drivers/infiniband/core/cma.c

   1 /*
   2  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
   3  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
   4  * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
   5  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/completion.h>
  37 #include <linux/in.h>
  38 #include <linux/in6.h>
  39 #include <linux/mutex.h>
  40 #include <linux/random.h>
  41 #include <linux/igmp.h>
  42 #include <linux/idr.h>
  43 #include <linux/inetdevice.h>
  44 #include <linux/slab.h>
  45 #include <linux/module.h>
  46 #include <net/route.h>
  47
  48 #include <net/net_namespace.h>
  49 #include <net/netns/generic.h>
  50 #include <net/tcp.h>
  51 #include <net/ipv6.h>
  52 #include <net/ip_fib.h>
  53 #include <net/ip6_route.h>
  54
  55 #include <rdma/rdma_cm.h>
  56 #include <rdma/rdma_cm_ib.h>
  57 #include <rdma/rdma_netlink.h>
  58 #include <rdma/ib.h>
  59 #include <rdma/ib_cache.h>
  60 #include <rdma/ib_cm.h>
  61 #include <rdma/ib_sa.h>
  62 #include <rdma/iw_cm.h>
  63
  64 #include "core_priv.h"
  65 #include "cma_priv.h"
  66
  67 MODULE_AUTHOR("Sean Hefty");
  68 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  69 MODULE_LICENSE("Dual BSD/GPL");
  70
  71 #define CMA_CM_RESPONSE_TIMEOUT 20
  72 #define CMA_QUERY_CLASSPORT_INFO_TIMEOUT 3000
  73 #define CMA_MAX_CM_RETRIES 15
  74 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  75 #define CMA_IBOE_PACKET_LIFETIME 18
  76 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
  77
  78 static const char * const cma_events[] = {
  79         [RDMA_CM_EVENT_ADDR_RESOLVED]    = "address resolved",
  80         [RDMA_CM_EVENT_ADDR_ERROR]       = "address error",
  81         [RDMA_CM_EVENT_ROUTE_RESOLVED]   = "route resolved ",
  82         [RDMA_CM_EVENT_ROUTE_ERROR]      = "route error",
  83         [RDMA_CM_EVENT_CONNECT_REQUEST]  = "connect request",
  84         [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
  85         [RDMA_CM_EVENT_CONNECT_ERROR]    = "connect error",
  86         [RDMA_CM_EVENT_UNREACHABLE]      = "unreachable",
  87         [RDMA_CM_EVENT_REJECTED]         = "rejected",
  88         [RDMA_CM_EVENT_ESTABLISHED]      = "established",
  89         [RDMA_CM_EVENT_DISCONNECTED]     = "disconnected",
  90         [RDMA_CM_EVENT_DEVICE_REMOVAL]   = "device removal",
  91         [RDMA_CM_EVENT_MULTICAST_JOIN]   = "multicast join",
  92         [RDMA_CM_EVENT_MULTICAST_ERROR]  = "multicast error",
  93         [RDMA_CM_EVENT_ADDR_CHANGE]      = "address change",
  94         [RDMA_CM_EVENT_TIMEWAIT_EXIT]    = "timewait exit",
  95 };
  96
  97 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
  98 {
  99         size_t index = event;
 100
 101         return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
 102                         cma_events[index] : "unrecognized event";
 103 }
 104 EXPORT_SYMBOL(rdma_event_msg);
 105
 106 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
 107                                                 int reason)
 108 {
 109         if (rdma_ib_or_roce(id->device, id->port_num))
 110                 return ibcm_reject_msg(reason);
 111
 112         if (rdma_protocol_iwarp(id->device, id->port_num))
 113                 return iwcm_reject_msg(reason);
 114
 115         WARN_ON_ONCE(1);
 116         return "unrecognized transport";
 117 }
 118 EXPORT_SYMBOL(rdma_reject_msg);
 119
 120 bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
 121 {
 122         if (rdma_ib_or_roce(id->device, id->port_num))
 123                 return reason == IB_CM_REJ_CONSUMER_DEFINED;
 124
 125         if (rdma_protocol_iwarp(id->device, id->port_num))
 126                 return reason == -ECONNREFUSED;
 127
 128         WARN_ON_ONCE(1);
 129         return false;
 130 }
 131 EXPORT_SYMBOL(rdma_is_consumer_reject);
 132
 133 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
 134                                       struct rdma_cm_event *ev, u8 *data_len)
 135 {
 136         const void *p;
 137
 138         if (rdma_is_consumer_reject(id, ev->status)) {
 139                 *data_len = ev->param.conn.private_data_len;
 140                 p = ev->param.conn.private_data;
 141         } else {
 142                 *data_len = 0;
 143                 p = NULL;
 144         }
 145         return p;
 146 }
 147 EXPORT_SYMBOL(rdma_consumer_reject_data);
 148
 149 /**
 150  * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
 151  * @id: Communication Identifier
 152  */
 153 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
 154 {
 155         struct rdma_id_private *id_priv;
 156
 157         id_priv = container_of(id, struct rdma_id_private, id);
 158         if (id->device->node_type == RDMA_NODE_RNIC)
 159                 return id_priv->cm_id.iw;
 160         return NULL;
 161 }
 162 EXPORT_SYMBOL(rdma_iw_cm_id);
 163
 164 /**
 165  * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
 166  * @res: rdma resource tracking entry pointer
 167  */
 168 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
 169 {
 170         struct rdma_id_private *id_priv =
 171                 container_of(res, struct rdma_id_private, res);
 172
 173         return &id_priv->id;
 174 }
 175 EXPORT_SYMBOL(rdma_res_to_id);
 176
 177 static void cma_add_one(struct ib_device *device);
 178 static void cma_remove_one(struct ib_device *device, void *client_data);
 179
 180 static struct ib_client cma_client = {
 181         .name   = "cma",
 182         .add    = cma_add_one,
 183         .remove = cma_remove_one
 184 };
 185
 186 static struct ib_sa_client sa_client;
 187 static LIST_HEAD(dev_list);
 188 static LIST_HEAD(listen_any_list);
 189 static DEFINE_MUTEX(lock);
 190 static struct workqueue_struct *cma_wq;
 191 static unsigned int cma_pernet_id;
 192
 193 struct cma_pernet {
 194         struct idr tcp_ps;
 195         struct idr udp_ps;
 196         struct idr ipoib_ps;
 197         struct idr ib_ps;
 198 };
 199
 200 static struct cma_pernet *cma_pernet(struct net *net)
 201 {
 202         return net_generic(net, cma_pernet_id);
 203 }
 204
 205 static struct idr *cma_pernet_idr(struct net *net, enum rdma_ucm_port_space ps)
 206 {
 207         struct cma_pernet *pernet = cma_pernet(net);
 208
 209         switch (ps) {
 210         case RDMA_PS_TCP:
 211                 return &pernet->tcp_ps;
 212         case RDMA_PS_UDP:
 213                 return &pernet->udp_ps;
 214         case RDMA_PS_IPOIB:
 215                 return &pernet->ipoib_ps;
 216         case RDMA_PS_IB:
 217                 return &pernet->ib_ps;
 218         default:
 219                 return NULL;
 220         }
 221 }
 222
 223 struct cma_device {
 224         struct list_head        list;
 225         struct ib_device        *device;
 226         struct completion       comp;
 227         atomic_t                refcount;
 228         struct list_head        id_list;
 229         enum ib_gid_type        *default_gid_type;
 230         u8                      *default_roce_tos;
 231 };
 232
 233 struct rdma_bind_list {
 234         enum rdma_ucm_port_space ps;
 235         struct hlist_head       owners;
 236         unsigned short          port;
 237 };
 238
 239 struct class_port_info_context {
 240         struct ib_class_port_info       *class_port_info;
 241         struct ib_device                *device;
 242         struct completion               done;
 243         struct ib_sa_query              *sa_query;
 244         u8                              port_num;
 245 };
 246
 247 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
 248                         struct rdma_bind_list *bind_list, int snum)
 249 {
 250         struct idr *idr = cma_pernet_idr(net, ps);
 251
 252         return idr_alloc(idr, bind_list, snum, snum + 1, GFP_KERNEL);
 253 }
 254
 255 static struct rdma_bind_list *cma_ps_find(struct net *net,
 256                                           enum rdma_ucm_port_space ps, int snum)
 257 {
 258         struct idr *idr = cma_pernet_idr(net, ps);
 259
 260         return idr_find(idr, snum);
 261 }
 262
 263 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
 264                           int snum)
 265 {
 266         struct idr *idr = cma_pernet_idr(net, ps);
 267
 268         idr_remove(idr, snum);
 269 }
 270
 271 enum {
 272         CMA_OPTION_AFONLY,
 273 };
 274
 275 void cma_ref_dev(struct cma_device *cma_dev)
 276 {
 277         atomic_inc(&cma_dev->refcount);
 278 }
 279
 280 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter  filter,
 281                                              void               *cookie)
 282 {
 283         struct cma_device *cma_dev;
 284         struct cma_device *found_cma_dev = NULL;
 285
 286         mutex_lock(&lock);
 287
 288         list_for_each_entry(cma_dev, &dev_list, list)
 289                 if (filter(cma_dev->device, cookie)) {
 290                         found_cma_dev = cma_dev;
 291                         break;
 292                 }
 293
 294         if (found_cma_dev)
 295                 cma_ref_dev(found_cma_dev);
 296         mutex_unlock(&lock);
 297         return found_cma_dev;
 298 }
 299
 300 int cma_get_default_gid_type(struct cma_device *cma_dev,
 301                              unsigned int port)
 302 {
 303         if (!rdma_is_port_valid(cma_dev->device, port))
 304                 return -EINVAL;
 305
 306         return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
 307 }
 308
 309 int cma_set_default_gid_type(struct cma_device *cma_dev,
 310                              unsigned int port,
 311                              enum ib_gid_type default_gid_type)
 312 {
 313         unsigned long supported_gids;
 314
 315         if (!rdma_is_port_valid(cma_dev->device, port))
 316                 return -EINVAL;
 317
 318         supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
 319
 320         if (!(supported_gids & 1 << default_gid_type))
 321                 return -EINVAL;
 322
 323         cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
 324                 default_gid_type;
 325
 326         return 0;
 327 }
 328
 329 int cma_get_default_roce_tos(struct cma_device *cma_dev, unsigned int port)
 330 {
 331         if (!rdma_is_port_valid(cma_dev->device, port))
 332                 return -EINVAL;
 333
 334         return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
 335 }
 336
 337 int cma_set_default_roce_tos(struct cma_device *cma_dev, unsigned int port,
 338                              u8 default_roce_tos)
 339 {
 340         if (!rdma_is_port_valid(cma_dev->device, port))
 341                 return -EINVAL;
 342
 343         cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
 344                  default_roce_tos;
 345
 346         return 0;
 347 }
 348 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
 349 {
 350         return cma_dev->device;
 351 }
 352
 353 /*
 354  * Device removal can occur at anytime, so we need extra handling to
 355  * serialize notifying the user of device removal with other callbacks.
 356  * We do this by disabling removal notification while a callback is in process,
 357  * and reporting it after the callback completes.
 358  */
 359
 360 struct cma_multicast {
 361         struct rdma_id_private *id_priv;
 362         union {
 363                 struct ib_sa_multicast *ib;
 364         } multicast;
 365         struct list_head        list;
 366         void                    *context;
 367         struct sockaddr_storage addr;
 368         struct kref             mcref;
 369         u8                      join_state;
 370 };
 371
 372 struct cma_work {
 373         struct work_struct      work;
 374         struct rdma_id_private  *id;
 375         enum rdma_cm_state      old_state;
 376         enum rdma_cm_state      new_state;
 377         struct rdma_cm_event    event;
 378 };
 379
 380 struct cma_ndev_work {
 381         struct work_struct      work;
 382         struct rdma_id_private  *id;
 383         struct rdma_cm_event    event;
 384 };
 385
 386 struct iboe_mcast_work {
 387         struct work_struct       work;
 388         struct rdma_id_private  *id;
 389         struct cma_multicast    *mc;
 390 };
 391
 392 union cma_ip_addr {
 393         struct in6_addr ip6;
 394         struct {
 395                 __be32 pad[3];
 396                 __be32 addr;
 397         } ip4;
 398 };
 399
 400 struct cma_hdr {
 401         u8 cma_version;
 402         u8 ip_version;  /* IP version: 7:4 */
 403         __be16 port;
 404         union cma_ip_addr src_addr;
 405         union cma_ip_addr dst_addr;
 406 };
 407
 408 #define CMA_VERSION 0x00
 409
 410 struct cma_req_info {
 411         struct sockaddr_storage listen_addr_storage;
 412         struct sockaddr_storage src_addr_storage;
 413         struct ib_device *device;
 414         union ib_gid local_gid;
 415         __be64 service_id;
 416         int port;
 417         bool has_gid;
 418         u16 pkey;
 419 };
 420
 421 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 422 {
 423         unsigned long flags;
 424         int ret;
 425
 426         spin_lock_irqsave(&id_priv->lock, flags);
 427         ret = (id_priv->state == comp);
 428         spin_unlock_irqrestore(&id_priv->lock, flags);
 429         return ret;
 430 }
 431
 432 static int cma_comp_exch(struct rdma_id_private *id_priv,
 433                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 434 {
 435         unsigned long flags;
 436         int ret;
 437
 438         spin_lock_irqsave(&id_priv->lock, flags);
 439         if ((ret = (id_priv->state == comp)))
 440                 id_priv->state = exch;
 441         spin_unlock_irqrestore(&id_priv->lock, flags);
 442         return ret;
 443 }
 444
 445 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 446                                    enum rdma_cm_state exch)
 447 {
 448         unsigned long flags;
 449         enum rdma_cm_state old;
 450
 451         spin_lock_irqsave(&id_priv->lock, flags);
 452         old = id_priv->state;
 453         id_priv->state = exch;
 454         spin_unlock_irqrestore(&id_priv->lock, flags);
 455         return old;
 456 }
 457
 458 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
 459 {
 460         return hdr->ip_version >> 4;
 461 }
 462
 463 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 464 {
 465         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 466 }
 467
 468 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
 469 {
 470         struct in_device *in_dev = NULL;
 471
 472         if (ndev) {
 473                 rtnl_lock();
 474                 in_dev = __in_dev_get_rtnl(ndev);
 475                 if (in_dev) {
 476                         if (join)
 477                                 ip_mc_inc_group(in_dev,
 478                                                 *(__be32 *)(mgid->raw + 12));
 479                         else
 480                                 ip_mc_dec_group(in_dev,
 481                                                 *(__be32 *)(mgid->raw + 12));
 482                 }
 483                 rtnl_unlock();
 484         }
 485         return (in_dev) ? 0 : -ENODEV;
 486 }
 487
 488 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
 489                                struct cma_device *cma_dev)
 490 {
 491         cma_ref_dev(cma_dev);
 492         id_priv->cma_dev = cma_dev;
 493         id_priv->id.device = cma_dev->device;
 494         id_priv->id.route.addr.dev_addr.transport =
 495                 rdma_node_get_transport(cma_dev->device->node_type);
 496         list_add_tail(&id_priv->list, &cma_dev->id_list);
 497         rdma_restrack_add(&id_priv->res);
 498 }
 499
 500 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 501                               struct cma_device *cma_dev)
 502 {
 503         _cma_attach_to_dev(id_priv, cma_dev);
 504         id_priv->gid_type =
 505                 cma_dev->default_gid_type[id_priv->id.port_num -
 506                                           rdma_start_port(cma_dev->device)];
 507 }
 508
 509 void cma_deref_dev(struct cma_device *cma_dev)
 510 {
 511         if (atomic_dec_and_test(&cma_dev->refcount))
 512                 complete(&cma_dev->comp);
 513 }
 514
 515 static inline void release_mc(struct kref *kref)
 516 {
 517         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 518
 519         kfree(mc->multicast.ib);
 520         kfree(mc);
 521 }
 522
 523 static void cma_release_dev(struct rdma_id_private *id_priv)
 524 {
 525         mutex_lock(&lock);
 526         list_del(&id_priv->list);
 527         cma_deref_dev(id_priv->cma_dev);
 528         id_priv->cma_dev = NULL;
 529         mutex_unlock(&lock);
 530 }
 531
 532 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 533 {
 534         return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
 535 }
 536
 537 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 538 {
 539         return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
 540 }
 541
 542 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 543 {
 544         return id_priv->id.route.addr.src_addr.ss_family;
 545 }
 546
 547 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 548 {
 549         struct ib_sa_mcmember_rec rec;
 550         int ret = 0;
 551
 552         if (id_priv->qkey) {
 553                 if (qkey && id_priv->qkey != qkey)
 554                         return -EINVAL;
 555                 return 0;
 556         }
 557
 558         if (qkey) {
 559                 id_priv->qkey = qkey;
 560                 return 0;
 561         }
 562
 563         switch (id_priv->id.ps) {
 564         case RDMA_PS_UDP:
 565         case RDMA_PS_IB:
 566                 id_priv->qkey = RDMA_UDP_QKEY;
 567                 break;
 568         case RDMA_PS_IPOIB:
 569                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 570                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 571                                              id_priv->id.port_num, &rec.mgid,
 572                                              &rec);
 573                 if (!ret)
 574                         id_priv->qkey = be32_to_cpu(rec.qkey);
 575                 break;
 576         default:
 577                 break;
 578         }
 579         return ret;
 580 }
 581
 582 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 583 {
 584         dev_addr->dev_type = ARPHRD_INFINIBAND;
 585         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 586         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 587 }
 588
 589 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 590 {
 591         int ret;
 592
 593         if (addr->sa_family != AF_IB) {
 594                 ret = rdma_translate_ip(addr, dev_addr);
 595         } else {
 596                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 597                 ret = 0;
 598         }
 599
 600         return ret;
 601 }
 602
 603 static const struct ib_gid_attr *
 604 cma_validate_port(struct ib_device *device, u8 port,
 605                   enum ib_gid_type gid_type,
 606                   union ib_gid *gid,
 607                   struct rdma_id_private *id_priv)
 608 {
 609         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 610         int bound_if_index = dev_addr->bound_dev_if;
 611         const struct ib_gid_attr *sgid_attr;
 612         int dev_type = dev_addr->dev_type;
 613         struct net_device *ndev = NULL;
 614
 615         if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
 616                 return ERR_PTR(-ENODEV);
 617
 618         if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
 619                 return ERR_PTR(-ENODEV);
 620
 621         if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
 622                 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
 623                 if (!ndev)
 624                         return ERR_PTR(-ENODEV);
 625         } else {
 626                 gid_type = IB_GID_TYPE_IB;
 627         }
 628
 629         sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
 630         if (ndev)
 631                 dev_put(ndev);
 632         return sgid_attr;
 633 }
 634
 635 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
 636                                const struct ib_gid_attr *sgid_attr)
 637 {
 638         WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
 639         id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
 640 }
 641
 642 static int cma_acquire_dev(struct rdma_id_private *id_priv,
 643                            const struct rdma_id_private *listen_id_priv)
 644 {
 645         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 646         const struct ib_gid_attr *sgid_attr;
 647         struct cma_device *cma_dev;
 648         union ib_gid gid, iboe_gid, *gidp;
 649         enum ib_gid_type gid_type;
 650         int ret = -ENODEV;
 651         u8 port;
 652
 653         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 654             id_priv->id.ps == RDMA_PS_IPOIB)
 655                 return -EINVAL;
 656
 657         mutex_lock(&lock);
 658         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 659                     &iboe_gid);
 660
 661         memcpy(&gid, dev_addr->src_dev_addr +
 662                rdma_addr_gid_offset(dev_addr), sizeof gid);
 663
 664         if (listen_id_priv) {
 665                 cma_dev = listen_id_priv->cma_dev;
 666                 port = listen_id_priv->id.port_num;
 667                 gidp = rdma_protocol_roce(cma_dev->device, port) ?
 668                        &iboe_gid : &gid;
 669                 gid_type = listen_id_priv->gid_type;
 670                 sgid_attr = cma_validate_port(cma_dev->device, port,
 671                                               gid_type, gidp, id_priv);
 672                 if (!IS_ERR(sgid_attr)) {
 673                         id_priv->id.port_num = port;
 674                         cma_bind_sgid_attr(id_priv, sgid_attr);
 675                         ret = 0;
 676                         goto out;
 677                 }
 678         }
 679
 680         list_for_each_entry(cma_dev, &dev_list, list) {
 681                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
 682                         if (listen_id_priv &&
 683                             listen_id_priv->cma_dev == cma_dev &&
 684                             listen_id_priv->id.port_num == port)
 685                                 continue;
 686
 687                         gidp = rdma_protocol_roce(cma_dev->device, port) ?
 688                                &iboe_gid : &gid;
 689                         gid_type = cma_dev->default_gid_type[port - 1];
 690                         sgid_attr = cma_validate_port(cma_dev->device, port,
 691                                                       gid_type, gidp, id_priv);
 692                         if (!IS_ERR(sgid_attr)) {
 693                                 id_priv->id.port_num = port;
 694                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 695                                 ret = 0;
 696                                 goto out;
 697                         }
 698                 }
 699         }
 700
 701 out:
 702         if (!ret)
 703                 cma_attach_to_dev(id_priv, cma_dev);
 704
 705         mutex_unlock(&lock);
 706         return ret;
 707 }
 708
 709 /*
 710  * Select the source IB device and address to reach the destination IB address.
 711  */
 712 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 713 {
 714         struct cma_device *cma_dev, *cur_dev;
 715         struct sockaddr_ib *addr;
 716         union ib_gid gid, sgid, *dgid;
 717         u16 pkey, index;
 718         u8 p;
 719         enum ib_port_state port_state;
 720         int i;
 721
 722         cma_dev = NULL;
 723         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 724         dgid = (union ib_gid *) &addr->sib_addr;
 725         pkey = ntohs(addr->sib_pkey);
 726
 727         mutex_lock(&lock);
 728         list_for_each_entry(cur_dev, &dev_list, list) {
 729                 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
 730                         if (!rdma_cap_af_ib(cur_dev->device, p))
 731                                 continue;
 732
 733                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 734                                 continue;
 735
 736                         if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
 737                                 continue;
 738                         for (i = 0; !rdma_query_gid(cur_dev->device,
 739                                                     p, i, &gid);
 740                              i++) {
 741                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 742                                         cma_dev = cur_dev;
 743                                         sgid = gid;
 744                                         id_priv->id.port_num = p;
 745                                         goto found;
 746                                 }
 747
 748                                 if (!cma_dev && (gid.global.subnet_prefix ==
 749                                     dgid->global.subnet_prefix) &&
 750                                     port_state == IB_PORT_ACTIVE) {
 751                                         cma_dev = cur_dev;
 752                                         sgid = gid;
 753                                         id_priv->id.port_num = p;
 754                                         goto found;
 755                                 }
 756                         }
 757                 }
 758         }
 759         mutex_unlock(&lock);
 760         return -ENODEV;
 761
 762 found:
 763         cma_attach_to_dev(id_priv, cma_dev);
 764         mutex_unlock(&lock);
 765         addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
 766         memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
 767         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 768         return 0;
 769 }
 770
 771 static void cma_deref_id(struct rdma_id_private *id_priv)
 772 {
 773         if (atomic_dec_and_test(&id_priv->refcount))
 774                 complete(&id_priv->comp);
 775 }
 776
 777 struct rdma_cm_id *__rdma_create_id(struct net *net,
 778                                     rdma_cm_event_handler event_handler,
 779                                     void *context, enum rdma_ucm_port_space ps,
 780                                     enum ib_qp_type qp_type, const char *caller)
 781 {
 782         struct rdma_id_private *id_priv;
 783
 784         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 785         if (!id_priv)
 786                 return ERR_PTR(-ENOMEM);
 787
 788         if (caller)
 789                 id_priv->res.kern_name = caller;
 790         else
 791                 rdma_restrack_set_task(&id_priv->res, current);
 792         id_priv->res.type = RDMA_RESTRACK_CM_ID;
 793         id_priv->state = RDMA_CM_IDLE;
 794         id_priv->id.context = context;
 795         id_priv->id.event_handler = event_handler;
 796         id_priv->id.ps = ps;
 797         id_priv->id.qp_type = qp_type;
 798         id_priv->tos_set = false;
 799         id_priv->gid_type = IB_GID_TYPE_IB;
 800         spin_lock_init(&id_priv->lock);
 801         mutex_init(&id_priv->qp_mutex);
 802         init_completion(&id_priv->comp);
 803         atomic_set(&id_priv->refcount, 1);
 804         mutex_init(&id_priv->handler_mutex);
 805         INIT_LIST_HEAD(&id_priv->listen_list);
 806         INIT_LIST_HEAD(&id_priv->mc_list);
 807         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 808         id_priv->id.route.addr.dev_addr.net = get_net(net);
 809         id_priv->seq_num &= 0x00ffffff;
 810
 811         return &id_priv->id;
 812 }
 813 EXPORT_SYMBOL(__rdma_create_id);
 814
 815 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 816 {
 817         struct ib_qp_attr qp_attr;
 818         int qp_attr_mask, ret;
 819
 820         qp_attr.qp_state = IB_QPS_INIT;
 821         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 822         if (ret)
 823                 return ret;
 824
 825         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 826         if (ret)
 827                 return ret;
 828
 829         qp_attr.qp_state = IB_QPS_RTR;
 830         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 831         if (ret)
 832                 return ret;
 833
 834         qp_attr.qp_state = IB_QPS_RTS;
 835         qp_attr.sq_psn = 0;
 836         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 837
 838         return ret;
 839 }
 840
 841 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 842 {
 843         struct ib_qp_attr qp_attr;
 844         int qp_attr_mask, ret;
 845
 846         qp_attr.qp_state = IB_QPS_INIT;
 847         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 848         if (ret)
 849                 return ret;
 850
 851         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 852 }
 853
 854 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 855                    struct ib_qp_init_attr *qp_init_attr)
 856 {
 857         struct rdma_id_private *id_priv;
 858         struct ib_qp *qp;
 859         int ret;
 860
 861         id_priv = container_of(id, struct rdma_id_private, id);
 862         if (id->device != pd->device)
 863                 return -EINVAL;
 864
 865         qp_init_attr->port_num = id->port_num;
 866         qp = ib_create_qp(pd, qp_init_attr);
 867         if (IS_ERR(qp))
 868                 return PTR_ERR(qp);
 869
 870         if (id->qp_type == IB_QPT_UD)
 871                 ret = cma_init_ud_qp(id_priv, qp);
 872         else
 873                 ret = cma_init_conn_qp(id_priv, qp);
 874         if (ret)
 875                 goto err;
 876
 877         id->qp = qp;
 878         id_priv->qp_num = qp->qp_num;
 879         id_priv->srq = (qp->srq != NULL);
 880         return 0;
 881 err:
 882         ib_destroy_qp(qp);
 883         return ret;
 884 }
 885 EXPORT_SYMBOL(rdma_create_qp);
 886
 887 void rdma_destroy_qp(struct rdma_cm_id *id)
 888 {
 889         struct rdma_id_private *id_priv;
 890
 891         id_priv = container_of(id, struct rdma_id_private, id);
 892         mutex_lock(&id_priv->qp_mutex);
 893         ib_destroy_qp(id_priv->id.qp);
 894         id_priv->id.qp = NULL;
 895         mutex_unlock(&id_priv->qp_mutex);
 896 }
 897 EXPORT_SYMBOL(rdma_destroy_qp);
 898
 899 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 900                              struct rdma_conn_param *conn_param)
 901 {
 902         struct ib_qp_attr qp_attr;
 903         int qp_attr_mask, ret;
 904
 905         mutex_lock(&id_priv->qp_mutex);
 906         if (!id_priv->id.qp) {
 907                 ret = 0;
 908                 goto out;
 909         }
 910
 911         /* Need to update QP attributes from default values. */
 912         qp_attr.qp_state = IB_QPS_INIT;
 913         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 914         if (ret)
 915                 goto out;
 916
 917         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 918         if (ret)
 919                 goto out;
 920
 921         qp_attr.qp_state = IB_QPS_RTR;
 922         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 923         if (ret)
 924                 goto out;
 925
 926         BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
 927
 928         if (conn_param)
 929                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 930         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 931 out:
 932         mutex_unlock(&id_priv->qp_mutex);
 933         return ret;
 934 }
 935
 936 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 937                              struct rdma_conn_param *conn_param)
 938 {
 939         struct ib_qp_attr qp_attr;
 940         int qp_attr_mask, ret;
 941
 942         mutex_lock(&id_priv->qp_mutex);
 943         if (!id_priv->id.qp) {
 944                 ret = 0;
 945                 goto out;
 946         }
 947
 948         qp_attr.qp_state = IB_QPS_RTS;
 949         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 950         if (ret)
 951                 goto out;
 952
 953         if (conn_param)
 954                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 955         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 956 out:
 957         mutex_unlock(&id_priv->qp_mutex);
 958         return ret;
 959 }
 960
 961 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 962 {
 963         struct ib_qp_attr qp_attr;
 964         int ret;
 965
 966         mutex_lock(&id_priv->qp_mutex);
 967         if (!id_priv->id.qp) {
 968                 ret = 0;
 969                 goto out;
 970         }
 971
 972         qp_attr.qp_state = IB_QPS_ERR;
 973         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 974 out:
 975         mutex_unlock(&id_priv->qp_mutex);
 976         return ret;
 977 }
 978
 979 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 980                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 981 {
 982         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 983         int ret;
 984         u16 pkey;
 985
 986         if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
 987                 pkey = 0xffff;
 988         else
 989                 pkey = ib_addr_get_pkey(dev_addr);
 990
 991         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 992                                   pkey, &qp_attr->pkey_index);
 993         if (ret)
 994                 return ret;
 995
 996         qp_attr->port_num = id_priv->id.port_num;
 997         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 998
 999         if (id_priv->id.qp_type == IB_QPT_UD) {
1000                 ret = cma_set_qkey(id_priv, 0);
1001                 if (ret)
1002                         return ret;
1003
1004                 qp_attr->qkey = id_priv->qkey;
1005                 *qp_attr_mask |= IB_QP_QKEY;
1006         } else {
1007                 qp_attr->qp_access_flags = 0;
1008                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1009         }
1010         return 0;
1011 }
1012
1013 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1014                        int *qp_attr_mask)
1015 {
1016         struct rdma_id_private *id_priv;
1017         int ret = 0;
1018
1019         id_priv = container_of(id, struct rdma_id_private, id);
1020         if (rdma_cap_ib_cm(id->device, id->port_num)) {
1021                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1022                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1023                 else
1024                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1025                                                  qp_attr_mask);
1026
1027                 if (qp_attr->qp_state == IB_QPS_RTR)
1028                         qp_attr->rq_psn = id_priv->seq_num;
1029         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1030                 if (!id_priv->cm_id.iw) {
1031                         qp_attr->qp_access_flags = 0;
1032                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1033                 } else
1034                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1035                                                  qp_attr_mask);
1036                 qp_attr->port_num = id_priv->id.port_num;
1037                 *qp_attr_mask |= IB_QP_PORT;
1038         } else
1039                 ret = -ENOSYS;
1040
1041         return ret;
1042 }
1043 EXPORT_SYMBOL(rdma_init_qp_attr);
1044
1045 static inline bool cma_zero_addr(const struct sockaddr *addr)
1046 {
1047         switch (addr->sa_family) {
1048         case AF_INET:
1049                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1050         case AF_INET6:
1051                 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1052         case AF_IB:
1053                 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1054         default:
1055                 return false;
1056         }
1057 }
1058
1059 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1060 {
1061         switch (addr->sa_family) {
1062         case AF_INET:
1063                 return ipv4_is_loopback(
1064                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1065         case AF_INET6:
1066                 return ipv6_addr_loopback(
1067                         &((struct sockaddr_in6 *)addr)->sin6_addr);
1068         case AF_IB:
1069                 return ib_addr_loopback(
1070                         &((struct sockaddr_ib *)addr)->sib_addr);
1071         default:
1072                 return false;
1073         }
1074 }
1075
1076 static inline bool cma_any_addr(const struct sockaddr *addr)
1077 {
1078         return cma_zero_addr(addr) || cma_loopback_addr(addr);
1079 }
1080
1081 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1082 {
1083         if (src->sa_family != dst->sa_family)
1084                 return -1;
1085
1086         switch (src->sa_family) {
1087         case AF_INET:
1088                 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1089                        ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1090         case AF_INET6: {
1091                 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1092                 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1093                 bool link_local;
1094
1095                 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1096                                           &dst_addr6->sin6_addr))
1097                         return 1;
1098                 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1099                              IPV6_ADDR_LINKLOCAL;
1100                 /* Link local must match their scope_ids */
1101                 return link_local ? (src_addr6->sin6_scope_id !=
1102                                      dst_addr6->sin6_scope_id) :
1103                                     0;
1104         }
1105
1106         default:
1107                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1108                                    &((struct sockaddr_ib *) dst)->sib_addr);
1109         }
1110 }
1111
1112 static __be16 cma_port(const struct sockaddr *addr)
1113 {
1114         struct sockaddr_ib *sib;
1115
1116         switch (addr->sa_family) {
1117         case AF_INET:
1118                 return ((struct sockaddr_in *) addr)->sin_port;
1119         case AF_INET6:
1120                 return ((struct sockaddr_in6 *) addr)->sin6_port;
1121         case AF_IB:
1122                 sib = (struct sockaddr_ib *) addr;
1123                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1124                                     be64_to_cpu(sib->sib_sid_mask)));
1125         default:
1126                 return 0;
1127         }
1128 }
1129
1130 static inline int cma_any_port(const struct sockaddr *addr)
1131 {
1132         return !cma_port(addr);
1133 }
1134
1135 static void cma_save_ib_info(struct sockaddr *src_addr,
1136                              struct sockaddr *dst_addr,
1137                              const struct rdma_cm_id *listen_id,
1138                              const struct sa_path_rec *path)
1139 {
1140         struct sockaddr_ib *listen_ib, *ib;
1141
1142         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1143         if (src_addr) {
1144                 ib = (struct sockaddr_ib *)src_addr;
1145                 ib->sib_family = AF_IB;
1146                 if (path) {
1147                         ib->sib_pkey = path->pkey;
1148                         ib->sib_flowinfo = path->flow_label;
1149                         memcpy(&ib->sib_addr, &path->sgid, 16);
1150                         ib->sib_sid = path->service_id;
1151                         ib->sib_scope_id = 0;
1152                 } else {
1153                         ib->sib_pkey = listen_ib->sib_pkey;
1154                         ib->sib_flowinfo = listen_ib->sib_flowinfo;
1155                         ib->sib_addr = listen_ib->sib_addr;
1156                         ib->sib_sid = listen_ib->sib_sid;
1157                         ib->sib_scope_id = listen_ib->sib_scope_id;
1158                 }
1159                 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1160         }
1161         if (dst_addr) {
1162                 ib = (struct sockaddr_ib *)dst_addr;
1163                 ib->sib_family = AF_IB;
1164                 if (path) {
1165                         ib->sib_pkey = path->pkey;
1166                         ib->sib_flowinfo = path->flow_label;
1167                         memcpy(&ib->sib_addr, &path->dgid, 16);
1168                 }
1169         }
1170 }
1171
1172 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1173                               struct sockaddr_in *dst_addr,
1174                               struct cma_hdr *hdr,
1175                               __be16 local_port)
1176 {
1177         if (src_addr) {
1178                 *src_addr = (struct sockaddr_in) {
1179                         .sin_family = AF_INET,
1180                         .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1181                         .sin_port = local_port,
1182                 };
1183         }
1184
1185         if (dst_addr) {
1186                 *dst_addr = (struct sockaddr_in) {
1187                         .sin_family = AF_INET,
1188                         .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1189                         .sin_port = hdr->port,
1190                 };
1191         }
1192 }
1193
1194 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1195                               struct sockaddr_in6 *dst_addr,
1196                               struct cma_hdr *hdr,
1197                               __be16 local_port)
1198 {
1199         if (src_addr) {
1200                 *src_addr = (struct sockaddr_in6) {
1201                         .sin6_family = AF_INET6,
1202                         .sin6_addr = hdr->dst_addr.ip6,
1203                         .sin6_port = local_port,
1204                 };
1205         }
1206
1207         if (dst_addr) {
1208                 *dst_addr = (struct sockaddr_in6) {
1209                         .sin6_family = AF_INET6,
1210                         .sin6_addr = hdr->src_addr.ip6,
1211                         .sin6_port = hdr->port,
1212                 };
1213         }
1214 }
1215
1216 static u16 cma_port_from_service_id(__be64 service_id)
1217 {
1218         return (u16)be64_to_cpu(service_id);
1219 }
1220
1221 static int cma_save_ip_info(struct sockaddr *src_addr,
1222                             struct sockaddr *dst_addr,
1223                             const struct ib_cm_event *ib_event,
1224                             __be64 service_id)
1225 {
1226         struct cma_hdr *hdr;
1227         __be16 port;
1228
1229         hdr = ib_event->private_data;
1230         if (hdr->cma_version != CMA_VERSION)
1231                 return -EINVAL;
1232
1233         port = htons(cma_port_from_service_id(service_id));
1234
1235         switch (cma_get_ip_ver(hdr)) {
1236         case 4:
1237                 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1238                                   (struct sockaddr_in *)dst_addr, hdr, port);
1239                 break;
1240         case 6:
1241                 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1242                                   (struct sockaddr_in6 *)dst_addr, hdr, port);
1243                 break;
1244         default:
1245                 return -EAFNOSUPPORT;
1246         }
1247
1248         return 0;
1249 }
1250
1251 static int cma_save_net_info(struct sockaddr *src_addr,
1252                              struct sockaddr *dst_addr,
1253                              const struct rdma_cm_id *listen_id,
1254                              const struct ib_cm_event *ib_event,
1255                              sa_family_t sa_family, __be64 service_id)
1256 {
1257         if (sa_family == AF_IB) {
1258                 if (ib_event->event == IB_CM_REQ_RECEIVED)
1259                         cma_save_ib_info(src_addr, dst_addr, listen_id,
1260                                          ib_event->param.req_rcvd.primary_path);
1261                 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1262                         cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1263                 return 0;
1264         }
1265
1266         return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1267 }
1268
1269 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1270                              struct cma_req_info *req)
1271 {
1272         const struct ib_cm_req_event_param *req_param =
1273                 &ib_event->param.req_rcvd;
1274         const struct ib_cm_sidr_req_event_param *sidr_param =
1275                 &ib_event->param.sidr_req_rcvd;
1276
1277         switch (ib_event->event) {
1278         case IB_CM_REQ_RECEIVED:
1279                 req->device     = req_param->listen_id->device;
1280                 req->port       = req_param->port;
1281                 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1282                        sizeof(req->local_gid));
1283                 req->has_gid    = true;
1284                 req->service_id = req_param->primary_path->service_id;
1285                 req->pkey       = be16_to_cpu(req_param->primary_path->pkey);
1286                 if (req->pkey != req_param->bth_pkey)
1287                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1288                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1289                                             req_param->bth_pkey, req->pkey);
1290                 break;
1291         case IB_CM_SIDR_REQ_RECEIVED:
1292                 req->device     = sidr_param->listen_id->device;
1293                 req->port       = sidr_param->port;
1294                 req->has_gid    = false;
1295                 req->service_id = sidr_param->service_id;
1296                 req->pkey       = sidr_param->pkey;
1297                 if (req->pkey != sidr_param->bth_pkey)
1298                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1299                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1300                                             sidr_param->bth_pkey, req->pkey);
1301                 break;
1302         default:
1303                 return -EINVAL;
1304         }
1305
1306         return 0;
1307 }
1308
1309 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1310                                   const struct sockaddr_in *dst_addr,
1311                                   const struct sockaddr_in *src_addr)
1312 {
1313         __be32 daddr = dst_addr->sin_addr.s_addr,
1314                saddr = src_addr->sin_addr.s_addr;
1315         struct fib_result res;
1316         struct flowi4 fl4;
1317         int err;
1318         bool ret;
1319
1320         if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1321             ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1322             ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1323             ipv4_is_loopback(saddr))
1324                 return false;
1325
1326         memset(&fl4, 0, sizeof(fl4));
1327         fl4.flowi4_iif = net_dev->ifindex;
1328         fl4.daddr = daddr;
1329         fl4.saddr = saddr;
1330
1331         rcu_read_lock();
1332         err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1333         ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1334         rcu_read_unlock();
1335
1336         return ret;
1337 }
1338
1339 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1340                                   const struct sockaddr_in6 *dst_addr,
1341                                   const struct sockaddr_in6 *src_addr)
1342 {
1343 #if IS_ENABLED(CONFIG_IPV6)
1344         const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1345                            IPV6_ADDR_LINKLOCAL;
1346         struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1347                                          &src_addr->sin6_addr, net_dev->ifindex,
1348                                          NULL, strict);
1349         bool ret;
1350
1351         if (!rt)
1352                 return false;
1353
1354         ret = rt->rt6i_idev->dev == net_dev;
1355         ip6_rt_put(rt);
1356
1357         return ret;
1358 #else
1359         return false;
1360 #endif
1361 }
1362
1363 static bool validate_net_dev(struct net_device *net_dev,
1364                              const struct sockaddr *daddr,
1365                              const struct sockaddr *saddr)
1366 {
1367         const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1368         const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1369         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1370         const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1371
1372         switch (daddr->sa_family) {
1373         case AF_INET:
1374                 return saddr->sa_family == AF_INET &&
1375                        validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1376
1377         case AF_INET6:
1378                 return saddr->sa_family == AF_INET6 &&
1379                        validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1380
1381         default:
1382                 return false;
1383         }
1384 }
1385
1386 static struct net_device *
1387 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1388 {
1389         const struct ib_gid_attr *sgid_attr = NULL;
1390
1391         if (ib_event->event == IB_CM_REQ_RECEIVED)
1392                 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1393         else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1394                 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1395
1396         if (!sgid_attr)
1397                 return NULL;
1398         dev_hold(sgid_attr->ndev);
1399         return sgid_attr->ndev;
1400 }
1401
1402 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1403                                           struct cma_req_info *req)
1404 {
1405         struct sockaddr *listen_addr =
1406                         (struct sockaddr *)&req->listen_addr_storage;
1407         struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1408         struct net_device *net_dev;
1409         const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1410         int err;
1411
1412         err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1413                                req->service_id);
1414         if (err)
1415                 return ERR_PTR(err);
1416
1417         if (rdma_protocol_roce(req->device, req->port))
1418                 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1419         else
1420                 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1421                                                    req->pkey,
1422                                                    gid, listen_addr);
1423         if (!net_dev)
1424                 return ERR_PTR(-ENODEV);
1425
1426         return net_dev;
1427 }
1428
1429 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1430 {
1431         return (be64_to_cpu(service_id) >> 16) & 0xffff;
1432 }
1433
1434 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1435                                    const struct cma_hdr *hdr)
1436 {
1437         struct sockaddr *addr = cma_src_addr(id_priv);
1438         __be32 ip4_addr;
1439         struct in6_addr ip6_addr;
1440
1441         if (cma_any_addr(addr) && !id_priv->afonly)
1442                 return true;
1443
1444         switch (addr->sa_family) {
1445         case AF_INET:
1446                 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1447                 if (cma_get_ip_ver(hdr) != 4)
1448                         return false;
1449                 if (!cma_any_addr(addr) &&
1450                     hdr->dst_addr.ip4.addr != ip4_addr)
1451                         return false;
1452                 break;
1453         case AF_INET6:
1454                 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1455                 if (cma_get_ip_ver(hdr) != 6)
1456                         return false;
1457                 if (!cma_any_addr(addr) &&
1458                     memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1459                         return false;
1460                 break;
1461         case AF_IB:
1462                 return true;
1463         default:
1464                 return false;
1465         }
1466
1467         return true;
1468 }
1469
1470 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1471 {
1472         struct ib_device *device = id->device;
1473         const int port_num = id->port_num ?: rdma_start_port(device);
1474
1475         return rdma_protocol_roce(device, port_num);
1476 }
1477
1478 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1479                               const struct net_device *net_dev,
1480                               u8 port_num)
1481 {
1482         const struct rdma_addr *addr = &id->route.addr;
1483
1484         if (!net_dev)
1485                 /* This request is an AF_IB request */
1486                 return (!id->port_num || id->port_num == port_num) &&
1487                        (addr->src_addr.ss_family == AF_IB);
1488
1489         /*
1490          * Net namespaces must match, and if the listner is listening
1491          * on a specific netdevice than netdevice must match as well.
1492          */
1493         if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1494             (!!addr->dev_addr.bound_dev_if ==
1495              (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1496                 return true;
1497         else
1498                 return false;
1499 }
1500
1501 static struct rdma_id_private *cma_find_listener(
1502                 const struct rdma_bind_list *bind_list,
1503                 const struct ib_cm_id *cm_id,
1504                 const struct ib_cm_event *ib_event,
1505                 const struct cma_req_info *req,
1506                 const struct net_device *net_dev)
1507 {
1508         struct rdma_id_private *id_priv, *id_priv_dev;
1509
1510         lockdep_assert_held(&lock);
1511
1512         if (!bind_list)
1513                 return ERR_PTR(-EINVAL);
1514
1515         hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1516                 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1517                         if (id_priv->id.device == cm_id->device &&
1518                             cma_match_net_dev(&id_priv->id, net_dev, req->port))
1519                                 return id_priv;
1520                         list_for_each_entry(id_priv_dev,
1521                                             &id_priv->listen_list,
1522                                             listen_list) {
1523                                 if (id_priv_dev->id.device == cm_id->device &&
1524                                     cma_match_net_dev(&id_priv_dev->id, net_dev, req->port))
1525                                         return id_priv_dev;
1526                         }
1527                 }
1528         }
1529
1530         return ERR_PTR(-EINVAL);
1531 }
1532
1533 static struct rdma_id_private *
1534 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1535                      const struct ib_cm_event *ib_event,
1536                      struct net_device **net_dev)
1537 {
1538         struct cma_req_info req;
1539         struct rdma_bind_list *bind_list;
1540         struct rdma_id_private *id_priv;
1541         int err;
1542
1543         err = cma_save_req_info(ib_event, &req);
1544         if (err)
1545                 return ERR_PTR(err);
1546
1547         *net_dev = cma_get_net_dev(ib_event, &req);
1548         if (IS_ERR(*net_dev)) {
1549                 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1550                         /* Assuming the protocol is AF_IB */
1551                         *net_dev = NULL;
1552                 } else {
1553                         return ERR_CAST(*net_dev);
1554                 }
1555         }
1556
1557         mutex_lock(&lock);
1558         /*
1559          * Net namespace might be getting deleted while route lookup,
1560          * cm_id lookup is in progress. Therefore, perform netdevice
1561          * validation, cm_id lookup under rcu lock.
1562          * RCU lock along with netdevice state check, synchronizes with
1563          * netdevice migrating to different net namespace and also avoids
1564          * case where net namespace doesn't get deleted while lookup is in
1565          * progress.
1566          * If the device state is not IFF_UP, its properties such as ifindex
1567          * and nd_net cannot be trusted to remain valid without rcu lock.
1568          * net/core/dev.c change_net_namespace() ensures to synchronize with
1569          * ongoing operations on net device after device is closed using
1570          * synchronize_net().
1571          */
1572         rcu_read_lock();
1573         if (*net_dev) {
1574                 /*
1575                  * If netdevice is down, it is likely that it is administratively
1576                  * down or it might be migrating to different namespace.
1577                  * In that case avoid further processing, as the net namespace
1578                  * or ifindex may change.
1579                  */
1580                 if (((*net_dev)->flags & IFF_UP) == 0) {
1581                         id_priv = ERR_PTR(-EHOSTUNREACH);
1582                         goto err;
1583                 }
1584
1585                 if (!validate_net_dev(*net_dev,
1586                                  (struct sockaddr *)&req.listen_addr_storage,
1587                                  (struct sockaddr *)&req.src_addr_storage)) {
1588                         id_priv = ERR_PTR(-EHOSTUNREACH);
1589                         goto err;
1590                 }
1591         }
1592
1593         bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1594                                 rdma_ps_from_service_id(req.service_id),
1595                                 cma_port_from_service_id(req.service_id));
1596         id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev);
1597 err:
1598         rcu_read_unlock();
1599         mutex_unlock(&lock);
1600         if (IS_ERR(id_priv) && *net_dev) {
1601                 dev_put(*net_dev);
1602                 *net_dev = NULL;
1603         }
1604         return id_priv;
1605 }
1606
1607 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1608 {
1609         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1610 }
1611
1612 static void cma_cancel_route(struct rdma_id_private *id_priv)
1613 {
1614         if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1615                 if (id_priv->query)
1616                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1617         }
1618 }
1619
1620 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1621 {
1622         struct rdma_id_private *dev_id_priv;
1623
1624         /*
1625          * Remove from listen_any_list to prevent added devices from spawning
1626          * additional listen requests.
1627          */
1628         mutex_lock(&lock);
1629         list_del(&id_priv->list);
1630
1631         while (!list_empty(&id_priv->listen_list)) {
1632                 dev_id_priv = list_entry(id_priv->listen_list.next,
1633                                          struct rdma_id_private, listen_list);
1634                 /* sync with device removal to avoid duplicate destruction */
1635                 list_del_init(&dev_id_priv->list);
1636                 list_del(&dev_id_priv->listen_list);
1637                 mutex_unlock(&lock);
1638
1639                 rdma_destroy_id(&dev_id_priv->id);
1640                 mutex_lock(&lock);
1641         }
1642         mutex_unlock(&lock);
1643 }
1644
1645 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1646                                  enum rdma_cm_state state)
1647 {
1648         switch (state) {
1649         case RDMA_CM_ADDR_QUERY:
1650                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1651                 break;
1652         case RDMA_CM_ROUTE_QUERY:
1653                 cma_cancel_route(id_priv);
1654                 break;
1655         case RDMA_CM_LISTEN:
1656                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1657                         cma_cancel_listens(id_priv);
1658                 break;
1659         default:
1660                 break;
1661         }
1662 }
1663
1664 static void cma_release_port(struct rdma_id_private *id_priv)
1665 {
1666         struct rdma_bind_list *bind_list = id_priv->bind_list;
1667         struct net *net = id_priv->id.route.addr.dev_addr.net;
1668
1669         if (!bind_list)
1670                 return;
1671
1672         mutex_lock(&lock);
1673         hlist_del(&id_priv->node);
1674         if (hlist_empty(&bind_list->owners)) {
1675                 cma_ps_remove(net, bind_list->ps, bind_list->port);
1676                 kfree(bind_list);
1677         }
1678         mutex_unlock(&lock);
1679 }
1680
1681 static void destroy_mc(struct rdma_id_private *id_priv,
1682                        struct cma_multicast *mc)
1683 {
1684         if (rdma_cap_ib_mcast(id_priv->id.device, id_priv->id.port_num)) {
1685                 ib_sa_free_multicast(mc->multicast.ib);
1686                 kfree(mc);
1687                 return;
1688         }
1689
1690         if (rdma_protocol_roce(id_priv->id.device,
1691                                       id_priv->id.port_num)) {
1692                 struct rdma_dev_addr *dev_addr =
1693                         &id_priv->id.route.addr.dev_addr;
1694                 struct net_device *ndev = NULL;
1695
1696                 if (dev_addr->bound_dev_if)
1697                         ndev = dev_get_by_index(dev_addr->net,
1698                                                 dev_addr->bound_dev_if);
1699                 if (ndev) {
1700                         cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid, false);
1701                         dev_put(ndev);
1702                 }
1703                 kref_put(&mc->mcref, release_mc);
1704         }
1705 }
1706
1707 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1708 {
1709         struct cma_multicast *mc;
1710
1711         while (!list_empty(&id_priv->mc_list)) {
1712                 mc = list_first_entry(&id_priv->mc_list, struct cma_multicast,
1713                                       list);
1714                 list_del(&mc->list);
1715                 destroy_mc(id_priv, mc);
1716         }
1717 }
1718
1719 void rdma_destroy_id(struct rdma_cm_id *id)
1720 {
1721         struct rdma_id_private *id_priv;
1722         enum rdma_cm_state state;
1723
1724         id_priv = container_of(id, struct rdma_id_private, id);
1725         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1726         cma_cancel_operation(id_priv, state);
1727
1728         /*
1729          * Wait for any active callback to finish.  New callbacks will find
1730          * the id_priv state set to destroying and abort.
1731          */
1732         mutex_lock(&id_priv->handler_mutex);
1733         mutex_unlock(&id_priv->handler_mutex);
1734
1735         rdma_restrack_del(&id_priv->res);
1736         if (id_priv->cma_dev) {
1737                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1738                         if (id_priv->cm_id.ib)
1739                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1740                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1741                         if (id_priv->cm_id.iw)
1742                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1743                 }
1744                 cma_leave_mc_groups(id_priv);
1745                 cma_release_dev(id_priv);
1746         }
1747
1748         cma_release_port(id_priv);
1749         cma_deref_id(id_priv);
1750         wait_for_completion(&id_priv->comp);
1751
1752         if (id_priv->internal_id)
1753                 cma_deref_id(id_priv->id.context);
1754
1755         kfree(id_priv->id.route.path_rec);
1756
1757         if (id_priv->id.route.addr.dev_addr.sgid_attr)
1758                 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
1759
1760         put_net(id_priv->id.route.addr.dev_addr.net);
1761         kfree(id_priv);
1762 }
1763 EXPORT_SYMBOL(rdma_destroy_id);
1764
1765 static int cma_rep_recv(struct rdma_id_private *id_priv)
1766 {
1767         int ret;
1768
1769         ret = cma_modify_qp_rtr(id_priv, NULL);
1770         if (ret)
1771                 goto reject;
1772
1773         ret = cma_modify_qp_rts(id_priv, NULL);
1774         if (ret)
1775                 goto reject;
1776
1777         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1778         if (ret)
1779                 goto reject;
1780
1781         return 0;
1782 reject:
1783         pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1784         cma_modify_qp_err(id_priv);
1785         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1786                        NULL, 0, NULL, 0);
1787         return ret;
1788 }
1789
1790 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1791                                    const struct ib_cm_rep_event_param *rep_data,
1792                                    void *private_data)
1793 {
1794         event->param.conn.private_data = private_data;
1795         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1796         event->param.conn.responder_resources = rep_data->responder_resources;
1797         event->param.conn.initiator_depth = rep_data->initiator_depth;
1798         event->param.conn.flow_control = rep_data->flow_control;
1799         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1800         event->param.conn.srq = rep_data->srq;
1801         event->param.conn.qp_num = rep_data->remote_qpn;
1802 }
1803
1804 static int cma_ib_handler(struct ib_cm_id *cm_id,
1805                           const struct ib_cm_event *ib_event)
1806 {
1807         struct rdma_id_private *id_priv = cm_id->context;
1808         struct rdma_cm_event event = {};
1809         int ret = 0;
1810
1811         mutex_lock(&id_priv->handler_mutex);
1812         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1813              id_priv->state != RDMA_CM_CONNECT) ||
1814             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1815              id_priv->state != RDMA_CM_DISCONNECT))
1816                 goto out;
1817
1818         switch (ib_event->event) {
1819         case IB_CM_REQ_ERROR:
1820         case IB_CM_REP_ERROR:
1821                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1822                 event.status = -ETIMEDOUT;
1823                 break;
1824         case IB_CM_REP_RECEIVED:
1825                 if (cma_comp(id_priv, RDMA_CM_CONNECT) &&
1826                     (id_priv->id.qp_type != IB_QPT_UD))
1827                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1828                 if (id_priv->id.qp) {
1829                         event.status = cma_rep_recv(id_priv);
1830                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1831                                                      RDMA_CM_EVENT_ESTABLISHED;
1832                 } else {
1833                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1834                 }
1835                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1836                                        ib_event->private_data);
1837                 break;
1838         case IB_CM_RTU_RECEIVED:
1839         case IB_CM_USER_ESTABLISHED:
1840                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1841                 break;
1842         case IB_CM_DREQ_ERROR:
1843                 event.status = -ETIMEDOUT; /* fall through */
1844         case IB_CM_DREQ_RECEIVED:
1845         case IB_CM_DREP_RECEIVED:
1846                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1847                                    RDMA_CM_DISCONNECT))
1848                         goto out;
1849                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1850                 break;
1851         case IB_CM_TIMEWAIT_EXIT:
1852                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1853                 break;
1854         case IB_CM_MRA_RECEIVED:
1855                 /* ignore event */
1856                 goto out;
1857         case IB_CM_REJ_RECEIVED:
1858                 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
1859                                                                                 ib_event->param.rej_rcvd.reason));
1860                 cma_modify_qp_err(id_priv);
1861                 event.status = ib_event->param.rej_rcvd.reason;
1862                 event.event = RDMA_CM_EVENT_REJECTED;
1863                 event.param.conn.private_data = ib_event->private_data;
1864                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1865                 break;
1866         default:
1867                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
1868                        ib_event->event);
1869                 goto out;
1870         }
1871
1872         ret = id_priv->id.event_handler(&id_priv->id, &event);
1873         if (ret) {
1874                 /* Destroy the CM ID by returning a non-zero value. */
1875                 id_priv->cm_id.ib = NULL;
1876                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1877                 mutex_unlock(&id_priv->handler_mutex);
1878                 rdma_destroy_id(&id_priv->id);
1879                 return ret;
1880         }
1881 out:
1882         mutex_unlock(&id_priv->handler_mutex);
1883         return ret;
1884 }
1885
1886 static struct rdma_id_private *
1887 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
1888                    const struct ib_cm_event *ib_event,
1889                    struct net_device *net_dev)
1890 {
1891         struct rdma_id_private *listen_id_priv;
1892         struct rdma_id_private *id_priv;
1893         struct rdma_cm_id *id;
1894         struct rdma_route *rt;
1895         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1896         struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
1897         const __be64 service_id =
1898                 ib_event->param.req_rcvd.primary_path->service_id;
1899         int ret;
1900
1901         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
1902         id = __rdma_create_id(listen_id->route.addr.dev_addr.net,
1903                             listen_id->event_handler, listen_id->context,
1904                             listen_id->ps, ib_event->param.req_rcvd.qp_type,
1905                             listen_id_priv->res.kern_name);
1906         if (IS_ERR(id))
1907                 return NULL;
1908
1909         id_priv = container_of(id, struct rdma_id_private, id);
1910         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1911                               (struct sockaddr *)&id->route.addr.dst_addr,
1912                               listen_id, ib_event, ss_family, service_id))
1913                 goto err;
1914
1915         rt = &id->route;
1916         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1917         rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
1918                                      GFP_KERNEL);
1919         if (!rt->path_rec)
1920                 goto err;
1921
1922         rt->path_rec[0] = *path;
1923         if (rt->num_paths == 2)
1924                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1925
1926         if (net_dev) {
1927                 rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL);
1928         } else {
1929                 if (!cma_protocol_roce(listen_id) &&
1930                     cma_any_addr(cma_src_addr(id_priv))) {
1931                         rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1932                         rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1933                         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1934                 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
1935                         ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1936                         if (ret)
1937                                 goto err;
1938                 }
1939         }
1940         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1941
1942         id_priv->state = RDMA_CM_CONNECT;
1943         return id_priv;
1944
1945 err:
1946         rdma_destroy_id(id);
1947         return NULL;
1948 }
1949
1950 static struct rdma_id_private *
1951 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
1952                   const struct ib_cm_event *ib_event,
1953                   struct net_device *net_dev)
1954 {
1955         const struct rdma_id_private *listen_id_priv;
1956         struct rdma_id_private *id_priv;
1957         struct rdma_cm_id *id;
1958         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1959         struct net *net = listen_id->route.addr.dev_addr.net;
1960         int ret;
1961
1962         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
1963         id = __rdma_create_id(net, listen_id->event_handler, listen_id->context,
1964                               listen_id->ps, IB_QPT_UD,
1965                               listen_id_priv->res.kern_name);
1966         if (IS_ERR(id))
1967                 return NULL;
1968
1969         id_priv = container_of(id, struct rdma_id_private, id);
1970         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1971                               (struct sockaddr *)&id->route.addr.dst_addr,
1972                               listen_id, ib_event, ss_family,
1973                               ib_event->param.sidr_req_rcvd.service_id))
1974                 goto err;
1975
1976         if (net_dev) {
1977                 rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL);
1978         } else {
1979                 if (!cma_any_addr(cma_src_addr(id_priv))) {
1980                         ret = cma_translate_addr(cma_src_addr(id_priv),
1981                                                  &id->route.addr.dev_addr);
1982                         if (ret)
1983                                 goto err;
1984                 }
1985         }
1986
1987         id_priv->state = RDMA_CM_CONNECT;
1988         return id_priv;
1989 err:
1990         rdma_destroy_id(id);
1991         return NULL;
1992 }
1993
1994 static void cma_set_req_event_data(struct rdma_cm_event *event,
1995                                    const struct ib_cm_req_event_param *req_data,
1996                                    void *private_data, int offset)
1997 {
1998         event->param.conn.private_data = private_data + offset;
1999         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2000         event->param.conn.responder_resources = req_data->responder_resources;
2001         event->param.conn.initiator_depth = req_data->initiator_depth;
2002         event->param.conn.flow_control = req_data->flow_control;
2003         event->param.conn.retry_count = req_data->retry_count;
2004         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2005         event->param.conn.srq = req_data->srq;
2006         event->param.conn.qp_num = req_data->remote_qpn;
2007 }
2008
2009 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2010                                     const struct ib_cm_event *ib_event)
2011 {
2012         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2013                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2014                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2015                  (id->qp_type == IB_QPT_UD)) ||
2016                 (!id->qp_type));
2017 }
2018
2019 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2020                               const struct ib_cm_event *ib_event)
2021 {
2022         struct rdma_id_private *listen_id, *conn_id = NULL;
2023         struct rdma_cm_event event = {};
2024         struct net_device *net_dev;
2025         u8 offset;
2026         int ret;
2027
2028         listen_id = cma_ib_id_from_event(cm_id, ib_event, &net_dev);
2029         if (IS_ERR(listen_id))
2030                 return PTR_ERR(listen_id);
2031
2032         if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2033                 ret = -EINVAL;
2034                 goto net_dev_put;
2035         }
2036
2037         mutex_lock(&listen_id->handler_mutex);
2038         if (listen_id->state != RDMA_CM_LISTEN) {
2039                 ret = -ECONNABORTED;
2040                 goto err1;
2041         }
2042
2043         offset = cma_user_data_offset(listen_id);
2044         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2045         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2046                 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2047                 event.param.ud.private_data = ib_event->private_data + offset;
2048                 event.param.ud.private_data_len =
2049                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2050         } else {
2051                 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2052                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2053                                        ib_event->private_data, offset);
2054         }
2055         if (!conn_id) {
2056                 ret = -ENOMEM;
2057                 goto err1;
2058         }
2059
2060         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2061         ret = cma_acquire_dev(conn_id, listen_id);
2062         if (ret)
2063                 goto err2;
2064
2065         conn_id->cm_id.ib = cm_id;
2066         cm_id->context = conn_id;
2067         cm_id->cm_handler = cma_ib_handler;
2068
2069         /*
2070          * Protect against the user destroying conn_id from another thread
2071          * until we're done accessing it.
2072          */
2073         atomic_inc(&conn_id->refcount);
2074         ret = conn_id->id.event_handler(&conn_id->id, &event);
2075         if (ret)
2076                 goto err3;
2077         /*
2078          * Acquire mutex to prevent user executing rdma_destroy_id()
2079          * while we're accessing the cm_id.
2080          */
2081         mutex_lock(&lock);
2082         if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
2083             (conn_id->id.qp_type != IB_QPT_UD))
2084                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2085         mutex_unlock(&lock);
2086         mutex_unlock(&conn_id->handler_mutex);
2087         mutex_unlock(&listen_id->handler_mutex);
2088         cma_deref_id(conn_id);
2089         if (net_dev)
2090                 dev_put(net_dev);
2091         return 0;
2092
2093 err3:
2094         cma_deref_id(conn_id);
2095         /* Destroy the CM ID by returning a non-zero value. */
2096         conn_id->cm_id.ib = NULL;
2097 err2:
2098         cma_exch(conn_id, RDMA_CM_DESTROYING);
2099         mutex_unlock(&conn_id->handler_mutex);
2100 err1:
2101         mutex_unlock(&listen_id->handler_mutex);
2102         if (conn_id)
2103                 rdma_destroy_id(&conn_id->id);
2104
2105 net_dev_put:
2106         if (net_dev)
2107                 dev_put(net_dev);
2108
2109         return ret;
2110 }
2111
2112 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2113 {
2114         if (addr->sa_family == AF_IB)
2115                 return ((struct sockaddr_ib *) addr)->sib_sid;
2116
2117         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2118 }
2119 EXPORT_SYMBOL(rdma_get_service_id);
2120
2121 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2122                     union ib_gid *dgid)
2123 {
2124         struct rdma_addr *addr = &cm_id->route.addr;
2125
2126         if (!cm_id->device) {
2127                 if (sgid)
2128                         memset(sgid, 0, sizeof(*sgid));
2129                 if (dgid)
2130                         memset(dgid, 0, sizeof(*dgid));
2131                 return;
2132         }
2133
2134         if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2135                 if (sgid)
2136                         rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2137                 if (dgid)
2138                         rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2139         } else {
2140                 if (sgid)
2141                         rdma_addr_get_sgid(&addr->dev_addr, sgid);
2142                 if (dgid)
2143                         rdma_addr_get_dgid(&addr->dev_addr, dgid);
2144         }
2145 }
2146 EXPORT_SYMBOL(rdma_read_gids);
2147
2148 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2149 {
2150         struct rdma_id_private *id_priv = iw_id->context;
2151         struct rdma_cm_event event = {};
2152         int ret = 0;
2153         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2154         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2155
2156         mutex_lock(&id_priv->handler_mutex);
2157         if (id_priv->state != RDMA_CM_CONNECT)
2158                 goto out;
2159
2160         switch (iw_event->event) {
2161         case IW_CM_EVENT_CLOSE:
2162                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2163                 break;
2164         case IW_CM_EVENT_CONNECT_REPLY:
2165                 memcpy(cma_src_addr(id_priv), laddr,
2166                        rdma_addr_size(laddr));
2167                 memcpy(cma_dst_addr(id_priv), raddr,
2168                        rdma_addr_size(raddr));
2169                 switch (iw_event->status) {
2170                 case 0:
2171                         event.event = RDMA_CM_EVENT_ESTABLISHED;
2172                         event.param.conn.initiator_depth = iw_event->ird;
2173                         event.param.conn.responder_resources = iw_event->ord;
2174                         break;
2175                 case -ECONNRESET:
2176                 case -ECONNREFUSED:
2177                         event.event = RDMA_CM_EVENT_REJECTED;
2178                         break;
2179                 case -ETIMEDOUT:
2180                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2181                         break;
2182                 default:
2183                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2184                         break;
2185                 }
2186                 break;
2187         case IW_CM_EVENT_ESTABLISHED:
2188                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2189                 event.param.conn.initiator_depth = iw_event->ird;
2190                 event.param.conn.responder_resources = iw_event->ord;
2191                 break;
2192         default:
2193                 goto out;
2194         }
2195
2196         event.status = iw_event->status;
2197         event.param.conn.private_data = iw_event->private_data;
2198         event.param.conn.private_data_len = iw_event->private_data_len;
2199         ret = id_priv->id.event_handler(&id_priv->id, &event);
2200         if (ret) {
2201                 /* Destroy the CM ID by returning a non-zero value. */
2202                 id_priv->cm_id.iw = NULL;
2203                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2204                 mutex_unlock(&id_priv->handler_mutex);
2205                 rdma_destroy_id(&id_priv->id);
2206                 return ret;
2207         }
2208
2209 out:
2210         mutex_unlock(&id_priv->handler_mutex);
2211         return ret;
2212 }
2213
2214 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2215                                struct iw_cm_event *iw_event)
2216 {
2217         struct rdma_cm_id *new_cm_id;
2218         struct rdma_id_private *listen_id, *conn_id;
2219         struct rdma_cm_event event = {};
2220         int ret = -ECONNABORTED;
2221         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2222         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2223
2224         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2225         event.param.conn.private_data = iw_event->private_data;
2226         event.param.conn.private_data_len = iw_event->private_data_len;
2227         event.param.conn.initiator_depth = iw_event->ird;
2228         event.param.conn.responder_resources = iw_event->ord;
2229
2230         listen_id = cm_id->context;
2231
2232         mutex_lock(&listen_id->handler_mutex);
2233         if (listen_id->state != RDMA_CM_LISTEN)
2234                 goto out;
2235
2236         /* Create a new RDMA id for the new IW CM ID */
2237         new_cm_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2238                                      listen_id->id.event_handler,
2239                                      listen_id->id.context,
2240                                      RDMA_PS_TCP, IB_QPT_RC,
2241                                      listen_id->res.kern_name);
2242         if (IS_ERR(new_cm_id)) {
2243                 ret = -ENOMEM;
2244                 goto out;
2245         }
2246         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2247         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2248         conn_id->state = RDMA_CM_CONNECT;
2249
2250         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2251         if (ret) {
2252                 mutex_unlock(&conn_id->handler_mutex);
2253                 rdma_destroy_id(new_cm_id);
2254                 goto out;
2255         }
2256
2257         ret = cma_acquire_dev(conn_id, listen_id);
2258         if (ret) {
2259                 mutex_unlock(&conn_id->handler_mutex);
2260                 rdma_destroy_id(new_cm_id);
2261                 goto out;
2262         }
2263
2264         conn_id->cm_id.iw = cm_id;
2265         cm_id->context = conn_id;
2266         cm_id->cm_handler = cma_iw_handler;
2267
2268         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2269         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2270
2271         /*
2272          * Protect against the user destroying conn_id from another thread
2273          * until we're done accessing it.
2274          */
2275         atomic_inc(&conn_id->refcount);
2276         ret = conn_id->id.event_handler(&conn_id->id, &event);
2277         if (ret) {
2278                 /* User wants to destroy the CM ID */
2279                 conn_id->cm_id.iw = NULL;
2280                 cma_exch(conn_id, RDMA_CM_DESTROYING);
2281                 mutex_unlock(&conn_id->handler_mutex);
2282                 mutex_unlock(&listen_id->handler_mutex);
2283                 cma_deref_id(conn_id);
2284                 rdma_destroy_id(&conn_id->id);
2285                 return ret;
2286         }
2287
2288         mutex_unlock(&conn_id->handler_mutex);
2289         cma_deref_id(conn_id);
2290
2291 out:
2292         mutex_unlock(&listen_id->handler_mutex);
2293         return ret;
2294 }
2295
2296 static int cma_ib_listen(struct rdma_id_private *id_priv)
2297 {
2298         struct sockaddr *addr;
2299         struct ib_cm_id *id;
2300         __be64 svc_id;
2301
2302         addr = cma_src_addr(id_priv);
2303         svc_id = rdma_get_service_id(&id_priv->id, addr);
2304         id = ib_cm_insert_listen(id_priv->id.device,
2305                                  cma_ib_req_handler, svc_id);
2306         if (IS_ERR(id))
2307                 return PTR_ERR(id);
2308         id_priv->cm_id.ib = id;
2309
2310         return 0;
2311 }
2312
2313 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2314 {
2315         int ret;
2316         struct iw_cm_id *id;
2317
2318         id = iw_create_cm_id(id_priv->id.device,
2319                              iw_conn_req_handler,
2320                              id_priv);
2321         if (IS_ERR(id))
2322                 return PTR_ERR(id);
2323
2324         id->tos = id_priv->tos;
2325         id_priv->cm_id.iw = id;
2326
2327         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2328                rdma_addr_size(cma_src_addr(id_priv)));
2329
2330         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2331
2332         if (ret) {
2333                 iw_destroy_cm_id(id_priv->cm_id.iw);
2334                 id_priv->cm_id.iw = NULL;
2335         }
2336
2337         return ret;
2338 }
2339
2340 static int cma_listen_handler(struct rdma_cm_id *id,
2341                               struct rdma_cm_event *event)
2342 {
2343         struct rdma_id_private *id_priv = id->context;
2344
2345         id->context = id_priv->id.context;
2346         id->event_handler = id_priv->id.event_handler;
2347         return id_priv->id.event_handler(id, event);
2348 }
2349
2350 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2351                               struct cma_device *cma_dev)
2352 {
2353         struct rdma_id_private *dev_id_priv;
2354         struct rdma_cm_id *id;
2355         struct net *net = id_priv->id.route.addr.dev_addr.net;
2356         int ret;
2357
2358         lockdep_assert_held(&lock);
2359
2360         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2361                 return;
2362
2363         id = __rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2364                               id_priv->id.qp_type, id_priv->res.kern_name);
2365         if (IS_ERR(id))
2366                 return;
2367
2368         dev_id_priv = container_of(id, struct rdma_id_private, id);
2369
2370         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2371         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2372                rdma_addr_size(cma_src_addr(id_priv)));
2373
2374         _cma_attach_to_dev(dev_id_priv, cma_dev);
2375         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2376         atomic_inc(&id_priv->refcount);
2377         dev_id_priv->internal_id = 1;
2378         dev_id_priv->afonly = id_priv->afonly;
2379
2380         ret = rdma_listen(id, id_priv->backlog);
2381         if (ret)
2382                 pr_warn("RDMA CMA: cma_listen_on_dev, error %d, listening on device %s\n",
2383                         ret, cma_dev->device->name);
2384 }
2385
2386 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2387 {
2388         struct cma_device *cma_dev;
2389
2390         mutex_lock(&lock);
2391         list_add_tail(&id_priv->list, &listen_any_list);
2392         list_for_each_entry(cma_dev, &dev_list, list)
2393                 cma_listen_on_dev(id_priv, cma_dev);
2394         mutex_unlock(&lock);
2395 }
2396
2397 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2398 {
2399         struct rdma_id_private *id_priv;
2400
2401         id_priv = container_of(id, struct rdma_id_private, id);
2402         id_priv->tos = (u8) tos;
2403         id_priv->tos_set = true;
2404 }
2405 EXPORT_SYMBOL(rdma_set_service_type);
2406
2407 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2408                               void *context)
2409 {
2410         struct cma_work *work = context;
2411         struct rdma_route *route;
2412
2413         route = &work->id->id.route;
2414
2415         if (!status) {
2416                 route->num_paths = 1;
2417                 *route->path_rec = *path_rec;
2418         } else {
2419                 work->old_state = RDMA_CM_ROUTE_QUERY;
2420                 work->new_state = RDMA_CM_ADDR_RESOLVED;
2421                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2422                 work->event.status = status;
2423                 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2424                                      status);
2425         }
2426
2427         queue_work(cma_wq, &work->work);
2428 }
2429
2430 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
2431                               struct cma_work *work)
2432 {
2433         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2434         struct sa_path_rec path_rec;
2435         ib_sa_comp_mask comp_mask;
2436         struct sockaddr_in6 *sin6;
2437         struct sockaddr_ib *sib;
2438
2439         memset(&path_rec, 0, sizeof path_rec);
2440
2441         if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2442                 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2443         else
2444                 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2445         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2446         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2447         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2448         path_rec.numb_path = 1;
2449         path_rec.reversible = 1;
2450         path_rec.service_id = rdma_get_service_id(&id_priv->id,
2451                                                   cma_dst_addr(id_priv));
2452
2453         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2454                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2455                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2456
2457         switch (cma_family(id_priv)) {
2458         case AF_INET:
2459                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2460                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2461                 break;
2462         case AF_INET6:
2463                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2464                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2465                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2466                 break;
2467         case AF_IB:
2468                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2469                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2470                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2471                 break;
2472         }
2473
2474         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2475                                                id_priv->id.port_num, &path_rec,
2476                                                comp_mask, timeout_ms,
2477                                                GFP_KERNEL, cma_query_handler,
2478                                                work, &id_priv->query);
2479
2480         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2481 }
2482
2483 static void cma_work_handler(struct work_struct *_work)
2484 {
2485         struct cma_work *work = container_of(_work, struct cma_work, work);
2486         struct rdma_id_private *id_priv = work->id;
2487         int destroy = 0;
2488
2489         mutex_lock(&id_priv->handler_mutex);
2490         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2491                 goto out;
2492
2493         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2494                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2495                 destroy = 1;
2496         }
2497 out:
2498         mutex_unlock(&id_priv->handler_mutex);
2499         cma_deref_id(id_priv);
2500         if (destroy)
2501                 rdma_destroy_id(&id_priv->id);
2502         kfree(work);
2503 }
2504
2505 static void cma_ndev_work_handler(struct work_struct *_work)
2506 {
2507         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2508         struct rdma_id_private *id_priv = work->id;
2509         int destroy = 0;
2510
2511         mutex_lock(&id_priv->handler_mutex);
2512         if (id_priv->state == RDMA_CM_DESTROYING ||
2513             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2514                 goto out;
2515
2516         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2517                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2518                 destroy = 1;
2519         }
2520
2521 out:
2522         mutex_unlock(&id_priv->handler_mutex);
2523         cma_deref_id(id_priv);
2524         if (destroy)
2525                 rdma_destroy_id(&id_priv->id);
2526         kfree(work);
2527 }
2528
2529 static void cma_init_resolve_route_work(struct cma_work *work,
2530                                         struct rdma_id_private *id_priv)
2531 {
2532         work->id = id_priv;
2533         INIT_WORK(&work->work, cma_work_handler);
2534         work->old_state = RDMA_CM_ROUTE_QUERY;
2535         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2536         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2537 }
2538
2539 static void cma_init_resolve_addr_work(struct cma_work *work,
2540                                        struct rdma_id_private *id_priv)
2541 {
2542         work->id = id_priv;
2543         INIT_WORK(&work->work, cma_work_handler);
2544         work->old_state = RDMA_CM_ADDR_QUERY;
2545         work->new_state = RDMA_CM_ADDR_RESOLVED;
2546         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2547 }
2548
2549 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
2550 {
2551         struct rdma_route *route = &id_priv->id.route;
2552         struct cma_work *work;
2553         int ret;
2554
2555         work = kzalloc(sizeof *work, GFP_KERNEL);
2556         if (!work)
2557                 return -ENOMEM;
2558
2559         cma_init_resolve_route_work(work, id_priv);
2560
2561         if (!route->path_rec)
2562                 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2563         if (!route->path_rec) {
2564                 ret = -ENOMEM;
2565                 goto err1;
2566         }
2567
2568         ret = cma_query_ib_route(id_priv, timeout_ms, work);
2569         if (ret)
2570                 goto err2;
2571
2572         return 0;
2573 err2:
2574         kfree(route->path_rec);
2575         route->path_rec = NULL;
2576 err1:
2577         kfree(work);
2578         return ret;
2579 }
2580
2581 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2582                                            unsigned long supported_gids,
2583                                            enum ib_gid_type default_gid)
2584 {
2585         if ((network_type == RDMA_NETWORK_IPV4 ||
2586              network_type == RDMA_NETWORK_IPV6) &&
2587             test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2588                 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2589
2590         return default_gid;
2591 }
2592
2593 /*
2594  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2595  * path record type based on GID type.
2596  * It also sets up other L2 fields which includes destination mac address
2597  * netdev ifindex, of the path record.
2598  * It returns the netdev of the bound interface for this path record entry.
2599  */
2600 static struct net_device *
2601 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2602 {
2603         struct rdma_route *route = &id_priv->id.route;
2604         enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2605         struct rdma_addr *addr = &route->addr;
2606         unsigned long supported_gids;
2607         struct net_device *ndev;
2608
2609         if (!addr->dev_addr.bound_dev_if)
2610                 return NULL;
2611
2612         ndev = dev_get_by_index(addr->dev_addr.net,
2613                                 addr->dev_addr.bound_dev_if);
2614         if (!ndev)
2615                 return NULL;
2616
2617         supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2618                                                     id_priv->id.port_num);
2619         gid_type = cma_route_gid_type(addr->dev_addr.network,
2620                                       supported_gids,
2621                                       id_priv->gid_type);
2622         /* Use the hint from IP Stack to select GID Type */
2623         if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2624                 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2625         route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2626
2627         route->path_rec->roce.route_resolved = true;
2628         sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2629         return ndev;
2630 }
2631
2632 int rdma_set_ib_path(struct rdma_cm_id *id,
2633                      struct sa_path_rec *path_rec)
2634 {
2635         struct rdma_id_private *id_priv;
2636         struct net_device *ndev;
2637         int ret;
2638
2639         id_priv = container_of(id, struct rdma_id_private, id);
2640         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2641                            RDMA_CM_ROUTE_RESOLVED))
2642                 return -EINVAL;
2643
2644         id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2645                                      GFP_KERNEL);
2646         if (!id->route.path_rec) {
2647                 ret = -ENOMEM;
2648                 goto err;
2649         }
2650
2651         if (rdma_protocol_roce(id->device, id->port_num)) {
2652                 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2653                 if (!ndev) {
2654                         ret = -ENODEV;
2655                         goto err_free;
2656                 }
2657                 dev_put(ndev);
2658         }
2659
2660         id->route.num_paths = 1;
2661         return 0;
2662
2663 err_free:
2664         kfree(id->route.path_rec);
2665         id->route.path_rec = NULL;
2666 err:
2667         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2668         return ret;
2669 }
2670 EXPORT_SYMBOL(rdma_set_ib_path);
2671
2672 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2673 {
2674         struct cma_work *work;
2675
2676         work = kzalloc(sizeof *work, GFP_KERNEL);
2677         if (!work)
2678                 return -ENOMEM;
2679
2680         cma_init_resolve_route_work(work, id_priv);
2681         queue_work(cma_wq, &work->work);
2682         return 0;
2683 }
2684
2685 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2686 {
2687         int prio;
2688         struct net_device *dev;
2689
2690         prio = rt_tos2priority(tos);
2691         dev = is_vlan_dev(ndev) ? vlan_dev_real_dev(ndev) : ndev;
2692         if (dev->num_tc)
2693                 return netdev_get_prio_tc_map(dev, prio);
2694
2695 #if IS_ENABLED(CONFIG_VLAN_8021Q)
2696         if (is_vlan_dev(ndev))
2697                 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
2698                         VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2699 #endif
2700         return 0;
2701 }
2702
2703 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2704 {
2705         struct rdma_route *route = &id_priv->id.route;
2706         struct rdma_addr *addr = &route->addr;
2707         struct cma_work *work;
2708         int ret;
2709         struct net_device *ndev;
2710
2711         u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
2712                                         rdma_start_port(id_priv->cma_dev->device)];
2713         u8 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
2714
2715
2716         work = kzalloc(sizeof *work, GFP_KERNEL);
2717         if (!work)
2718                 return -ENOMEM;
2719
2720         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2721         if (!route->path_rec) {
2722                 ret = -ENOMEM;
2723                 goto err1;
2724         }
2725
2726         route->num_paths = 1;
2727
2728         ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2729         if (!ndev) {
2730                 ret = -ENODEV;
2731                 goto err2;
2732         }
2733
2734         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2735                     &route->path_rec->sgid);
2736         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2737                     &route->path_rec->dgid);
2738
2739         if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2740                 /* TODO: get the hoplimit from the inet/inet6 device */
2741                 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2742         else
2743                 route->path_rec->hop_limit = 1;
2744         route->path_rec->reversible = 1;
2745         route->path_rec->pkey = cpu_to_be16(0xffff);
2746         route->path_rec->mtu_selector = IB_SA_EQ;
2747         route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
2748         route->path_rec->traffic_class = tos;
2749         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2750         route->path_rec->rate_selector = IB_SA_EQ;
2751         route->path_rec->rate = iboe_get_rate(ndev);
2752         dev_put(ndev);
2753         route->path_rec->packet_life_time_selector = IB_SA_EQ;
2754         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2755         if (!route->path_rec->mtu) {
2756                 ret = -EINVAL;
2757                 goto err2;
2758         }
2759
2760         cma_init_resolve_route_work(work, id_priv);
2761         queue_work(cma_wq, &work->work);
2762
2763         return 0;
2764
2765 err2:
2766         kfree(route->path_rec);
2767         route->path_rec = NULL;
2768         route->num_paths = 0;
2769 err1:
2770         kfree(work);
2771         return ret;
2772 }
2773
2774 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2775 {
2776         struct rdma_id_private *id_priv;
2777         int ret;
2778
2779         id_priv = container_of(id, struct rdma_id_private, id);
2780         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2781                 return -EINVAL;
2782
2783         atomic_inc(&id_priv->refcount);
2784         if (rdma_cap_ib_sa(id->device, id->port_num))
2785                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2786         else if (rdma_protocol_roce(id->device, id->port_num))
2787                 ret = cma_resolve_iboe_route(id_priv);
2788         else if (rdma_protocol_iwarp(id->device, id->port_num))
2789                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2790         else
2791                 ret = -ENOSYS;
2792
2793         if (ret)
2794                 goto err;
2795
2796         return 0;
2797 err:
2798         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2799         cma_deref_id(id_priv);
2800         return ret;
2801 }
2802 EXPORT_SYMBOL(rdma_resolve_route);
2803
2804 static void cma_set_loopback(struct sockaddr *addr)
2805 {
2806         switch (addr->sa_family) {
2807         case AF_INET:
2808                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2809                 break;
2810         case AF_INET6:
2811                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2812                               0, 0, 0, htonl(1));
2813                 break;
2814         default:
2815                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2816                             0, 0, 0, htonl(1));
2817                 break;
2818         }
2819 }
2820
2821 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2822 {
2823         struct cma_device *cma_dev, *cur_dev;
2824         union ib_gid gid;
2825         enum ib_port_state port_state;
2826         u16 pkey;
2827         int ret;
2828         u8 p;
2829
2830         cma_dev = NULL;
2831         mutex_lock(&lock);
2832         list_for_each_entry(cur_dev, &dev_list, list) {
2833                 if (cma_family(id_priv) == AF_IB &&
2834                     !rdma_cap_ib_cm(cur_dev->device, 1))
2835                         continue;
2836
2837                 if (!cma_dev)
2838                         cma_dev = cur_dev;
2839
2840                 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2841                         if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
2842                             port_state == IB_PORT_ACTIVE) {
2843                                 cma_dev = cur_dev;
2844                                 goto port_found;
2845                         }
2846                 }
2847         }
2848
2849         if (!cma_dev) {
2850                 ret = -ENODEV;
2851                 goto out;
2852         }
2853
2854         p = 1;
2855
2856 port_found:
2857         ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
2858         if (ret)
2859                 goto out;
2860
2861         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2862         if (ret)
2863                 goto out;
2864
2865         id_priv->id.route.addr.dev_addr.dev_type =
2866                 (rdma_protocol_ib(cma_dev->device, p)) ?
2867                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2868
2869         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2870         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2871         id_priv->id.port_num = p;
2872         cma_attach_to_dev(id_priv, cma_dev);
2873         cma_set_loopback(cma_src_addr(id_priv));
2874 out:
2875         mutex_unlock(&lock);
2876         return ret;
2877 }
2878
2879 static void addr_handler(int status, struct sockaddr *src_addr,
2880                          struct rdma_dev_addr *dev_addr, void *context)
2881 {
2882         struct rdma_id_private *id_priv = context;
2883         struct rdma_cm_event event = {};
2884         struct sockaddr *addr;
2885         struct sockaddr_storage old_addr;
2886
2887         mutex_lock(&id_priv->handler_mutex);
2888         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2889                            RDMA_CM_ADDR_RESOLVED))
2890                 goto out;
2891
2892         /*
2893          * Store the previous src address, so that if we fail to acquire
2894          * matching rdma device, old address can be restored back, which helps
2895          * to cancel the cma listen operation correctly.
2896          */
2897         addr = cma_src_addr(id_priv);
2898         memcpy(&old_addr, addr, rdma_addr_size(addr));
2899         memcpy(addr, src_addr, rdma_addr_size(src_addr));
2900         if (!status && !id_priv->cma_dev) {
2901                 status = cma_acquire_dev(id_priv, NULL);
2902                 if (status)
2903                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
2904                                              status);
2905         } else if (status) {
2906                 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
2907         }
2908
2909         if (status) {
2910                 memcpy(addr, &old_addr,
2911                        rdma_addr_size((struct sockaddr *)&old_addr));
2912                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2913                                    RDMA_CM_ADDR_BOUND))
2914                         goto out;
2915                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2916                 event.status = status;
2917         } else
2918                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2919
2920         if (id_priv->id.event_handler(&id_priv->id, &event)) {
2921                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2922                 mutex_unlock(&id_priv->handler_mutex);
2923                 cma_deref_id(id_priv);
2924                 rdma_destroy_id(&id_priv->id);
2925                 return;
2926         }
2927 out:
2928         mutex_unlock(&id_priv->handler_mutex);
2929         cma_deref_id(id_priv);
2930 }
2931
2932 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2933 {
2934         struct cma_work *work;
2935         union ib_gid gid;
2936         int ret;
2937
2938         work = kzalloc(sizeof *work, GFP_KERNEL);
2939         if (!work)
2940                 return -ENOMEM;
2941
2942         if (!id_priv->cma_dev) {
2943                 ret = cma_bind_loopback(id_priv);
2944                 if (ret)
2945                         goto err;
2946         }
2947
2948         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2949         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2950
2951         cma_init_resolve_addr_work(work, id_priv);
2952         queue_work(cma_wq, &work->work);
2953         return 0;
2954 err:
2955         kfree(work);
2956         return ret;
2957 }
2958
2959 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2960 {
2961         struct cma_work *work;
2962         int ret;
2963
2964         work = kzalloc(sizeof *work, GFP_KERNEL);
2965         if (!work)
2966                 return -ENOMEM;
2967
2968         if (!id_priv->cma_dev) {
2969                 ret = cma_resolve_ib_dev(id_priv);
2970                 if (ret)
2971                         goto err;
2972         }
2973
2974         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2975                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2976
2977         cma_init_resolve_addr_work(work, id_priv);
2978         queue_work(cma_wq, &work->work);
2979         return 0;
2980 err:
2981         kfree(work);
2982         return ret;
2983 }
2984
2985 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2986                          const struct sockaddr *dst_addr)
2987 {
2988         if (!src_addr || !src_addr->sa_family) {
2989                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2990                 src_addr->sa_family = dst_addr->sa_family;
2991                 if (IS_ENABLED(CONFIG_IPV6) &&
2992                     dst_addr->sa_family == AF_INET6) {
2993                         struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
2994                         struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
2995                         src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
2996                         if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
2997                                 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
2998                 } else if (dst_addr->sa_family == AF_IB) {
2999                         ((struct sockaddr_ib *) src_addr)->sib_pkey =
3000                                 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
3001                 }
3002         }
3003         return rdma_bind_addr(id, src_addr);
3004 }
3005
3006 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3007                       const struct sockaddr *dst_addr, int timeout_ms)
3008 {
3009         struct rdma_id_private *id_priv;
3010         int ret;
3011
3012         id_priv = container_of(id, struct rdma_id_private, id);
3013         if (id_priv->state == RDMA_CM_IDLE) {
3014                 ret = cma_bind_addr(id, src_addr, dst_addr);
3015                 if (ret)
3016                         return ret;
3017         }
3018
3019         if (cma_family(id_priv) != dst_addr->sa_family)
3020                 return -EINVAL;
3021
3022         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
3023                 return -EINVAL;
3024
3025         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3026         atomic_inc(&id_priv->refcount);
3027         if (cma_any_addr(dst_addr)) {
3028                 ret = cma_resolve_loopback(id_priv);
3029         } else {
3030                 if (dst_addr->sa_family == AF_IB) {
3031                         ret = cma_resolve_ib_addr(id_priv);
3032                 } else {
3033                         ret = rdma_resolve_ip(cma_src_addr(id_priv),
3034                                               dst_addr, &id->route.addr.dev_addr,
3035                                               timeout_ms, addr_handler, id_priv);
3036                 }
3037         }
3038         if (ret)
3039                 goto err;
3040
3041         return 0;
3042 err:
3043         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3044         cma_deref_id(id_priv);
3045         return ret;
3046 }
3047 EXPORT_SYMBOL(rdma_resolve_addr);
3048
3049 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3050 {
3051         struct rdma_id_private *id_priv;
3052         unsigned long flags;
3053         int ret;
3054
3055         id_priv = container_of(id, struct rdma_id_private, id);
3056         spin_lock_irqsave(&id_priv->lock, flags);
3057         if (reuse || id_priv->state == RDMA_CM_IDLE) {
3058                 id_priv->reuseaddr = reuse;
3059                 ret = 0;
3060         } else {
3061                 ret = -EINVAL;
3062         }
3063         spin_unlock_irqrestore(&id_priv->lock, flags);
3064         return ret;
3065 }
3066 EXPORT_SYMBOL(rdma_set_reuseaddr);
3067
3068 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3069 {
3070         struct rdma_id_private *id_priv;
3071         unsigned long flags;
3072         int ret;
3073
3074         id_priv = container_of(id, struct rdma_id_private, id);
3075         spin_lock_irqsave(&id_priv->lock, flags);
3076         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3077                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3078                 id_priv->afonly = afonly;
3079                 ret = 0;
3080         } else {
3081                 ret = -EINVAL;
3082         }
3083         spin_unlock_irqrestore(&id_priv->lock, flags);
3084         return ret;
3085 }
3086 EXPORT_SYMBOL(rdma_set_afonly);
3087
3088 static void cma_bind_port(struct rdma_bind_list *bind_list,
3089                           struct rdma_id_private *id_priv)
3090 {
3091         struct sockaddr *addr;
3092         struct sockaddr_ib *sib;
3093         u64 sid, mask;
3094         __be16 port;
3095
3096         lockdep_assert_held(&lock);
3097
3098         addr = cma_src_addr(id_priv);
3099         port = htons(bind_list->port);
3100
3101         switch (addr->sa_family) {
3102         case AF_INET:
3103                 ((struct sockaddr_in *) addr)->sin_port = port;
3104                 break;
3105         case AF_INET6:
3106                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3107                 break;
3108         case AF_IB:
3109                 sib = (struct sockaddr_ib *) addr;
3110                 sid = be64_to_cpu(sib->sib_sid);
3111                 mask = be64_to_cpu(sib->sib_sid_mask);
3112                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3113                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3114                 break;
3115         }
3116         id_priv->bind_list = bind_list;
3117         hlist_add_head(&id_priv->node, &bind_list->owners);
3118 }
3119
3120 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3121                           struct rdma_id_private *id_priv, unsigned short snum)
3122 {
3123         struct rdma_bind_list *bind_list;
3124         int ret;
3125
3126         lockdep_assert_held(&lock);
3127
3128         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3129         if (!bind_list)
3130                 return -ENOMEM;
3131
3132         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3133                            snum);
3134         if (ret < 0)
3135                 goto err;
3136
3137         bind_list->ps = ps;
3138         bind_list->port = (unsigned short)ret;
3139         cma_bind_port(bind_list, id_priv);
3140         return 0;
3141 err:
3142         kfree(bind_list);
3143         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3144 }
3145
3146 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3147                               struct rdma_id_private *id_priv)
3148 {
3149         struct rdma_id_private *cur_id;
3150         struct sockaddr  *daddr = cma_dst_addr(id_priv);
3151         struct sockaddr  *saddr = cma_src_addr(id_priv);
3152         __be16 dport = cma_port(daddr);
3153
3154         lockdep_assert_held(&lock);
3155
3156         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3157                 struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3158                 struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3159                 __be16 cur_dport = cma_port(cur_daddr);
3160
3161                 if (id_priv == cur_id)
3162                         continue;
3163
3164                 /* different dest port -> unique */
3165                 if (!cma_any_port(daddr) &&
3166                     !cma_any_port(cur_daddr) &&
3167                     (dport != cur_dport))
3168                         continue;
3169
3170                 /* different src address -> unique */
3171                 if (!cma_any_addr(saddr) &&
3172                     !cma_any_addr(cur_saddr) &&
3173                     cma_addr_cmp(saddr, cur_saddr))
3174                         continue;
3175
3176                 /* different dst address -> unique */
3177                 if (!cma_any_addr(daddr) &&
3178                     !cma_any_addr(cur_daddr) &&
3179                     cma_addr_cmp(daddr, cur_daddr))
3180                         continue;
3181
3182                 return -EADDRNOTAVAIL;
3183         }
3184         return 0;
3185 }
3186
3187 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3188                               struct rdma_id_private *id_priv)
3189 {
3190         static unsigned int last_used_port;
3191         int low, high, remaining;
3192         unsigned int rover;
3193         struct net *net = id_priv->id.route.addr.dev_addr.net;
3194
3195         lockdep_assert_held(&lock);
3196
3197         inet_get_local_port_range(net, &low, &high);
3198         remaining = (high - low) + 1;
3199         rover = prandom_u32() % remaining + low;
3200 retry:
3201         if (last_used_port != rover) {
3202                 struct rdma_bind_list *bind_list;
3203                 int ret;
3204
3205                 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3206
3207                 if (!bind_list) {
3208                         ret = cma_alloc_port(ps, id_priv, rover);
3209                 } else {
3210                         ret = cma_port_is_unique(bind_list, id_priv);
3211                         if (!ret)
3212                                 cma_bind_port(bind_list, id_priv);
3213                 }
3214                 /*
3215                  * Remember previously used port number in order to avoid
3216                  * re-using same port immediately after it is closed.
3217                  */
3218                 if (!ret)
3219                         last_used_port = rover;
3220                 if (ret != -EADDRNOTAVAIL)
3221                         return ret;
3222         }
3223         if (--remaining) {
3224                 rover++;
3225                 if ((rover < low) || (rover > high))
3226                         rover = low;
3227                 goto retry;
3228         }
3229         return -EADDRNOTAVAIL;
3230 }
3231
3232 /*
3233  * Check that the requested port is available.  This is called when trying to
3234  * bind to a specific port, or when trying to listen on a bound port.  In
3235  * the latter case, the provided id_priv may already be on the bind_list, but
3236  * we still need to check that it's okay to start listening.
3237  */
3238 static int cma_check_port(struct rdma_bind_list *bind_list,
3239                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
3240 {
3241         struct rdma_id_private *cur_id;
3242         struct sockaddr *addr, *cur_addr;
3243
3244         lockdep_assert_held(&lock);
3245
3246         addr = cma_src_addr(id_priv);
3247         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3248                 if (id_priv == cur_id)
3249                         continue;
3250
3251                 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
3252                     cur_id->reuseaddr)
3253                         continue;
3254
3255                 cur_addr = cma_src_addr(cur_id);
3256                 if (id_priv->afonly && cur_id->afonly &&
3257                     (addr->sa_family != cur_addr->sa_family))
3258                         continue;
3259
3260                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3261                         return -EADDRNOTAVAIL;
3262
3263                 if (!cma_addr_cmp(addr, cur_addr))
3264                         return -EADDRINUSE;
3265         }
3266         return 0;
3267 }
3268
3269 static int cma_use_port(enum rdma_ucm_port_space ps,
3270                         struct rdma_id_private *id_priv)
3271 {
3272         struct rdma_bind_list *bind_list;
3273         unsigned short snum;
3274         int ret;
3275
3276         lockdep_assert_held(&lock);
3277
3278         snum = ntohs(cma_port(cma_src_addr(id_priv)));
3279         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3280                 return -EACCES;
3281
3282         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3283         if (!bind_list) {
3284                 ret = cma_alloc_port(ps, id_priv, snum);
3285         } else {
3286                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3287                 if (!ret)
3288                         cma_bind_port(bind_list, id_priv);
3289         }
3290         return ret;
3291 }
3292
3293 static int cma_bind_listen(struct rdma_id_private *id_priv)
3294 {
3295         struct rdma_bind_list *bind_list = id_priv->bind_list;
3296         int ret = 0;
3297
3298         mutex_lock(&lock);
3299         if (bind_list->owners.first->next)
3300                 ret = cma_check_port(bind_list, id_priv, 0);
3301         mutex_unlock(&lock);
3302         return ret;
3303 }
3304
3305 static enum rdma_ucm_port_space
3306 cma_select_inet_ps(struct rdma_id_private *id_priv)
3307 {
3308         switch (id_priv->id.ps) {
3309         case RDMA_PS_TCP:
3310         case RDMA_PS_UDP:
3311         case RDMA_PS_IPOIB:
3312         case RDMA_PS_IB:
3313                 return id_priv->id.ps;
3314         default:
3315
3316                 return 0;
3317         }
3318 }
3319
3320 static enum rdma_ucm_port_space
3321 cma_select_ib_ps(struct rdma_id_private *id_priv)
3322 {
3323         enum rdma_ucm_port_space ps = 0;
3324         struct sockaddr_ib *sib;
3325         u64 sid_ps, mask, sid;
3326
3327         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3328         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3329         sid = be64_to_cpu(sib->sib_sid) & mask;
3330
3331         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3332                 sid_ps = RDMA_IB_IP_PS_IB;
3333                 ps = RDMA_PS_IB;
3334         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3335                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3336                 sid_ps = RDMA_IB_IP_PS_TCP;
3337                 ps = RDMA_PS_TCP;
3338         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3339                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3340                 sid_ps = RDMA_IB_IP_PS_UDP;
3341                 ps = RDMA_PS_UDP;
3342         }
3343
3344         if (ps) {
3345                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3346                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3347                                                 be64_to_cpu(sib->sib_sid_mask));
3348         }
3349         return ps;
3350 }
3351
3352 static int cma_get_port(struct rdma_id_private *id_priv)
3353 {
3354         enum rdma_ucm_port_space ps;
3355         int ret;
3356
3357         if (cma_family(id_priv) != AF_IB)
3358                 ps = cma_select_inet_ps(id_priv);
3359         else
3360                 ps = cma_select_ib_ps(id_priv);
3361         if (!ps)
3362                 return -EPROTONOSUPPORT;
3363
3364         mutex_lock(&lock);
3365         if (cma_any_port(cma_src_addr(id_priv)))
3366                 ret = cma_alloc_any_port(ps, id_priv);
3367         else
3368                 ret = cma_use_port(ps, id_priv);
3369         mutex_unlock(&lock);
3370
3371         return ret;
3372 }
3373
3374 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3375                                struct sockaddr *addr)
3376 {
3377 #if IS_ENABLED(CONFIG_IPV6)
3378         struct sockaddr_in6 *sin6;
3379
3380         if (addr->sa_family != AF_INET6)
3381                 return 0;
3382
3383         sin6 = (struct sockaddr_in6 *) addr;
3384
3385         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3386                 return 0;
3387
3388         if (!sin6->sin6_scope_id)
3389                         return -EINVAL;
3390
3391         dev_addr->bound_dev_if = sin6->sin6_scope_id;
3392 #endif
3393         return 0;
3394 }
3395
3396 int rdma_listen(struct rdma_cm_id *id, int backlog)
3397 {
3398         struct rdma_id_private *id_priv;
3399         int ret;
3400
3401         id_priv = container_of(id, struct rdma_id_private, id);
3402         if (id_priv->state == RDMA_CM_IDLE) {
3403                 id->route.addr.src_addr.ss_family = AF_INET;
3404                 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3405                 if (ret)
3406                         return ret;
3407         }
3408
3409         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3410                 return -EINVAL;
3411
3412         if (id_priv->reuseaddr) {
3413                 ret = cma_bind_listen(id_priv);
3414                 if (ret)
3415                         goto err;
3416         }
3417
3418         id_priv->backlog = backlog;
3419         if (id->device) {
3420                 if (rdma_cap_ib_cm(id->device, 1)) {
3421                         ret = cma_ib_listen(id_priv);
3422                         if (ret)
3423                                 goto err;
3424                 } else if (rdma_cap_iw_cm(id->device, 1)) {
3425                         ret = cma_iw_listen(id_priv, backlog);
3426                         if (ret)
3427                                 goto err;
3428                 } else {
3429                         ret = -ENOSYS;
3430                         goto err;
3431                 }
3432         } else
3433                 cma_listen_on_all(id_priv);
3434
3435         return 0;
3436 err:
3437         id_priv->backlog = 0;
3438         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3439         return ret;
3440 }
3441 EXPORT_SYMBOL(rdma_listen);
3442
3443 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3444 {
3445         struct rdma_id_private *id_priv;
3446         int ret;
3447         struct sockaddr  *daddr;
3448
3449         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3450             addr->sa_family != AF_IB)
3451                 return -EAFNOSUPPORT;
3452
3453         id_priv = container_of(id, struct rdma_id_private, id);
3454         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3455                 return -EINVAL;
3456
3457         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3458         if (ret)
3459                 goto err1;
3460
3461         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3462         if (!cma_any_addr(addr)) {
3463                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3464                 if (ret)
3465                         goto err1;
3466
3467                 ret = cma_acquire_dev(id_priv, NULL);
3468                 if (ret)
3469                         goto err1;
3470         }
3471
3472         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3473                 if (addr->sa_family == AF_INET)
3474                         id_priv->afonly = 1;
3475 #if IS_ENABLED(CONFIG_IPV6)
3476                 else if (addr->sa_family == AF_INET6) {
3477                         struct net *net = id_priv->id.route.addr.dev_addr.net;
3478
3479                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
3480                 }
3481 #endif
3482         }
3483         daddr = cma_dst_addr(id_priv);
3484         daddr->sa_family = addr->sa_family;
3485
3486         ret = cma_get_port(id_priv);
3487         if (ret)
3488                 goto err2;
3489
3490         return 0;
3491 err2:
3492         rdma_restrack_del(&id_priv->res);
3493         if (id_priv->cma_dev)
3494                 cma_release_dev(id_priv);
3495 err1:
3496         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3497         return ret;
3498 }
3499 EXPORT_SYMBOL(rdma_bind_addr);
3500
3501 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3502 {
3503         struct cma_hdr *cma_hdr;
3504
3505         cma_hdr = hdr;
3506         cma_hdr->cma_version = CMA_VERSION;
3507         if (cma_family(id_priv) == AF_INET) {
3508                 struct sockaddr_in *src4, *dst4;
3509
3510                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3511                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3512
3513                 cma_set_ip_ver(cma_hdr, 4);
3514                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3515                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3516                 cma_hdr->port = src4->sin_port;
3517         } else if (cma_family(id_priv) == AF_INET6) {
3518                 struct sockaddr_in6 *src6, *dst6;
3519
3520                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3521                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3522
3523                 cma_set_ip_ver(cma_hdr, 6);
3524                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3525                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3526                 cma_hdr->port = src6->sin6_port;
3527         }
3528         return 0;
3529 }
3530
3531 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3532                                 const struct ib_cm_event *ib_event)
3533 {
3534         struct rdma_id_private *id_priv = cm_id->context;
3535         struct rdma_cm_event event = {};
3536         const struct ib_cm_sidr_rep_event_param *rep =
3537                                 &ib_event->param.sidr_rep_rcvd;
3538         int ret = 0;
3539
3540         mutex_lock(&id_priv->handler_mutex);
3541         if (id_priv->state != RDMA_CM_CONNECT)
3542                 goto out;
3543
3544         switch (ib_event->event) {
3545         case IB_CM_SIDR_REQ_ERROR:
3546                 event.event = RDMA_CM_EVENT_UNREACHABLE;
3547                 event.status = -ETIMEDOUT;
3548                 break;
3549         case IB_CM_SIDR_REP_RECEIVED:
3550                 event.param.ud.private_data = ib_event->private_data;
3551                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3552                 if (rep->status != IB_SIDR_SUCCESS) {
3553                         event.event = RDMA_CM_EVENT_UNREACHABLE;
3554                         event.status = ib_event->param.sidr_rep_rcvd.status;
3555                         pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3556                                              event.status);
3557                         break;
3558                 }
3559                 ret = cma_set_qkey(id_priv, rep->qkey);
3560                 if (ret) {
3561                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3562                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
3563                         event.status = ret;
3564                         break;
3565                 }
3566                 ib_init_ah_attr_from_path(id_priv->id.device,
3567                                           id_priv->id.port_num,
3568                                           id_priv->id.route.path_rec,
3569                                           &event.param.ud.ah_attr,
3570                                           rep->sgid_attr);
3571                 event.param.ud.qp_num = rep->qpn;
3572                 event.param.ud.qkey = rep->qkey;
3573                 event.event = RDMA_CM_EVENT_ESTABLISHED;
3574                 event.status = 0;
3575                 break;
3576         default:
3577                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3578                        ib_event->event);
3579                 goto out;
3580         }
3581
3582         ret = id_priv->id.event_handler(&id_priv->id, &event);
3583
3584         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
3585         if (ret) {
3586                 /* Destroy the CM ID by returning a non-zero value. */
3587                 id_priv->cm_id.ib = NULL;
3588                 cma_exch(id_priv, RDMA_CM_DESTROYING);
3589                 mutex_unlock(&id_priv->handler_mutex);
3590                 rdma_destroy_id(&id_priv->id);
3591                 return ret;
3592         }
3593 out:
3594         mutex_unlock(&id_priv->handler_mutex);
3595         return ret;
3596 }
3597
3598 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3599                               struct rdma_conn_param *conn_param)
3600 {
3601         struct ib_cm_sidr_req_param req;
3602         struct ib_cm_id *id;
3603         void *private_data;
3604         u8 offset;
3605         int ret;
3606
3607         memset(&req, 0, sizeof req);
3608         offset = cma_user_data_offset(id_priv);
3609         req.private_data_len = offset + conn_param->private_data_len;
3610         if (req.private_data_len < conn_param->private_data_len)
3611                 return -EINVAL;
3612
3613         if (req.private_data_len) {
3614                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3615                 if (!private_data)
3616                         return -ENOMEM;
3617         } else {
3618                 private_data = NULL;
3619         }
3620
3621         if (conn_param->private_data && conn_param->private_data_len)
3622                 memcpy(private_data + offset, conn_param->private_data,
3623                        conn_param->private_data_len);
3624
3625         if (private_data) {
3626                 ret = cma_format_hdr(private_data, id_priv);
3627                 if (ret)
3628                         goto out;
3629                 req.private_data = private_data;
3630         }
3631
3632         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3633                              id_priv);
3634         if (IS_ERR(id)) {
3635                 ret = PTR_ERR(id);
3636                 goto out;
3637         }
3638         id_priv->cm_id.ib = id;
3639
3640         req.path = id_priv->id.route.path_rec;
3641         req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3642         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3643         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3644         req.max_cm_retries = CMA_MAX_CM_RETRIES;
3645
3646         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3647         if (ret) {
3648                 ib_destroy_cm_id(id_priv->cm_id.ib);
3649                 id_priv->cm_id.ib = NULL;
3650         }
3651 out:
3652         kfree(private_data);
3653         return ret;
3654 }
3655
3656 static int cma_connect_ib(struct rdma_id_private *id_priv,
3657                           struct rdma_conn_param *conn_param)
3658 {
3659         struct ib_cm_req_param req;
3660         struct rdma_route *route;
3661         void *private_data;
3662         struct ib_cm_id *id;
3663         u8 offset;
3664         int 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(private_data + offset, conn_param->private_data,
3682                        conn_param->private_data_len);
3683
3684         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3685         if (IS_ERR(id)) {
3686                 ret = PTR_ERR(id);
3687                 goto out;
3688         }
3689         id_priv->cm_id.ib = id;
3690
3691         route = &id_priv->id.route;
3692         if (private_data) {
3693                 ret = cma_format_hdr(private_data, id_priv);
3694                 if (ret)
3695                         goto out;
3696                 req.private_data = private_data;
3697         }
3698
3699         req.primary_path = &route->path_rec[0];
3700         if (route->num_paths == 2)
3701                 req.alternate_path = &route->path_rec[1];
3702
3703         req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3704         /* Alternate path SGID attribute currently unsupported */
3705         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3706         req.qp_num = id_priv->qp_num;
3707         req.qp_type = id_priv->id.qp_type;
3708         req.starting_psn = id_priv->seq_num;
3709         req.responder_resources = conn_param->responder_resources;
3710         req.initiator_depth = conn_param->initiator_depth;
3711         req.flow_control = conn_param->flow_control;
3712         req.retry_count = min_t(u8, 7, conn_param->retry_count);
3713         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3714         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3715         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3716         req.max_cm_retries = CMA_MAX_CM_RETRIES;
3717         req.srq = id_priv->srq ? 1 : 0;
3718
3719         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3720 out:
3721         if (ret && !IS_ERR(id)) {
3722                 ib_destroy_cm_id(id);
3723                 id_priv->cm_id.ib = NULL;
3724         }
3725
3726         kfree(private_data);
3727         return ret;
3728 }
3729
3730 static int cma_connect_iw(struct rdma_id_private *id_priv,
3731                           struct rdma_conn_param *conn_param)
3732 {
3733         struct iw_cm_id *cm_id;
3734         int ret;
3735         struct iw_cm_conn_param iw_param;
3736
3737         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3738         if (IS_ERR(cm_id))
3739                 return PTR_ERR(cm_id);
3740
3741         cm_id->tos = id_priv->tos;
3742         id_priv->cm_id.iw = cm_id;
3743
3744         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3745                rdma_addr_size(cma_src_addr(id_priv)));
3746         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3747                rdma_addr_size(cma_dst_addr(id_priv)));
3748
3749         ret = cma_modify_qp_rtr(id_priv, conn_param);
3750         if (ret)
3751                 goto out;
3752
3753         if (conn_param) {
3754                 iw_param.ord = conn_param->initiator_depth;
3755                 iw_param.ird = conn_param->responder_resources;
3756                 iw_param.private_data = conn_param->private_data;
3757                 iw_param.private_data_len = conn_param->private_data_len;
3758                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3759         } else {
3760                 memset(&iw_param, 0, sizeof iw_param);
3761                 iw_param.qpn = id_priv->qp_num;
3762         }
3763         ret = iw_cm_connect(cm_id, &iw_param);
3764 out:
3765         if (ret) {
3766                 iw_destroy_cm_id(cm_id);
3767                 id_priv->cm_id.iw = NULL;
3768         }
3769         return ret;
3770 }
3771
3772 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3773 {
3774         struct rdma_id_private *id_priv;
3775         int ret;
3776
3777         id_priv = container_of(id, struct rdma_id_private, id);
3778         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3779                 return -EINVAL;
3780
3781         if (!id->qp) {
3782                 id_priv->qp_num = conn_param->qp_num;
3783                 id_priv->srq = conn_param->srq;
3784         }
3785
3786         if (rdma_cap_ib_cm(id->device, id->port_num)) {
3787                 if (id->qp_type == IB_QPT_UD)
3788                         ret = cma_resolve_ib_udp(id_priv, conn_param);
3789                 else
3790                         ret = cma_connect_ib(id_priv, conn_param);
3791         } else if (rdma_cap_iw_cm(id->device, id->port_num))
3792                 ret = cma_connect_iw(id_priv, conn_param);
3793         else
3794                 ret = -ENOSYS;
3795         if (ret)
3796                 goto err;
3797
3798         return 0;
3799 err:
3800         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3801         return ret;
3802 }
3803 EXPORT_SYMBOL(rdma_connect);
3804
3805 static int cma_accept_ib(struct rdma_id_private *id_priv,
3806                          struct rdma_conn_param *conn_param)
3807 {
3808         struct ib_cm_rep_param rep;
3809         int ret;
3810
3811         ret = cma_modify_qp_rtr(id_priv, conn_param);
3812         if (ret)
3813                 goto out;
3814
3815         ret = cma_modify_qp_rts(id_priv, conn_param);
3816         if (ret)
3817                 goto out;
3818
3819         memset(&rep, 0, sizeof rep);
3820         rep.qp_num = id_priv->qp_num;
3821         rep.starting_psn = id_priv->seq_num;
3822         rep.private_data = conn_param->private_data;
3823         rep.private_data_len = conn_param->private_data_len;
3824         rep.responder_resources = conn_param->responder_resources;
3825         rep.initiator_depth = conn_param->initiator_depth;
3826         rep.failover_accepted = 0;
3827         rep.flow_control = conn_param->flow_control;
3828         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3829         rep.srq = id_priv->srq ? 1 : 0;
3830
3831         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3832 out:
3833         return ret;
3834 }
3835
3836 static int cma_accept_iw(struct rdma_id_private *id_priv,
3837                   struct rdma_conn_param *conn_param)
3838 {
3839         struct iw_cm_conn_param iw_param;
3840         int ret;
3841
3842         if (!conn_param)
3843                 return -EINVAL;
3844
3845         ret = cma_modify_qp_rtr(id_priv, conn_param);
3846         if (ret)
3847                 return ret;
3848
3849         iw_param.ord = conn_param->initiator_depth;
3850         iw_param.ird = conn_param->responder_resources;
3851         iw_param.private_data = conn_param->private_data;
3852         iw_param.private_data_len = conn_param->private_data_len;
3853         if (id_priv->id.qp) {
3854                 iw_param.qpn = id_priv->qp_num;
3855         } else
3856                 iw_param.qpn = conn_param->qp_num;
3857
3858         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3859 }
3860
3861 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3862                              enum ib_cm_sidr_status status, u32 qkey,
3863                              const void *private_data, int private_data_len)
3864 {
3865         struct ib_cm_sidr_rep_param rep;
3866         int ret;
3867
3868         memset(&rep, 0, sizeof rep);
3869         rep.status = status;
3870         if (status == IB_SIDR_SUCCESS) {
3871                 ret = cma_set_qkey(id_priv, qkey);
3872                 if (ret)
3873                         return ret;
3874                 rep.qp_num = id_priv->qp_num;
3875                 rep.qkey = id_priv->qkey;
3876         }
3877         rep.private_data = private_data;
3878         rep.private_data_len = private_data_len;
3879
3880         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3881 }
3882
3883 int __rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
3884                   const char *caller)
3885 {
3886         struct rdma_id_private *id_priv;
3887         int ret;
3888
3889         id_priv = container_of(id, struct rdma_id_private, id);
3890
3891         if (caller)
3892                 id_priv->res.kern_name = caller;
3893         else
3894                 rdma_restrack_set_task(&id_priv->res, current);
3895
3896         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3897                 return -EINVAL;
3898
3899         if (!id->qp && conn_param) {
3900                 id_priv->qp_num = conn_param->qp_num;
3901                 id_priv->srq = conn_param->srq;
3902         }
3903
3904         if (rdma_cap_ib_cm(id->device, id->port_num)) {
3905                 if (id->qp_type == IB_QPT_UD) {
3906                         if (conn_param)
3907                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3908                                                         conn_param->qkey,
3909                                                         conn_param->private_data,
3910                                                         conn_param->private_data_len);
3911                         else
3912                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3913                                                         0, NULL, 0);
3914                 } else {
3915                         if (conn_param)
3916                                 ret = cma_accept_ib(id_priv, conn_param);
3917                         else
3918                                 ret = cma_rep_recv(id_priv);
3919                 }
3920         } else if (rdma_cap_iw_cm(id->device, id->port_num))
3921                 ret = cma_accept_iw(id_priv, conn_param);
3922         else
3923                 ret = -ENOSYS;
3924
3925         if (ret)
3926                 goto reject;
3927
3928         return 0;
3929 reject:
3930         cma_modify_qp_err(id_priv);
3931         rdma_reject(id, NULL, 0);
3932         return ret;
3933 }
3934 EXPORT_SYMBOL(__rdma_accept);
3935
3936 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3937 {
3938         struct rdma_id_private *id_priv;
3939         int ret;
3940
3941         id_priv = container_of(id, struct rdma_id_private, id);
3942         if (!id_priv->cm_id.ib)
3943                 return -EINVAL;
3944
3945         switch (id->device->node_type) {
3946         case RDMA_NODE_IB_CA:
3947                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3948                 break;
3949         default:
3950                 ret = 0;
3951                 break;
3952         }
3953         return ret;
3954 }
3955 EXPORT_SYMBOL(rdma_notify);
3956
3957 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3958                 u8 private_data_len)
3959 {
3960         struct rdma_id_private *id_priv;
3961         int ret;
3962
3963         id_priv = container_of(id, struct rdma_id_private, id);
3964         if (!id_priv->cm_id.ib)
3965                 return -EINVAL;
3966
3967         if (rdma_cap_ib_cm(id->device, id->port_num)) {
3968                 if (id->qp_type == IB_QPT_UD)
3969                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3970                                                 private_data, private_data_len);
3971                 else
3972                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
3973                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
3974                                              0, private_data, private_data_len);
3975         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3976                 ret = iw_cm_reject(id_priv->cm_id.iw,
3977                                    private_data, private_data_len);
3978         } else
3979                 ret = -ENOSYS;
3980
3981         return ret;
3982 }
3983 EXPORT_SYMBOL(rdma_reject);
3984
3985 int rdma_disconnect(struct rdma_cm_id *id)
3986 {
3987         struct rdma_id_private *id_priv;
3988         int ret;
3989
3990         id_priv = container_of(id, struct rdma_id_private, id);
3991         if (!id_priv->cm_id.ib)
3992                 return -EINVAL;
3993
3994         if (rdma_cap_ib_cm(id->device, id->port_num)) {
3995                 ret = cma_modify_qp_err(id_priv);
3996                 if (ret)
3997                         goto out;
3998                 /* Initiate or respond to a disconnect. */
3999                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
4000                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
4001         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4002                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4003         } else
4004                 ret = -EINVAL;
4005
4006 out:
4007         return ret;
4008 }
4009 EXPORT_SYMBOL(rdma_disconnect);
4010
4011 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4012 {
4013         struct rdma_id_private *id_priv;
4014         struct cma_multicast *mc = multicast->context;
4015         struct rdma_cm_event event = {};
4016         int ret = 0;
4017
4018         id_priv = mc->id_priv;
4019         mutex_lock(&id_priv->handler_mutex);
4020         if (id_priv->state != RDMA_CM_ADDR_BOUND &&
4021             id_priv->state != RDMA_CM_ADDR_RESOLVED)
4022                 goto out;
4023
4024         if (!status)
4025                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4026         else
4027                 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4028                                      status);
4029         event.status = status;
4030         event.param.ud.private_data = mc->context;
4031         if (!status) {
4032                 struct rdma_dev_addr *dev_addr =
4033                         &id_priv->id.route.addr.dev_addr;
4034                 struct net_device *ndev =
4035                         dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4036                 enum ib_gid_type gid_type =
4037                         id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4038                         rdma_start_port(id_priv->cma_dev->device)];
4039
4040                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
4041                 ret = ib_init_ah_from_mcmember(id_priv->id.device,
4042                                                id_priv->id.port_num,
4043                                                &multicast->rec,
4044                                                ndev, gid_type,
4045                                                &event.param.ud.ah_attr);
4046                 if (ret)
4047                         event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4048
4049                 event.param.ud.qp_num = 0xFFFFFF;
4050                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4051                 if (ndev)
4052                         dev_put(ndev);
4053         } else
4054                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4055
4056         ret = id_priv->id.event_handler(&id_priv->id, &event);
4057
4058         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4059         if (ret) {
4060                 cma_exch(id_priv, RDMA_CM_DESTROYING);
4061                 mutex_unlock(&id_priv->handler_mutex);
4062                 rdma_destroy_id(&id_priv->id);
4063                 return 0;
4064         }
4065
4066 out:
4067         mutex_unlock(&id_priv->handler_mutex);
4068         return 0;
4069 }
4070
4071 static void cma_set_mgid(struct rdma_id_private *id_priv,
4072                          struct sockaddr *addr, union ib_gid *mgid)
4073 {
4074         unsigned char mc_map[MAX_ADDR_LEN];
4075         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4076         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4077         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4078
4079         if (cma_any_addr(addr)) {
4080                 memset(mgid, 0, sizeof *mgid);
4081         } else if ((addr->sa_family == AF_INET6) &&
4082                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4083                                                                  0xFF10A01B)) {
4084                 /* IPv6 address is an SA assigned MGID. */
4085                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4086         } else if (addr->sa_family == AF_IB) {
4087                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4088         } else if (addr->sa_family == AF_INET6) {
4089                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4090                 if (id_priv->id.ps == RDMA_PS_UDP)
4091                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4092                 *mgid = *(union ib_gid *) (mc_map + 4);
4093         } else {
4094                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4095                 if (id_priv->id.ps == RDMA_PS_UDP)
4096                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4097                 *mgid = *(union ib_gid *) (mc_map + 4);
4098         }
4099 }
4100
4101 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4102                                  struct cma_multicast *mc)
4103 {
4104         struct ib_sa_mcmember_rec rec;
4105         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4106         ib_sa_comp_mask comp_mask;
4107         int ret;
4108
4109         ib_addr_get_mgid(dev_addr, &rec.mgid);
4110         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4111                                      &rec.mgid, &rec);
4112         if (ret)
4113                 return ret;
4114
4115         ret = cma_set_qkey(id_priv, 0);
4116         if (ret)
4117                 return ret;
4118
4119         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4120         rec.qkey = cpu_to_be32(id_priv->qkey);
4121         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4122         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4123         rec.join_state = mc->join_state;
4124
4125         if ((rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) &&
4126             (!ib_sa_sendonly_fullmem_support(&sa_client,
4127                                              id_priv->id.device,
4128                                              id_priv->id.port_num))) {
4129                 pr_warn("RDMA CM: %s port %u Unable to multicast join\n"
4130                         "RDMA CM: SM doesn't support Send Only Full Member option\n",
4131                         id_priv->id.device->name, id_priv->id.port_num);
4132                 return -EOPNOTSUPP;
4133         }
4134
4135         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4136                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4137                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4138                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
4139                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4140
4141         if (id_priv->id.ps == RDMA_PS_IPOIB)
4142                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4143                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4144                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4145                              IB_SA_MCMEMBER_REC_MTU |
4146                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
4147
4148         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4149                                                 id_priv->id.port_num, &rec,
4150                                                 comp_mask, GFP_KERNEL,
4151                                                 cma_ib_mc_handler, mc);
4152         return PTR_ERR_OR_ZERO(mc->multicast.ib);
4153 }
4154
4155 static void iboe_mcast_work_handler(struct work_struct *work)
4156 {
4157         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
4158         struct cma_multicast *mc = mw->mc;
4159         struct ib_sa_multicast *m = mc->multicast.ib;
4160
4161         mc->multicast.ib->context = mc;
4162         cma_ib_mc_handler(0, m);
4163         kref_put(&mc->mcref, release_mc);
4164         kfree(mw);
4165 }
4166
4167 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4168                               enum ib_gid_type gid_type)
4169 {
4170         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4171         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4172
4173         if (cma_any_addr(addr)) {
4174                 memset(mgid, 0, sizeof *mgid);
4175         } else if (addr->sa_family == AF_INET6) {
4176                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4177         } else {
4178                 mgid->raw[0] =
4179                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4180                 mgid->raw[1] =
4181                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4182                 mgid->raw[2] = 0;
4183                 mgid->raw[3] = 0;
4184                 mgid->raw[4] = 0;
4185                 mgid->raw[5] = 0;
4186                 mgid->raw[6] = 0;
4187                 mgid->raw[7] = 0;
4188                 mgid->raw[8] = 0;
4189                 mgid->raw[9] = 0;
4190                 mgid->raw[10] = 0xff;
4191                 mgid->raw[11] = 0xff;
4192                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4193         }
4194 }
4195
4196 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4197                                    struct cma_multicast *mc)
4198 {
4199         struct iboe_mcast_work *work;
4200         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4201         int err = 0;
4202         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4203         struct net_device *ndev = NULL;
4204         enum ib_gid_type gid_type;
4205         bool send_only;
4206
4207         send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4208
4209         if (cma_zero_addr((struct sockaddr *)&mc->addr))
4210                 return -EINVAL;
4211
4212         work = kzalloc(sizeof *work, GFP_KERNEL);
4213         if (!work)
4214                 return -ENOMEM;
4215
4216         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
4217         if (!mc->multicast.ib) {
4218                 err = -ENOMEM;
4219                 goto out1;
4220         }
4221
4222         gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4223                    rdma_start_port(id_priv->cma_dev->device)];
4224         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid, gid_type);
4225
4226         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
4227         if (id_priv->id.ps == RDMA_PS_UDP)
4228                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4229
4230         if (dev_addr->bound_dev_if)
4231                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4232         if (!ndev) {
4233                 err = -ENODEV;
4234                 goto out2;
4235         }
4236         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
4237         mc->multicast.ib->rec.hop_limit = 1;
4238         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
4239
4240         if (addr->sa_family == AF_INET) {
4241                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4242                         mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4243                         if (!send_only) {
4244                                 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
4245                                                     true);
4246                         }
4247                 }
4248         } else {
4249                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4250                         err = -ENOTSUPP;
4251         }
4252         dev_put(ndev);
4253         if (err || !mc->multicast.ib->rec.mtu) {
4254                 if (!err)
4255                         err = -EINVAL;
4256                 goto out2;
4257         }
4258         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4259                     &mc->multicast.ib->rec.port_gid);
4260         work->id = id_priv;
4261         work->mc = mc;
4262         INIT_WORK(&work->work, iboe_mcast_work_handler);
4263         kref_get(&mc->mcref);
4264         queue_work(cma_wq, &work->work);
4265
4266         return 0;
4267
4268 out2:
4269         kfree(mc->multicast.ib);
4270 out1:
4271         kfree(work);
4272         return err;
4273 }
4274
4275 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4276                         u8 join_state, void *context)
4277 {
4278         struct rdma_id_private *id_priv;
4279         struct cma_multicast *mc;
4280         int ret;
4281
4282         /* Not supported for kernel QPs */
4283         if (WARN_ON(id->qp))
4284                 return -EINVAL;
4285
4286         if (!id->device)
4287                 return -EINVAL;
4288
4289         id_priv = container_of(id, struct rdma_id_private, id);
4290         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4291             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4292                 return -EINVAL;
4293
4294         mc = kmalloc(sizeof *mc, GFP_KERNEL);
4295         if (!mc)
4296                 return -ENOMEM;
4297
4298         memcpy(&mc->addr, addr, rdma_addr_size(addr));
4299         mc->context = context;
4300         mc->id_priv = id_priv;
4301         mc->join_state = join_state;
4302
4303         if (rdma_protocol_roce(id->device, id->port_num)) {
4304                 kref_init(&mc->mcref);
4305                 ret = cma_iboe_join_multicast(id_priv, mc);
4306                 if (ret)
4307                         goto out_err;
4308         } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4309                 ret = cma_join_ib_multicast(id_priv, mc);
4310                 if (ret)
4311                         goto out_err;
4312         } else {
4313                 ret = -ENOSYS;
4314                 goto out_err;
4315         }
4316
4317         spin_lock(&id_priv->lock);
4318         list_add(&mc->list, &id_priv->mc_list);
4319         spin_unlock(&id_priv->lock);
4320
4321         return 0;
4322 out_err:
4323         kfree(mc);
4324         return ret;
4325 }
4326 EXPORT_SYMBOL(rdma_join_multicast);
4327
4328 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4329 {
4330         struct rdma_id_private *id_priv;
4331         struct cma_multicast *mc;
4332
4333         id_priv = container_of(id, struct rdma_id_private, id);
4334         spin_lock_irq(&id_priv->lock);
4335         list_for_each_entry(mc, &id_priv->mc_list, list) {
4336                 if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
4337                         continue;
4338                 list_del(&mc->list);
4339                 spin_unlock_irq(&id_priv->lock);
4340
4341                 WARN_ON(id_priv->cma_dev->device != id->device);
4342                 destroy_mc(id_priv, mc);
4343                 return;
4344         }
4345         spin_unlock_irq(&id_priv->lock);
4346 }
4347 EXPORT_SYMBOL(rdma_leave_multicast);
4348
4349 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4350 {
4351         struct rdma_dev_addr *dev_addr;
4352         struct cma_ndev_work *work;
4353
4354         dev_addr = &id_priv->id.route.addr.dev_addr;
4355
4356         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4357             (net_eq(dev_net(ndev), dev_addr->net)) &&
4358             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4359                 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4360                         ndev->name, &id_priv->id);
4361                 work = kzalloc(sizeof *work, GFP_KERNEL);
4362                 if (!work)
4363                         return -ENOMEM;
4364
4365                 INIT_WORK(&work->work, cma_ndev_work_handler);
4366                 work->id = id_priv;
4367                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4368                 atomic_inc(&id_priv->refcount);
4369                 queue_work(cma_wq, &work->work);
4370         }
4371
4372         return 0;
4373 }
4374
4375 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4376                                void *ptr)
4377 {
4378         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4379         struct cma_device *cma_dev;
4380         struct rdma_id_private *id_priv;
4381         int ret = NOTIFY_DONE;
4382
4383         if (event != NETDEV_BONDING_FAILOVER)
4384                 return NOTIFY_DONE;
4385
4386         if (!netif_is_bond_master(ndev))
4387                 return NOTIFY_DONE;
4388
4389         mutex_lock(&lock);
4390         list_for_each_entry(cma_dev, &dev_list, list)
4391                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4392                         ret = cma_netdev_change(ndev, id_priv);
4393                         if (ret)
4394                                 goto out;
4395                 }
4396
4397 out:
4398         mutex_unlock(&lock);
4399         return ret;
4400 }
4401
4402 static struct notifier_block cma_nb = {
4403         .notifier_call = cma_netdev_callback
4404 };
4405
4406 static void cma_add_one(struct ib_device *device)
4407 {
4408         struct cma_device *cma_dev;
4409         struct rdma_id_private *id_priv;
4410         unsigned int i;
4411         unsigned long supported_gids = 0;
4412
4413         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4414         if (!cma_dev)
4415                 return;
4416
4417         cma_dev->device = device;
4418         cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4419                                             sizeof(*cma_dev->default_gid_type),
4420                                             GFP_KERNEL);
4421         if (!cma_dev->default_gid_type)
4422                 goto free_cma_dev;
4423
4424         cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4425                                             sizeof(*cma_dev->default_roce_tos),
4426                                             GFP_KERNEL);
4427         if (!cma_dev->default_roce_tos)
4428                 goto free_gid_type;
4429
4430         for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
4431                 supported_gids = roce_gid_type_mask_support(device, i);
4432                 WARN_ON(!supported_gids);
4433                 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4434                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
4435                                 CMA_PREFERRED_ROCE_GID_TYPE;
4436                 else
4437                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
4438                                 find_first_bit(&supported_gids, BITS_PER_LONG);
4439                 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4440         }
4441
4442         init_completion(&cma_dev->comp);
4443         atomic_set(&cma_dev->refcount, 1);
4444         INIT_LIST_HEAD(&cma_dev->id_list);
4445         ib_set_client_data(device, &cma_client, cma_dev);
4446
4447         mutex_lock(&lock);
4448         list_add_tail(&cma_dev->list, &dev_list);
4449         list_for_each_entry(id_priv, &listen_any_list, list)
4450                 cma_listen_on_dev(id_priv, cma_dev);
4451         mutex_unlock(&lock);
4452
4453         return;
4454
4455 free_gid_type:
4456         kfree(cma_dev->default_gid_type);
4457
4458 free_cma_dev:
4459         kfree(cma_dev);
4460
4461         return;
4462 }
4463
4464 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4465 {
4466         struct rdma_cm_event event = {};
4467         enum rdma_cm_state state;
4468         int ret = 0;
4469
4470         /* Record that we want to remove the device */
4471         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4472         if (state == RDMA_CM_DESTROYING)
4473                 return 0;
4474
4475         cma_cancel_operation(id_priv, state);
4476         mutex_lock(&id_priv->handler_mutex);
4477
4478         /* Check for destruction from another callback. */
4479         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4480                 goto out;
4481
4482         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4483         ret = id_priv->id.event_handler(&id_priv->id, &event);
4484 out:
4485         mutex_unlock(&id_priv->handler_mutex);
4486         return ret;
4487 }
4488
4489 static void cma_process_remove(struct cma_device *cma_dev)
4490 {
4491         struct rdma_id_private *id_priv;
4492         int ret;
4493
4494         mutex_lock(&lock);
4495         while (!list_empty(&cma_dev->id_list)) {
4496                 id_priv = list_entry(cma_dev->id_list.next,
4497                                      struct rdma_id_private, list);
4498
4499                 list_del(&id_priv->listen_list);
4500                 list_del_init(&id_priv->list);
4501                 atomic_inc(&id_priv->refcount);
4502                 mutex_unlock(&lock);
4503
4504                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4505                 cma_deref_id(id_priv);
4506                 if (ret)
4507                         rdma_destroy_id(&id_priv->id);
4508
4509                 mutex_lock(&lock);
4510         }
4511         mutex_unlock(&lock);
4512
4513         cma_deref_dev(cma_dev);
4514         wait_for_completion(&cma_dev->comp);
4515 }
4516
4517 static void cma_remove_one(struct ib_device *device, void *client_data)
4518 {
4519         struct cma_device *cma_dev = client_data;
4520
4521         if (!cma_dev)
4522                 return;
4523
4524         mutex_lock(&lock);
4525         list_del(&cma_dev->list);
4526         mutex_unlock(&lock);
4527
4528         cma_process_remove(cma_dev);
4529         kfree(cma_dev->default_roce_tos);
4530         kfree(cma_dev->default_gid_type);
4531         kfree(cma_dev);
4532 }
4533
4534 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
4535 {
4536         struct nlmsghdr *nlh;
4537         struct rdma_cm_id_stats *id_stats;
4538         struct rdma_id_private *id_priv;
4539         struct rdma_cm_id *id = NULL;
4540         struct cma_device *cma_dev;
4541         int i_dev = 0, i_id = 0;
4542
4543         /*
4544          * We export all of the IDs as a sequence of messages.  Each
4545          * ID gets its own netlink message.
4546          */
4547         mutex_lock(&lock);
4548
4549         list_for_each_entry(cma_dev, &dev_list, list) {
4550                 if (i_dev < cb->args[0]) {
4551                         i_dev++;
4552                         continue;
4553                 }
4554
4555                 i_id = 0;
4556                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4557                         if (i_id < cb->args[1]) {
4558                                 i_id++;
4559                                 continue;
4560                         }
4561
4562                         id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
4563                                                 sizeof *id_stats, RDMA_NL_RDMA_CM,
4564                                                 RDMA_NL_RDMA_CM_ID_STATS,
4565                                                 NLM_F_MULTI);
4566                         if (!id_stats)
4567                                 goto out;
4568
4569                         memset(id_stats, 0, sizeof *id_stats);
4570                         id = &id_priv->id;
4571                         id_stats->node_type = id->route.addr.dev_addr.dev_type;
4572                         id_stats->port_num = id->port_num;
4573                         id_stats->bound_dev_if =
4574                                 id->route.addr.dev_addr.bound_dev_if;
4575
4576                         if (ibnl_put_attr(skb, nlh,
4577                                           rdma_addr_size(cma_src_addr(id_priv)),
4578                                           cma_src_addr(id_priv),
4579                                           RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
4580                                 goto out;
4581                         if (ibnl_put_attr(skb, nlh,
4582                                           rdma_addr_size(cma_dst_addr(id_priv)),
4583                                           cma_dst_addr(id_priv),
4584                                           RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
4585                                 goto out;
4586
4587                         id_stats->pid   = task_pid_vnr(id_priv->res.task);
4588                         id_stats->port_space    = id->ps;
4589                         id_stats->cm_state      = id_priv->state;
4590                         id_stats->qp_num        = id_priv->qp_num;
4591                         id_stats->qp_type       = id->qp_type;
4592
4593                         i_id++;
4594                         nlmsg_end(skb, nlh);
4595                 }
4596
4597                 cb->args[1] = 0;
4598                 i_dev++;
4599         }
4600
4601 out:
4602         mutex_unlock(&lock);
4603         cb->args[0] = i_dev;
4604         cb->args[1] = i_id;
4605
4606         return skb->len;
4607 }
4608
4609 static const struct rdma_nl_cbs cma_cb_table[RDMA_NL_RDMA_CM_NUM_OPS] = {
4610         [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats},
4611 };
4612
4613 static int cma_init_net(struct net *net)
4614 {
4615         struct cma_pernet *pernet = cma_pernet(net);
4616
4617         idr_init(&pernet->tcp_ps);
4618         idr_init(&pernet->udp_ps);
4619         idr_init(&pernet->ipoib_ps);
4620         idr_init(&pernet->ib_ps);
4621
4622         return 0;
4623 }
4624
4625 static void cma_exit_net(struct net *net)
4626 {
4627         struct cma_pernet *pernet = cma_pernet(net);
4628
4629         idr_destroy(&pernet->tcp_ps);
4630         idr_destroy(&pernet->udp_ps);
4631         idr_destroy(&pernet->ipoib_ps);
4632         idr_destroy(&pernet->ib_ps);
4633 }
4634
4635 static struct pernet_operations cma_pernet_operations = {
4636         .init = cma_init_net,
4637         .exit = cma_exit_net,
4638         .id = &cma_pernet_id,
4639         .size = sizeof(struct cma_pernet),
4640 };
4641
4642 static int __init cma_init(void)
4643 {
4644         int ret;
4645
4646         /*
4647          * There is a rare lock ordering dependency in cma_netdev_callback()
4648          * that only happens when bonding is enabled. Teach lockdep that rtnl
4649          * must never be nested under lock so it can find these without having
4650          * to test with bonding.
4651          */
4652         if (IS_ENABLED(CONFIG_LOCKDEP)) {
4653                 rtnl_lock();
4654                 mutex_lock(&lock);
4655                 mutex_unlock(&lock);
4656                 rtnl_unlock();
4657         }
4658
4659         cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4660         if (!cma_wq)
4661                 return -ENOMEM;
4662
4663         ret = register_pernet_subsys(&cma_pernet_operations);
4664         if (ret)
4665                 goto err_wq;
4666
4667         ib_sa_register_client(&sa_client);
4668         register_netdevice_notifier(&cma_nb);
4669
4670         ret = ib_register_client(&cma_client);
4671         if (ret)
4672                 goto err;
4673
4674         rdma_nl_register(RDMA_NL_RDMA_CM, cma_cb_table);
4675         cma_configfs_init();
4676
4677         return 0;
4678
4679 err:
4680         unregister_netdevice_notifier(&cma_nb);
4681         ib_sa_unregister_client(&sa_client);
4682         unregister_pernet_subsys(&cma_pernet_operations);
4683 err_wq:
4684         destroy_workqueue(cma_wq);
4685         return ret;
4686 }
4687
4688 static void __exit cma_cleanup(void)
4689 {
4690         cma_configfs_exit();
4691         rdma_nl_unregister(RDMA_NL_RDMA_CM);
4692         ib_unregister_client(&cma_client);
4693         unregister_netdevice_notifier(&cma_nb);
4694         ib_sa_unregister_client(&sa_client);
4695         unregister_pernet_subsys(&cma_pernet_operations);
4696         destroy_workqueue(cma_wq);
4697 }
4698
4699 MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_RDMA_CM, 1);
4700
4701 module_init(cma_init);
4702 module_exit(cma_cleanup);