1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0 3 * 4 * Copyright (c) 2005 Voltaire Inc. All rights reserved. 5 * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved. 6 * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved. 7 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved. 8 * 9 * This software is available to you under a choice of one of two 10 * licenses. You may choose to be licensed under the terms of the GNU 11 * General Public License (GPL) Version 2, available from the file 12 * COPYING in the main directory of this source tree, or the 13 * OpenIB.org BSD license below: 14 * 15 * Redistribution and use in source and binary forms, with or 16 * without modification, are permitted provided that the following 17 * conditions are met: 18 * 19 * - Redistributions of source code must retain the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer. 22 * 23 * - Redistributions in binary form must reproduce the above 24 * copyright notice, this list of conditions and the following 25 * disclaimer in the documentation and/or other materials 26 * provided with the distribution. 27 * 28 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 29 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 30 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 31 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 32 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 33 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 34 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 35 * SOFTWARE. 36 */ 37 38 #include <sys/cdefs.h> 39 __FBSDID("$FreeBSD$"); 40 41 #define LINUXKPI_PARAM_PREFIX ibcore_ 42 43 #include "opt_inet.h" 44 #include "opt_inet6.h" 45 46 #include <linux/completion.h> 47 #include <linux/in.h> 48 #include <linux/in6.h> 49 #include <linux/mutex.h> 50 #include <linux/random.h> 51 #include <linux/idr.h> 52 #include <linux/slab.h> 53 #include <linux/module.h> 54 #include <net/route.h> 55 #include <net/route/nhop.h> 56 57 #include <net/tcp.h> 58 #include <net/ipv6.h> 59 60 #include <netinet/in_fib.h> 61 62 #include <netinet6/in6_fib.h> 63 #include <netinet6/scope6_var.h> 64 #include <netinet6/ip6_var.h> 65 66 #include <rdma/rdma_cm.h> 67 #include <rdma/rdma_cm_ib.h> 68 #include <rdma/rdma_sdp.h> 69 #include <rdma/ib.h> 70 #include <rdma/ib_addr.h> 71 #include <rdma/ib_cache.h> 72 #include <rdma/ib_cm.h> 73 #include <rdma/ib_sa.h> 74 #include <rdma/iw_cm.h> 75 76 #include <sys/priv.h> 77 78 #include "core_priv.h" 79 80 MODULE_AUTHOR("Sean Hefty"); 81 MODULE_DESCRIPTION("Generic RDMA CM Agent"); 82 MODULE_LICENSE("Dual BSD/GPL"); 83 84 #define CMA_CM_RESPONSE_TIMEOUT 20 85 #define CMA_QUERY_CLASSPORT_INFO_TIMEOUT 3000 86 #define CMA_MAX_CM_RETRIES 15 87 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24) 88 #define CMA_IBOE_PACKET_LIFETIME 18 89 90 static const char * const cma_events[] = { 91 [RDMA_CM_EVENT_ADDR_RESOLVED] = "address resolved", 92 [RDMA_CM_EVENT_ADDR_ERROR] = "address error", 93 [RDMA_CM_EVENT_ROUTE_RESOLVED] = "route resolved ", 94 [RDMA_CM_EVENT_ROUTE_ERROR] = "route error", 95 [RDMA_CM_EVENT_CONNECT_REQUEST] = "connect request", 96 [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response", 97 [RDMA_CM_EVENT_CONNECT_ERROR] = "connect error", 98 [RDMA_CM_EVENT_UNREACHABLE] = "unreachable", 99 [RDMA_CM_EVENT_REJECTED] = "rejected", 100 [RDMA_CM_EVENT_ESTABLISHED] = "established", 101 [RDMA_CM_EVENT_DISCONNECTED] = "disconnected", 102 [RDMA_CM_EVENT_DEVICE_REMOVAL] = "device removal", 103 [RDMA_CM_EVENT_MULTICAST_JOIN] = "multicast join", 104 [RDMA_CM_EVENT_MULTICAST_ERROR] = "multicast error", 105 [RDMA_CM_EVENT_ADDR_CHANGE] = "address change", 106 [RDMA_CM_EVENT_TIMEWAIT_EXIT] = "timewait exit", 107 }; 108 109 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event) 110 { 111 size_t index = event; 112 113 return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ? 114 cma_events[index] : "unrecognized event"; 115 } 116 EXPORT_SYMBOL(rdma_event_msg); 117 118 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id, 119 int reason) 120 { 121 if (rdma_ib_or_roce(id->device, id->port_num)) 122 return ibcm_reject_msg(reason); 123 124 if (rdma_protocol_iwarp(id->device, id->port_num)) 125 return iwcm_reject_msg(reason); 126 127 WARN_ON_ONCE(1); 128 return "unrecognized transport"; 129 } 130 EXPORT_SYMBOL(rdma_reject_msg); 131 132 static int cma_check_linklocal(struct rdma_dev_addr *, struct sockaddr *); 133 static void cma_add_one(struct ib_device *device); 134 static void cma_remove_one(struct ib_device *device, void *client_data); 135 static enum rdma_port_space rdma_ps_from_service_id(__be64 service_id); 136 137 static struct ib_client cma_client = { 138 .name = "cma", 139 .add = cma_add_one, 140 .remove = cma_remove_one 141 }; 142 143 static struct ib_sa_client sa_client; 144 static struct rdma_addr_client addr_client; 145 static LIST_HEAD(dev_list); 146 static LIST_HEAD(listen_any_list); 147 static DEFINE_MUTEX(lock); 148 static struct workqueue_struct *cma_wq; 149 150 struct cma_pernet { 151 struct idr tcp_ps; 152 struct idr udp_ps; 153 struct idr ipoib_ps; 154 struct idr ib_ps; 155 struct idr sdp_ps; 156 }; 157 158 VNET_DEFINE(struct cma_pernet, cma_pernet); 159 160 static struct cma_pernet *cma_pernet_ptr(struct vnet *vnet) 161 { 162 struct cma_pernet *retval; 163 164 CURVNET_SET_QUIET(vnet); 165 retval = &VNET(cma_pernet); 166 CURVNET_RESTORE(); 167 168 return (retval); 169 } 170 171 static struct idr *cma_pernet_idr(struct vnet *net, enum rdma_port_space ps) 172 { 173 struct cma_pernet *pernet = cma_pernet_ptr(net); 174 175 switch (ps) { 176 case RDMA_PS_TCP: 177 return &pernet->tcp_ps; 178 case RDMA_PS_UDP: 179 return &pernet->udp_ps; 180 case RDMA_PS_IPOIB: 181 return &pernet->ipoib_ps; 182 case RDMA_PS_IB: 183 return &pernet->ib_ps; 184 case RDMA_PS_SDP: 185 return &pernet->sdp_ps; 186 default: 187 return NULL; 188 } 189 } 190 191 struct cma_device { 192 struct list_head list; 193 struct ib_device *device; 194 struct completion comp; 195 atomic_t refcount; 196 struct list_head id_list; 197 struct sysctl_ctx_list sysctl_ctx; 198 enum ib_gid_type *default_gid_type; 199 }; 200 201 struct rdma_bind_list { 202 enum rdma_port_space ps; 203 struct hlist_head owners; 204 unsigned short port; 205 }; 206 207 struct class_port_info_context { 208 struct ib_class_port_info *class_port_info; 209 struct ib_device *device; 210 struct completion done; 211 struct ib_sa_query *sa_query; 212 u8 port_num; 213 }; 214 215 static int cma_ps_alloc(struct vnet *vnet, enum rdma_port_space ps, 216 struct rdma_bind_list *bind_list, int snum) 217 { 218 struct idr *idr = cma_pernet_idr(vnet, ps); 219 220 return idr_alloc(idr, bind_list, snum, snum + 1, GFP_KERNEL); 221 } 222 223 static struct rdma_bind_list *cma_ps_find(struct vnet *net, 224 enum rdma_port_space ps, int snum) 225 { 226 struct idr *idr = cma_pernet_idr(net, ps); 227 228 return idr_find(idr, snum); 229 } 230 231 static void cma_ps_remove(struct vnet *net, enum rdma_port_space ps, int snum) 232 { 233 struct idr *idr = cma_pernet_idr(net, ps); 234 235 idr_remove(idr, snum); 236 } 237 238 enum { 239 CMA_OPTION_AFONLY, 240 }; 241 242 void cma_ref_dev(struct cma_device *cma_dev) 243 { 244 atomic_inc(&cma_dev->refcount); 245 } 246 247 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter, 248 void *cookie) 249 { 250 struct cma_device *cma_dev; 251 struct cma_device *found_cma_dev = NULL; 252 253 mutex_lock(&lock); 254 255 list_for_each_entry(cma_dev, &dev_list, list) 256 if (filter(cma_dev->device, cookie)) { 257 found_cma_dev = cma_dev; 258 break; 259 } 260 261 if (found_cma_dev) 262 cma_ref_dev(found_cma_dev); 263 mutex_unlock(&lock); 264 return found_cma_dev; 265 } 266 267 int cma_get_default_gid_type(struct cma_device *cma_dev, 268 unsigned int port) 269 { 270 if (!rdma_is_port_valid(cma_dev->device, port)) 271 return -EINVAL; 272 273 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)]; 274 } 275 276 int cma_set_default_gid_type(struct cma_device *cma_dev, 277 unsigned int port, 278 enum ib_gid_type default_gid_type) 279 { 280 unsigned long supported_gids; 281 282 if (!rdma_is_port_valid(cma_dev->device, port)) 283 return -EINVAL; 284 285 supported_gids = roce_gid_type_mask_support(cma_dev->device, port); 286 287 if (!(supported_gids & 1 << default_gid_type)) 288 return -EINVAL; 289 290 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] = 291 default_gid_type; 292 293 return 0; 294 } 295 296 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev) 297 { 298 return cma_dev->device; 299 } 300 301 /* 302 * Device removal can occur at anytime, so we need extra handling to 303 * serialize notifying the user of device removal with other callbacks. 304 * We do this by disabling removal notification while a callback is in process, 305 * and reporting it after the callback completes. 306 */ 307 struct rdma_id_private { 308 struct rdma_cm_id id; 309 310 struct rdma_bind_list *bind_list; 311 struct hlist_node node; 312 struct list_head list; /* listen_any_list or cma_device.list */ 313 struct list_head listen_list; /* per device listens */ 314 struct cma_device *cma_dev; 315 struct list_head mc_list; 316 317 int internal_id; 318 enum rdma_cm_state state; 319 spinlock_t lock; 320 struct mutex qp_mutex; 321 322 struct completion comp; 323 atomic_t refcount; 324 struct mutex handler_mutex; 325 326 int backlog; 327 int timeout_ms; 328 struct ib_sa_query *query; 329 int query_id; 330 union { 331 struct ib_cm_id *ib; 332 struct iw_cm_id *iw; 333 } cm_id; 334 335 u32 seq_num; 336 u32 qkey; 337 u32 qp_num; 338 pid_t owner; 339 u32 options; 340 u8 srq; 341 u8 tos; 342 u8 timeout_set:1; 343 u8 reuseaddr; 344 u8 afonly; 345 u8 timeout; 346 enum ib_gid_type gid_type; 347 }; 348 349 struct cma_multicast { 350 struct rdma_id_private *id_priv; 351 union { 352 struct ib_sa_multicast *ib; 353 } multicast; 354 struct list_head list; 355 void *context; 356 struct sockaddr_storage addr; 357 struct kref mcref; 358 bool igmp_joined; 359 u8 join_state; 360 }; 361 362 struct cma_work { 363 struct work_struct work; 364 struct rdma_id_private *id; 365 enum rdma_cm_state old_state; 366 enum rdma_cm_state new_state; 367 struct rdma_cm_event event; 368 }; 369 370 struct cma_ndev_work { 371 struct work_struct work; 372 struct rdma_id_private *id; 373 struct rdma_cm_event event; 374 }; 375 376 struct iboe_mcast_work { 377 struct work_struct work; 378 struct rdma_id_private *id; 379 struct cma_multicast *mc; 380 }; 381 382 struct cma_hdr { 383 u8 cma_version; 384 u8 ip_version; /* IP version: 7:4 */ 385 __be16 port; 386 union cma_ip_addr src_addr; 387 union cma_ip_addr dst_addr; 388 }; 389 390 #define CMA_VERSION 0x00 391 #define SDP_MAJ_VERSION 0x2 392 393 struct cma_req_info { 394 struct ib_device *device; 395 int port; 396 union ib_gid local_gid; 397 __be64 service_id; 398 u16 pkey; 399 bool has_gid:1; 400 }; 401 402 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp) 403 { 404 unsigned long flags; 405 int ret; 406 407 spin_lock_irqsave(&id_priv->lock, flags); 408 ret = (id_priv->state == comp); 409 spin_unlock_irqrestore(&id_priv->lock, flags); 410 return ret; 411 } 412 413 static int cma_comp_exch(struct rdma_id_private *id_priv, 414 enum rdma_cm_state comp, enum rdma_cm_state exch) 415 { 416 unsigned long flags; 417 int ret; 418 419 spin_lock_irqsave(&id_priv->lock, flags); 420 if ((ret = (id_priv->state == comp))) 421 id_priv->state = exch; 422 spin_unlock_irqrestore(&id_priv->lock, flags); 423 return ret; 424 } 425 426 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv, 427 enum rdma_cm_state exch) 428 { 429 unsigned long flags; 430 enum rdma_cm_state old; 431 432 spin_lock_irqsave(&id_priv->lock, flags); 433 old = id_priv->state; 434 id_priv->state = exch; 435 spin_unlock_irqrestore(&id_priv->lock, flags); 436 return old; 437 } 438 439 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr) 440 { 441 return hdr->ip_version >> 4; 442 } 443 444 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver) 445 { 446 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF); 447 } 448 449 static inline u8 sdp_get_majv(u8 sdp_version) 450 { 451 return sdp_version >> 4; 452 } 453 454 static inline u8 sdp_get_ip_ver(const struct sdp_hh *hh) 455 { 456 return hh->ipv_cap >> 4; 457 } 458 459 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver) 460 { 461 hh->ipv_cap = (ip_ver << 4) | (hh->ipv_cap & 0xF); 462 } 463 464 static int cma_igmp_send(if_t ndev, const union ib_gid *mgid, bool join) 465 { 466 int retval; 467 468 if (ndev) { 469 union rdma_sockaddr addr; 470 471 rdma_gid2ip(&addr._sockaddr, mgid); 472 473 CURVNET_SET_QUIET(if_getvnet(ndev)); 474 if (join) 475 retval = -if_addmulti(ndev, &addr._sockaddr, NULL); 476 else 477 retval = -if_delmulti(ndev, &addr._sockaddr); 478 CURVNET_RESTORE(); 479 } else { 480 retval = -ENODEV; 481 } 482 return retval; 483 } 484 485 static void _cma_attach_to_dev(struct rdma_id_private *id_priv, 486 struct cma_device *cma_dev) 487 { 488 cma_ref_dev(cma_dev); 489 id_priv->cma_dev = cma_dev; 490 id_priv->gid_type = 0; 491 id_priv->id.device = cma_dev->device; 492 id_priv->id.route.addr.dev_addr.transport = 493 rdma_node_get_transport(cma_dev->device->node_type); 494 list_add_tail(&id_priv->list, &cma_dev->id_list); 495 } 496 497 static void cma_attach_to_dev(struct rdma_id_private *id_priv, 498 struct cma_device *cma_dev) 499 { 500 _cma_attach_to_dev(id_priv, cma_dev); 501 id_priv->gid_type = 502 cma_dev->default_gid_type[id_priv->id.port_num - 503 rdma_start_port(cma_dev->device)]; 504 } 505 506 void cma_deref_dev(struct cma_device *cma_dev) 507 { 508 if (atomic_dec_and_test(&cma_dev->refcount)) 509 complete(&cma_dev->comp); 510 } 511 512 static inline void release_mc(struct kref *kref) 513 { 514 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref); 515 516 kfree(mc->multicast.ib); 517 kfree(mc); 518 } 519 520 static void cma_release_dev(struct rdma_id_private *id_priv) 521 { 522 mutex_lock(&lock); 523 list_del(&id_priv->list); 524 cma_deref_dev(id_priv->cma_dev); 525 id_priv->cma_dev = NULL; 526 mutex_unlock(&lock); 527 } 528 529 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv) 530 { 531 return (struct sockaddr *) &id_priv->id.route.addr.src_addr; 532 } 533 534 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv) 535 { 536 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr; 537 } 538 539 static inline unsigned short cma_family(struct rdma_id_private *id_priv) 540 { 541 return id_priv->id.route.addr.src_addr.ss_family; 542 } 543 544 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey) 545 { 546 struct ib_sa_mcmember_rec rec; 547 int ret = 0; 548 549 if (id_priv->qkey) { 550 if (qkey && id_priv->qkey != qkey) 551 return -EINVAL; 552 return 0; 553 } 554 555 if (qkey) { 556 id_priv->qkey = qkey; 557 return 0; 558 } 559 560 switch (id_priv->id.ps) { 561 case RDMA_PS_UDP: 562 case RDMA_PS_IB: 563 id_priv->qkey = RDMA_UDP_QKEY; 564 break; 565 case RDMA_PS_IPOIB: 566 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid); 567 ret = ib_sa_get_mcmember_rec(id_priv->id.device, 568 id_priv->id.port_num, &rec.mgid, 569 &rec); 570 if (!ret) 571 id_priv->qkey = be32_to_cpu(rec.qkey); 572 break; 573 default: 574 break; 575 } 576 return ret; 577 } 578 579 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr) 580 { 581 dev_addr->dev_type = ARPHRD_INFINIBAND; 582 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr); 583 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey)); 584 } 585 586 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr) 587 { 588 int ret; 589 590 if (addr->sa_family != AF_IB) { 591 ret = rdma_translate_ip(addr, dev_addr); 592 } else { 593 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr); 594 ret = 0; 595 } 596 597 return ret; 598 } 599 600 static inline int cma_validate_port(struct ib_device *device, u8 port, 601 enum ib_gid_type gid_type, 602 union ib_gid *gid, 603 const struct rdma_dev_addr *dev_addr) 604 { 605 const int dev_type = dev_addr->dev_type; 606 if_t ndev; 607 int ret = -ENODEV; 608 609 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port)) 610 return ret; 611 612 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port)) 613 return ret; 614 615 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) { 616 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if); 617 } else { 618 ndev = NULL; 619 gid_type = IB_GID_TYPE_IB; 620 } 621 622 ret = ib_find_cached_gid_by_port(device, gid, gid_type, port, 623 ndev, NULL); 624 625 if (ndev) 626 dev_put(ndev); 627 628 return ret; 629 } 630 631 static int cma_acquire_dev(struct rdma_id_private *id_priv, 632 struct rdma_id_private *listen_id_priv) 633 { 634 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 635 struct cma_device *cma_dev; 636 union ib_gid gid, iboe_gid, *gidp; 637 int ret = -ENODEV; 638 u8 port; 639 640 if (dev_addr->dev_type != ARPHRD_INFINIBAND && 641 id_priv->id.ps == RDMA_PS_IPOIB) 642 return -EINVAL; 643 644 mutex_lock(&lock); 645 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr, 646 &iboe_gid); 647 648 memcpy(&gid, dev_addr->src_dev_addr + 649 rdma_addr_gid_offset(dev_addr), sizeof gid); 650 651 if (listen_id_priv) { 652 cma_dev = listen_id_priv->cma_dev; 653 port = listen_id_priv->id.port_num; 654 655 if (rdma_is_port_valid(cma_dev->device, port)) { 656 gidp = rdma_protocol_roce(cma_dev->device, port) ? 657 &iboe_gid : &gid; 658 659 ret = cma_validate_port(cma_dev->device, port, 660 rdma_protocol_ib(cma_dev->device, port) ? 661 IB_GID_TYPE_IB : 662 listen_id_priv->gid_type, gidp, dev_addr); 663 if (!ret) { 664 id_priv->id.port_num = port; 665 goto out; 666 } 667 } 668 } 669 670 list_for_each_entry(cma_dev, &dev_list, list) { 671 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) { 672 if (listen_id_priv && 673 listen_id_priv->cma_dev == cma_dev && 674 listen_id_priv->id.port_num == port) 675 continue; 676 677 gidp = rdma_protocol_roce(cma_dev->device, port) ? 678 &iboe_gid : &gid; 679 680 ret = cma_validate_port(cma_dev->device, port, 681 rdma_protocol_ib(cma_dev->device, port) ? 682 IB_GID_TYPE_IB : 683 cma_dev->default_gid_type[port - 1], 684 gidp, dev_addr); 685 if (!ret) { 686 id_priv->id.port_num = port; 687 goto out; 688 } 689 } 690 } 691 692 out: 693 if (!ret) 694 cma_attach_to_dev(id_priv, cma_dev); 695 696 mutex_unlock(&lock); 697 return ret; 698 } 699 700 /* 701 * Select the source IB device and address to reach the destination IB address. 702 */ 703 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv) 704 { 705 struct cma_device *cma_dev, *cur_dev; 706 struct sockaddr_ib *addr; 707 union ib_gid gid, sgid, *dgid; 708 u16 pkey, index; 709 u8 p; 710 int i; 711 712 cma_dev = NULL; 713 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv); 714 dgid = (union ib_gid *) &addr->sib_addr; 715 pkey = ntohs(addr->sib_pkey); 716 717 list_for_each_entry(cur_dev, &dev_list, list) { 718 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) { 719 if (!rdma_cap_af_ib(cur_dev->device, p)) 720 continue; 721 722 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index)) 723 continue; 724 725 for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i, 726 &gid, NULL); 727 i++) { 728 if (!memcmp(&gid, dgid, sizeof(gid))) { 729 cma_dev = cur_dev; 730 sgid = gid; 731 id_priv->id.port_num = p; 732 goto found; 733 } 734 735 if (!cma_dev && (gid.global.subnet_prefix == 736 dgid->global.subnet_prefix)) { 737 cma_dev = cur_dev; 738 sgid = gid; 739 id_priv->id.port_num = p; 740 } 741 } 742 } 743 } 744 745 if (!cma_dev) 746 return -ENODEV; 747 748 found: 749 cma_attach_to_dev(id_priv, cma_dev); 750 addr = (struct sockaddr_ib *) cma_src_addr(id_priv); 751 memcpy(&addr->sib_addr, &sgid, sizeof sgid); 752 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr); 753 return 0; 754 } 755 756 static void cma_deref_id(struct rdma_id_private *id_priv) 757 { 758 if (atomic_dec_and_test(&id_priv->refcount)) 759 complete(&id_priv->comp); 760 } 761 762 struct rdma_cm_id *rdma_create_id(struct vnet *net, 763 rdma_cm_event_handler event_handler, 764 void *context, enum rdma_port_space ps, 765 enum ib_qp_type qp_type) 766 { 767 struct rdma_id_private *id_priv; 768 769 #ifdef VIMAGE 770 if (net == NULL) 771 return ERR_PTR(-EINVAL); 772 #endif 773 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL); 774 if (!id_priv) 775 return ERR_PTR(-ENOMEM); 776 777 id_priv->owner = task_pid_nr(current); 778 id_priv->state = RDMA_CM_IDLE; 779 id_priv->id.context = context; 780 id_priv->id.event_handler = event_handler; 781 id_priv->id.ps = ps; 782 id_priv->id.qp_type = qp_type; 783 id_priv->timeout_set = false; 784 spin_lock_init(&id_priv->lock); 785 mutex_init(&id_priv->qp_mutex); 786 init_completion(&id_priv->comp); 787 atomic_set(&id_priv->refcount, 1); 788 mutex_init(&id_priv->handler_mutex); 789 INIT_LIST_HEAD(&id_priv->listen_list); 790 INIT_LIST_HEAD(&id_priv->mc_list); 791 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num); 792 id_priv->id.route.addr.dev_addr.net = net; 793 794 return &id_priv->id; 795 } 796 EXPORT_SYMBOL(rdma_create_id); 797 798 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp) 799 { 800 struct ib_qp_attr qp_attr; 801 int qp_attr_mask, ret; 802 803 qp_attr.qp_state = IB_QPS_INIT; 804 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 805 if (ret) 806 return ret; 807 808 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 809 if (ret) 810 return ret; 811 812 qp_attr.qp_state = IB_QPS_RTR; 813 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE); 814 if (ret) 815 return ret; 816 817 qp_attr.qp_state = IB_QPS_RTS; 818 qp_attr.sq_psn = 0; 819 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN); 820 821 return ret; 822 } 823 824 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp) 825 { 826 struct ib_qp_attr qp_attr; 827 int qp_attr_mask, ret; 828 829 qp_attr.qp_state = IB_QPS_INIT; 830 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 831 if (ret) 832 return ret; 833 834 return ib_modify_qp(qp, &qp_attr, qp_attr_mask); 835 } 836 837 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd, 838 struct ib_qp_init_attr *qp_init_attr) 839 { 840 struct rdma_id_private *id_priv; 841 struct ib_qp *qp; 842 int ret; 843 844 id_priv = container_of(id, struct rdma_id_private, id); 845 if (id->device != pd->device) 846 return -EINVAL; 847 848 qp_init_attr->port_num = id->port_num; 849 qp = ib_create_qp(pd, qp_init_attr); 850 if (IS_ERR(qp)) 851 return PTR_ERR(qp); 852 853 if (id->qp_type == IB_QPT_UD) 854 ret = cma_init_ud_qp(id_priv, qp); 855 else 856 ret = cma_init_conn_qp(id_priv, qp); 857 if (ret) 858 goto err; 859 860 id->qp = qp; 861 id_priv->qp_num = qp->qp_num; 862 id_priv->srq = (qp->srq != NULL); 863 return 0; 864 err: 865 ib_destroy_qp(qp); 866 return ret; 867 } 868 EXPORT_SYMBOL(rdma_create_qp); 869 870 void rdma_destroy_qp(struct rdma_cm_id *id) 871 { 872 struct rdma_id_private *id_priv; 873 874 id_priv = container_of(id, struct rdma_id_private, id); 875 mutex_lock(&id_priv->qp_mutex); 876 ib_destroy_qp(id_priv->id.qp); 877 id_priv->id.qp = NULL; 878 mutex_unlock(&id_priv->qp_mutex); 879 } 880 EXPORT_SYMBOL(rdma_destroy_qp); 881 882 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv, 883 struct rdma_conn_param *conn_param) 884 { 885 struct ib_qp_attr qp_attr; 886 int qp_attr_mask, ret; 887 union ib_gid sgid; 888 889 mutex_lock(&id_priv->qp_mutex); 890 if (!id_priv->id.qp) { 891 ret = 0; 892 goto out; 893 } 894 895 /* Need to update QP attributes from default values. */ 896 qp_attr.qp_state = IB_QPS_INIT; 897 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 898 if (ret) 899 goto out; 900 901 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 902 if (ret) 903 goto out; 904 905 qp_attr.qp_state = IB_QPS_RTR; 906 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 907 if (ret) 908 goto out; 909 910 ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num, 911 qp_attr.ah_attr.grh.sgid_index, &sgid, NULL); 912 if (ret) 913 goto out; 914 915 BUG_ON(id_priv->cma_dev->device != id_priv->id.device); 916 917 if (conn_param) 918 qp_attr.max_dest_rd_atomic = conn_param->responder_resources; 919 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 920 out: 921 mutex_unlock(&id_priv->qp_mutex); 922 return ret; 923 } 924 925 static int cma_modify_qp_rts(struct rdma_id_private *id_priv, 926 struct rdma_conn_param *conn_param) 927 { 928 struct ib_qp_attr qp_attr; 929 int qp_attr_mask, ret; 930 931 mutex_lock(&id_priv->qp_mutex); 932 if (!id_priv->id.qp) { 933 ret = 0; 934 goto out; 935 } 936 937 qp_attr.qp_state = IB_QPS_RTS; 938 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask); 939 if (ret) 940 goto out; 941 942 if (conn_param) 943 qp_attr.max_rd_atomic = conn_param->initiator_depth; 944 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask); 945 out: 946 mutex_unlock(&id_priv->qp_mutex); 947 return ret; 948 } 949 950 static int cma_modify_qp_err(struct rdma_id_private *id_priv) 951 { 952 struct ib_qp_attr qp_attr; 953 int ret; 954 955 mutex_lock(&id_priv->qp_mutex); 956 if (!id_priv->id.qp) { 957 ret = 0; 958 goto out; 959 } 960 961 qp_attr.qp_state = IB_QPS_ERR; 962 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE); 963 out: 964 mutex_unlock(&id_priv->qp_mutex); 965 return ret; 966 } 967 968 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv, 969 struct ib_qp_attr *qp_attr, int *qp_attr_mask) 970 { 971 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 972 int ret; 973 u16 pkey; 974 975 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num)) 976 pkey = 0xffff; 977 else 978 pkey = ib_addr_get_pkey(dev_addr); 979 980 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num, 981 pkey, &qp_attr->pkey_index); 982 if (ret) 983 return ret; 984 985 qp_attr->port_num = id_priv->id.port_num; 986 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT; 987 988 if (id_priv->id.qp_type == IB_QPT_UD) { 989 ret = cma_set_qkey(id_priv, 0); 990 if (ret) 991 return ret; 992 993 qp_attr->qkey = id_priv->qkey; 994 *qp_attr_mask |= IB_QP_QKEY; 995 } else { 996 qp_attr->qp_access_flags = 0; 997 *qp_attr_mask |= IB_QP_ACCESS_FLAGS; 998 } 999 return 0; 1000 } 1001 1002 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr, 1003 int *qp_attr_mask) 1004 { 1005 struct rdma_id_private *id_priv; 1006 int ret = 0; 1007 1008 id_priv = container_of(id, struct rdma_id_private, id); 1009 if (rdma_cap_ib_cm(id->device, id->port_num)) { 1010 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD)) 1011 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask); 1012 else 1013 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr, 1014 qp_attr_mask); 1015 1016 if (qp_attr->qp_state == IB_QPS_RTR) 1017 qp_attr->rq_psn = id_priv->seq_num; 1018 } else if (rdma_cap_iw_cm(id->device, id->port_num)) { 1019 if (!id_priv->cm_id.iw) { 1020 qp_attr->qp_access_flags = 0; 1021 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS; 1022 } else 1023 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr, 1024 qp_attr_mask); 1025 qp_attr->port_num = id_priv->id.port_num; 1026 *qp_attr_mask |= IB_QP_PORT; 1027 } else 1028 ret = -ENOSYS; 1029 1030 if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set) 1031 qp_attr->timeout = id_priv->timeout; 1032 1033 return ret; 1034 } 1035 EXPORT_SYMBOL(rdma_init_qp_attr); 1036 1037 static inline int cma_zero_addr(struct sockaddr *addr) 1038 { 1039 switch (addr->sa_family) { 1040 case AF_INET: 1041 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr); 1042 case AF_INET6: 1043 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr); 1044 case AF_IB: 1045 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr); 1046 default: 1047 return 0; 1048 } 1049 } 1050 1051 static inline int cma_loopback_addr(struct sockaddr *addr) 1052 { 1053 switch (addr->sa_family) { 1054 #ifdef INET 1055 /* 1056 * ipv4_is_loopback() requires an inet variable via vnet, 1057 * not present if INET is not included. 1058 */ 1059 case AF_INET: 1060 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr); 1061 #endif 1062 #ifdef INET6 1063 case AF_INET6: 1064 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr); 1065 #endif 1066 case AF_IB: 1067 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr); 1068 default: 1069 return 0; 1070 } 1071 } 1072 1073 static inline bool cma_any_addr(struct vnet *vnet, struct sockaddr *addr) 1074 { 1075 bool ret; 1076 1077 CURVNET_SET_QUIET(vnet); 1078 ret = cma_zero_addr(addr) || cma_loopback_addr(addr); 1079 CURVNET_RESTORE(); 1080 1081 return (ret); 1082 } 1083 1084 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst) 1085 { 1086 if (src->sa_family != dst->sa_family) 1087 return -1; 1088 1089 switch (src->sa_family) { 1090 case AF_INET: 1091 return ((struct sockaddr_in *) src)->sin_addr.s_addr != 1092 ((struct sockaddr_in *) dst)->sin_addr.s_addr; 1093 case AF_INET6: 1094 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr, 1095 &((struct sockaddr_in6 *) dst)->sin6_addr); 1096 default: 1097 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr, 1098 &((struct sockaddr_ib *) dst)->sib_addr); 1099 } 1100 } 1101 1102 static __be16 cma_port(struct sockaddr *addr) 1103 { 1104 struct sockaddr_ib *sib; 1105 1106 switch (addr->sa_family) { 1107 case AF_INET: 1108 return ((struct sockaddr_in *) addr)->sin_port; 1109 case AF_INET6: 1110 return ((struct sockaddr_in6 *) addr)->sin6_port; 1111 case AF_IB: 1112 sib = (struct sockaddr_ib *) addr; 1113 return htons((u16) (be64_to_cpu(sib->sib_sid) & 1114 be64_to_cpu(sib->sib_sid_mask))); 1115 default: 1116 return 0; 1117 } 1118 } 1119 1120 static inline int cma_any_port(struct sockaddr *addr) 1121 { 1122 return !cma_port(addr); 1123 } 1124 1125 static void cma_save_ib_info(struct sockaddr *src_addr, 1126 struct sockaddr *dst_addr, 1127 struct rdma_cm_id *listen_id, 1128 struct ib_sa_path_rec *path) 1129 { 1130 struct sockaddr_ib *listen_ib, *ib; 1131 1132 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr; 1133 if (src_addr) { 1134 ib = (struct sockaddr_ib *)src_addr; 1135 ib->sib_family = AF_IB; 1136 if (path) { 1137 ib->sib_pkey = path->pkey; 1138 ib->sib_flowinfo = path->flow_label; 1139 memcpy(&ib->sib_addr, &path->sgid, 16); 1140 ib->sib_sid = path->service_id; 1141 ib->sib_scope_id = 0; 1142 } else { 1143 ib->sib_pkey = listen_ib->sib_pkey; 1144 ib->sib_flowinfo = listen_ib->sib_flowinfo; 1145 ib->sib_addr = listen_ib->sib_addr; 1146 ib->sib_sid = listen_ib->sib_sid; 1147 ib->sib_scope_id = listen_ib->sib_scope_id; 1148 } 1149 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL); 1150 } 1151 if (dst_addr) { 1152 ib = (struct sockaddr_ib *)dst_addr; 1153 ib->sib_family = AF_IB; 1154 if (path) { 1155 ib->sib_pkey = path->pkey; 1156 ib->sib_flowinfo = path->flow_label; 1157 memcpy(&ib->sib_addr, &path->dgid, 16); 1158 } 1159 } 1160 } 1161 1162 static void cma_save_ip4_info(struct sockaddr_in *src_addr, 1163 struct sockaddr_in *dst_addr, 1164 struct cma_hdr *hdr, 1165 __be16 local_port) 1166 { 1167 if (src_addr) { 1168 *src_addr = (struct sockaddr_in) { 1169 .sin_len = sizeof(struct sockaddr_in), 1170 .sin_family = AF_INET, 1171 .sin_addr.s_addr = hdr->dst_addr.ip4.addr, 1172 .sin_port = local_port, 1173 }; 1174 } 1175 1176 if (dst_addr) { 1177 *dst_addr = (struct sockaddr_in) { 1178 .sin_len = sizeof(struct sockaddr_in), 1179 .sin_family = AF_INET, 1180 .sin_addr.s_addr = hdr->src_addr.ip4.addr, 1181 .sin_port = hdr->port, 1182 }; 1183 } 1184 } 1185 1186 static void cma_ip6_clear_scope_id(struct in6_addr *addr) 1187 { 1188 /* make sure link local scope ID gets zeroed */ 1189 if (IN6_IS_SCOPE_LINKLOCAL(addr) || 1190 IN6_IS_ADDR_MC_INTFACELOCAL(addr)) { 1191 /* use byte-access to be alignment safe */ 1192 addr->s6_addr[2] = 0; 1193 addr->s6_addr[3] = 0; 1194 } 1195 } 1196 1197 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr, 1198 struct sockaddr_in6 *dst_addr, 1199 struct cma_hdr *hdr, 1200 __be16 local_port) 1201 { 1202 if (src_addr) { 1203 *src_addr = (struct sockaddr_in6) { 1204 .sin6_len = sizeof(struct sockaddr_in6), 1205 .sin6_family = AF_INET6, 1206 .sin6_addr = hdr->dst_addr.ip6, 1207 .sin6_port = local_port, 1208 }; 1209 cma_ip6_clear_scope_id(&src_addr->sin6_addr); 1210 } 1211 1212 if (dst_addr) { 1213 *dst_addr = (struct sockaddr_in6) { 1214 .sin6_len = sizeof(struct sockaddr_in6), 1215 .sin6_family = AF_INET6, 1216 .sin6_addr = hdr->src_addr.ip6, 1217 .sin6_port = hdr->port, 1218 }; 1219 cma_ip6_clear_scope_id(&dst_addr->sin6_addr); 1220 } 1221 } 1222 1223 static u16 cma_port_from_service_id(__be64 service_id) 1224 { 1225 return (u16)be64_to_cpu(service_id); 1226 } 1227 1228 static int sdp_save_ip_info(struct sockaddr *src_addr, 1229 struct sockaddr *dst_addr, 1230 const struct sdp_hh *hdr, 1231 __be64 service_id) 1232 { 1233 __be16 local_port; 1234 1235 BUG_ON(src_addr == NULL || dst_addr == NULL); 1236 1237 if (sdp_get_majv(hdr->majv_minv) != SDP_MAJ_VERSION) 1238 return -EINVAL; 1239 1240 local_port = htons(cma_port_from_service_id(service_id)); 1241 1242 switch (sdp_get_ip_ver(hdr)) { 1243 case 4: { 1244 struct sockaddr_in *s4, *d4; 1245 1246 s4 = (void *)src_addr; 1247 d4 = (void *)dst_addr; 1248 1249 *s4 = (struct sockaddr_in) { 1250 .sin_len = sizeof(*s4), 1251 .sin_family = AF_INET, 1252 .sin_addr.s_addr = hdr->dst_addr.ip4.addr, 1253 .sin_port = local_port, 1254 }; 1255 *d4 = (struct sockaddr_in) { 1256 .sin_len = sizeof(*d4), 1257 .sin_family = AF_INET, 1258 .sin_addr.s_addr = hdr->src_addr.ip4.addr, 1259 .sin_port = hdr->port, 1260 }; 1261 break; 1262 } 1263 case 6: { 1264 struct sockaddr_in6 *s6, *d6; 1265 1266 s6 = (void *)src_addr; 1267 d6 = (void *)dst_addr; 1268 1269 *s6 = (struct sockaddr_in6) { 1270 .sin6_len = sizeof(*s6), 1271 .sin6_family = AF_INET6, 1272 .sin6_addr = hdr->dst_addr.ip6, 1273 .sin6_port = local_port, 1274 }; 1275 *d6 = (struct sockaddr_in6) { 1276 .sin6_len = sizeof(*d6), 1277 .sin6_family = AF_INET6, 1278 .sin6_addr = hdr->src_addr.ip6, 1279 .sin6_port = hdr->port, 1280 }; 1281 cma_ip6_clear_scope_id(&s6->sin6_addr); 1282 cma_ip6_clear_scope_id(&d6->sin6_addr); 1283 break; 1284 } 1285 default: 1286 return -EAFNOSUPPORT; 1287 } 1288 1289 return 0; 1290 } 1291 1292 static int cma_save_ip_info(struct sockaddr *src_addr, 1293 struct sockaddr *dst_addr, 1294 struct ib_cm_event *ib_event, 1295 __be64 service_id) 1296 { 1297 struct cma_hdr *hdr; 1298 __be16 port; 1299 1300 if (rdma_ps_from_service_id(service_id) == RDMA_PS_SDP) 1301 return sdp_save_ip_info(src_addr, dst_addr, 1302 ib_event->private_data, service_id); 1303 1304 hdr = ib_event->private_data; 1305 if (hdr->cma_version != CMA_VERSION) 1306 return -EINVAL; 1307 1308 port = htons(cma_port_from_service_id(service_id)); 1309 1310 switch (cma_get_ip_ver(hdr)) { 1311 case 4: 1312 cma_save_ip4_info((struct sockaddr_in *)src_addr, 1313 (struct sockaddr_in *)dst_addr, hdr, port); 1314 break; 1315 case 6: 1316 cma_save_ip6_info((struct sockaddr_in6 *)src_addr, 1317 (struct sockaddr_in6 *)dst_addr, hdr, port); 1318 break; 1319 default: 1320 return -EAFNOSUPPORT; 1321 } 1322 1323 return 0; 1324 } 1325 1326 static int cma_save_net_info(struct sockaddr *src_addr, 1327 struct sockaddr *dst_addr, 1328 struct rdma_cm_id *listen_id, 1329 struct ib_cm_event *ib_event, 1330 sa_family_t sa_family, __be64 service_id) 1331 { 1332 if (sa_family == AF_IB) { 1333 if (ib_event->event == IB_CM_REQ_RECEIVED) 1334 cma_save_ib_info(src_addr, dst_addr, listen_id, 1335 ib_event->param.req_rcvd.primary_path); 1336 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) 1337 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL); 1338 return 0; 1339 } 1340 1341 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id); 1342 } 1343 1344 static int cma_save_req_info(const struct ib_cm_event *ib_event, 1345 struct cma_req_info *req) 1346 { 1347 const struct ib_cm_req_event_param *req_param = 1348 &ib_event->param.req_rcvd; 1349 const struct ib_cm_sidr_req_event_param *sidr_param = 1350 &ib_event->param.sidr_req_rcvd; 1351 1352 switch (ib_event->event) { 1353 case IB_CM_REQ_RECEIVED: 1354 req->device = req_param->listen_id->device; 1355 req->port = req_param->port; 1356 memcpy(&req->local_gid, &req_param->primary_path->sgid, 1357 sizeof(req->local_gid)); 1358 req->has_gid = true; 1359 req->service_id = req_param->primary_path->service_id; 1360 req->pkey = be16_to_cpu(req_param->primary_path->pkey); 1361 if (req->pkey != req_param->bth_pkey) 1362 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n" 1363 "RDMA CMA: in the future this may cause the request to be dropped\n", 1364 req_param->bth_pkey, req->pkey); 1365 break; 1366 case IB_CM_SIDR_REQ_RECEIVED: 1367 req->device = sidr_param->listen_id->device; 1368 req->port = sidr_param->port; 1369 req->has_gid = false; 1370 req->service_id = sidr_param->service_id; 1371 req->pkey = sidr_param->pkey; 1372 if (req->pkey != sidr_param->bth_pkey) 1373 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n" 1374 "RDMA CMA: in the future this may cause the request to be dropped\n", 1375 sidr_param->bth_pkey, req->pkey); 1376 break; 1377 default: 1378 return -EINVAL; 1379 } 1380 1381 return 0; 1382 } 1383 1384 #ifdef INET 1385 static bool validate_ipv4_net_dev_addr(struct vnet *vnet, 1386 const __be32 saddr, const __be32 daddr) 1387 { 1388 bool ret; 1389 CURVNET_SET(vnet); 1390 ret = ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) || 1391 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) || 1392 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) || 1393 ipv4_is_loopback(saddr); 1394 CURVNET_RESTORE(); 1395 return (ret); 1396 } 1397 #endif 1398 1399 static bool validate_ipv4_net_dev(if_t net_dev, 1400 const struct sockaddr_in *dst_addr, 1401 const struct sockaddr_in *src_addr) 1402 { 1403 #ifdef INET 1404 __be32 daddr = dst_addr->sin_addr.s_addr, 1405 saddr = src_addr->sin_addr.s_addr; 1406 if_t dst_dev; 1407 struct nhop_object *nh; 1408 bool ret; 1409 1410 if (validate_ipv4_net_dev_addr(if_getvnet(net_dev), saddr, daddr)) 1411 return false; 1412 1413 dst_dev = ip_ifp_find(if_getvnet(net_dev), daddr); 1414 if (dst_dev != net_dev) { 1415 if (dst_dev != NULL) 1416 dev_put(dst_dev); 1417 return false; 1418 } 1419 dev_put(dst_dev); 1420 1421 /* 1422 * Check for loopback. 1423 */ 1424 if (saddr == daddr) 1425 return true; 1426 1427 CURVNET_SET(if_getvnet(net_dev)); 1428 nh = fib4_lookup(RT_DEFAULT_FIB, src_addr->sin_addr, 0, NHR_NONE, 0); 1429 if (nh != NULL) 1430 ret = (nh->nh_ifp == net_dev); 1431 else 1432 ret = false; 1433 CURVNET_RESTORE(); 1434 return ret; 1435 #else 1436 return false; 1437 #endif 1438 } 1439 1440 static bool validate_ipv6_net_dev(if_t net_dev, 1441 const struct sockaddr_in6 *dst_addr, 1442 const struct sockaddr_in6 *src_addr) 1443 { 1444 #ifdef INET6 1445 struct sockaddr_in6 src_tmp = *src_addr; 1446 struct sockaddr_in6 dst_tmp = *dst_addr; 1447 if_t dst_dev; 1448 struct nhop_object *nh; 1449 bool ret; 1450 1451 dst_dev = ip6_ifp_find(if_getvnet(net_dev), dst_tmp.sin6_addr, 1452 if_getindex(net_dev)); 1453 if (dst_dev != net_dev) { 1454 if (dst_dev != NULL) 1455 dev_put(dst_dev); 1456 return false; 1457 } 1458 dev_put(dst_dev); 1459 1460 CURVNET_SET(if_getvnet(net_dev)); 1461 1462 /* 1463 * Make sure the scope ID gets embedded. 1464 */ 1465 src_tmp.sin6_scope_id = if_getindex(net_dev); 1466 sa6_embedscope(&src_tmp, 0); 1467 1468 dst_tmp.sin6_scope_id = if_getindex(net_dev); 1469 sa6_embedscope(&dst_tmp, 0); 1470 1471 /* 1472 * Check for loopback after scope ID 1473 * has been embedded: 1474 */ 1475 if (memcmp(&src_tmp.sin6_addr, &dst_tmp.sin6_addr, 1476 sizeof(dst_tmp.sin6_addr)) == 0) { 1477 ret = true; 1478 } else { 1479 /* non-loopback case */ 1480 nh = fib6_lookup(RT_DEFAULT_FIB, &src_addr->sin6_addr, 1481 if_getindex(net_dev), NHR_NONE, 0); 1482 if (nh != NULL) 1483 ret = (nh->nh_ifp == net_dev); 1484 else 1485 ret = false; 1486 } 1487 CURVNET_RESTORE(); 1488 return ret; 1489 #else 1490 return false; 1491 #endif 1492 } 1493 1494 static bool validate_net_dev(if_t net_dev, 1495 const struct sockaddr *daddr, 1496 const struct sockaddr *saddr) 1497 { 1498 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr; 1499 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr; 1500 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr; 1501 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr; 1502 1503 switch (daddr->sa_family) { 1504 case AF_INET: 1505 return saddr->sa_family == AF_INET && 1506 validate_ipv4_net_dev(net_dev, daddr4, saddr4); 1507 1508 case AF_INET6: 1509 return saddr->sa_family == AF_INET6 && 1510 validate_ipv6_net_dev(net_dev, daddr6, saddr6); 1511 1512 default: 1513 return false; 1514 } 1515 } 1516 1517 static if_t 1518 roce_get_net_dev_by_cm_event(struct ib_device *device, u8 port_num, 1519 const struct ib_cm_event *ib_event) 1520 { 1521 struct ib_gid_attr sgid_attr; 1522 union ib_gid sgid; 1523 int err = -EINVAL; 1524 1525 if (ib_event->event == IB_CM_REQ_RECEIVED) { 1526 err = ib_get_cached_gid(device, port_num, 1527 ib_event->param.req_rcvd.ppath_sgid_index, &sgid, &sgid_attr); 1528 } else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) { 1529 err = ib_get_cached_gid(device, port_num, 1530 ib_event->param.sidr_req_rcvd.sgid_index, &sgid, &sgid_attr); 1531 } 1532 if (err) 1533 return (NULL); 1534 return (sgid_attr.ndev); 1535 } 1536 1537 static if_t cma_get_net_dev(struct ib_cm_event *ib_event, 1538 const struct cma_req_info *req) 1539 { 1540 struct sockaddr_storage listen_addr_storage, src_addr_storage; 1541 struct sockaddr *listen_addr = (struct sockaddr *)&listen_addr_storage, 1542 *src_addr = (struct sockaddr *)&src_addr_storage; 1543 if_t net_dev; 1544 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL; 1545 struct epoch_tracker et; 1546 int err; 1547 1548 err = cma_save_ip_info(listen_addr, src_addr, ib_event, 1549 req->service_id); 1550 if (err) 1551 return ERR_PTR(err); 1552 1553 if (rdma_protocol_roce(req->device, req->port)) { 1554 net_dev = roce_get_net_dev_by_cm_event(req->device, req->port, 1555 ib_event); 1556 } else { 1557 net_dev = ib_get_net_dev_by_params(req->device, req->port, 1558 req->pkey, 1559 gid, listen_addr); 1560 } 1561 if (!net_dev) 1562 return ERR_PTR(-ENODEV); 1563 1564 NET_EPOCH_ENTER(et); 1565 if (!validate_net_dev(net_dev, listen_addr, src_addr)) { 1566 NET_EPOCH_EXIT(et); 1567 dev_put(net_dev); 1568 return ERR_PTR(-EHOSTUNREACH); 1569 } 1570 NET_EPOCH_EXIT(et); 1571 1572 return net_dev; 1573 } 1574 1575 static enum rdma_port_space rdma_ps_from_service_id(__be64 service_id) 1576 { 1577 return (be64_to_cpu(service_id) >> 16) & 0xffff; 1578 } 1579 1580 static bool sdp_match_private_data(struct rdma_id_private *id_priv, 1581 const struct sdp_hh *hdr, 1582 struct sockaddr *addr) 1583 { 1584 __be32 ip4_addr; 1585 struct in6_addr ip6_addr; 1586 struct vnet *vnet = id_priv->id.route.addr.dev_addr.net; 1587 1588 switch (addr->sa_family) { 1589 case AF_INET: 1590 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr; 1591 if (sdp_get_ip_ver(hdr) != 4) 1592 return false; 1593 if (!cma_any_addr(vnet, addr) && 1594 hdr->dst_addr.ip4.addr != ip4_addr) 1595 return false; 1596 break; 1597 case AF_INET6: 1598 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr; 1599 if (sdp_get_ip_ver(hdr) != 6) 1600 return false; 1601 cma_ip6_clear_scope_id(&ip6_addr); 1602 if (!cma_any_addr(vnet, addr) && 1603 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr))) 1604 return false; 1605 break; 1606 case AF_IB: 1607 return true; 1608 default: 1609 return false; 1610 } 1611 1612 return true; 1613 } 1614 1615 static bool cma_match_private_data(struct rdma_id_private *id_priv, 1616 const void *vhdr) 1617 { 1618 const struct cma_hdr *hdr = vhdr; 1619 struct sockaddr *addr = cma_src_addr(id_priv); 1620 struct vnet *vnet = id_priv->id.route.addr.dev_addr.net; 1621 __be32 ip4_addr; 1622 struct in6_addr ip6_addr; 1623 1624 if (cma_any_addr(vnet, addr) && !id_priv->afonly) 1625 return true; 1626 1627 if (id_priv->id.ps == RDMA_PS_SDP) 1628 return sdp_match_private_data(id_priv, vhdr, addr); 1629 1630 switch (addr->sa_family) { 1631 case AF_INET: 1632 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr; 1633 if (cma_get_ip_ver(hdr) != 4) 1634 return false; 1635 if (!cma_any_addr(vnet, addr) && 1636 hdr->dst_addr.ip4.addr != ip4_addr) 1637 return false; 1638 break; 1639 case AF_INET6: 1640 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr; 1641 if (cma_get_ip_ver(hdr) != 6) 1642 return false; 1643 cma_ip6_clear_scope_id(&ip6_addr); 1644 if (!cma_any_addr(vnet, addr) && 1645 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr))) 1646 return false; 1647 break; 1648 case AF_IB: 1649 return true; 1650 default: 1651 return false; 1652 } 1653 1654 return true; 1655 } 1656 1657 static bool cma_protocol_roce_dev_port(struct ib_device *device, int port_num) 1658 { 1659 enum rdma_link_layer ll = rdma_port_get_link_layer(device, port_num); 1660 enum rdma_transport_type transport = 1661 rdma_node_get_transport(device->node_type); 1662 1663 return ll == IB_LINK_LAYER_ETHERNET && transport == RDMA_TRANSPORT_IB; 1664 } 1665 1666 static bool cma_protocol_roce(const struct rdma_cm_id *id) 1667 { 1668 struct ib_device *device = id->device; 1669 const int port_num = id->port_num ?: rdma_start_port(device); 1670 1671 return cma_protocol_roce_dev_port(device, port_num); 1672 } 1673 1674 static bool cma_match_net_dev(const struct rdma_cm_id *id, 1675 const if_t net_dev, 1676 u8 port_num) 1677 { 1678 const struct rdma_addr *addr = &id->route.addr; 1679 1680 if (!net_dev) { 1681 if (id->port_num && id->port_num != port_num) 1682 return false; 1683 1684 if (id->ps == RDMA_PS_SDP) { 1685 if (addr->src_addr.ss_family == AF_INET || 1686 addr->src_addr.ss_family == AF_INET6) 1687 return true; 1688 return false; 1689 } 1690 /* This request is an AF_IB request or a RoCE request */ 1691 return addr->src_addr.ss_family == AF_IB || 1692 cma_protocol_roce_dev_port(id->device, port_num); 1693 } 1694 1695 return !addr->dev_addr.bound_dev_if || 1696 (net_eq(dev_net(net_dev), addr->dev_addr.net) && 1697 addr->dev_addr.bound_dev_if == if_getindex(net_dev)); 1698 } 1699 1700 static struct rdma_id_private *cma_find_listener( 1701 const struct rdma_bind_list *bind_list, 1702 const struct ib_cm_id *cm_id, 1703 const struct ib_cm_event *ib_event, 1704 const struct cma_req_info *req, 1705 const if_t net_dev) 1706 { 1707 struct rdma_id_private *id_priv, *id_priv_dev; 1708 1709 if (!bind_list) 1710 return ERR_PTR(-EINVAL); 1711 1712 hlist_for_each_entry(id_priv, &bind_list->owners, node) { 1713 if (cma_match_private_data(id_priv, ib_event->private_data)) { 1714 if (id_priv->id.device == cm_id->device && 1715 cma_match_net_dev(&id_priv->id, net_dev, req->port)) 1716 return id_priv; 1717 list_for_each_entry(id_priv_dev, 1718 &id_priv->listen_list, 1719 listen_list) { 1720 if (id_priv_dev->id.device == cm_id->device && 1721 cma_match_net_dev(&id_priv_dev->id, net_dev, req->port)) 1722 return id_priv_dev; 1723 } 1724 } 1725 } 1726 1727 return ERR_PTR(-EINVAL); 1728 } 1729 1730 static struct rdma_id_private *cma_id_from_event(struct ib_cm_id *cm_id, 1731 struct ib_cm_event *ib_event, 1732 if_t *net_dev) 1733 { 1734 struct cma_req_info req; 1735 struct rdma_bind_list *bind_list; 1736 struct rdma_id_private *id_priv; 1737 int err; 1738 1739 err = cma_save_req_info(ib_event, &req); 1740 if (err) 1741 return ERR_PTR(err); 1742 1743 if (rdma_ps_from_service_id(cm_id->service_id) == RDMA_PS_SDP) { 1744 *net_dev = NULL; 1745 goto there_is_no_net_dev; 1746 } 1747 1748 *net_dev = cma_get_net_dev(ib_event, &req); 1749 if (IS_ERR(*net_dev)) { 1750 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) { 1751 /* Assuming the protocol is AF_IB */ 1752 *net_dev = NULL; 1753 } else { 1754 return ERR_CAST(*net_dev); 1755 } 1756 } 1757 1758 there_is_no_net_dev: 1759 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net, 1760 rdma_ps_from_service_id(req.service_id), 1761 cma_port_from_service_id(req.service_id)); 1762 id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev); 1763 if (IS_ERR(id_priv) && *net_dev) { 1764 dev_put(*net_dev); 1765 *net_dev = NULL; 1766 } 1767 1768 return id_priv; 1769 } 1770 1771 static inline int cma_user_data_offset(struct rdma_id_private *id_priv) 1772 { 1773 if (cma_family(id_priv) == AF_IB) 1774 return 0; 1775 if (id_priv->id.ps == RDMA_PS_SDP) 1776 return 0; 1777 return sizeof(struct cma_hdr); 1778 } 1779 1780 static void cma_cancel_route(struct rdma_id_private *id_priv) 1781 { 1782 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) { 1783 if (id_priv->query) 1784 ib_sa_cancel_query(id_priv->query_id, id_priv->query); 1785 } 1786 } 1787 1788 static void cma_cancel_listens(struct rdma_id_private *id_priv) 1789 { 1790 struct rdma_id_private *dev_id_priv; 1791 1792 /* 1793 * Remove from listen_any_list to prevent added devices from spawning 1794 * additional listen requests. 1795 */ 1796 mutex_lock(&lock); 1797 list_del(&id_priv->list); 1798 1799 while (!list_empty(&id_priv->listen_list)) { 1800 dev_id_priv = list_entry(id_priv->listen_list.next, 1801 struct rdma_id_private, listen_list); 1802 /* sync with device removal to avoid duplicate destruction */ 1803 list_del_init(&dev_id_priv->list); 1804 list_del(&dev_id_priv->listen_list); 1805 mutex_unlock(&lock); 1806 1807 rdma_destroy_id(&dev_id_priv->id); 1808 mutex_lock(&lock); 1809 } 1810 mutex_unlock(&lock); 1811 } 1812 1813 static void cma_cancel_operation(struct rdma_id_private *id_priv, 1814 enum rdma_cm_state state) 1815 { 1816 struct vnet *vnet = id_priv->id.route.addr.dev_addr.net; 1817 1818 switch (state) { 1819 case RDMA_CM_ADDR_QUERY: 1820 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr); 1821 break; 1822 case RDMA_CM_ROUTE_QUERY: 1823 cma_cancel_route(id_priv); 1824 break; 1825 case RDMA_CM_LISTEN: 1826 if (cma_any_addr(vnet, cma_src_addr(id_priv)) && !id_priv->cma_dev) 1827 cma_cancel_listens(id_priv); 1828 break; 1829 default: 1830 break; 1831 } 1832 } 1833 1834 static void cma_release_port(struct rdma_id_private *id_priv) 1835 { 1836 struct rdma_bind_list *bind_list = id_priv->bind_list; 1837 struct vnet *net = id_priv->id.route.addr.dev_addr.net; 1838 1839 if (!bind_list) 1840 return; 1841 1842 mutex_lock(&lock); 1843 hlist_del(&id_priv->node); 1844 if (hlist_empty(&bind_list->owners)) { 1845 cma_ps_remove(net, bind_list->ps, bind_list->port); 1846 kfree(bind_list); 1847 } 1848 mutex_unlock(&lock); 1849 } 1850 1851 static void cma_leave_mc_groups(struct rdma_id_private *id_priv) 1852 { 1853 struct cma_multicast *mc; 1854 1855 while (!list_empty(&id_priv->mc_list)) { 1856 mc = container_of(id_priv->mc_list.next, 1857 struct cma_multicast, list); 1858 list_del(&mc->list); 1859 if (rdma_cap_ib_mcast(id_priv->cma_dev->device, 1860 id_priv->id.port_num)) { 1861 ib_sa_free_multicast(mc->multicast.ib); 1862 kfree(mc); 1863 } else { 1864 if (mc->igmp_joined) { 1865 struct rdma_dev_addr *dev_addr = 1866 &id_priv->id.route.addr.dev_addr; 1867 if_t ndev = NULL; 1868 1869 if (dev_addr->bound_dev_if) 1870 ndev = dev_get_by_index(dev_addr->net, 1871 dev_addr->bound_dev_if); 1872 if (ndev) { 1873 cma_igmp_send(ndev, 1874 &mc->multicast.ib->rec.mgid, 1875 false); 1876 dev_put(ndev); 1877 } 1878 } 1879 kref_put(&mc->mcref, release_mc); 1880 } 1881 } 1882 } 1883 1884 void rdma_destroy_id(struct rdma_cm_id *id) 1885 { 1886 struct rdma_id_private *id_priv; 1887 enum rdma_cm_state state; 1888 1889 id_priv = container_of(id, struct rdma_id_private, id); 1890 state = cma_exch(id_priv, RDMA_CM_DESTROYING); 1891 cma_cancel_operation(id_priv, state); 1892 1893 /* 1894 * Wait for any active callback to finish. New callbacks will find 1895 * the id_priv state set to destroying and abort. 1896 */ 1897 mutex_lock(&id_priv->handler_mutex); 1898 mutex_unlock(&id_priv->handler_mutex); 1899 1900 if (id_priv->cma_dev) { 1901 if (rdma_cap_ib_cm(id_priv->id.device, 1)) { 1902 if (id_priv->cm_id.ib) 1903 ib_destroy_cm_id(id_priv->cm_id.ib); 1904 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) { 1905 if (id_priv->cm_id.iw) 1906 iw_destroy_cm_id(id_priv->cm_id.iw); 1907 } 1908 cma_leave_mc_groups(id_priv); 1909 cma_release_dev(id_priv); 1910 } 1911 1912 cma_release_port(id_priv); 1913 cma_deref_id(id_priv); 1914 wait_for_completion(&id_priv->comp); 1915 1916 if (id_priv->internal_id) 1917 cma_deref_id(id_priv->id.context); 1918 1919 kfree(id_priv->id.route.path_rec); 1920 kfree(id_priv); 1921 } 1922 EXPORT_SYMBOL(rdma_destroy_id); 1923 1924 static int cma_rep_recv(struct rdma_id_private *id_priv) 1925 { 1926 int ret; 1927 1928 ret = cma_modify_qp_rtr(id_priv, NULL); 1929 if (ret) 1930 goto reject; 1931 1932 ret = cma_modify_qp_rts(id_priv, NULL); 1933 if (ret) 1934 goto reject; 1935 1936 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0); 1937 if (ret) 1938 goto reject; 1939 1940 return 0; 1941 reject: 1942 cma_modify_qp_err(id_priv); 1943 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED, 1944 NULL, 0, NULL, 0); 1945 return ret; 1946 } 1947 1948 static int sdp_verify_rep(const struct sdp_hah *data) 1949 { 1950 if (sdp_get_majv(data->majv_minv) != SDP_MAJ_VERSION) 1951 return -EINVAL; 1952 return 0; 1953 } 1954 1955 static void cma_set_rep_event_data(struct rdma_cm_event *event, 1956 struct ib_cm_rep_event_param *rep_data, 1957 void *private_data) 1958 { 1959 event->param.conn.private_data = private_data; 1960 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE; 1961 event->param.conn.responder_resources = rep_data->responder_resources; 1962 event->param.conn.initiator_depth = rep_data->initiator_depth; 1963 event->param.conn.flow_control = rep_data->flow_control; 1964 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count; 1965 event->param.conn.srq = rep_data->srq; 1966 event->param.conn.qp_num = rep_data->remote_qpn; 1967 } 1968 1969 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event) 1970 { 1971 struct rdma_id_private *id_priv = cm_id->context; 1972 struct rdma_cm_event event; 1973 int ret = 0; 1974 1975 mutex_lock(&id_priv->handler_mutex); 1976 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT && 1977 id_priv->state != RDMA_CM_CONNECT) || 1978 (ib_event->event == IB_CM_TIMEWAIT_EXIT && 1979 id_priv->state != RDMA_CM_DISCONNECT)) 1980 goto out; 1981 1982 memset(&event, 0, sizeof event); 1983 switch (ib_event->event) { 1984 case IB_CM_REQ_ERROR: 1985 case IB_CM_REP_ERROR: 1986 event.event = RDMA_CM_EVENT_UNREACHABLE; 1987 event.status = -ETIMEDOUT; 1988 break; 1989 case IB_CM_REP_RECEIVED: 1990 if (id_priv->id.ps == RDMA_PS_SDP) { 1991 event.status = sdp_verify_rep(ib_event->private_data); 1992 if (event.status) 1993 event.event = RDMA_CM_EVENT_CONNECT_ERROR; 1994 else 1995 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE; 1996 } else { 1997 if (id_priv->id.qp) { 1998 event.status = cma_rep_recv(id_priv); 1999 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR : 2000 RDMA_CM_EVENT_ESTABLISHED; 2001 } else { 2002 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE; 2003 } 2004 } 2005 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd, 2006 ib_event->private_data); 2007 break; 2008 case IB_CM_RTU_RECEIVED: 2009 case IB_CM_USER_ESTABLISHED: 2010 event.event = RDMA_CM_EVENT_ESTABLISHED; 2011 break; 2012 case IB_CM_DREQ_ERROR: 2013 event.status = -ETIMEDOUT; /* fall through */ 2014 case IB_CM_DREQ_RECEIVED: 2015 case IB_CM_DREP_RECEIVED: 2016 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT, 2017 RDMA_CM_DISCONNECT)) 2018 goto out; 2019 event.event = RDMA_CM_EVENT_DISCONNECTED; 2020 break; 2021 case IB_CM_TIMEWAIT_EXIT: 2022 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT; 2023 break; 2024 case IB_CM_MRA_RECEIVED: 2025 /* ignore event */ 2026 goto out; 2027 case IB_CM_REJ_RECEIVED: 2028 cma_modify_qp_err(id_priv); 2029 event.status = ib_event->param.rej_rcvd.reason; 2030 event.event = RDMA_CM_EVENT_REJECTED; 2031 event.param.conn.private_data = ib_event->private_data; 2032 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE; 2033 break; 2034 default: 2035 pr_err("RDMA CMA: unexpected IB CM event: %d\n", 2036 ib_event->event); 2037 goto out; 2038 } 2039 2040 ret = id_priv->id.event_handler(&id_priv->id, &event); 2041 if (ret) { 2042 /* Destroy the CM ID by returning a non-zero value. */ 2043 id_priv->cm_id.ib = NULL; 2044 cma_exch(id_priv, RDMA_CM_DESTROYING); 2045 mutex_unlock(&id_priv->handler_mutex); 2046 rdma_destroy_id(&id_priv->id); 2047 return ret; 2048 } 2049 out: 2050 mutex_unlock(&id_priv->handler_mutex); 2051 return ret; 2052 } 2053 2054 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id, 2055 struct ib_cm_event *ib_event, 2056 if_t net_dev) 2057 { 2058 struct rdma_id_private *id_priv; 2059 struct rdma_cm_id *id; 2060 struct rdma_route *rt; 2061 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family; 2062 const __be64 service_id = 2063 ib_event->param.req_rcvd.primary_path->service_id; 2064 struct vnet *vnet = listen_id->route.addr.dev_addr.net; 2065 int ret; 2066 2067 id = rdma_create_id(vnet, 2068 listen_id->event_handler, listen_id->context, 2069 listen_id->ps, ib_event->param.req_rcvd.qp_type); 2070 if (IS_ERR(id)) 2071 return NULL; 2072 2073 id_priv = container_of(id, struct rdma_id_private, id); 2074 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr, 2075 (struct sockaddr *)&id->route.addr.dst_addr, 2076 listen_id, ib_event, ss_family, service_id)) 2077 goto err; 2078 2079 rt = &id->route; 2080 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1; 2081 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths, 2082 GFP_KERNEL); 2083 if (!rt->path_rec) 2084 goto err; 2085 2086 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path; 2087 if (rt->num_paths == 2) 2088 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path; 2089 2090 if (net_dev) { 2091 ret = rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL); 2092 if (ret) 2093 goto err; 2094 } else { 2095 if (!cma_protocol_roce(listen_id) && 2096 cma_any_addr(vnet, cma_src_addr(id_priv))) { 2097 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND; 2098 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid); 2099 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey)); 2100 } else if (!cma_any_addr(vnet, cma_src_addr(id_priv))) { 2101 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr); 2102 if (ret) 2103 goto err; 2104 } 2105 } 2106 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid); 2107 2108 id_priv->state = RDMA_CM_CONNECT; 2109 return id_priv; 2110 2111 err: 2112 rdma_destroy_id(id); 2113 return NULL; 2114 } 2115 2116 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id, 2117 struct ib_cm_event *ib_event, 2118 if_t net_dev) 2119 { 2120 struct rdma_id_private *id_priv; 2121 struct rdma_cm_id *id; 2122 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family; 2123 struct vnet *vnet = listen_id->route.addr.dev_addr.net; 2124 int ret; 2125 2126 id = rdma_create_id(vnet, listen_id->event_handler, listen_id->context, 2127 listen_id->ps, IB_QPT_UD); 2128 if (IS_ERR(id)) 2129 return NULL; 2130 2131 id_priv = container_of(id, struct rdma_id_private, id); 2132 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr, 2133 (struct sockaddr *)&id->route.addr.dst_addr, 2134 listen_id, ib_event, ss_family, 2135 ib_event->param.sidr_req_rcvd.service_id)) 2136 goto err; 2137 2138 if (net_dev) { 2139 ret = rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL); 2140 if (ret) 2141 goto err; 2142 } else { 2143 if (!cma_any_addr(vnet, cma_src_addr(id_priv))) { 2144 ret = cma_translate_addr(cma_src_addr(id_priv), 2145 &id->route.addr.dev_addr); 2146 if (ret) 2147 goto err; 2148 } 2149 } 2150 2151 id_priv->state = RDMA_CM_CONNECT; 2152 return id_priv; 2153 err: 2154 rdma_destroy_id(id); 2155 return NULL; 2156 } 2157 2158 static void cma_set_req_event_data(struct rdma_cm_event *event, 2159 struct ib_cm_req_event_param *req_data, 2160 void *private_data, int offset) 2161 { 2162 event->param.conn.private_data = (char *)private_data + offset; 2163 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset; 2164 event->param.conn.responder_resources = req_data->responder_resources; 2165 event->param.conn.initiator_depth = req_data->initiator_depth; 2166 event->param.conn.flow_control = req_data->flow_control; 2167 event->param.conn.retry_count = req_data->retry_count; 2168 event->param.conn.rnr_retry_count = req_data->rnr_retry_count; 2169 event->param.conn.srq = req_data->srq; 2170 event->param.conn.qp_num = req_data->remote_qpn; 2171 } 2172 2173 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event) 2174 { 2175 return (((ib_event->event == IB_CM_REQ_RECEIVED) && 2176 (ib_event->param.req_rcvd.qp_type == id->qp_type)) || 2177 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) && 2178 (id->qp_type == IB_QPT_UD)) || 2179 (!id->qp_type)); 2180 } 2181 2182 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event) 2183 { 2184 struct rdma_id_private *listen_id, *conn_id = NULL; 2185 struct rdma_cm_event event; 2186 if_t net_dev; 2187 int offset, ret; 2188 2189 listen_id = cma_id_from_event(cm_id, ib_event, &net_dev); 2190 if (IS_ERR(listen_id)) 2191 return PTR_ERR(listen_id); 2192 2193 if (!cma_check_req_qp_type(&listen_id->id, ib_event)) { 2194 ret = -EINVAL; 2195 goto net_dev_put; 2196 } 2197 2198 mutex_lock(&listen_id->handler_mutex); 2199 if (listen_id->state != RDMA_CM_LISTEN) { 2200 ret = -ECONNABORTED; 2201 goto err1; 2202 } 2203 2204 memset(&event, 0, sizeof event); 2205 offset = cma_user_data_offset(listen_id); 2206 event.event = RDMA_CM_EVENT_CONNECT_REQUEST; 2207 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) { 2208 conn_id = cma_new_udp_id(&listen_id->id, ib_event, net_dev); 2209 event.param.ud.private_data = (char *)ib_event->private_data + offset; 2210 event.param.ud.private_data_len = 2211 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset; 2212 } else { 2213 conn_id = cma_new_conn_id(&listen_id->id, ib_event, net_dev); 2214 cma_set_req_event_data(&event, &ib_event->param.req_rcvd, 2215 ib_event->private_data, offset); 2216 } 2217 if (!conn_id) { 2218 ret = -ENOMEM; 2219 goto err1; 2220 } 2221 2222 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING); 2223 ret = cma_acquire_dev(conn_id, listen_id); 2224 if (ret) 2225 goto err2; 2226 2227 conn_id->cm_id.ib = cm_id; 2228 cm_id->context = conn_id; 2229 cm_id->cm_handler = cma_ib_handler; 2230 2231 /* 2232 * Protect against the user destroying conn_id from another thread 2233 * until we're done accessing it. 2234 */ 2235 atomic_inc(&conn_id->refcount); 2236 ret = conn_id->id.event_handler(&conn_id->id, &event); 2237 if (ret) 2238 goto err3; 2239 /* 2240 * Acquire mutex to prevent user executing rdma_destroy_id() 2241 * while we're accessing the cm_id. 2242 */ 2243 mutex_lock(&lock); 2244 if (cma_comp(conn_id, RDMA_CM_CONNECT) && 2245 (conn_id->id.qp_type != IB_QPT_UD)) 2246 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0); 2247 mutex_unlock(&lock); 2248 mutex_unlock(&conn_id->handler_mutex); 2249 mutex_unlock(&listen_id->handler_mutex); 2250 cma_deref_id(conn_id); 2251 if (net_dev) 2252 dev_put(net_dev); 2253 return 0; 2254 2255 err3: 2256 cma_deref_id(conn_id); 2257 /* Destroy the CM ID by returning a non-zero value. */ 2258 conn_id->cm_id.ib = NULL; 2259 err2: 2260 cma_exch(conn_id, RDMA_CM_DESTROYING); 2261 mutex_unlock(&conn_id->handler_mutex); 2262 err1: 2263 mutex_unlock(&listen_id->handler_mutex); 2264 if (conn_id) 2265 rdma_destroy_id(&conn_id->id); 2266 2267 net_dev_put: 2268 if (net_dev) 2269 dev_put(net_dev); 2270 2271 return ret; 2272 } 2273 2274 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr) 2275 { 2276 if (addr->sa_family == AF_IB) 2277 return ((struct sockaddr_ib *) addr)->sib_sid; 2278 2279 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr))); 2280 } 2281 EXPORT_SYMBOL(rdma_get_service_id); 2282 2283 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event) 2284 { 2285 struct rdma_id_private *id_priv = iw_id->context; 2286 struct rdma_cm_event event; 2287 int ret = 0; 2288 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr; 2289 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr; 2290 2291 mutex_lock(&id_priv->handler_mutex); 2292 if (id_priv->state != RDMA_CM_CONNECT) 2293 goto out; 2294 2295 memset(&event, 0, sizeof event); 2296 switch (iw_event->event) { 2297 case IW_CM_EVENT_CLOSE: 2298 event.event = RDMA_CM_EVENT_DISCONNECTED; 2299 break; 2300 case IW_CM_EVENT_CONNECT_REPLY: 2301 memcpy(cma_src_addr(id_priv), laddr, 2302 rdma_addr_size(laddr)); 2303 memcpy(cma_dst_addr(id_priv), raddr, 2304 rdma_addr_size(raddr)); 2305 switch (iw_event->status) { 2306 case 0: 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 case -ECONNRESET: 2312 case -ECONNREFUSED: 2313 event.event = RDMA_CM_EVENT_REJECTED; 2314 break; 2315 case -ETIMEDOUT: 2316 event.event = RDMA_CM_EVENT_UNREACHABLE; 2317 break; 2318 default: 2319 event.event = RDMA_CM_EVENT_CONNECT_ERROR; 2320 break; 2321 } 2322 break; 2323 case IW_CM_EVENT_ESTABLISHED: 2324 event.event = RDMA_CM_EVENT_ESTABLISHED; 2325 event.param.conn.initiator_depth = iw_event->ird; 2326 event.param.conn.responder_resources = iw_event->ord; 2327 break; 2328 default: 2329 BUG_ON(1); 2330 } 2331 2332 event.status = iw_event->status; 2333 event.param.conn.private_data = iw_event->private_data; 2334 event.param.conn.private_data_len = iw_event->private_data_len; 2335 ret = id_priv->id.event_handler(&id_priv->id, &event); 2336 if (ret) { 2337 /* Destroy the CM ID by returning a non-zero value. */ 2338 id_priv->cm_id.iw = NULL; 2339 cma_exch(id_priv, RDMA_CM_DESTROYING); 2340 mutex_unlock(&id_priv->handler_mutex); 2341 rdma_destroy_id(&id_priv->id); 2342 return ret; 2343 } 2344 2345 out: 2346 mutex_unlock(&id_priv->handler_mutex); 2347 return ret; 2348 } 2349 2350 static int iw_conn_req_handler(struct iw_cm_id *cm_id, 2351 struct iw_cm_event *iw_event) 2352 { 2353 struct rdma_cm_id *new_cm_id; 2354 struct rdma_id_private *listen_id, *conn_id; 2355 struct rdma_cm_event event; 2356 int ret = -ECONNABORTED; 2357 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr; 2358 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr; 2359 2360 listen_id = cm_id->context; 2361 2362 mutex_lock(&listen_id->handler_mutex); 2363 if (listen_id->state != RDMA_CM_LISTEN) 2364 goto out; 2365 2366 /* Create a new RDMA id for the new IW CM ID */ 2367 new_cm_id = rdma_create_id(listen_id->id.route.addr.dev_addr.net, 2368 listen_id->id.event_handler, 2369 listen_id->id.context, 2370 RDMA_PS_TCP, IB_QPT_RC); 2371 if (IS_ERR(new_cm_id)) { 2372 ret = -ENOMEM; 2373 goto out; 2374 } 2375 conn_id = container_of(new_cm_id, struct rdma_id_private, id); 2376 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING); 2377 conn_id->state = RDMA_CM_CONNECT; 2378 2379 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr); 2380 if (ret) { 2381 mutex_unlock(&conn_id->handler_mutex); 2382 rdma_destroy_id(new_cm_id); 2383 goto out; 2384 } 2385 2386 ret = cma_acquire_dev(conn_id, listen_id); 2387 if (ret) { 2388 mutex_unlock(&conn_id->handler_mutex); 2389 rdma_destroy_id(new_cm_id); 2390 goto out; 2391 } 2392 2393 conn_id->cm_id.iw = cm_id; 2394 cm_id->context = conn_id; 2395 cm_id->cm_handler = cma_iw_handler; 2396 2397 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr)); 2398 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr)); 2399 2400 memset(&event, 0, sizeof event); 2401 event.event = RDMA_CM_EVENT_CONNECT_REQUEST; 2402 event.param.conn.private_data = iw_event->private_data; 2403 event.param.conn.private_data_len = iw_event->private_data_len; 2404 event.param.conn.initiator_depth = iw_event->ird; 2405 event.param.conn.responder_resources = iw_event->ord; 2406 2407 /* 2408 * Protect against the user destroying conn_id from another thread 2409 * until we're done accessing it. 2410 */ 2411 atomic_inc(&conn_id->refcount); 2412 ret = conn_id->id.event_handler(&conn_id->id, &event); 2413 if (ret) { 2414 /* User wants to destroy the CM ID */ 2415 conn_id->cm_id.iw = NULL; 2416 cma_exch(conn_id, RDMA_CM_DESTROYING); 2417 mutex_unlock(&conn_id->handler_mutex); 2418 cma_deref_id(conn_id); 2419 rdma_destroy_id(&conn_id->id); 2420 goto out; 2421 } 2422 2423 mutex_unlock(&conn_id->handler_mutex); 2424 cma_deref_id(conn_id); 2425 2426 out: 2427 mutex_unlock(&listen_id->handler_mutex); 2428 return ret; 2429 } 2430 2431 static int cma_ib_listen(struct rdma_id_private *id_priv) 2432 { 2433 struct sockaddr *addr; 2434 struct ib_cm_id *id; 2435 __be64 svc_id; 2436 2437 addr = cma_src_addr(id_priv); 2438 svc_id = rdma_get_service_id(&id_priv->id, addr); 2439 id = ib_cm_insert_listen(id_priv->id.device, cma_req_handler, svc_id); 2440 if (IS_ERR(id)) 2441 return PTR_ERR(id); 2442 id_priv->cm_id.ib = id; 2443 2444 return 0; 2445 } 2446 2447 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog) 2448 { 2449 int ret; 2450 struct iw_cm_id *id; 2451 2452 id = iw_create_cm_id(id_priv->id.device, 2453 iw_conn_req_handler, 2454 id_priv); 2455 if (IS_ERR(id)) 2456 return PTR_ERR(id); 2457 2458 id->tos = id_priv->tos; 2459 id_priv->cm_id.iw = id; 2460 2461 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv), 2462 rdma_addr_size(cma_src_addr(id_priv))); 2463 2464 ret = iw_cm_listen(id_priv->cm_id.iw, backlog); 2465 2466 if (ret) { 2467 iw_destroy_cm_id(id_priv->cm_id.iw); 2468 id_priv->cm_id.iw = NULL; 2469 } 2470 2471 return ret; 2472 } 2473 2474 static int cma_listen_handler(struct rdma_cm_id *id, 2475 struct rdma_cm_event *event) 2476 { 2477 struct rdma_id_private *id_priv = id->context; 2478 2479 id->context = id_priv->id.context; 2480 id->event_handler = id_priv->id.event_handler; 2481 return id_priv->id.event_handler(id, event); 2482 } 2483 2484 static void cma_listen_on_dev(struct rdma_id_private *id_priv, 2485 struct cma_device *cma_dev) 2486 { 2487 struct rdma_id_private *dev_id_priv; 2488 struct rdma_cm_id *id; 2489 struct vnet *net = id_priv->id.route.addr.dev_addr.net; 2490 int ret; 2491 2492 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1)) 2493 return; 2494 2495 id = rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps, 2496 id_priv->id.qp_type); 2497 if (IS_ERR(id)) 2498 return; 2499 2500 dev_id_priv = container_of(id, struct rdma_id_private, id); 2501 2502 dev_id_priv->state = RDMA_CM_ADDR_BOUND; 2503 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv), 2504 rdma_addr_size(cma_src_addr(id_priv))); 2505 2506 _cma_attach_to_dev(dev_id_priv, cma_dev); 2507 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list); 2508 atomic_inc(&id_priv->refcount); 2509 dev_id_priv->internal_id = 1; 2510 dev_id_priv->afonly = id_priv->afonly; 2511 2512 ret = rdma_listen(id, id_priv->backlog); 2513 if (ret) 2514 pr_warn("RDMA CMA: cma_listen_on_dev, error %d, listening on device %s\n", 2515 ret, cma_dev->device->name); 2516 } 2517 2518 static void cma_listen_on_all(struct rdma_id_private *id_priv) 2519 { 2520 struct cma_device *cma_dev; 2521 2522 mutex_lock(&lock); 2523 list_add_tail(&id_priv->list, &listen_any_list); 2524 list_for_each_entry(cma_dev, &dev_list, list) 2525 cma_listen_on_dev(id_priv, cma_dev); 2526 mutex_unlock(&lock); 2527 } 2528 2529 void rdma_set_service_type(struct rdma_cm_id *id, int tos) 2530 { 2531 struct rdma_id_private *id_priv; 2532 2533 id_priv = container_of(id, struct rdma_id_private, id); 2534 id_priv->tos = (u8) tos; 2535 } 2536 EXPORT_SYMBOL(rdma_set_service_type); 2537 2538 /** 2539 * rdma_set_ack_timeout() - Set the ack timeout of QP associated 2540 * with a connection identifier. 2541 * @id: Communication identifier to associated with service type. 2542 * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec. 2543 * 2544 * This function should be called before rdma_connect() on active side, 2545 * and on passive side before rdma_accept(). It is applicable to primary 2546 * path only. The timeout will affect the local side of the QP, it is not 2547 * negotiated with remote side and zero disables the timer. 2548 * 2549 * Return: 0 for success 2550 */ 2551 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout) 2552 { 2553 struct rdma_id_private *id_priv; 2554 2555 if (id->qp_type != IB_QPT_RC) 2556 return -EINVAL; 2557 2558 id_priv = container_of(id, struct rdma_id_private, id); 2559 id_priv->timeout = timeout; 2560 id_priv->timeout_set = true; 2561 2562 return 0; 2563 } 2564 EXPORT_SYMBOL(rdma_set_ack_timeout); 2565 2566 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec, 2567 void *context) 2568 { 2569 struct cma_work *work = context; 2570 struct rdma_route *route; 2571 2572 route = &work->id->id.route; 2573 2574 if (!status) { 2575 route->num_paths = 1; 2576 *route->path_rec = *path_rec; 2577 } else { 2578 work->old_state = RDMA_CM_ROUTE_QUERY; 2579 work->new_state = RDMA_CM_ADDR_RESOLVED; 2580 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR; 2581 work->event.status = status; 2582 } 2583 2584 queue_work(cma_wq, &work->work); 2585 } 2586 2587 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms, 2588 struct cma_work *work) 2589 { 2590 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 2591 struct ib_sa_path_rec path_rec; 2592 ib_sa_comp_mask comp_mask; 2593 struct sockaddr_in6 *sin6; 2594 struct sockaddr_ib *sib; 2595 2596 memset(&path_rec, 0, sizeof path_rec); 2597 rdma_addr_get_sgid(dev_addr, &path_rec.sgid); 2598 rdma_addr_get_dgid(dev_addr, &path_rec.dgid); 2599 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr)); 2600 path_rec.numb_path = 1; 2601 path_rec.reversible = 1; 2602 path_rec.service_id = rdma_get_service_id(&id_priv->id, 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 (id_priv->id.event_handler(&id_priv->id, &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 int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms) 2657 { 2658 struct rdma_route *route = &id_priv->id.route; 2659 struct cma_work *work; 2660 int ret; 2661 2662 work = kzalloc(sizeof *work, GFP_KERNEL); 2663 if (!work) 2664 return -ENOMEM; 2665 2666 work->id = id_priv; 2667 INIT_WORK(&work->work, cma_work_handler); 2668 work->old_state = RDMA_CM_ROUTE_QUERY; 2669 work->new_state = RDMA_CM_ROUTE_RESOLVED; 2670 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 2671 2672 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL); 2673 if (!route->path_rec) { 2674 ret = -ENOMEM; 2675 goto err1; 2676 } 2677 2678 ret = cma_query_ib_route(id_priv, timeout_ms, work); 2679 if (ret) 2680 goto err2; 2681 2682 return 0; 2683 err2: 2684 kfree(route->path_rec); 2685 route->path_rec = NULL; 2686 err1: 2687 kfree(work); 2688 return ret; 2689 } 2690 2691 int rdma_set_ib_paths(struct rdma_cm_id *id, 2692 struct ib_sa_path_rec *path_rec, int num_paths) 2693 { 2694 struct rdma_id_private *id_priv; 2695 int ret; 2696 2697 id_priv = container_of(id, struct rdma_id_private, id); 2698 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, 2699 RDMA_CM_ROUTE_RESOLVED)) 2700 return -EINVAL; 2701 2702 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths, 2703 GFP_KERNEL); 2704 if (!id->route.path_rec) { 2705 ret = -ENOMEM; 2706 goto err; 2707 } 2708 2709 id->route.num_paths = num_paths; 2710 return 0; 2711 err: 2712 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED); 2713 return ret; 2714 } 2715 EXPORT_SYMBOL(rdma_set_ib_paths); 2716 2717 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms) 2718 { 2719 struct cma_work *work; 2720 2721 work = kzalloc(sizeof *work, GFP_KERNEL); 2722 if (!work) 2723 return -ENOMEM; 2724 2725 work->id = id_priv; 2726 INIT_WORK(&work->work, cma_work_handler); 2727 work->old_state = RDMA_CM_ROUTE_QUERY; 2728 work->new_state = RDMA_CM_ROUTE_RESOLVED; 2729 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 2730 queue_work(cma_wq, &work->work); 2731 return 0; 2732 } 2733 2734 static int iboe_tos_to_sl(if_t ndev, int tos) 2735 { 2736 /* get service level, SL, from IPv4 type of service, TOS */ 2737 int sl = (tos >> 5) & 0x7; 2738 2739 /* final mappings are done by the vendor specific drivers */ 2740 return sl; 2741 } 2742 2743 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type, 2744 unsigned long supported_gids, 2745 enum ib_gid_type default_gid) 2746 { 2747 if ((network_type == RDMA_NETWORK_IPV4 || 2748 network_type == RDMA_NETWORK_IPV6) && 2749 test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids)) 2750 return IB_GID_TYPE_ROCE_UDP_ENCAP; 2751 2752 return default_gid; 2753 } 2754 2755 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv) 2756 { 2757 struct rdma_route *route = &id_priv->id.route; 2758 struct rdma_addr *addr = &route->addr; 2759 struct cma_work *work; 2760 int ret; 2761 if_t ndev = NULL; 2762 2763 2764 work = kzalloc(sizeof *work, GFP_KERNEL); 2765 if (!work) 2766 return -ENOMEM; 2767 2768 work->id = id_priv; 2769 INIT_WORK(&work->work, cma_work_handler); 2770 2771 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL); 2772 if (!route->path_rec) { 2773 ret = -ENOMEM; 2774 goto err1; 2775 } 2776 2777 route->num_paths = 1; 2778 2779 if (addr->dev_addr.bound_dev_if) { 2780 unsigned long supported_gids; 2781 2782 ndev = dev_get_by_index(addr->dev_addr.net, 2783 addr->dev_addr.bound_dev_if); 2784 if (!ndev) { 2785 ret = -ENODEV; 2786 goto err2; 2787 } 2788 2789 route->path_rec->net = if_getvnet(ndev); 2790 route->path_rec->ifindex = if_getindex(ndev); 2791 supported_gids = roce_gid_type_mask_support(id_priv->id.device, 2792 id_priv->id.port_num); 2793 route->path_rec->gid_type = 2794 cma_route_gid_type(addr->dev_addr.network, 2795 supported_gids, 2796 id_priv->gid_type); 2797 } 2798 if (!ndev) { 2799 ret = -ENODEV; 2800 goto err2; 2801 } 2802 2803 memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN); 2804 2805 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr, 2806 &route->path_rec->sgid); 2807 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr, 2808 &route->path_rec->dgid); 2809 2810 /* Use the hint from IP Stack to select GID Type */ 2811 if (route->path_rec->gid_type < ib_network_to_gid_type(addr->dev_addr.network)) 2812 route->path_rec->gid_type = ib_network_to_gid_type(addr->dev_addr.network); 2813 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB) 2814 /* TODO: get the hoplimit from the inet/inet6 device */ 2815 route->path_rec->hop_limit = addr->dev_addr.hoplimit; 2816 else 2817 route->path_rec->hop_limit = 1; 2818 route->path_rec->reversible = 1; 2819 route->path_rec->pkey = cpu_to_be16(0xffff); 2820 route->path_rec->mtu_selector = IB_SA_EQ; 2821 route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos); 2822 route->path_rec->traffic_class = id_priv->tos; 2823 route->path_rec->mtu = iboe_get_mtu(if_getmtu(ndev)); 2824 route->path_rec->rate_selector = IB_SA_EQ; 2825 route->path_rec->rate = iboe_get_rate(ndev); 2826 dev_put(ndev); 2827 route->path_rec->packet_life_time_selector = IB_SA_EQ; 2828 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME; 2829 if (!route->path_rec->mtu) { 2830 ret = -EINVAL; 2831 goto err2; 2832 } 2833 2834 work->old_state = RDMA_CM_ROUTE_QUERY; 2835 work->new_state = RDMA_CM_ROUTE_RESOLVED; 2836 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 2837 work->event.status = 0; 2838 2839 queue_work(cma_wq, &work->work); 2840 2841 return 0; 2842 2843 err2: 2844 kfree(route->path_rec); 2845 route->path_rec = NULL; 2846 err1: 2847 kfree(work); 2848 return ret; 2849 } 2850 2851 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms) 2852 { 2853 struct rdma_id_private *id_priv; 2854 int ret; 2855 2856 id_priv = container_of(id, struct rdma_id_private, id); 2857 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY)) 2858 return -EINVAL; 2859 2860 atomic_inc(&id_priv->refcount); 2861 if (rdma_cap_ib_sa(id->device, id->port_num)) 2862 ret = cma_resolve_ib_route(id_priv, timeout_ms); 2863 else if (rdma_protocol_roce(id->device, id->port_num)) 2864 ret = cma_resolve_iboe_route(id_priv); 2865 else if (rdma_protocol_iwarp(id->device, id->port_num)) 2866 ret = cma_resolve_iw_route(id_priv, timeout_ms); 2867 else 2868 ret = -ENOSYS; 2869 2870 if (ret) 2871 goto err; 2872 2873 return 0; 2874 err: 2875 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED); 2876 cma_deref_id(id_priv); 2877 return ret; 2878 } 2879 EXPORT_SYMBOL(rdma_resolve_route); 2880 2881 static void cma_set_loopback(struct sockaddr *addr) 2882 { 2883 switch (addr->sa_family) { 2884 case AF_INET: 2885 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK); 2886 break; 2887 case AF_INET6: 2888 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr, 2889 0, 0, 0, htonl(1)); 2890 break; 2891 default: 2892 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr, 2893 0, 0, 0, htonl(1)); 2894 break; 2895 } 2896 } 2897 2898 static int cma_bind_loopback(struct rdma_id_private *id_priv) 2899 { 2900 struct cma_device *cma_dev, *cur_dev; 2901 struct ib_port_attr port_attr; 2902 union ib_gid gid; 2903 u16 pkey; 2904 int ret; 2905 u8 p; 2906 2907 cma_dev = NULL; 2908 mutex_lock(&lock); 2909 list_for_each_entry(cur_dev, &dev_list, list) { 2910 if (cma_family(id_priv) == AF_IB && 2911 !rdma_cap_ib_cm(cur_dev->device, 1)) 2912 continue; 2913 2914 if (!cma_dev) 2915 cma_dev = cur_dev; 2916 2917 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) { 2918 if (!ib_query_port(cur_dev->device, p, &port_attr) && 2919 port_attr.state == IB_PORT_ACTIVE) { 2920 cma_dev = cur_dev; 2921 goto port_found; 2922 } 2923 } 2924 } 2925 2926 if (!cma_dev) { 2927 ret = -ENODEV; 2928 goto out; 2929 } 2930 2931 p = 1; 2932 2933 port_found: 2934 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid, NULL); 2935 if (ret) 2936 goto out; 2937 2938 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey); 2939 if (ret) 2940 goto out; 2941 2942 id_priv->id.route.addr.dev_addr.dev_type = 2943 (rdma_protocol_ib(cma_dev->device, p)) ? 2944 ARPHRD_INFINIBAND : ARPHRD_ETHER; 2945 2946 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid); 2947 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey); 2948 id_priv->id.port_num = p; 2949 cma_attach_to_dev(id_priv, cma_dev); 2950 cma_set_loopback(cma_src_addr(id_priv)); 2951 out: 2952 mutex_unlock(&lock); 2953 return ret; 2954 } 2955 2956 static void addr_handler(int status, struct sockaddr *src_addr, 2957 struct rdma_dev_addr *dev_addr, void *context) 2958 { 2959 struct rdma_id_private *id_priv = context; 2960 struct rdma_cm_event event; 2961 2962 memset(&event, 0, sizeof event); 2963 mutex_lock(&id_priv->handler_mutex); 2964 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, 2965 RDMA_CM_ADDR_RESOLVED)) 2966 goto out; 2967 2968 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr)); 2969 if (!status && !id_priv->cma_dev) 2970 status = cma_acquire_dev(id_priv, NULL); 2971 2972 if (status) { 2973 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, 2974 RDMA_CM_ADDR_BOUND)) 2975 goto out; 2976 event.event = RDMA_CM_EVENT_ADDR_ERROR; 2977 event.status = status; 2978 } else 2979 event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 2980 2981 if (id_priv->id.event_handler(&id_priv->id, &event)) { 2982 cma_exch(id_priv, RDMA_CM_DESTROYING); 2983 mutex_unlock(&id_priv->handler_mutex); 2984 cma_deref_id(id_priv); 2985 rdma_destroy_id(&id_priv->id); 2986 return; 2987 } 2988 out: 2989 mutex_unlock(&id_priv->handler_mutex); 2990 cma_deref_id(id_priv); 2991 } 2992 2993 static int cma_resolve_loopback(struct rdma_id_private *id_priv) 2994 { 2995 struct cma_work *work; 2996 union ib_gid gid; 2997 int ret; 2998 2999 work = kzalloc(sizeof *work, GFP_KERNEL); 3000 if (!work) 3001 return -ENOMEM; 3002 3003 if (!id_priv->cma_dev) { 3004 ret = cma_bind_loopback(id_priv); 3005 if (ret) 3006 goto err; 3007 } 3008 3009 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid); 3010 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid); 3011 3012 work->id = id_priv; 3013 INIT_WORK(&work->work, cma_work_handler); 3014 work->old_state = RDMA_CM_ADDR_QUERY; 3015 work->new_state = RDMA_CM_ADDR_RESOLVED; 3016 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 3017 queue_work(cma_wq, &work->work); 3018 return 0; 3019 err: 3020 kfree(work); 3021 return ret; 3022 } 3023 3024 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv) 3025 { 3026 struct cma_work *work; 3027 int ret; 3028 3029 work = kzalloc(sizeof *work, GFP_KERNEL); 3030 if (!work) 3031 return -ENOMEM; 3032 3033 if (!id_priv->cma_dev) { 3034 ret = cma_resolve_ib_dev(id_priv); 3035 if (ret) 3036 goto err; 3037 } 3038 3039 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *) 3040 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr)); 3041 3042 work->id = id_priv; 3043 INIT_WORK(&work->work, cma_work_handler); 3044 work->old_state = RDMA_CM_ADDR_QUERY; 3045 work->new_state = RDMA_CM_ADDR_RESOLVED; 3046 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED; 3047 queue_work(cma_wq, &work->work); 3048 return 0; 3049 err: 3050 kfree(work); 3051 return ret; 3052 } 3053 3054 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr, 3055 struct sockaddr *dst_addr) 3056 { 3057 if (!src_addr || !src_addr->sa_family) { 3058 src_addr = (struct sockaddr *) &id->route.addr.src_addr; 3059 src_addr->sa_family = dst_addr->sa_family; 3060 if (dst_addr->sa_family == AF_INET6) { 3061 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr; 3062 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr; 3063 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id; 3064 if (IN6_IS_SCOPE_LINKLOCAL(&dst_addr6->sin6_addr) || 3065 IN6_IS_ADDR_MC_INTFACELOCAL(&dst_addr6->sin6_addr)) 3066 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id; 3067 } else if (dst_addr->sa_family == AF_IB) { 3068 ((struct sockaddr_ib *) src_addr)->sib_pkey = 3069 ((struct sockaddr_ib *) dst_addr)->sib_pkey; 3070 } 3071 } 3072 return rdma_bind_addr(id, src_addr); 3073 } 3074 3075 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr, 3076 struct sockaddr *dst_addr, int timeout_ms) 3077 { 3078 struct rdma_id_private *id_priv; 3079 struct vnet *vnet = id->route.addr.dev_addr.net; 3080 int ret; 3081 3082 id_priv = container_of(id, struct rdma_id_private, id); 3083 if (id_priv->state == RDMA_CM_IDLE) { 3084 ret = cma_bind_addr(id, src_addr, dst_addr); 3085 if (ret) 3086 return ret; 3087 } 3088 3089 if (cma_family(id_priv) != dst_addr->sa_family) 3090 return -EINVAL; 3091 3092 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) 3093 return -EINVAL; 3094 3095 atomic_inc(&id_priv->refcount); 3096 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr)); 3097 if (cma_any_addr(vnet, dst_addr)) { 3098 ret = cma_resolve_loopback(id_priv); 3099 } else { 3100 if (dst_addr->sa_family == AF_IB) { 3101 ret = cma_resolve_ib_addr(id_priv); 3102 } else { 3103 ret = cma_check_linklocal(&id->route.addr.dev_addr, dst_addr); 3104 if (ret) 3105 goto err; 3106 3107 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv), 3108 dst_addr, &id->route.addr.dev_addr, 3109 timeout_ms, addr_handler, id_priv); 3110 } 3111 } 3112 if (ret) 3113 goto err; 3114 3115 return 0; 3116 err: 3117 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND); 3118 cma_deref_id(id_priv); 3119 return ret; 3120 } 3121 EXPORT_SYMBOL(rdma_resolve_addr); 3122 3123 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse) 3124 { 3125 struct rdma_id_private *id_priv; 3126 unsigned long flags; 3127 int ret; 3128 3129 id_priv = container_of(id, struct rdma_id_private, id); 3130 spin_lock_irqsave(&id_priv->lock, flags); 3131 if (reuse || id_priv->state == RDMA_CM_IDLE) { 3132 id_priv->reuseaddr = reuse; 3133 ret = 0; 3134 } else { 3135 ret = -EINVAL; 3136 } 3137 spin_unlock_irqrestore(&id_priv->lock, flags); 3138 return ret; 3139 } 3140 EXPORT_SYMBOL(rdma_set_reuseaddr); 3141 3142 int rdma_set_afonly(struct rdma_cm_id *id, int afonly) 3143 { 3144 struct rdma_id_private *id_priv; 3145 unsigned long flags; 3146 int ret; 3147 3148 id_priv = container_of(id, struct rdma_id_private, id); 3149 spin_lock_irqsave(&id_priv->lock, flags); 3150 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) { 3151 id_priv->options |= (1 << CMA_OPTION_AFONLY); 3152 id_priv->afonly = afonly; 3153 ret = 0; 3154 } else { 3155 ret = -EINVAL; 3156 } 3157 spin_unlock_irqrestore(&id_priv->lock, flags); 3158 return ret; 3159 } 3160 EXPORT_SYMBOL(rdma_set_afonly); 3161 3162 static void cma_bind_port(struct rdma_bind_list *bind_list, 3163 struct rdma_id_private *id_priv) 3164 { 3165 struct sockaddr *addr; 3166 struct sockaddr_ib *sib; 3167 u64 sid, mask; 3168 __be16 port; 3169 3170 addr = cma_src_addr(id_priv); 3171 port = htons(bind_list->port); 3172 3173 switch (addr->sa_family) { 3174 case AF_INET: 3175 ((struct sockaddr_in *) addr)->sin_port = port; 3176 break; 3177 case AF_INET6: 3178 ((struct sockaddr_in6 *) addr)->sin6_port = port; 3179 break; 3180 case AF_IB: 3181 sib = (struct sockaddr_ib *) addr; 3182 sid = be64_to_cpu(sib->sib_sid); 3183 mask = be64_to_cpu(sib->sib_sid_mask); 3184 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port)); 3185 sib->sib_sid_mask = cpu_to_be64(~0ULL); 3186 break; 3187 } 3188 id_priv->bind_list = bind_list; 3189 hlist_add_head(&id_priv->node, &bind_list->owners); 3190 } 3191 3192 static int cma_alloc_port(enum rdma_port_space ps, 3193 struct rdma_id_private *id_priv, unsigned short snum) 3194 { 3195 struct rdma_bind_list *bind_list; 3196 int ret; 3197 3198 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL); 3199 if (!bind_list) 3200 return -ENOMEM; 3201 3202 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list, 3203 snum); 3204 if (ret < 0) 3205 goto err; 3206 3207 bind_list->ps = ps; 3208 bind_list->port = (unsigned short)ret; 3209 cma_bind_port(bind_list, id_priv); 3210 return 0; 3211 err: 3212 kfree(bind_list); 3213 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret; 3214 } 3215 3216 static int cma_alloc_any_port(enum rdma_port_space ps, 3217 struct rdma_id_private *id_priv) 3218 { 3219 static unsigned int last_used_port; 3220 int low, high, remaining; 3221 unsigned int rover; 3222 struct vnet *net = id_priv->id.route.addr.dev_addr.net; 3223 u32 rand; 3224 3225 inet_get_local_port_range(net, &low, &high); 3226 remaining = (high - low) + 1; 3227 get_random_bytes(&rand, sizeof(rand)); 3228 rover = rand % remaining + low; 3229 retry: 3230 if (last_used_port != rover && 3231 !cma_ps_find(net, ps, (unsigned short)rover)) { 3232 int ret = cma_alloc_port(ps, id_priv, rover); 3233 /* 3234 * Remember previously used port number in order to avoid 3235 * re-using same port immediately after it is closed. 3236 */ 3237 if (!ret) 3238 last_used_port = rover; 3239 if (ret != -EADDRNOTAVAIL) 3240 return ret; 3241 } 3242 if (--remaining) { 3243 rover++; 3244 if ((rover < low) || (rover > high)) 3245 rover = low; 3246 goto retry; 3247 } 3248 return -EADDRNOTAVAIL; 3249 } 3250 3251 /* 3252 * Check that the requested port is available. This is called when trying to 3253 * bind to a specific port, or when trying to listen on a bound port. In 3254 * the latter case, the provided id_priv may already be on the bind_list, but 3255 * we still need to check that it's okay to start listening. 3256 */ 3257 static int cma_check_port(struct rdma_bind_list *bind_list, 3258 struct rdma_id_private *id_priv, uint8_t reuseaddr) 3259 { 3260 struct rdma_id_private *cur_id; 3261 struct sockaddr *addr, *cur_addr; 3262 struct vnet *vnet; 3263 3264 addr = cma_src_addr(id_priv); 3265 hlist_for_each_entry(cur_id, &bind_list->owners, node) { 3266 if (id_priv == cur_id) 3267 continue; 3268 3269 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr && 3270 cur_id->reuseaddr) 3271 continue; 3272 3273 cur_addr = cma_src_addr(cur_id); 3274 if (id_priv->afonly && cur_id->afonly && 3275 (addr->sa_family != cur_addr->sa_family)) 3276 continue; 3277 3278 vnet = cur_id->id.route.addr.dev_addr.net; 3279 if (cma_any_addr(vnet, addr) || cma_any_addr(vnet, cur_addr)) 3280 return -EADDRNOTAVAIL; 3281 3282 if (!cma_addr_cmp(addr, cur_addr)) 3283 return -EADDRINUSE; 3284 } 3285 return 0; 3286 } 3287 3288 static int cma_use_port(enum rdma_port_space ps, 3289 struct rdma_id_private *id_priv) 3290 { 3291 struct rdma_bind_list *bind_list; 3292 unsigned short snum; 3293 int ret; 3294 3295 snum = ntohs(cma_port(cma_src_addr(id_priv))); 3296 if (snum < IPPORT_RESERVED && 3297 priv_check(curthread, PRIV_NETINET_BINDANY) != 0) 3298 return -EACCES; 3299 3300 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum); 3301 if (!bind_list) { 3302 ret = cma_alloc_port(ps, id_priv, snum); 3303 } else { 3304 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr); 3305 if (!ret) 3306 cma_bind_port(bind_list, id_priv); 3307 } 3308 return ret; 3309 } 3310 3311 static int cma_bind_listen(struct rdma_id_private *id_priv) 3312 { 3313 struct rdma_bind_list *bind_list = id_priv->bind_list; 3314 int ret = 0; 3315 3316 mutex_lock(&lock); 3317 if (bind_list->owners.first->next) 3318 ret = cma_check_port(bind_list, id_priv, 0); 3319 mutex_unlock(&lock); 3320 return ret; 3321 } 3322 3323 static enum rdma_port_space cma_select_inet_ps( 3324 struct rdma_id_private *id_priv) 3325 { 3326 switch (id_priv->id.ps) { 3327 case RDMA_PS_TCP: 3328 case RDMA_PS_UDP: 3329 case RDMA_PS_IPOIB: 3330 case RDMA_PS_IB: 3331 case RDMA_PS_SDP: 3332 return id_priv->id.ps; 3333 default: 3334 3335 return 0; 3336 } 3337 } 3338 3339 static enum rdma_port_space cma_select_ib_ps(struct rdma_id_private *id_priv) 3340 { 3341 enum rdma_port_space ps = 0; 3342 struct sockaddr_ib *sib; 3343 u64 sid_ps, mask, sid; 3344 3345 sib = (struct sockaddr_ib *) cma_src_addr(id_priv); 3346 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK; 3347 sid = be64_to_cpu(sib->sib_sid) & mask; 3348 3349 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) { 3350 sid_ps = RDMA_IB_IP_PS_IB; 3351 ps = RDMA_PS_IB; 3352 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) && 3353 (sid == (RDMA_IB_IP_PS_TCP & mask))) { 3354 sid_ps = RDMA_IB_IP_PS_TCP; 3355 ps = RDMA_PS_TCP; 3356 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) && 3357 (sid == (RDMA_IB_IP_PS_UDP & mask))) { 3358 sid_ps = RDMA_IB_IP_PS_UDP; 3359 ps = RDMA_PS_UDP; 3360 } 3361 3362 if (ps) { 3363 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib))); 3364 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK | 3365 be64_to_cpu(sib->sib_sid_mask)); 3366 } 3367 return ps; 3368 } 3369 3370 static int cma_get_port(struct rdma_id_private *id_priv) 3371 { 3372 enum rdma_port_space ps; 3373 int ret; 3374 3375 if (cma_family(id_priv) != AF_IB) 3376 ps = cma_select_inet_ps(id_priv); 3377 else 3378 ps = cma_select_ib_ps(id_priv); 3379 if (!ps) 3380 return -EPROTONOSUPPORT; 3381 3382 mutex_lock(&lock); 3383 if (cma_any_port(cma_src_addr(id_priv))) 3384 ret = cma_alloc_any_port(ps, id_priv); 3385 else 3386 ret = cma_use_port(ps, id_priv); 3387 mutex_unlock(&lock); 3388 3389 return ret; 3390 } 3391 3392 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr, 3393 struct sockaddr *addr) 3394 { 3395 #ifdef INET6 3396 struct sockaddr_in6 sin6; 3397 3398 if (addr->sa_family != AF_INET6) 3399 return 0; 3400 3401 sin6 = *(struct sockaddr_in6 *)addr; 3402 3403 if (IN6_IS_SCOPE_LINKLOCAL(&sin6.sin6_addr) || 3404 IN6_IS_ADDR_MC_INTFACELOCAL(&sin6.sin6_addr)) { 3405 bool failure; 3406 3407 CURVNET_SET_QUIET(dev_addr->net); 3408 failure = sa6_recoverscope(&sin6) || sin6.sin6_scope_id == 0; 3409 CURVNET_RESTORE(); 3410 3411 /* check if IPv6 scope ID is not set */ 3412 if (failure) 3413 return -EINVAL; 3414 dev_addr->bound_dev_if = sin6.sin6_scope_id; 3415 } 3416 #endif 3417 return 0; 3418 } 3419 3420 int rdma_listen(struct rdma_cm_id *id, int backlog) 3421 { 3422 struct rdma_id_private *id_priv; 3423 int ret; 3424 3425 id_priv = container_of(id, struct rdma_id_private, id); 3426 if (id_priv->state == RDMA_CM_IDLE) { 3427 id->route.addr.src_addr.ss_family = AF_INET; 3428 ret = rdma_bind_addr(id, cma_src_addr(id_priv)); 3429 if (ret) 3430 return ret; 3431 } 3432 3433 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) 3434 return -EINVAL; 3435 3436 if (id_priv->reuseaddr) { 3437 ret = cma_bind_listen(id_priv); 3438 if (ret) 3439 goto err; 3440 } 3441 3442 id_priv->backlog = backlog; 3443 if (id->device) { 3444 if (rdma_cap_ib_cm(id->device, 1)) { 3445 ret = cma_ib_listen(id_priv); 3446 if (ret) 3447 goto err; 3448 } else if (rdma_cap_iw_cm(id->device, 1)) { 3449 ret = cma_iw_listen(id_priv, backlog); 3450 if (ret) 3451 goto err; 3452 } else { 3453 ret = -ENOSYS; 3454 goto err; 3455 } 3456 } else 3457 cma_listen_on_all(id_priv); 3458 3459 return 0; 3460 err: 3461 id_priv->backlog = 0; 3462 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND); 3463 return ret; 3464 } 3465 EXPORT_SYMBOL(rdma_listen); 3466 3467 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr) 3468 { 3469 struct rdma_id_private *id_priv; 3470 struct vnet *vnet = id->route.addr.dev_addr.net; 3471 int ret; 3472 3473 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 && 3474 addr->sa_family != AF_IB) 3475 return -EAFNOSUPPORT; 3476 3477 id_priv = container_of(id, struct rdma_id_private, id); 3478 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND)) 3479 return -EINVAL; 3480 3481 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr); 3482 if (ret) 3483 goto err1; 3484 3485 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr)); 3486 if (!cma_any_addr(vnet, addr)) { 3487 ret = cma_translate_addr(addr, &id->route.addr.dev_addr); 3488 if (ret) 3489 goto err1; 3490 3491 ret = cma_acquire_dev(id_priv, NULL); 3492 if (ret) 3493 goto err1; 3494 } 3495 3496 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) { 3497 if (addr->sa_family == AF_INET) 3498 id_priv->afonly = 1; 3499 #ifdef INET6 3500 else if (addr->sa_family == AF_INET6) { 3501 CURVNET_SET_QUIET(vnet); 3502 id_priv->afonly = V_ip6_v6only; 3503 CURVNET_RESTORE(); 3504 } 3505 #endif 3506 } 3507 ret = cma_get_port(id_priv); 3508 if (ret) 3509 goto err2; 3510 3511 return 0; 3512 err2: 3513 if (id_priv->cma_dev) 3514 cma_release_dev(id_priv); 3515 err1: 3516 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE); 3517 return ret; 3518 } 3519 EXPORT_SYMBOL(rdma_bind_addr); 3520 3521 static int sdp_format_hdr(struct sdp_hh *sdp_hdr, struct rdma_id_private *id_priv) 3522 { 3523 /* 3524 * XXXCEM: CMA just sets the version itself rather than relying on 3525 * passed in packet to have the major version set. Should we? 3526 */ 3527 if (sdp_get_majv(sdp_hdr->majv_minv) != SDP_MAJ_VERSION) 3528 return -EINVAL; 3529 3530 if (cma_family(id_priv) == AF_INET) { 3531 struct sockaddr_in *src4, *dst4; 3532 3533 src4 = (struct sockaddr_in *) cma_src_addr(id_priv); 3534 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv); 3535 3536 sdp_set_ip_ver(sdp_hdr, 4); 3537 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr; 3538 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr; 3539 sdp_hdr->port = src4->sin_port; 3540 } else if (cma_family(id_priv) == AF_INET6) { 3541 struct sockaddr_in6 *src6, *dst6; 3542 3543 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv); 3544 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv); 3545 3546 sdp_set_ip_ver(sdp_hdr, 6); 3547 sdp_hdr->src_addr.ip6 = src6->sin6_addr; 3548 sdp_hdr->dst_addr.ip6 = dst6->sin6_addr; 3549 sdp_hdr->port = src6->sin6_port; 3550 cma_ip6_clear_scope_id(&sdp_hdr->src_addr.ip6); 3551 cma_ip6_clear_scope_id(&sdp_hdr->dst_addr.ip6); 3552 } else 3553 return -EAFNOSUPPORT; 3554 return 0; 3555 } 3556 3557 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv) 3558 { 3559 struct cma_hdr *cma_hdr; 3560 3561 if (id_priv->id.ps == RDMA_PS_SDP) 3562 return sdp_format_hdr(hdr, id_priv); 3563 3564 cma_hdr = hdr; 3565 cma_hdr->cma_version = CMA_VERSION; 3566 if (cma_family(id_priv) == AF_INET) { 3567 struct sockaddr_in *src4, *dst4; 3568 3569 src4 = (struct sockaddr_in *) cma_src_addr(id_priv); 3570 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv); 3571 3572 cma_set_ip_ver(cma_hdr, 4); 3573 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr; 3574 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr; 3575 cma_hdr->port = src4->sin_port; 3576 } else if (cma_family(id_priv) == AF_INET6) { 3577 struct sockaddr_in6 *src6, *dst6; 3578 3579 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv); 3580 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv); 3581 3582 cma_set_ip_ver(cma_hdr, 6); 3583 cma_hdr->src_addr.ip6 = src6->sin6_addr; 3584 cma_hdr->dst_addr.ip6 = dst6->sin6_addr; 3585 cma_hdr->port = src6->sin6_port; 3586 cma_ip6_clear_scope_id(&cma_hdr->src_addr.ip6); 3587 cma_ip6_clear_scope_id(&cma_hdr->dst_addr.ip6); 3588 } 3589 return 0; 3590 } 3591 3592 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id, 3593 struct ib_cm_event *ib_event) 3594 { 3595 struct rdma_id_private *id_priv = cm_id->context; 3596 struct rdma_cm_event event; 3597 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd; 3598 int ret = 0; 3599 3600 mutex_lock(&id_priv->handler_mutex); 3601 if (id_priv->state != RDMA_CM_CONNECT) 3602 goto out; 3603 3604 memset(&event, 0, sizeof event); 3605 switch (ib_event->event) { 3606 case IB_CM_SIDR_REQ_ERROR: 3607 event.event = RDMA_CM_EVENT_UNREACHABLE; 3608 event.status = -ETIMEDOUT; 3609 break; 3610 case IB_CM_SIDR_REP_RECEIVED: 3611 event.param.ud.private_data = ib_event->private_data; 3612 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE; 3613 if (rep->status != IB_SIDR_SUCCESS) { 3614 event.event = RDMA_CM_EVENT_UNREACHABLE; 3615 event.status = ib_event->param.sidr_rep_rcvd.status; 3616 break; 3617 } 3618 ret = cma_set_qkey(id_priv, rep->qkey); 3619 if (ret) { 3620 event.event = RDMA_CM_EVENT_ADDR_ERROR; 3621 event.status = ret; 3622 break; 3623 } 3624 ret = ib_init_ah_from_path(id_priv->id.device, 3625 id_priv->id.port_num, 3626 id_priv->id.route.path_rec, 3627 &event.param.ud.ah_attr); 3628 if (ret) { 3629 event.event = RDMA_CM_EVENT_ADDR_ERROR; 3630 event.status = ret; 3631 break; 3632 } 3633 event.param.ud.qp_num = rep->qpn; 3634 event.param.ud.qkey = rep->qkey; 3635 event.event = RDMA_CM_EVENT_ESTABLISHED; 3636 event.status = 0; 3637 break; 3638 default: 3639 pr_err("RDMA CMA: unexpected IB CM event: %d\n", 3640 ib_event->event); 3641 goto out; 3642 } 3643 3644 ret = id_priv->id.event_handler(&id_priv->id, &event); 3645 if (ret) { 3646 /* Destroy the CM ID by returning a non-zero value. */ 3647 id_priv->cm_id.ib = NULL; 3648 cma_exch(id_priv, RDMA_CM_DESTROYING); 3649 mutex_unlock(&id_priv->handler_mutex); 3650 rdma_destroy_id(&id_priv->id); 3651 return ret; 3652 } 3653 out: 3654 mutex_unlock(&id_priv->handler_mutex); 3655 return ret; 3656 } 3657 3658 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv, 3659 struct rdma_conn_param *conn_param) 3660 { 3661 struct ib_cm_sidr_req_param req; 3662 struct ib_cm_id *id; 3663 void *private_data; 3664 int offset, ret; 3665 3666 memset(&req, 0, sizeof req); 3667 offset = cma_user_data_offset(id_priv); 3668 req.private_data_len = offset + conn_param->private_data_len; 3669 if (req.private_data_len < conn_param->private_data_len) 3670 return -EINVAL; 3671 3672 if (req.private_data_len) { 3673 private_data = kzalloc(req.private_data_len, GFP_ATOMIC); 3674 if (!private_data) 3675 return -ENOMEM; 3676 } else { 3677 private_data = NULL; 3678 } 3679 3680 if (conn_param->private_data && conn_param->private_data_len) 3681 memcpy((char *)private_data + offset, conn_param->private_data, 3682 conn_param->private_data_len); 3683 3684 if (private_data) { 3685 ret = cma_format_hdr(private_data, id_priv); 3686 if (ret) 3687 goto out; 3688 req.private_data = private_data; 3689 } 3690 3691 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler, 3692 id_priv); 3693 if (IS_ERR(id)) { 3694 ret = PTR_ERR(id); 3695 goto out; 3696 } 3697 id_priv->cm_id.ib = id; 3698 3699 req.path = id_priv->id.route.path_rec; 3700 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv)); 3701 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8); 3702 req.max_cm_retries = CMA_MAX_CM_RETRIES; 3703 3704 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req); 3705 if (ret) { 3706 ib_destroy_cm_id(id_priv->cm_id.ib); 3707 id_priv->cm_id.ib = NULL; 3708 } 3709 out: 3710 kfree(private_data); 3711 return ret; 3712 } 3713 3714 static int cma_connect_ib(struct rdma_id_private *id_priv, 3715 struct rdma_conn_param *conn_param) 3716 { 3717 struct ib_cm_req_param req; 3718 struct rdma_route *route; 3719 void *private_data; 3720 struct ib_cm_id *id; 3721 int offset, ret; 3722 3723 memset(&req, 0, sizeof req); 3724 offset = cma_user_data_offset(id_priv); 3725 req.private_data_len = offset + conn_param->private_data_len; 3726 if (req.private_data_len < conn_param->private_data_len) 3727 return -EINVAL; 3728 3729 if (req.private_data_len) { 3730 private_data = kzalloc(req.private_data_len, GFP_ATOMIC); 3731 if (!private_data) 3732 return -ENOMEM; 3733 } else { 3734 private_data = NULL; 3735 } 3736 3737 if (conn_param->private_data && conn_param->private_data_len) 3738 memcpy((char *)private_data + offset, conn_param->private_data, 3739 conn_param->private_data_len); 3740 3741 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv); 3742 if (IS_ERR(id)) { 3743 ret = PTR_ERR(id); 3744 goto out; 3745 } 3746 id_priv->cm_id.ib = id; 3747 3748 route = &id_priv->id.route; 3749 if (private_data) { 3750 ret = cma_format_hdr(private_data, id_priv); 3751 if (ret) 3752 goto out; 3753 req.private_data = private_data; 3754 } 3755 3756 req.primary_path = &route->path_rec[0]; 3757 if (route->num_paths == 2) 3758 req.alternate_path = &route->path_rec[1]; 3759 3760 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv)); 3761 req.qp_num = id_priv->qp_num; 3762 req.qp_type = id_priv->id.qp_type; 3763 req.starting_psn = id_priv->seq_num; 3764 req.responder_resources = conn_param->responder_resources; 3765 req.initiator_depth = conn_param->initiator_depth; 3766 req.flow_control = conn_param->flow_control; 3767 req.retry_count = min_t(u8, 7, conn_param->retry_count); 3768 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count); 3769 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT; 3770 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT; 3771 req.max_cm_retries = CMA_MAX_CM_RETRIES; 3772 req.srq = id_priv->srq ? 1 : 0; 3773 3774 ret = ib_send_cm_req(id_priv->cm_id.ib, &req); 3775 out: 3776 if (ret && !IS_ERR(id)) { 3777 ib_destroy_cm_id(id); 3778 id_priv->cm_id.ib = NULL; 3779 } 3780 3781 kfree(private_data); 3782 return ret; 3783 } 3784 3785 static int cma_connect_iw(struct rdma_id_private *id_priv, 3786 struct rdma_conn_param *conn_param) 3787 { 3788 struct iw_cm_id *cm_id; 3789 int ret; 3790 struct iw_cm_conn_param iw_param; 3791 3792 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv); 3793 if (IS_ERR(cm_id)) 3794 return PTR_ERR(cm_id); 3795 3796 cm_id->tos = id_priv->tos; 3797 id_priv->cm_id.iw = cm_id; 3798 3799 memcpy(&cm_id->local_addr, cma_src_addr(id_priv), 3800 rdma_addr_size(cma_src_addr(id_priv))); 3801 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv), 3802 rdma_addr_size(cma_dst_addr(id_priv))); 3803 3804 ret = cma_modify_qp_rtr(id_priv, conn_param); 3805 if (ret) 3806 goto out; 3807 3808 if (conn_param) { 3809 iw_param.ord = conn_param->initiator_depth; 3810 iw_param.ird = conn_param->responder_resources; 3811 iw_param.private_data = conn_param->private_data; 3812 iw_param.private_data_len = conn_param->private_data_len; 3813 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num; 3814 } else { 3815 memset(&iw_param, 0, sizeof iw_param); 3816 iw_param.qpn = id_priv->qp_num; 3817 } 3818 ret = iw_cm_connect(cm_id, &iw_param); 3819 out: 3820 if (ret) { 3821 iw_destroy_cm_id(cm_id); 3822 id_priv->cm_id.iw = NULL; 3823 } 3824 return ret; 3825 } 3826 3827 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param) 3828 { 3829 struct rdma_id_private *id_priv; 3830 int ret; 3831 3832 id_priv = container_of(id, struct rdma_id_private, id); 3833 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT)) 3834 return -EINVAL; 3835 3836 if (!id->qp) { 3837 id_priv->qp_num = conn_param->qp_num; 3838 id_priv->srq = conn_param->srq; 3839 } 3840 3841 if (rdma_cap_ib_cm(id->device, id->port_num)) { 3842 if (id->qp_type == IB_QPT_UD) 3843 ret = cma_resolve_ib_udp(id_priv, conn_param); 3844 else 3845 ret = cma_connect_ib(id_priv, conn_param); 3846 } else if (rdma_cap_iw_cm(id->device, id->port_num)) 3847 ret = cma_connect_iw(id_priv, conn_param); 3848 else 3849 ret = -ENOSYS; 3850 if (ret) 3851 goto err; 3852 3853 return 0; 3854 err: 3855 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED); 3856 return ret; 3857 } 3858 EXPORT_SYMBOL(rdma_connect); 3859 3860 static int cma_accept_ib(struct rdma_id_private *id_priv, 3861 struct rdma_conn_param *conn_param) 3862 { 3863 struct ib_cm_rep_param rep; 3864 int ret; 3865 3866 ret = cma_modify_qp_rtr(id_priv, conn_param); 3867 if (ret) 3868 goto out; 3869 3870 ret = cma_modify_qp_rts(id_priv, conn_param); 3871 if (ret) 3872 goto out; 3873 3874 memset(&rep, 0, sizeof rep); 3875 rep.qp_num = id_priv->qp_num; 3876 rep.starting_psn = id_priv->seq_num; 3877 rep.private_data = conn_param->private_data; 3878 rep.private_data_len = conn_param->private_data_len; 3879 rep.responder_resources = conn_param->responder_resources; 3880 rep.initiator_depth = conn_param->initiator_depth; 3881 rep.failover_accepted = 0; 3882 rep.flow_control = conn_param->flow_control; 3883 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count); 3884 rep.srq = id_priv->srq ? 1 : 0; 3885 3886 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep); 3887 out: 3888 return ret; 3889 } 3890 3891 static int cma_accept_iw(struct rdma_id_private *id_priv, 3892 struct rdma_conn_param *conn_param) 3893 { 3894 struct iw_cm_conn_param iw_param; 3895 int ret; 3896 3897 ret = cma_modify_qp_rtr(id_priv, conn_param); 3898 if (ret) 3899 return ret; 3900 3901 iw_param.ord = conn_param->initiator_depth; 3902 iw_param.ird = conn_param->responder_resources; 3903 iw_param.private_data = conn_param->private_data; 3904 iw_param.private_data_len = conn_param->private_data_len; 3905 if (id_priv->id.qp) { 3906 iw_param.qpn = id_priv->qp_num; 3907 } else 3908 iw_param.qpn = conn_param->qp_num; 3909 3910 return iw_cm_accept(id_priv->cm_id.iw, &iw_param); 3911 } 3912 3913 static int cma_send_sidr_rep(struct rdma_id_private *id_priv, 3914 enum ib_cm_sidr_status status, u32 qkey, 3915 const void *private_data, int private_data_len) 3916 { 3917 struct ib_cm_sidr_rep_param rep; 3918 int ret; 3919 3920 memset(&rep, 0, sizeof rep); 3921 rep.status = status; 3922 if (status == IB_SIDR_SUCCESS) { 3923 ret = cma_set_qkey(id_priv, qkey); 3924 if (ret) 3925 return ret; 3926 rep.qp_num = id_priv->qp_num; 3927 rep.qkey = id_priv->qkey; 3928 } 3929 rep.private_data = private_data; 3930 rep.private_data_len = private_data_len; 3931 3932 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep); 3933 } 3934 3935 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param) 3936 { 3937 struct rdma_id_private *id_priv; 3938 int ret; 3939 3940 id_priv = container_of(id, struct rdma_id_private, id); 3941 3942 id_priv->owner = task_pid_nr(current); 3943 3944 if (!cma_comp(id_priv, RDMA_CM_CONNECT)) 3945 return -EINVAL; 3946 3947 if (!id->qp && conn_param) { 3948 id_priv->qp_num = conn_param->qp_num; 3949 id_priv->srq = conn_param->srq; 3950 } 3951 3952 if (rdma_cap_ib_cm(id->device, id->port_num)) { 3953 if (id->qp_type == IB_QPT_UD) { 3954 if (conn_param) 3955 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS, 3956 conn_param->qkey, 3957 conn_param->private_data, 3958 conn_param->private_data_len); 3959 else 3960 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS, 3961 0, NULL, 0); 3962 } else { 3963 if (conn_param) 3964 ret = cma_accept_ib(id_priv, conn_param); 3965 else 3966 ret = cma_rep_recv(id_priv); 3967 } 3968 } else if (rdma_cap_iw_cm(id->device, id->port_num)) 3969 ret = cma_accept_iw(id_priv, conn_param); 3970 else 3971 ret = -ENOSYS; 3972 3973 if (ret) 3974 goto reject; 3975 3976 return 0; 3977 reject: 3978 cma_modify_qp_err(id_priv); 3979 rdma_reject(id, NULL, 0); 3980 return ret; 3981 } 3982 EXPORT_SYMBOL(rdma_accept); 3983 3984 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event) 3985 { 3986 struct rdma_id_private *id_priv; 3987 int ret; 3988 3989 id_priv = container_of(id, struct rdma_id_private, id); 3990 if (!id_priv->cm_id.ib) 3991 return -EINVAL; 3992 3993 switch (id->device->node_type) { 3994 case RDMA_NODE_IB_CA: 3995 ret = ib_cm_notify(id_priv->cm_id.ib, event); 3996 break; 3997 default: 3998 ret = 0; 3999 break; 4000 } 4001 return ret; 4002 } 4003 EXPORT_SYMBOL(rdma_notify); 4004 4005 int rdma_reject(struct rdma_cm_id *id, const void *private_data, 4006 u8 private_data_len) 4007 { 4008 struct rdma_id_private *id_priv; 4009 int ret; 4010 4011 id_priv = container_of(id, struct rdma_id_private, id); 4012 if (!id_priv->cm_id.ib) 4013 return -EINVAL; 4014 4015 if (rdma_cap_ib_cm(id->device, id->port_num)) { 4016 if (id->qp_type == IB_QPT_UD) 4017 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0, 4018 private_data, private_data_len); 4019 else 4020 ret = ib_send_cm_rej(id_priv->cm_id.ib, 4021 IB_CM_REJ_CONSUMER_DEFINED, NULL, 4022 0, private_data, private_data_len); 4023 } else if (rdma_cap_iw_cm(id->device, id->port_num)) { 4024 ret = iw_cm_reject(id_priv->cm_id.iw, 4025 private_data, private_data_len); 4026 } else 4027 ret = -ENOSYS; 4028 4029 return ret; 4030 } 4031 EXPORT_SYMBOL(rdma_reject); 4032 4033 int rdma_disconnect(struct rdma_cm_id *id) 4034 { 4035 struct rdma_id_private *id_priv; 4036 int ret; 4037 4038 id_priv = container_of(id, struct rdma_id_private, id); 4039 if (!id_priv->cm_id.ib) 4040 return -EINVAL; 4041 4042 if (rdma_cap_ib_cm(id->device, id->port_num)) { 4043 ret = cma_modify_qp_err(id_priv); 4044 if (ret) 4045 goto out; 4046 /* Initiate or respond to a disconnect. */ 4047 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) 4048 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0); 4049 } else if (rdma_cap_iw_cm(id->device, id->port_num)) { 4050 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0); 4051 } else 4052 ret = -EINVAL; 4053 4054 out: 4055 return ret; 4056 } 4057 EXPORT_SYMBOL(rdma_disconnect); 4058 4059 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast) 4060 { 4061 struct rdma_id_private *id_priv; 4062 struct cma_multicast *mc = multicast->context; 4063 struct rdma_cm_event event; 4064 int ret = 0; 4065 4066 id_priv = mc->id_priv; 4067 mutex_lock(&id_priv->handler_mutex); 4068 if (id_priv->state != RDMA_CM_ADDR_BOUND && 4069 id_priv->state != RDMA_CM_ADDR_RESOLVED) 4070 goto out; 4071 4072 if (!status) 4073 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey)); 4074 mutex_lock(&id_priv->qp_mutex); 4075 if (!status && id_priv->id.qp) 4076 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid, 4077 be16_to_cpu(multicast->rec.mlid)); 4078 mutex_unlock(&id_priv->qp_mutex); 4079 4080 memset(&event, 0, sizeof event); 4081 event.status = status; 4082 event.param.ud.private_data = mc->context; 4083 if (!status) { 4084 struct rdma_dev_addr *dev_addr = 4085 &id_priv->id.route.addr.dev_addr; 4086 if_t ndev = 4087 dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if); 4088 enum ib_gid_type gid_type = 4089 id_priv->cma_dev->default_gid_type[id_priv->id.port_num - 4090 rdma_start_port(id_priv->cma_dev->device)]; 4091 4092 event.event = RDMA_CM_EVENT_MULTICAST_JOIN; 4093 ret = ib_init_ah_from_mcmember(id_priv->id.device, 4094 id_priv->id.port_num, 4095 &multicast->rec, 4096 ndev, gid_type, 4097 &event.param.ud.ah_attr); 4098 if (ret) 4099 event.event = RDMA_CM_EVENT_MULTICAST_ERROR; 4100 4101 event.param.ud.qp_num = 0xFFFFFF; 4102 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey); 4103 if (ndev) 4104 dev_put(ndev); 4105 } else 4106 event.event = RDMA_CM_EVENT_MULTICAST_ERROR; 4107 4108 ret = id_priv->id.event_handler(&id_priv->id, &event); 4109 if (ret) { 4110 cma_exch(id_priv, RDMA_CM_DESTROYING); 4111 mutex_unlock(&id_priv->handler_mutex); 4112 rdma_destroy_id(&id_priv->id); 4113 return 0; 4114 } 4115 4116 out: 4117 mutex_unlock(&id_priv->handler_mutex); 4118 return 0; 4119 } 4120 4121 static void cma_set_mgid(struct rdma_id_private *id_priv, 4122 struct sockaddr *addr, union ib_gid *mgid) 4123 { 4124 unsigned char mc_map[MAX_ADDR_LEN]; 4125 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 4126 struct sockaddr_in *sin = (struct sockaddr_in *) addr; 4127 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr; 4128 4129 if (cma_any_addr(dev_addr->net, addr)) { 4130 memset(mgid, 0, sizeof *mgid); 4131 } else if ((addr->sa_family == AF_INET6) && 4132 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) == 4133 0xFF10A01B)) { 4134 /* IPv6 address is an SA assigned MGID. */ 4135 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid); 4136 } else if (addr->sa_family == AF_IB) { 4137 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid); 4138 } else if (addr->sa_family == AF_INET6) { 4139 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map); 4140 if (id_priv->id.ps == RDMA_PS_UDP) 4141 mc_map[7] = 0x01; /* Use RDMA CM signature */ 4142 *mgid = *(union ib_gid *) (mc_map + 4); 4143 } else { 4144 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map); 4145 if (id_priv->id.ps == RDMA_PS_UDP) 4146 mc_map[7] = 0x01; /* Use RDMA CM signature */ 4147 *mgid = *(union ib_gid *) (mc_map + 4); 4148 } 4149 } 4150 4151 static void cma_query_sa_classport_info_cb(int status, 4152 struct ib_class_port_info *rec, 4153 void *context) 4154 { 4155 struct class_port_info_context *cb_ctx = context; 4156 4157 WARN_ON(!context); 4158 4159 if (status || !rec) { 4160 pr_debug("RDMA CM: %s port %u failed query ClassPortInfo status: %d\n", 4161 cb_ctx->device->name, cb_ctx->port_num, status); 4162 goto out; 4163 } 4164 4165 memcpy(cb_ctx->class_port_info, rec, sizeof(struct ib_class_port_info)); 4166 4167 out: 4168 complete(&cb_ctx->done); 4169 } 4170 4171 static int cma_query_sa_classport_info(struct ib_device *device, u8 port_num, 4172 struct ib_class_port_info *class_port_info) 4173 { 4174 struct class_port_info_context *cb_ctx; 4175 int ret; 4176 4177 cb_ctx = kmalloc(sizeof(*cb_ctx), GFP_KERNEL); 4178 if (!cb_ctx) 4179 return -ENOMEM; 4180 4181 cb_ctx->device = device; 4182 cb_ctx->class_port_info = class_port_info; 4183 cb_ctx->port_num = port_num; 4184 init_completion(&cb_ctx->done); 4185 4186 ret = ib_sa_classport_info_rec_query(&sa_client, device, port_num, 4187 CMA_QUERY_CLASSPORT_INFO_TIMEOUT, 4188 GFP_KERNEL, cma_query_sa_classport_info_cb, 4189 cb_ctx, &cb_ctx->sa_query); 4190 if (ret < 0) { 4191 pr_err("RDMA CM: %s port %u failed to send ClassPortInfo query, ret: %d\n", 4192 device->name, port_num, ret); 4193 goto out; 4194 } 4195 4196 wait_for_completion(&cb_ctx->done); 4197 4198 out: 4199 kfree(cb_ctx); 4200 return ret; 4201 } 4202 4203 static int cma_join_ib_multicast(struct rdma_id_private *id_priv, 4204 struct cma_multicast *mc) 4205 { 4206 struct ib_sa_mcmember_rec rec; 4207 struct ib_class_port_info class_port_info; 4208 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 4209 ib_sa_comp_mask comp_mask; 4210 int ret; 4211 4212 ib_addr_get_mgid(dev_addr, &rec.mgid); 4213 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num, 4214 &rec.mgid, &rec); 4215 if (ret) 4216 return ret; 4217 4218 ret = cma_set_qkey(id_priv, 0); 4219 if (ret) 4220 return ret; 4221 4222 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid); 4223 rec.qkey = cpu_to_be32(id_priv->qkey); 4224 rdma_addr_get_sgid(dev_addr, &rec.port_gid); 4225 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr)); 4226 rec.join_state = mc->join_state; 4227 4228 if (rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) { 4229 ret = cma_query_sa_classport_info(id_priv->id.device, 4230 id_priv->id.port_num, 4231 &class_port_info); 4232 4233 if (ret) 4234 return ret; 4235 4236 if (!(ib_get_cpi_capmask2(&class_port_info) & 4237 IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT)) { 4238 pr_warn("RDMA CM: %s port %u Unable to multicast join\n" 4239 "RDMA CM: SM doesn't support Send Only Full Member option\n", 4240 id_priv->id.device->name, id_priv->id.port_num); 4241 return -EOPNOTSUPP; 4242 } 4243 } 4244 4245 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID | 4246 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE | 4247 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL | 4248 IB_SA_MCMEMBER_REC_FLOW_LABEL | 4249 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS; 4250 4251 if (id_priv->id.ps == RDMA_PS_IPOIB) 4252 comp_mask |= IB_SA_MCMEMBER_REC_RATE | 4253 IB_SA_MCMEMBER_REC_RATE_SELECTOR | 4254 IB_SA_MCMEMBER_REC_MTU_SELECTOR | 4255 IB_SA_MCMEMBER_REC_MTU | 4256 IB_SA_MCMEMBER_REC_HOP_LIMIT; 4257 4258 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device, 4259 id_priv->id.port_num, &rec, 4260 comp_mask, GFP_KERNEL, 4261 cma_ib_mc_handler, mc); 4262 return PTR_ERR_OR_ZERO(mc->multicast.ib); 4263 } 4264 4265 static void iboe_mcast_work_handler(struct work_struct *work) 4266 { 4267 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work); 4268 struct cma_multicast *mc = mw->mc; 4269 struct ib_sa_multicast *m = mc->multicast.ib; 4270 4271 mc->multicast.ib->context = mc; 4272 cma_ib_mc_handler(0, m); 4273 kref_put(&mc->mcref, release_mc); 4274 kfree(mw); 4275 } 4276 4277 static void cma_iboe_set_mgid(struct vnet *vnet, struct sockaddr *addr, 4278 union ib_gid *mgid, enum ib_gid_type gid_type) 4279 { 4280 struct sockaddr_in *sin = (struct sockaddr_in *)addr; 4281 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr; 4282 4283 if (cma_any_addr(vnet, addr)) { 4284 memset(mgid, 0, sizeof *mgid); 4285 } else if (addr->sa_family == AF_INET6) { 4286 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid); 4287 } else { 4288 mgid->raw[0] = 4289 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff; 4290 mgid->raw[1] = 4291 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e; 4292 mgid->raw[2] = 0; 4293 mgid->raw[3] = 0; 4294 mgid->raw[4] = 0; 4295 mgid->raw[5] = 0; 4296 mgid->raw[6] = 0; 4297 mgid->raw[7] = 0; 4298 mgid->raw[8] = 0; 4299 mgid->raw[9] = 0; 4300 mgid->raw[10] = 0xff; 4301 mgid->raw[11] = 0xff; 4302 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr; 4303 } 4304 } 4305 4306 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv, 4307 struct cma_multicast *mc) 4308 { 4309 struct iboe_mcast_work *work; 4310 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr; 4311 int err = 0; 4312 struct sockaddr *addr = (struct sockaddr *)&mc->addr; 4313 if_t ndev = NULL; 4314 enum ib_gid_type gid_type; 4315 bool send_only; 4316 4317 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN); 4318 4319 if (cma_zero_addr((struct sockaddr *)&mc->addr)) 4320 return -EINVAL; 4321 4322 work = kzalloc(sizeof *work, GFP_KERNEL); 4323 if (!work) 4324 return -ENOMEM; 4325 4326 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL); 4327 if (!mc->multicast.ib) { 4328 err = -ENOMEM; 4329 goto out1; 4330 } 4331 4332 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num - 4333 rdma_start_port(id_priv->cma_dev->device)]; 4334 cma_iboe_set_mgid(dev_addr->net, addr, &mc->multicast.ib->rec.mgid, gid_type); 4335 4336 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff); 4337 if (id_priv->id.ps == RDMA_PS_UDP) 4338 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY); 4339 4340 if (dev_addr->bound_dev_if) 4341 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if); 4342 if (!ndev) { 4343 err = -ENODEV; 4344 goto out2; 4345 } 4346 mc->multicast.ib->rec.rate = iboe_get_rate(ndev); 4347 mc->multicast.ib->rec.hop_limit = 1; 4348 mc->multicast.ib->rec.mtu = iboe_get_mtu(if_getmtu(ndev)); 4349 4350 if (addr->sa_family == AF_INET || addr->sa_family == AF_INET6) { 4351 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) { 4352 mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT; 4353 if (!send_only) { 4354 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid, 4355 true); 4356 if (!err) 4357 mc->igmp_joined = true; 4358 } 4359 } 4360 } else { 4361 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) 4362 err = -ENOTSUPP; 4363 } 4364 dev_put(ndev); 4365 if (err || !mc->multicast.ib->rec.mtu) { 4366 if (!err) 4367 err = -EINVAL; 4368 goto out2; 4369 } 4370 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr, 4371 &mc->multicast.ib->rec.port_gid); 4372 work->id = id_priv; 4373 work->mc = mc; 4374 INIT_WORK(&work->work, iboe_mcast_work_handler); 4375 kref_get(&mc->mcref); 4376 queue_work(cma_wq, &work->work); 4377 4378 return 0; 4379 4380 out2: 4381 kfree(mc->multicast.ib); 4382 out1: 4383 kfree(work); 4384 return err; 4385 } 4386 4387 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr, 4388 u8 join_state, void *context) 4389 { 4390 struct rdma_id_private *id_priv; 4391 struct cma_multicast *mc; 4392 int ret; 4393 4394 if (!id->device) 4395 return -EINVAL; 4396 4397 id_priv = container_of(id, struct rdma_id_private, id); 4398 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) && 4399 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED)) 4400 return -EINVAL; 4401 4402 mc = kmalloc(sizeof *mc, GFP_KERNEL); 4403 if (!mc) 4404 return -ENOMEM; 4405 4406 memcpy(&mc->addr, addr, rdma_addr_size(addr)); 4407 mc->context = context; 4408 mc->id_priv = id_priv; 4409 mc->igmp_joined = false; 4410 mc->join_state = join_state; 4411 spin_lock(&id_priv->lock); 4412 list_add(&mc->list, &id_priv->mc_list); 4413 spin_unlock(&id_priv->lock); 4414 4415 if (rdma_protocol_roce(id->device, id->port_num)) { 4416 kref_init(&mc->mcref); 4417 ret = cma_iboe_join_multicast(id_priv, mc); 4418 } else if (rdma_cap_ib_mcast(id->device, id->port_num)) 4419 ret = cma_join_ib_multicast(id_priv, mc); 4420 else 4421 ret = -ENOSYS; 4422 4423 if (ret) { 4424 spin_lock_irq(&id_priv->lock); 4425 list_del(&mc->list); 4426 spin_unlock_irq(&id_priv->lock); 4427 kfree(mc); 4428 } 4429 return ret; 4430 } 4431 EXPORT_SYMBOL(rdma_join_multicast); 4432 4433 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr) 4434 { 4435 struct rdma_id_private *id_priv; 4436 struct cma_multicast *mc; 4437 4438 id_priv = container_of(id, struct rdma_id_private, id); 4439 spin_lock_irq(&id_priv->lock); 4440 list_for_each_entry(mc, &id_priv->mc_list, list) { 4441 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) { 4442 list_del(&mc->list); 4443 spin_unlock_irq(&id_priv->lock); 4444 4445 if (id->qp) 4446 ib_detach_mcast(id->qp, 4447 &mc->multicast.ib->rec.mgid, 4448 be16_to_cpu(mc->multicast.ib->rec.mlid)); 4449 4450 BUG_ON(id_priv->cma_dev->device != id->device); 4451 4452 if (rdma_cap_ib_mcast(id->device, id->port_num)) { 4453 ib_sa_free_multicast(mc->multicast.ib); 4454 kfree(mc); 4455 } else if (rdma_protocol_roce(id->device, id->port_num)) { 4456 if (mc->igmp_joined) { 4457 struct rdma_dev_addr *dev_addr = 4458 &id->route.addr.dev_addr; 4459 if_t ndev = NULL; 4460 4461 if (dev_addr->bound_dev_if) 4462 ndev = dev_get_by_index(dev_addr->net, 4463 dev_addr->bound_dev_if); 4464 if (ndev) { 4465 cma_igmp_send(ndev, 4466 &mc->multicast.ib->rec.mgid, 4467 false); 4468 dev_put(ndev); 4469 } 4470 mc->igmp_joined = false; 4471 } 4472 kref_put(&mc->mcref, release_mc); 4473 } 4474 return; 4475 } 4476 } 4477 spin_unlock_irq(&id_priv->lock); 4478 } 4479 EXPORT_SYMBOL(rdma_leave_multicast); 4480 4481 static int 4482 sysctl_cma_default_roce_mode(SYSCTL_HANDLER_ARGS) 4483 { 4484 struct cma_device *cma_dev = arg1; 4485 const int port = arg2; 4486 char buf[64]; 4487 int error; 4488 4489 strlcpy(buf, ib_cache_gid_type_str( 4490 cma_get_default_gid_type(cma_dev, port)), sizeof(buf)); 4491 4492 error = sysctl_handle_string(oidp, buf, sizeof(buf), req); 4493 if (error != 0 || req->newptr == NULL) 4494 goto done; 4495 4496 error = ib_cache_gid_parse_type_str(buf); 4497 if (error < 0) { 4498 error = EINVAL; 4499 goto done; 4500 } 4501 4502 cma_set_default_gid_type(cma_dev, port, error); 4503 error = 0; 4504 done: 4505 return (error); 4506 } 4507 4508 static void cma_add_one(struct ib_device *device) 4509 { 4510 struct cma_device *cma_dev; 4511 struct rdma_id_private *id_priv; 4512 unsigned int i; 4513 4514 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL); 4515 if (!cma_dev) 4516 return; 4517 4518 sysctl_ctx_init(&cma_dev->sysctl_ctx); 4519 4520 cma_dev->device = device; 4521 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt, 4522 sizeof(*cma_dev->default_gid_type), 4523 GFP_KERNEL); 4524 if (!cma_dev->default_gid_type) { 4525 kfree(cma_dev); 4526 return; 4527 } 4528 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) { 4529 unsigned long supported_gids; 4530 unsigned int default_gid_type; 4531 4532 supported_gids = roce_gid_type_mask_support(device, i); 4533 4534 if (WARN_ON(!supported_gids)) { 4535 /* set something valid */ 4536 default_gid_type = 0; 4537 } else if (test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids)) { 4538 /* prefer RoCEv2, if supported */ 4539 default_gid_type = IB_GID_TYPE_ROCE_UDP_ENCAP; 4540 } else { 4541 default_gid_type = find_first_bit(&supported_gids, 4542 BITS_PER_LONG); 4543 } 4544 cma_dev->default_gid_type[i - rdma_start_port(device)] = 4545 default_gid_type; 4546 } 4547 4548 init_completion(&cma_dev->comp); 4549 atomic_set(&cma_dev->refcount, 1); 4550 INIT_LIST_HEAD(&cma_dev->id_list); 4551 ib_set_client_data(device, &cma_client, cma_dev); 4552 4553 mutex_lock(&lock); 4554 list_add_tail(&cma_dev->list, &dev_list); 4555 list_for_each_entry(id_priv, &listen_any_list, list) 4556 cma_listen_on_dev(id_priv, cma_dev); 4557 mutex_unlock(&lock); 4558 4559 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) { 4560 char buf[64]; 4561 4562 snprintf(buf, sizeof(buf), "default_roce_mode_port%d", i); 4563 4564 (void) SYSCTL_ADD_PROC(&cma_dev->sysctl_ctx, 4565 SYSCTL_CHILDREN(device->ports_parent->parent->oidp), 4566 OID_AUTO, buf, CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 4567 cma_dev, i, &sysctl_cma_default_roce_mode, "A", 4568 "Default RoCE mode. Valid values: IB/RoCE v1 and RoCE v2"); 4569 } 4570 } 4571 4572 static int cma_remove_id_dev(struct rdma_id_private *id_priv) 4573 { 4574 struct rdma_cm_event event; 4575 enum rdma_cm_state state; 4576 int ret = 0; 4577 4578 /* Record that we want to remove the device */ 4579 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL); 4580 if (state == RDMA_CM_DESTROYING) 4581 return 0; 4582 4583 cma_cancel_operation(id_priv, state); 4584 mutex_lock(&id_priv->handler_mutex); 4585 4586 /* Check for destruction from another callback. */ 4587 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL)) 4588 goto out; 4589 4590 memset(&event, 0, sizeof event); 4591 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL; 4592 ret = id_priv->id.event_handler(&id_priv->id, &event); 4593 out: 4594 mutex_unlock(&id_priv->handler_mutex); 4595 return ret; 4596 } 4597 4598 static void cma_process_remove(struct cma_device *cma_dev) 4599 { 4600 struct rdma_id_private *id_priv; 4601 int ret; 4602 4603 mutex_lock(&lock); 4604 while (!list_empty(&cma_dev->id_list)) { 4605 id_priv = list_entry(cma_dev->id_list.next, 4606 struct rdma_id_private, list); 4607 4608 list_del(&id_priv->listen_list); 4609 list_del_init(&id_priv->list); 4610 atomic_inc(&id_priv->refcount); 4611 mutex_unlock(&lock); 4612 4613 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv); 4614 cma_deref_id(id_priv); 4615 if (ret) 4616 rdma_destroy_id(&id_priv->id); 4617 4618 mutex_lock(&lock); 4619 } 4620 mutex_unlock(&lock); 4621 4622 cma_deref_dev(cma_dev); 4623 wait_for_completion(&cma_dev->comp); 4624 } 4625 4626 static void cma_remove_one(struct ib_device *device, void *client_data) 4627 { 4628 struct cma_device *cma_dev = client_data; 4629 4630 if (!cma_dev) 4631 return; 4632 4633 mutex_lock(&lock); 4634 list_del(&cma_dev->list); 4635 mutex_unlock(&lock); 4636 4637 cma_process_remove(cma_dev); 4638 sysctl_ctx_free(&cma_dev->sysctl_ctx); 4639 kfree(cma_dev->default_gid_type); 4640 kfree(cma_dev); 4641 } 4642 4643 static void cma_init_vnet(void *arg) 4644 { 4645 struct cma_pernet *pernet = &VNET(cma_pernet); 4646 4647 idr_init(&pernet->tcp_ps); 4648 idr_init(&pernet->udp_ps); 4649 idr_init(&pernet->ipoib_ps); 4650 idr_init(&pernet->ib_ps); 4651 idr_init(&pernet->sdp_ps); 4652 } 4653 VNET_SYSINIT(cma_init_vnet, SI_SUB_OFED_MODINIT - 1, SI_ORDER_FIRST, cma_init_vnet, NULL); 4654 4655 static void cma_destroy_vnet(void *arg) 4656 { 4657 struct cma_pernet *pernet = &VNET(cma_pernet); 4658 4659 idr_destroy(&pernet->tcp_ps); 4660 idr_destroy(&pernet->udp_ps); 4661 idr_destroy(&pernet->ipoib_ps); 4662 idr_destroy(&pernet->ib_ps); 4663 idr_destroy(&pernet->sdp_ps); 4664 } 4665 VNET_SYSUNINIT(cma_destroy_vnet, SI_SUB_OFED_MODINIT - 1, SI_ORDER_SECOND, cma_destroy_vnet, NULL); 4666 4667 static int __init cma_init(void) 4668 { 4669 int ret; 4670 4671 cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM); 4672 if (!cma_wq) 4673 return -ENOMEM; 4674 4675 ib_sa_register_client(&sa_client); 4676 rdma_addr_register_client(&addr_client); 4677 4678 ret = ib_register_client(&cma_client); 4679 if (ret) 4680 goto err; 4681 4682 cma_configfs_init(); 4683 4684 return 0; 4685 4686 err: 4687 rdma_addr_unregister_client(&addr_client); 4688 ib_sa_unregister_client(&sa_client); 4689 destroy_workqueue(cma_wq); 4690 return ret; 4691 } 4692 4693 static void __exit cma_cleanup(void) 4694 { 4695 cma_configfs_exit(); 4696 ib_unregister_client(&cma_client); 4697 rdma_addr_unregister_client(&addr_client); 4698 ib_sa_unregister_client(&sa_client); 4699 destroy_workqueue(cma_wq); 4700 } 4701 4702 module_init_order(cma_init, SI_ORDER_FOURTH); 4703 module_exit_order(cma_cleanup, SI_ORDER_FOURTH); 4704