state; 4685 int ret = 0; 4686 4687 /* Record that we want to remove the device */ 4688 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL); 4689 if (state == RDMA_CM_DESTROYING) 4690 return 0; 4691 4692 cma_cancel_operation(id_priv, state); 4693 mutex_lock(&id_priv->handler_mutex); 4694 4695 /* Check for destruction from another callback. */ 4696 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL)) 4697 goto out; 4698 4699 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL; 4700 ret = cma_cm_event_handler(id_priv, &event); 4701 out: 4702 mutex_unlock(&id_priv->handler_mutex); 4703 return ret; 4704 } 4705 4706 static void cma_process_remove(struct cma_device *cma_dev) 4707 { 4708 struct rdma_id_private *id_priv; 4709 int ret; 4710 4711 mutex_lock(&lock); 4712 while (!list_empty(&cma_dev->id_list)) { 4713 id_priv = list_entry(cma_dev->id_list.next, 4714 struct rdma_id_private, list); 4715 4716 list_del(&id_priv->listen_list); 4717 list_del_init(&id_priv->list); 4718 atomic_inc(&id_priv->refcount); 4719 mutex_unlock(&lock); 4720 4721 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv); 4722 cma_deref_id(id_priv); 4723 if (ret) 4724 rdma_destroy_id(&id_priv->id); 4725 4726 mutex_lock(&lock); 4727 } 4728 mutex_unlock(&lock); 4729 4730 cma_deref_dev(cma_dev); 4731 wait_for_completion(&cma_dev->comp); 4732 } 4733 4734 static void cma_remove_one(struct ib_device *device, void *client_data) 4735 { 4736 struct cma_device *cma_dev = client_data; 4737 4738 trace_cm_remove_one(device); 4739 4740 if (!cma_dev) 4741 return; 4742 4743 mutex_lock(&lock); 4744 list_del(&cma_dev->list); 4745 mutex_unlock(&lock); 4746 4747 cma_process_remove(cma_dev); 4748 kfree(cma_dev->default_roce_tos); 4749 kfree(cma_dev->default_gid_type); 4750 kfree(cma_dev); 4751 } 4752 4753 static int cma_init_net(struct net *net) 4754 { 4755 struct cma_pernet *pernet = cma_pernet(net); 4756 4757 xa_init(&pernet->tcp_ps); 4758 xa_init(&pernet->udp_ps); 4759 xa_init(&pernet->ipoib_ps); 4760 xa_init(&pernet->ib_ps); 4761 4762 return 0; 4763 } 4764 4765 static void cma_exit_net(struct net *net) 4766 { 4767 struct cma_pernet *pernet = cma_pernet(net); 4768 4769 WARN_ON(!xa_empty(&pernet->tcp_ps)); 4770 WARN_ON(!xa_empty(&pernet->udp_ps)); 4771 WARN_ON(!xa_empty(&pernet->ipoib_ps)); 4772 WARN_ON(!xa_empty(&pernet->ib_ps)); 4773 } 4774 4775 static struct pernet_operations cma_pernet_operations = { 4776 .init = cma_init_net, 4777 .exit = cma_exit_net, 4778 .id = &cma_pernet_id, 4779 .size = sizeof(struct cma_pernet), 4780 }; 4781 4782 static int __init cma_init(void) 4783 { 4784 int ret; 4785 4786 cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM); 4787 if (!cma_wq) 4788 return -ENOMEM; 4789 4790 ret = register_pernet_subsys(&cma_pernet_operations); 4791 if (ret) 4792 goto err_wq; 4793 4794 ib_sa_register_client(&sa_client); 4795 register_netdevice_notifier(&cma_nb); 4796 4797 ret = ib_register_client(&cma_client); 4798 if (ret) 4799 goto err; 4800 4801 ret = cma_configfs_init(); 4802 if (ret) 4803 goto err_ib; 4804 4805 return 0; 4806 4807 err_ib: 4808 ib_unregister_client(&cma_client); 4809 err: 4810 unregister_netdevice_notifier(&cma_nb); 4811 ib_sa_unregister_client(&sa_client); 4812 unregister_pernet_subsys(&cma_pernet_operations); 4813 err_wq: 4814 destroy_workqueue(cma_wq); 4815 return ret; 4816 } 4817 4818 static void __exit cma_cleanup(void) 4819 { 4820 cma_configfs_exit(); 4821 ib_unregister_client(&cma_client); 4822 unregister_netdevice_notifier(&cma_nb); 4823 ib_sa_unregister_client(&sa_client); 4824 unregister_pernet_subsys(&cma_pernet_operations); 4825 destroy_workqueue(cma_wq); 4826 } 4827 4828 module_init(cma_init); 4829 module_exit(cma_cleanup); 4830