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/idr.h>
  42 #include <linux/inetdevice.h>
  43 #include <linux/slab.h>
  44 #include <linux/module.h>
  45 #include <net/route.h>
  46
  47 #include <net/net_namespace.h>
  48 #include <net/netns/generic.h>
  49 #include <net/tcp.h>
  50 #include <net/ipv6.h>
  51 #include <net/ip_fib.h>
  52 #include <net/ip6_route.h>
  53
  54 #include <rdma/rdma_cm.h>
  55 #include <rdma/rdma_cm_ib.h>
  56 #include <rdma/rdma_netlink.h>
  57 #include <rdma/ib.h>
  58 #include <rdma/ib_cache.h>
  59 #include <rdma/ib_cm.h>
  60 #include <rdma/ib_sa.h>
  61 #include <rdma/iw_cm.h>
  62
  63 MODULE_AUTHOR("Sean Hefty");
  64 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  65 MODULE_LICENSE("Dual BSD/GPL");
  66
  67 #define CMA_CM_RESPONSE_TIMEOUT 20
  68 #define CMA_MAX_CM_RETRIES 15
  69 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  70 #define CMA_IBOE_PACKET_LIFETIME 18
  71
  72 static const char * const cma_events[] = {
  73         [RDMA_CM_EVENT_ADDR_RESOLVED]    = "address resolved",
  74         [RDMA_CM_EVENT_ADDR_ERROR]       = "address error",
  75         [RDMA_CM_EVENT_ROUTE_RESOLVED]   = "route resolved ",
  76         [RDMA_CM_EVENT_ROUTE_ERROR]      = "route error",
  77         [RDMA_CM_EVENT_CONNECT_REQUEST]  = "connect request",
  78         [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
  79         [RDMA_CM_EVENT_CONNECT_ERROR]    = "connect error",
  80         [RDMA_CM_EVENT_UNREACHABLE]      = "unreachable",
  81         [RDMA_CM_EVENT_REJECTED]         = "rejected",
  82         [RDMA_CM_EVENT_ESTABLISHED]      = "established",
  83         [RDMA_CM_EVENT_DISCONNECTED]     = "disconnected",
  84         [RDMA_CM_EVENT_DEVICE_REMOVAL]   = "device removal",
  85         [RDMA_CM_EVENT_MULTICAST_JOIN]   = "multicast join",
  86         [RDMA_CM_EVENT_MULTICAST_ERROR]  = "multicast error",
  87         [RDMA_CM_EVENT_ADDR_CHANGE]      = "address change",
  88         [RDMA_CM_EVENT_TIMEWAIT_EXIT]    = "timewait exit",
  89 };
  90
  91 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
  92 {
  93         size_t index = event;
  94
  95         return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
  96                         cma_events[index] : "unrecognized event";
  97 }
  98 EXPORT_SYMBOL(rdma_event_msg);
  99
 100 static void cma_add_one(struct ib_device *device);
 101 static void cma_remove_one(struct ib_device *device, void *client_data);
 102
 103 static struct ib_client cma_client = {
 104         .name   = "cma",
 105         .add    = cma_add_one,
 106         .remove = cma_remove_one
 107 };
 108
 109 static struct ib_sa_client sa_client;
 110 static struct rdma_addr_client addr_client;
 111 static LIST_HEAD(dev_list);
 112 static LIST_HEAD(listen_any_list);
 113 static DEFINE_MUTEX(lock);
 114 static struct workqueue_struct *cma_wq;
 115 static int cma_pernet_id;
 116
 117 struct cma_pernet {
 118         struct idr tcp_ps;
 119         struct idr udp_ps;
 120         struct idr ipoib_ps;
 121         struct idr ib_ps;
 122 };
 123
 124 static struct cma_pernet *cma_pernet(struct net *net)
 125 {
 126         return net_generic(net, cma_pernet_id);
 127 }
 128
 129 static struct idr *cma_pernet_idr(struct net *net, enum rdma_port_space ps)
 130 {
 131         struct cma_pernet *pernet = cma_pernet(net);
 132
 133         switch (ps) {
 134         case RDMA_PS_TCP:
 135                 return &pernet->tcp_ps;
 136         case RDMA_PS_UDP:
 137                 return &pernet->udp_ps;
 138         case RDMA_PS_IPOIB:
 139                 return &pernet->ipoib_ps;
 140         case RDMA_PS_IB:
 141                 return &pernet->ib_ps;
 142         default:
 143                 return NULL;
 144         }
 145 }
 146
 147 struct cma_device {
 148         struct list_head        list;
 149         struct ib_device        *device;
 150         struct completion       comp;
 151         atomic_t                refcount;
 152         struct list_head        id_list;
 153 };
 154
 155 struct rdma_bind_list {
 156         enum rdma_port_space    ps;
 157         struct hlist_head       owners;
 158         unsigned short          port;
 159 };
 160
 161 static int cma_ps_alloc(struct net *net, enum rdma_port_space ps,
 162                         struct rdma_bind_list *bind_list, int snum)
 163 {
 164         struct idr *idr = cma_pernet_idr(net, ps);
 165
 166         return idr_alloc(idr, bind_list, snum, snum + 1, GFP_KERNEL);
 167 }
 168
 169 static struct rdma_bind_list *cma_ps_find(struct net *net,
 170                                           enum rdma_port_space ps, int snum)
 171 {
 172         struct idr *idr = cma_pernet_idr(net, ps);
 173
 174         return idr_find(idr, snum);
 175 }
 176
 177 static void cma_ps_remove(struct net *net, enum rdma_port_space ps, int snum)
 178 {
 179         struct idr *idr = cma_pernet_idr(net, ps);
 180
 181         idr_remove(idr, snum);
 182 }
 183
 184 enum {
 185         CMA_OPTION_AFONLY,
 186 };
 187
 188 /*
 189  * Device removal can occur at anytime, so we need extra handling to
 190  * serialize notifying the user of device removal with other callbacks.
 191  * We do this by disabling removal notification while a callback is in process,
 192  * and reporting it after the callback completes.
 193  */
 194 struct rdma_id_private {
 195         struct rdma_cm_id       id;
 196
 197         struct rdma_bind_list   *bind_list;
 198         struct hlist_node       node;
 199         struct list_head        list; /* listen_any_list or cma_device.list */
 200         struct list_head        listen_list; /* per device listens */
 201         struct cma_device       *cma_dev;
 202         struct list_head        mc_list;
 203
 204         int                     internal_id;
 205         enum rdma_cm_state      state;
 206         spinlock_t              lock;
 207         struct mutex            qp_mutex;
 208
 209         struct completion       comp;
 210         atomic_t                refcount;
 211         struct mutex            handler_mutex;
 212
 213         int                     backlog;
 214         int                     timeout_ms;
 215         struct ib_sa_query      *query;
 216         int                     query_id;
 217         union {
 218                 struct ib_cm_id *ib;
 219                 struct iw_cm_id *iw;
 220         } cm_id;
 221
 222         u32                     seq_num;
 223         u32                     qkey;
 224         u32                     qp_num;
 225         pid_t                   owner;
 226         u32                     options;
 227         u8                      srq;
 228         u8                      tos;
 229         u8                      reuseaddr;
 230         u8                      afonly;
 231 };
 232
 233 struct cma_multicast {
 234         struct rdma_id_private *id_priv;
 235         union {
 236                 struct ib_sa_multicast *ib;
 237         } multicast;
 238         struct list_head        list;
 239         void                    *context;
 240         struct sockaddr_storage addr;
 241         struct kref             mcref;
 242 };
 243
 244 struct cma_work {
 245         struct work_struct      work;
 246         struct rdma_id_private  *id;
 247         enum rdma_cm_state      old_state;
 248         enum rdma_cm_state      new_state;
 249         struct rdma_cm_event    event;
 250 };
 251
 252 struct cma_ndev_work {
 253         struct work_struct      work;
 254         struct rdma_id_private  *id;
 255         struct rdma_cm_event    event;
 256 };
 257
 258 struct iboe_mcast_work {
 259         struct work_struct       work;
 260         struct rdma_id_private  *id;
 261         struct cma_multicast    *mc;
 262 };
 263
 264 union cma_ip_addr {
 265         struct in6_addr ip6;
 266         struct {
 267                 __be32 pad[3];
 268                 __be32 addr;
 269         } ip4;
 270 };
 271
 272 struct cma_hdr {
 273         u8 cma_version;
 274         u8 ip_version;  /* IP version: 7:4 */
 275         __be16 port;
 276         union cma_ip_addr src_addr;
 277         union cma_ip_addr dst_addr;
 278 };
 279
 280 #define CMA_VERSION 0x00
 281
 282 struct cma_req_info {
 283         struct ib_device *device;
 284         int port;
 285         union ib_gid local_gid;
 286         __be64 service_id;
 287         u16 pkey;
 288         bool has_gid:1;
 289 };
 290
 291 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 292 {
 293         unsigned long flags;
 294         int ret;
 295
 296         spin_lock_irqsave(&id_priv->lock, flags);
 297         ret = (id_priv->state == comp);
 298         spin_unlock_irqrestore(&id_priv->lock, flags);
 299         return ret;
 300 }
 301
 302 static int cma_comp_exch(struct rdma_id_private *id_priv,
 303                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 304 {
 305         unsigned long flags;
 306         int ret;
 307
 308         spin_lock_irqsave(&id_priv->lock, flags);
 309         if ((ret = (id_priv->state == comp)))
 310                 id_priv->state = exch;
 311         spin_unlock_irqrestore(&id_priv->lock, flags);
 312         return ret;
 313 }
 314
 315 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 316                                    enum rdma_cm_state exch)
 317 {
 318         unsigned long flags;
 319         enum rdma_cm_state old;
 320
 321         spin_lock_irqsave(&id_priv->lock, flags);
 322         old = id_priv->state;
 323         id_priv->state = exch;
 324         spin_unlock_irqrestore(&id_priv->lock, flags);
 325         return old;
 326 }
 327
 328 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
 329 {
 330         return hdr->ip_version >> 4;
 331 }
 332
 333 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 334 {
 335         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 336 }
 337
 338 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 339                               struct cma_device *cma_dev)
 340 {
 341         atomic_inc(&cma_dev->refcount);
 342         id_priv->cma_dev = cma_dev;
 343         id_priv->id.device = cma_dev->device;
 344         id_priv->id.route.addr.dev_addr.transport =
 345                 rdma_node_get_transport(cma_dev->device->node_type);
 346         list_add_tail(&id_priv->list, &cma_dev->id_list);
 347 }
 348
 349 static inline void cma_deref_dev(struct cma_device *cma_dev)
 350 {
 351         if (atomic_dec_and_test(&cma_dev->refcount))
 352                 complete(&cma_dev->comp);
 353 }
 354
 355 static inline void release_mc(struct kref *kref)
 356 {
 357         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 358
 359         kfree(mc->multicast.ib);
 360         kfree(mc);
 361 }
 362
 363 static void cma_release_dev(struct rdma_id_private *id_priv)
 364 {
 365         mutex_lock(&lock);
 366         list_del(&id_priv->list);
 367         cma_deref_dev(id_priv->cma_dev);
 368         id_priv->cma_dev = NULL;
 369         mutex_unlock(&lock);
 370 }
 371
 372 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 373 {
 374         return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
 375 }
 376
 377 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 378 {
 379         return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
 380 }
 381
 382 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 383 {
 384         return id_priv->id.route.addr.src_addr.ss_family;
 385 }
 386
 387 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 388 {
 389         struct ib_sa_mcmember_rec rec;
 390         int ret = 0;
 391
 392         if (id_priv->qkey) {
 393                 if (qkey && id_priv->qkey != qkey)
 394                         return -EINVAL;
 395                 return 0;
 396         }
 397
 398         if (qkey) {
 399                 id_priv->qkey = qkey;
 400                 return 0;
 401         }
 402
 403         switch (id_priv->id.ps) {
 404         case RDMA_PS_UDP:
 405         case RDMA_PS_IB:
 406                 id_priv->qkey = RDMA_UDP_QKEY;
 407                 break;
 408         case RDMA_PS_IPOIB:
 409                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 410                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 411                                              id_priv->id.port_num, &rec.mgid,
 412                                              &rec);
 413                 if (!ret)
 414                         id_priv->qkey = be32_to_cpu(rec.qkey);
 415                 break;
 416         default:
 417                 break;
 418         }
 419         return ret;
 420 }
 421
 422 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 423 {
 424         dev_addr->dev_type = ARPHRD_INFINIBAND;
 425         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 426         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 427 }
 428
 429 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 430 {
 431         int ret;
 432
 433         if (addr->sa_family != AF_IB) {
 434                 ret = rdma_translate_ip(addr, dev_addr, NULL);
 435         } else {
 436                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 437                 ret = 0;
 438         }
 439
 440         return ret;
 441 }
 442
 443 static inline int cma_validate_port(struct ib_device *device, u8 port,
 444                                       union ib_gid *gid, int dev_type,
 445                                       int bound_if_index)
 446 {
 447         int ret = -ENODEV;
 448         struct net_device *ndev = NULL;
 449
 450         if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
 451                 return ret;
 452
 453         if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
 454                 return ret;
 455
 456         if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port))
 457                 ndev = dev_get_by_index(&init_net, bound_if_index);
 458
 459         ret = ib_find_cached_gid_by_port(device, gid, port, ndev, NULL);
 460
 461         if (ndev)
 462                 dev_put(ndev);
 463
 464         return ret;
 465 }
 466
 467 static int cma_acquire_dev(struct rdma_id_private *id_priv,
 468                            struct rdma_id_private *listen_id_priv)
 469 {
 470         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 471         struct cma_device *cma_dev;
 472         union ib_gid gid, iboe_gid, *gidp;
 473         int ret = -ENODEV;
 474         u8 port;
 475
 476         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 477             id_priv->id.ps == RDMA_PS_IPOIB)
 478                 return -EINVAL;
 479
 480         mutex_lock(&lock);
 481         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 482                     &iboe_gid);
 483
 484         memcpy(&gid, dev_addr->src_dev_addr +
 485                rdma_addr_gid_offset(dev_addr), sizeof gid);
 486
 487         if (listen_id_priv) {
 488                 cma_dev = listen_id_priv->cma_dev;
 489                 port = listen_id_priv->id.port_num;
 490                 gidp = rdma_protocol_roce(cma_dev->device, port) ?
 491                        &iboe_gid : &gid;
 492
 493                 ret = cma_validate_port(cma_dev->device, port, gidp,
 494                                         dev_addr->dev_type,
 495                                         dev_addr->bound_dev_if);
 496                 if (!ret) {
 497                         id_priv->id.port_num = port;
 498                         goto out;
 499                 }
 500         }
 501
 502         list_for_each_entry(cma_dev, &dev_list, list) {
 503                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
 504                         if (listen_id_priv &&
 505                             listen_id_priv->cma_dev == cma_dev &&
 506                             listen_id_priv->id.port_num == port)
 507                                 continue;
 508
 509                         gidp = rdma_protocol_roce(cma_dev->device, port) ?
 510                                &iboe_gid : &gid;
 511
 512                         ret = cma_validate_port(cma_dev->device, port, gidp,
 513                                                 dev_addr->dev_type,
 514                                                 dev_addr->bound_dev_if);
 515                         if (!ret) {
 516                                 id_priv->id.port_num = port;
 517                                 goto out;
 518                         }
 519                 }
 520         }
 521
 522 out:
 523         if (!ret)
 524                 cma_attach_to_dev(id_priv, cma_dev);
 525
 526         mutex_unlock(&lock);
 527         return ret;
 528 }
 529
 530 /*
 531  * Select the source IB device and address to reach the destination IB address.
 532  */
 533 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 534 {
 535         struct cma_device *cma_dev, *cur_dev;
 536         struct sockaddr_ib *addr;
 537         union ib_gid gid, sgid, *dgid;
 538         u16 pkey, index;
 539         u8 p;
 540         int i;
 541
 542         cma_dev = NULL;
 543         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 544         dgid = (union ib_gid *) &addr->sib_addr;
 545         pkey = ntohs(addr->sib_pkey);
 546
 547         mutex_lock(&lock);
 548         list_for_each_entry(cur_dev, &dev_list, list) {
 549                 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
 550                         if (!rdma_cap_af_ib(cur_dev->device, p))
 551                                 continue;
 552
 553                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 554                                 continue;
 555
 556                         for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i,
 557                                                        &gid, NULL);
 558                              i++) {
 559                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 560                                         cma_dev = cur_dev;
 561                                         sgid = gid;
 562                                         id_priv->id.port_num = p;
 563                                         goto found;
 564                                 }
 565
 566                                 if (!cma_dev && (gid.global.subnet_prefix ==
 567                                                  dgid->global.subnet_prefix)) {
 568                                         cma_dev = cur_dev;
 569                                         sgid = gid;
 570                                         id_priv->id.port_num = p;
 571                                         goto found;
 572                                 }
 573                         }
 574                 }
 575         }
 576         mutex_unlock(&lock);
 577         return -ENODEV;
 578
 579 found:
 580         cma_attach_to_dev(id_priv, cma_dev);
 581         mutex_unlock(&lock);
 582         addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
 583         memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
 584         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 585         return 0;
 586 }
 587
 588 static void cma_deref_id(struct rdma_id_private *id_priv)
 589 {
 590         if (atomic_dec_and_test(&id_priv->refcount))
 591                 complete(&id_priv->comp);
 592 }
 593
 594 static int cma_disable_callback(struct rdma_id_private *id_priv,
 595                                 enum rdma_cm_state state)
 596 {
 597         mutex_lock(&id_priv->handler_mutex);
 598         if (id_priv->state != state) {
 599                 mutex_unlock(&id_priv->handler_mutex);
 600                 return -EINVAL;
 601         }
 602         return 0;
 603 }
 604
 605 struct rdma_cm_id *rdma_create_id(struct net *net,
 606                                   rdma_cm_event_handler event_handler,
 607                                   void *context, enum rdma_port_space ps,
 608                                   enum ib_qp_type qp_type)
 609 {
 610         struct rdma_id_private *id_priv;
 611
 612         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 613         if (!id_priv)
 614                 return ERR_PTR(-ENOMEM);
 615
 616         id_priv->owner = task_pid_nr(current);
 617         id_priv->state = RDMA_CM_IDLE;
 618         id_priv->id.context = context;
 619         id_priv->id.event_handler = event_handler;
 620         id_priv->id.ps = ps;
 621         id_priv->id.qp_type = qp_type;
 622         spin_lock_init(&id_priv->lock);
 623         mutex_init(&id_priv->qp_mutex);
 624         init_completion(&id_priv->comp);
 625         atomic_set(&id_priv->refcount, 1);
 626         mutex_init(&id_priv->handler_mutex);
 627         INIT_LIST_HEAD(&id_priv->listen_list);
 628         INIT_LIST_HEAD(&id_priv->mc_list);
 629         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 630         id_priv->id.route.addr.dev_addr.net = get_net(net);
 631         id_priv->seq_num &= 0x00ffffff;
 632
 633         return &id_priv->id;
 634 }
 635 EXPORT_SYMBOL(rdma_create_id);
 636
 637 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 638 {
 639         struct ib_qp_attr qp_attr;
 640         int qp_attr_mask, ret;
 641
 642         qp_attr.qp_state = IB_QPS_INIT;
 643         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 644         if (ret)
 645                 return ret;
 646
 647         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 648         if (ret)
 649                 return ret;
 650
 651         qp_attr.qp_state = IB_QPS_RTR;
 652         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 653         if (ret)
 654                 return ret;
 655
 656         qp_attr.qp_state = IB_QPS_RTS;
 657         qp_attr.sq_psn = 0;
 658         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 659
 660         return ret;
 661 }
 662
 663 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 664 {
 665         struct ib_qp_attr qp_attr;
 666         int qp_attr_mask, ret;
 667
 668         qp_attr.qp_state = IB_QPS_INIT;
 669         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 670         if (ret)
 671                 return ret;
 672
 673         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 674 }
 675
 676 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 677                    struct ib_qp_init_attr *qp_init_attr)
 678 {
 679         struct rdma_id_private *id_priv;
 680         struct ib_qp *qp;
 681         int ret;
 682
 683         id_priv = container_of(id, struct rdma_id_private, id);
 684         if (id->device != pd->device)
 685                 return -EINVAL;
 686
 687         qp = ib_create_qp(pd, qp_init_attr);
 688         if (IS_ERR(qp))
 689                 return PTR_ERR(qp);
 690
 691         if (id->qp_type == IB_QPT_UD)
 692                 ret = cma_init_ud_qp(id_priv, qp);
 693         else
 694                 ret = cma_init_conn_qp(id_priv, qp);
 695         if (ret)
 696                 goto err;
 697
 698         id->qp = qp;
 699         id_priv->qp_num = qp->qp_num;
 700         id_priv->srq = (qp->srq != NULL);
 701         return 0;
 702 err:
 703         ib_destroy_qp(qp);
 704         return ret;
 705 }
 706 EXPORT_SYMBOL(rdma_create_qp);
 707
 708 void rdma_destroy_qp(struct rdma_cm_id *id)
 709 {
 710         struct rdma_id_private *id_priv;
 711
 712         id_priv = container_of(id, struct rdma_id_private, id);
 713         mutex_lock(&id_priv->qp_mutex);
 714         ib_destroy_qp(id_priv->id.qp);
 715         id_priv->id.qp = NULL;
 716         mutex_unlock(&id_priv->qp_mutex);
 717 }
 718 EXPORT_SYMBOL(rdma_destroy_qp);
 719
 720 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 721                              struct rdma_conn_param *conn_param)
 722 {
 723         struct ib_qp_attr qp_attr;
 724         int qp_attr_mask, ret;
 725         union ib_gid sgid;
 726
 727         mutex_lock(&id_priv->qp_mutex);
 728         if (!id_priv->id.qp) {
 729                 ret = 0;
 730                 goto out;
 731         }
 732
 733         /* Need to update QP attributes from default values. */
 734         qp_attr.qp_state = IB_QPS_INIT;
 735         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 736         if (ret)
 737                 goto out;
 738
 739         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 740         if (ret)
 741                 goto out;
 742
 743         qp_attr.qp_state = IB_QPS_RTR;
 744         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 745         if (ret)
 746                 goto out;
 747
 748         ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num,
 749                            qp_attr.ah_attr.grh.sgid_index, &sgid, NULL);
 750         if (ret)
 751                 goto out;
 752
 753         BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
 754
 755         if (conn_param)
 756                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 757         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 758 out:
 759         mutex_unlock(&id_priv->qp_mutex);
 760         return ret;
 761 }
 762
 763 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 764                              struct rdma_conn_param *conn_param)
 765 {
 766         struct ib_qp_attr qp_attr;
 767         int qp_attr_mask, ret;
 768
 769         mutex_lock(&id_priv->qp_mutex);
 770         if (!id_priv->id.qp) {
 771                 ret = 0;
 772                 goto out;
 773         }
 774
 775         qp_attr.qp_state = IB_QPS_RTS;
 776         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 777         if (ret)
 778                 goto out;
 779
 780         if (conn_param)
 781                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 782         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 783 out:
 784         mutex_unlock(&id_priv->qp_mutex);
 785         return ret;
 786 }
 787
 788 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 789 {
 790         struct ib_qp_attr qp_attr;
 791         int ret;
 792
 793         mutex_lock(&id_priv->qp_mutex);
 794         if (!id_priv->id.qp) {
 795                 ret = 0;
 796                 goto out;
 797         }
 798
 799         qp_attr.qp_state = IB_QPS_ERR;
 800         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 801 out:
 802         mutex_unlock(&id_priv->qp_mutex);
 803         return ret;
 804 }
 805
 806 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 807                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 808 {
 809         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 810         int ret;
 811         u16 pkey;
 812
 813         if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
 814                 pkey = 0xffff;
 815         else
 816                 pkey = ib_addr_get_pkey(dev_addr);
 817
 818         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 819                                   pkey, &qp_attr->pkey_index);
 820         if (ret)
 821                 return ret;
 822
 823         qp_attr->port_num = id_priv->id.port_num;
 824         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 825
 826         if (id_priv->id.qp_type == IB_QPT_UD) {
 827                 ret = cma_set_qkey(id_priv, 0);
 828                 if (ret)
 829                         return ret;
 830
 831                 qp_attr->qkey = id_priv->qkey;
 832                 *qp_attr_mask |= IB_QP_QKEY;
 833         } else {
 834                 qp_attr->qp_access_flags = 0;
 835                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 836         }
 837         return 0;
 838 }
 839
 840 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 841                        int *qp_attr_mask)
 842 {
 843         struct rdma_id_private *id_priv;
 844         int ret = 0;
 845
 846         id_priv = container_of(id, struct rdma_id_private, id);
 847         if (rdma_cap_ib_cm(id->device, id->port_num)) {
 848                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
 849                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 850                 else
 851                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 852                                                  qp_attr_mask);
 853
 854                 if (qp_attr->qp_state == IB_QPS_RTR)
 855                         qp_attr->rq_psn = id_priv->seq_num;
 856         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
 857                 if (!id_priv->cm_id.iw) {
 858                         qp_attr->qp_access_flags = 0;
 859                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 860                 } else
 861                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 862                                                  qp_attr_mask);
 863                 qp_attr->port_num = id_priv->id.port_num;
 864                 *qp_attr_mask |= IB_QP_PORT;
 865         } else
 866                 ret = -ENOSYS;
 867
 868         return ret;
 869 }
 870 EXPORT_SYMBOL(rdma_init_qp_attr);
 871
 872 static inline int cma_zero_addr(struct sockaddr *addr)
 873 {
 874         switch (addr->sa_family) {
 875         case AF_INET:
 876                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
 877         case AF_INET6:
 878                 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
 879         case AF_IB:
 880                 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
 881         default:
 882                 return 0;
 883         }
 884 }
 885
 886 static inline int cma_loopback_addr(struct sockaddr *addr)
 887 {
 888         switch (addr->sa_family) {
 889         case AF_INET:
 890                 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
 891         case AF_INET6:
 892                 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
 893         case AF_IB:
 894                 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
 895         default:
 896                 return 0;
 897         }
 898 }
 899
 900 static inline int cma_any_addr(struct sockaddr *addr)
 901 {
 902         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 903 }
 904
 905 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
 906 {
 907         if (src->sa_family != dst->sa_family)
 908                 return -1;
 909
 910         switch (src->sa_family) {
 911         case AF_INET:
 912                 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
 913                        ((struct sockaddr_in *) dst)->sin_addr.s_addr;
 914         case AF_INET6:
 915                 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
 916                                      &((struct sockaddr_in6 *) dst)->sin6_addr);
 917         default:
 918                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
 919                                    &((struct sockaddr_ib *) dst)->sib_addr);
 920         }
 921 }
 922
 923 static __be16 cma_port(struct sockaddr *addr)
 924 {
 925         struct sockaddr_ib *sib;
 926
 927         switch (addr->sa_family) {
 928         case AF_INET:
 929                 return ((struct sockaddr_in *) addr)->sin_port;
 930         case AF_INET6:
 931                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 932         case AF_IB:
 933                 sib = (struct sockaddr_ib *) addr;
 934                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
 935                                     be64_to_cpu(sib->sib_sid_mask)));
 936         default:
 937                 return 0;
 938         }
 939 }
 940
 941 static inline int cma_any_port(struct sockaddr *addr)
 942 {
 943         return !cma_port(addr);
 944 }
 945
 946 static void cma_save_ib_info(struct sockaddr *src_addr,
 947                              struct sockaddr *dst_addr,
 948                              struct rdma_cm_id *listen_id,
 949                              struct ib_sa_path_rec *path)
 950 {
 951         struct sockaddr_ib *listen_ib, *ib;
 952
 953         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
 954         if (src_addr) {
 955                 ib = (struct sockaddr_ib *)src_addr;
 956                 ib->sib_family = AF_IB;
 957                 if (path) {
 958                         ib->sib_pkey = path->pkey;
 959                         ib->sib_flowinfo = path->flow_label;
 960                         memcpy(&ib->sib_addr, &path->sgid, 16);
 961                         ib->sib_sid = path->service_id;
 962                         ib->sib_scope_id = 0;
 963                 } else {
 964                         ib->sib_pkey = listen_ib->sib_pkey;
 965                         ib->sib_flowinfo = listen_ib->sib_flowinfo;
 966                         ib->sib_addr = listen_ib->sib_addr;
 967                         ib->sib_sid = listen_ib->sib_sid;
 968                         ib->sib_scope_id = listen_ib->sib_scope_id;
 969                 }
 970                 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
 971         }
 972         if (dst_addr) {
 973                 ib = (struct sockaddr_ib *)dst_addr;
 974                 ib->sib_family = AF_IB;
 975                 if (path) {
 976                         ib->sib_pkey = path->pkey;
 977                         ib->sib_flowinfo = path->flow_label;
 978                         memcpy(&ib->sib_addr, &path->dgid, 16);
 979                 }
 980         }
 981 }
 982
 983 static void cma_save_ip4_info(struct sockaddr *src_addr,
 984                               struct sockaddr *dst_addr,
 985                               struct cma_hdr *hdr,
 986                               __be16 local_port)
 987 {
 988         struct sockaddr_in *ip4;
 989
 990         if (src_addr) {
 991                 ip4 = (struct sockaddr_in *)src_addr;
 992                 ip4->sin_family = AF_INET;
 993                 ip4->sin_addr.s_addr = hdr->dst_addr.ip4.addr;
 994                 ip4->sin_port = local_port;
 995         }
 996
 997         if (dst_addr) {
 998                 ip4 = (struct sockaddr_in *)dst_addr;
 999                 ip4->sin_family = AF_INET;
1000                 ip4->sin_addr.s_addr = hdr->src_addr.ip4.addr;
1001                 ip4->sin_port = hdr->port;
1002         }
1003 }
1004
1005 static void cma_save_ip6_info(struct sockaddr *src_addr,
1006                               struct sockaddr *dst_addr,
1007                               struct cma_hdr *hdr,
1008                               __be16 local_port)
1009 {
1010         struct sockaddr_in6 *ip6;
1011
1012         if (src_addr) {
1013                 ip6 = (struct sockaddr_in6 *)src_addr;
1014                 ip6->sin6_family = AF_INET6;
1015                 ip6->sin6_addr = hdr->dst_addr.ip6;
1016                 ip6->sin6_port = local_port;
1017         }
1018
1019         if (dst_addr) {
1020                 ip6 = (struct sockaddr_in6 *)dst_addr;
1021                 ip6->sin6_family = AF_INET6;
1022                 ip6->sin6_addr = hdr->src_addr.ip6;
1023                 ip6->sin6_port = hdr->port;
1024         }
1025 }
1026
1027 static u16 cma_port_from_service_id(__be64 service_id)
1028 {
1029         return (u16)be64_to_cpu(service_id);
1030 }
1031
1032 static int cma_save_ip_info(struct sockaddr *src_addr,
1033                             struct sockaddr *dst_addr,
1034                             struct ib_cm_event *ib_event,
1035                             __be64 service_id)
1036 {
1037         struct cma_hdr *hdr;
1038         __be16 port;
1039
1040         hdr = ib_event->private_data;
1041         if (hdr->cma_version != CMA_VERSION)
1042                 return -EINVAL;
1043
1044         port = htons(cma_port_from_service_id(service_id));
1045
1046         switch (cma_get_ip_ver(hdr)) {
1047         case 4:
1048                 cma_save_ip4_info(src_addr, dst_addr, hdr, port);
1049                 break;
1050         case 6:
1051                 cma_save_ip6_info(src_addr, dst_addr, hdr, port);
1052                 break;
1053         default:
1054                 return -EAFNOSUPPORT;
1055         }
1056
1057         return 0;
1058 }
1059
1060 static int cma_save_net_info(struct sockaddr *src_addr,
1061                              struct sockaddr *dst_addr,
1062                              struct rdma_cm_id *listen_id,
1063                              struct ib_cm_event *ib_event,
1064                              sa_family_t sa_family, __be64 service_id)
1065 {
1066         if (sa_family == AF_IB) {
1067                 if (ib_event->event == IB_CM_REQ_RECEIVED)
1068                         cma_save_ib_info(src_addr, dst_addr, listen_id,
1069                                          ib_event->param.req_rcvd.primary_path);
1070                 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1071                         cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1072                 return 0;
1073         }
1074
1075         return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1076 }
1077
1078 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1079                              struct cma_req_info *req)
1080 {
1081         const struct ib_cm_req_event_param *req_param =
1082                 &ib_event->param.req_rcvd;
1083         const struct ib_cm_sidr_req_event_param *sidr_param =
1084                 &ib_event->param.sidr_req_rcvd;
1085
1086         switch (ib_event->event) {
1087         case IB_CM_REQ_RECEIVED:
1088                 req->device     = req_param->listen_id->device;
1089                 req->port       = req_param->port;
1090                 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1091                        sizeof(req->local_gid));
1092                 req->has_gid    = true;
1093                 req->service_id = req_param->primary_path->service_id;
1094                 req->pkey       = be16_to_cpu(req_param->primary_path->pkey);
1095                 break;
1096         case IB_CM_SIDR_REQ_RECEIVED:
1097                 req->device     = sidr_param->listen_id->device;
1098                 req->port       = sidr_param->port;
1099                 req->has_gid    = false;
1100                 req->service_id = sidr_param->service_id;
1101                 req->pkey       = sidr_param->pkey;
1102                 break;
1103         default:
1104                 return -EINVAL;
1105         }
1106
1107         return 0;
1108 }
1109
1110 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1111                                   const struct sockaddr_in *dst_addr,
1112                                   const struct sockaddr_in *src_addr)
1113 {
1114         __be32 daddr = dst_addr->sin_addr.s_addr,
1115                saddr = src_addr->sin_addr.s_addr;
1116         struct fib_result res;
1117         struct flowi4 fl4;
1118         int err;
1119         bool ret;
1120
1121         if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1122             ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1123             ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1124             ipv4_is_loopback(saddr))
1125                 return false;
1126
1127         memset(&fl4, 0, sizeof(fl4));
1128         fl4.flowi4_iif = net_dev->ifindex;
1129         fl4.daddr = daddr;
1130         fl4.saddr = saddr;
1131
1132         rcu_read_lock();
1133         err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1134         ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1135         rcu_read_unlock();
1136
1137         return ret;
1138 }
1139
1140 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1141                                   const struct sockaddr_in6 *dst_addr,
1142                                   const struct sockaddr_in6 *src_addr)
1143 {
1144 #if IS_ENABLED(CONFIG_IPV6)
1145         const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1146                            IPV6_ADDR_LINKLOCAL;
1147         struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1148                                          &src_addr->sin6_addr, net_dev->ifindex,
1149                                          strict);
1150         bool ret;
1151
1152         if (!rt)
1153                 return false;
1154
1155         ret = rt->rt6i_idev->dev == net_dev;
1156         ip6_rt_put(rt);
1157
1158         return ret;
1159 #else
1160         return false;
1161 #endif
1162 }
1163
1164 static bool validate_net_dev(struct net_device *net_dev,
1165                              const struct sockaddr *daddr,
1166                              const struct sockaddr *saddr)
1167 {
1168         const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1169         const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1170         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1171         const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1172
1173         switch (daddr->sa_family) {
1174         case AF_INET:
1175                 return saddr->sa_family == AF_INET &&
1176                        validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1177
1178         case AF_INET6:
1179                 return saddr->sa_family == AF_INET6 &&
1180                        validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1181
1182         default:
1183                 return false;
1184         }
1185 }
1186
1187 static struct net_device *cma_get_net_dev(struct ib_cm_event *ib_event,
1188                                           const struct cma_req_info *req)
1189 {
1190         struct sockaddr_storage listen_addr_storage, src_addr_storage;
1191         struct sockaddr *listen_addr = (struct sockaddr *)&listen_addr_storage,
1192                         *src_addr = (struct sockaddr *)&src_addr_storage;
1193         struct net_device *net_dev;
1194         const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1195         int err;
1196
1197         err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1198                                req->service_id);
1199         if (err)
1200                 return ERR_PTR(err);
1201
1202         net_dev = ib_get_net_dev_by_params(req->device, req->port, req->pkey,
1203                                            gid, listen_addr);
1204         if (!net_dev)
1205                 return ERR_PTR(-ENODEV);
1206
1207         if (!validate_net_dev(net_dev, listen_addr, src_addr)) {
1208                 dev_put(net_dev);
1209                 return ERR_PTR(-EHOSTUNREACH);
1210         }
1211
1212         return net_dev;
1213 }
1214
1215 static enum rdma_port_space rdma_ps_from_service_id(__be64 service_id)
1216 {
1217         return (be64_to_cpu(service_id) >> 16) & 0xffff;
1218 }
1219
1220 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1221                                    const struct cma_hdr *hdr)
1222 {
1223         struct sockaddr *addr = cma_src_addr(id_priv);
1224         __be32 ip4_addr;
1225         struct in6_addr ip6_addr;
1226
1227         if (cma_any_addr(addr) && !id_priv->afonly)
1228                 return true;
1229
1230         switch (addr->sa_family) {
1231         case AF_INET:
1232                 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1233                 if (cma_get_ip_ver(hdr) != 4)
1234                         return false;
1235                 if (!cma_any_addr(addr) &&
1236                     hdr->dst_addr.ip4.addr != ip4_addr)
1237                         return false;
1238                 break;
1239         case AF_INET6:
1240                 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1241                 if (cma_get_ip_ver(hdr) != 6)
1242                         return false;
1243                 if (!cma_any_addr(addr) &&
1244                     memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1245                         return false;
1246                 break;
1247         case AF_IB:
1248                 return true;
1249         default:
1250                 return false;
1251         }
1252
1253         return true;
1254 }
1255
1256 static bool cma_protocol_roce_dev_port(struct ib_device *device, int port_num)
1257 {
1258         enum rdma_link_layer ll = rdma_port_get_link_layer(device, port_num);
1259         enum rdma_transport_type transport =
1260                 rdma_node_get_transport(device->node_type);
1261
1262         return ll == IB_LINK_LAYER_ETHERNET && transport == RDMA_TRANSPORT_IB;
1263 }
1264
1265 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1266 {
1267         struct ib_device *device = id->device;
1268         const int port_num = id->port_num ?: rdma_start_port(device);
1269
1270         return cma_protocol_roce_dev_port(device, port_num);
1271 }
1272
1273 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1274                               const struct net_device *net_dev,
1275                               u8 port_num)
1276 {
1277         const struct rdma_addr *addr = &id->route.addr;
1278
1279         if (!net_dev)
1280                 /* This request is an AF_IB request or a RoCE request */
1281                 return (!id->port_num || id->port_num == port_num) &&
1282                        (addr->src_addr.ss_family == AF_IB ||
1283                         cma_protocol_roce_dev_port(id->device, port_num));
1284
1285         /*
1286          * Net namespaces must match, and if the listner is listening
1287          * on a specific netdevice than netdevice must match as well.
1288          */
1289         if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1290             (!!addr->dev_addr.bound_dev_if ==
1291              (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1292                 return true;
1293         else
1294                 return false;
1295 }
1296
1297 static struct rdma_id_private *cma_find_listener(
1298                 const struct rdma_bind_list *bind_list,
1299                 const struct ib_cm_id *cm_id,
1300                 const struct ib_cm_event *ib_event,
1301                 const struct cma_req_info *req,
1302                 const struct net_device *net_dev)
1303 {
1304         struct rdma_id_private *id_priv, *id_priv_dev;
1305
1306         if (!bind_list)
1307                 return ERR_PTR(-EINVAL);
1308
1309         hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1310                 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1311                         if (id_priv->id.device == cm_id->device &&
1312                             cma_match_net_dev(&id_priv->id, net_dev, req->port))
1313                                 return id_priv;
1314                         list_for_each_entry(id_priv_dev,
1315                                             &id_priv->listen_list,
1316                                             listen_list) {
1317                                 if (id_priv_dev->id.device == cm_id->device &&
1318                                     cma_match_net_dev(&id_priv_dev->id, net_dev, req->port))
1319                                         return id_priv_dev;
1320                         }
1321                 }
1322         }
1323
1324         return ERR_PTR(-EINVAL);
1325 }
1326
1327 static struct rdma_id_private *cma_id_from_event(struct ib_cm_id *cm_id,
1328                                                  struct ib_cm_event *ib_event,
1329                                                  struct net_device **net_dev)
1330 {
1331         struct cma_req_info req;
1332         struct rdma_bind_list *bind_list;
1333         struct rdma_id_private *id_priv;
1334         int err;
1335
1336         err = cma_save_req_info(ib_event, &req);
1337         if (err)
1338                 return ERR_PTR(err);
1339
1340         *net_dev = cma_get_net_dev(ib_event, &req);
1341         if (IS_ERR(*net_dev)) {
1342                 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1343                         /* Assuming the protocol is AF_IB */
1344                         *net_dev = NULL;
1345                 } else if (cma_protocol_roce_dev_port(req.device, req.port)) {
1346                         /* TODO find the net dev matching the request parameters
1347                          * through the RoCE GID table */
1348                         *net_dev = NULL;
1349                 } else {
1350                         return ERR_CAST(*net_dev);
1351                 }
1352         }
1353
1354         bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1355                                 rdma_ps_from_service_id(req.service_id),
1356                                 cma_port_from_service_id(req.service_id));
1357         id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev);
1358         if (IS_ERR(id_priv) && *net_dev) {
1359                 dev_put(*net_dev);
1360                 *net_dev = NULL;
1361         }
1362
1363         return id_priv;
1364 }
1365
1366 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1367 {
1368         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1369 }
1370
1371 static void cma_cancel_route(struct rdma_id_private *id_priv)
1372 {
1373         if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1374                 if (id_priv->query)
1375                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1376         }
1377 }
1378
1379 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1380 {
1381         struct rdma_id_private *dev_id_priv;
1382
1383         /*
1384          * Remove from listen_any_list to prevent added devices from spawning
1385          * additional listen requests.
1386          */
1387         mutex_lock(&lock);
1388         list_del(&id_priv->list);
1389
1390         while (!list_empty(&id_priv->listen_list)) {
1391                 dev_id_priv = list_entry(id_priv->listen_list.next,
1392                                          struct rdma_id_private, listen_list);
1393                 /* sync with device removal to avoid duplicate destruction */
1394                 list_del_init(&dev_id_priv->list);
1395                 list_del(&dev_id_priv->listen_list);
1396                 mutex_unlock(&lock);
1397
1398                 rdma_destroy_id(&dev_id_priv->id);
1399                 mutex_lock(&lock);
1400         }
1401         mutex_unlock(&lock);
1402 }
1403
1404 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1405                                  enum rdma_cm_state state)
1406 {
1407         switch (state) {
1408         case RDMA_CM_ADDR_QUERY:
1409                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1410                 break;
1411         case RDMA_CM_ROUTE_QUERY:
1412                 cma_cancel_route(id_priv);
1413                 break;
1414         case RDMA_CM_LISTEN:
1415                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1416                         cma_cancel_listens(id_priv);
1417                 break;
1418         default:
1419                 break;
1420         }
1421 }
1422
1423 static void cma_release_port(struct rdma_id_private *id_priv)
1424 {
1425         struct rdma_bind_list *bind_list = id_priv->bind_list;
1426         struct net *net = id_priv->id.route.addr.dev_addr.net;
1427
1428         if (!bind_list)
1429                 return;
1430
1431         mutex_lock(&lock);
1432         hlist_del(&id_priv->node);
1433         if (hlist_empty(&bind_list->owners)) {
1434                 cma_ps_remove(net, bind_list->ps, bind_list->port);
1435                 kfree(bind_list);
1436         }
1437         mutex_unlock(&lock);
1438 }
1439
1440 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1441 {
1442         struct cma_multicast *mc;
1443
1444         while (!list_empty(&id_priv->mc_list)) {
1445                 mc = container_of(id_priv->mc_list.next,
1446                                   struct cma_multicast, list);
1447                 list_del(&mc->list);
1448                 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1449                                       id_priv->id.port_num)) {
1450                         ib_sa_free_multicast(mc->multicast.ib);
1451                         kfree(mc);
1452                 } else
1453                         kref_put(&mc->mcref, release_mc);
1454         }
1455 }
1456
1457 void rdma_destroy_id(struct rdma_cm_id *id)
1458 {
1459         struct rdma_id_private *id_priv;
1460         enum rdma_cm_state state;
1461
1462         id_priv = container_of(id, struct rdma_id_private, id);
1463         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1464         cma_cancel_operation(id_priv, state);
1465
1466         /*
1467          * Wait for any active callback to finish.  New callbacks will find
1468          * the id_priv state set to destroying and abort.
1469          */
1470         mutex_lock(&id_priv->handler_mutex);
1471         mutex_unlock(&id_priv->handler_mutex);
1472
1473         if (id_priv->cma_dev) {
1474                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1475                         if (id_priv->cm_id.ib)
1476                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1477                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1478                         if (id_priv->cm_id.iw)
1479                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1480                 }
1481                 cma_leave_mc_groups(id_priv);
1482                 cma_release_dev(id_priv);
1483         }
1484
1485         cma_release_port(id_priv);
1486         cma_deref_id(id_priv);
1487         wait_for_completion(&id_priv->comp);
1488
1489         if (id_priv->internal_id)
1490                 cma_deref_id(id_priv->id.context);
1491
1492         kfree(id_priv->id.route.path_rec);
1493         put_net(id_priv->id.route.addr.dev_addr.net);
1494         kfree(id_priv);
1495 }
1496 EXPORT_SYMBOL(rdma_destroy_id);
1497
1498 static int cma_rep_recv(struct rdma_id_private *id_priv)
1499 {
1500         int ret;
1501
1502         ret = cma_modify_qp_rtr(id_priv, NULL);
1503         if (ret)
1504                 goto reject;
1505
1506         ret = cma_modify_qp_rts(id_priv, NULL);
1507         if (ret)
1508                 goto reject;
1509
1510         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1511         if (ret)
1512                 goto reject;
1513
1514         return 0;
1515 reject:
1516         cma_modify_qp_err(id_priv);
1517         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1518                        NULL, 0, NULL, 0);
1519         return ret;
1520 }
1521
1522 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1523                                    struct ib_cm_rep_event_param *rep_data,
1524                                    void *private_data)
1525 {
1526         event->param.conn.private_data = private_data;
1527         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1528         event->param.conn.responder_resources = rep_data->responder_resources;
1529         event->param.conn.initiator_depth = rep_data->initiator_depth;
1530         event->param.conn.flow_control = rep_data->flow_control;
1531         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1532         event->param.conn.srq = rep_data->srq;
1533         event->param.conn.qp_num = rep_data->remote_qpn;
1534 }
1535
1536 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1537 {
1538         struct rdma_id_private *id_priv = cm_id->context;
1539         struct rdma_cm_event event;
1540         int ret = 0;
1541
1542         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1543                 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1544             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1545                 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1546                 return 0;
1547
1548         memset(&event, 0, sizeof event);
1549         switch (ib_event->event) {
1550         case IB_CM_REQ_ERROR:
1551         case IB_CM_REP_ERROR:
1552                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1553                 event.status = -ETIMEDOUT;
1554                 break;
1555         case IB_CM_REP_RECEIVED:
1556                 if (id_priv->id.qp) {
1557                         event.status = cma_rep_recv(id_priv);
1558                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1559                                                      RDMA_CM_EVENT_ESTABLISHED;
1560                 } else {
1561                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1562                 }
1563                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1564                                        ib_event->private_data);
1565                 break;
1566         case IB_CM_RTU_RECEIVED:
1567         case IB_CM_USER_ESTABLISHED:
1568                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1569                 break;
1570         case IB_CM_DREQ_ERROR:
1571                 event.status = -ETIMEDOUT; /* fall through */
1572         case IB_CM_DREQ_RECEIVED:
1573         case IB_CM_DREP_RECEIVED:
1574                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1575                                    RDMA_CM_DISCONNECT))
1576                         goto out;
1577                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1578                 break;
1579         case IB_CM_TIMEWAIT_EXIT:
1580                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1581                 break;
1582         case IB_CM_MRA_RECEIVED:
1583                 /* ignore event */
1584                 goto out;
1585         case IB_CM_REJ_RECEIVED:
1586                 cma_modify_qp_err(id_priv);
1587                 event.status = ib_event->param.rej_rcvd.reason;
1588                 event.event = RDMA_CM_EVENT_REJECTED;
1589                 event.param.conn.private_data = ib_event->private_data;
1590                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1591                 break;
1592         default:
1593                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1594                        ib_event->event);
1595                 goto out;
1596         }
1597
1598         ret = id_priv->id.event_handler(&id_priv->id, &event);
1599         if (ret) {
1600                 /* Destroy the CM ID by returning a non-zero value. */
1601                 id_priv->cm_id.ib = NULL;
1602                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1603                 mutex_unlock(&id_priv->handler_mutex);
1604                 rdma_destroy_id(&id_priv->id);
1605                 return ret;
1606         }
1607 out:
1608         mutex_unlock(&id_priv->handler_mutex);
1609         return ret;
1610 }
1611
1612 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1613                                                struct ib_cm_event *ib_event,
1614                                                struct net_device *net_dev)
1615 {
1616         struct rdma_id_private *id_priv;
1617         struct rdma_cm_id *id;
1618         struct rdma_route *rt;
1619         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1620         const __be64 service_id =
1621                       ib_event->param.req_rcvd.primary_path->service_id;
1622         int ret;
1623
1624         id = rdma_create_id(listen_id->route.addr.dev_addr.net,
1625                             listen_id->event_handler, listen_id->context,
1626                             listen_id->ps, ib_event->param.req_rcvd.qp_type);
1627         if (IS_ERR(id))
1628                 return NULL;
1629
1630         id_priv = container_of(id, struct rdma_id_private, id);
1631         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1632                               (struct sockaddr *)&id->route.addr.dst_addr,
1633                               listen_id, ib_event, ss_family, service_id))
1634                 goto err;
1635
1636         rt = &id->route;
1637         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1638         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1639                                GFP_KERNEL);
1640         if (!rt->path_rec)
1641                 goto err;
1642
1643         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1644         if (rt->num_paths == 2)
1645                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1646
1647         if (net_dev) {
1648                 ret = rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL);
1649                 if (ret)
1650                         goto err;
1651         } else {
1652                 if (!cma_protocol_roce(listen_id) &&
1653                     cma_any_addr(cma_src_addr(id_priv))) {
1654                         rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1655                         rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1656                         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1657                 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
1658                         ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1659                         if (ret)
1660                                 goto err;
1661                 }
1662         }
1663         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1664
1665         id_priv->state = RDMA_CM_CONNECT;
1666         return id_priv;
1667
1668 err:
1669         rdma_destroy_id(id);
1670         return NULL;
1671 }
1672
1673 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1674                                               struct ib_cm_event *ib_event,
1675                                               struct net_device *net_dev)
1676 {
1677         struct rdma_id_private *id_priv;
1678         struct rdma_cm_id *id;
1679         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1680         struct net *net = listen_id->route.addr.dev_addr.net;
1681         int ret;
1682
1683         id = rdma_create_id(net, listen_id->event_handler, listen_id->context,
1684                             listen_id->ps, IB_QPT_UD);
1685         if (IS_ERR(id))
1686                 return NULL;
1687
1688         id_priv = container_of(id, struct rdma_id_private, id);
1689         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1690                               (struct sockaddr *)&id->route.addr.dst_addr,
1691                               listen_id, ib_event, ss_family,
1692                               ib_event->param.sidr_req_rcvd.service_id))
1693                 goto err;
1694
1695         if (net_dev) {
1696                 ret = rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL);
1697                 if (ret)
1698                         goto err;
1699         } else {
1700                 if (!cma_any_addr(cma_src_addr(id_priv))) {
1701                         ret = cma_translate_addr(cma_src_addr(id_priv),
1702                                                  &id->route.addr.dev_addr);
1703                         if (ret)
1704                                 goto err;
1705                 }
1706         }
1707
1708         id_priv->state = RDMA_CM_CONNECT;
1709         return id_priv;
1710 err:
1711         rdma_destroy_id(id);
1712         return NULL;
1713 }
1714
1715 static void cma_set_req_event_data(struct rdma_cm_event *event,
1716                                    struct ib_cm_req_event_param *req_data,
1717                                    void *private_data, int offset)
1718 {
1719         event->param.conn.private_data = private_data + offset;
1720         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1721         event->param.conn.responder_resources = req_data->responder_resources;
1722         event->param.conn.initiator_depth = req_data->initiator_depth;
1723         event->param.conn.flow_control = req_data->flow_control;
1724         event->param.conn.retry_count = req_data->retry_count;
1725         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1726         event->param.conn.srq = req_data->srq;
1727         event->param.conn.qp_num = req_data->remote_qpn;
1728 }
1729
1730 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1731 {
1732         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1733                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1734                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1735                  (id->qp_type == IB_QPT_UD)) ||
1736                 (!id->qp_type));
1737 }
1738
1739 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1740 {
1741         struct rdma_id_private *listen_id, *conn_id;
1742         struct rdma_cm_event event;
1743         struct net_device *net_dev;
1744         u8 offset;
1745         int ret;
1746
1747         listen_id = cma_id_from_event(cm_id, ib_event, &net_dev);
1748         if (IS_ERR(listen_id))
1749                 return PTR_ERR(listen_id);
1750
1751         if (!cma_check_req_qp_type(&listen_id->id, ib_event)) {
1752                 ret = -EINVAL;
1753                 goto net_dev_put;
1754         }
1755
1756         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN)) {
1757                 ret = -ECONNABORTED;
1758                 goto net_dev_put;
1759         }
1760
1761         memset(&event, 0, sizeof event);
1762         offset = cma_user_data_offset(listen_id);
1763         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1764         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1765                 conn_id = cma_new_udp_id(&listen_id->id, ib_event, net_dev);
1766                 event.param.ud.private_data = ib_event->private_data + offset;
1767                 event.param.ud.private_data_len =
1768                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1769         } else {
1770                 conn_id = cma_new_conn_id(&listen_id->id, ib_event, net_dev);
1771                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1772                                        ib_event->private_data, offset);
1773         }
1774         if (!conn_id) {
1775                 ret = -ENOMEM;
1776                 goto err1;
1777         }
1778
1779         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1780         ret = cma_acquire_dev(conn_id, listen_id);
1781         if (ret)
1782                 goto err2;
1783
1784         conn_id->cm_id.ib = cm_id;
1785         cm_id->context = conn_id;
1786         cm_id->cm_handler = cma_ib_handler;
1787
1788         /*
1789          * Protect against the user destroying conn_id from another thread
1790          * until we're done accessing it.
1791          */
1792         atomic_inc(&conn_id->refcount);
1793         ret = conn_id->id.event_handler(&conn_id->id, &event);
1794         if (ret)
1795                 goto err3;
1796         /*
1797          * Acquire mutex to prevent user executing rdma_destroy_id()
1798          * while we're accessing the cm_id.
1799          */
1800         mutex_lock(&lock);
1801         if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
1802             (conn_id->id.qp_type != IB_QPT_UD))
1803                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1804         mutex_unlock(&lock);
1805         mutex_unlock(&conn_id->handler_mutex);
1806         mutex_unlock(&listen_id->handler_mutex);
1807         cma_deref_id(conn_id);
1808         if (net_dev)
1809                 dev_put(net_dev);
1810         return 0;
1811
1812 err3:
1813         cma_deref_id(conn_id);
1814         /* Destroy the CM ID by returning a non-zero value. */
1815         conn_id->cm_id.ib = NULL;
1816 err2:
1817         cma_exch(conn_id, RDMA_CM_DESTROYING);
1818         mutex_unlock(&conn_id->handler_mutex);
1819 err1:
1820         mutex_unlock(&listen_id->handler_mutex);
1821         if (conn_id)
1822                 rdma_destroy_id(&conn_id->id);
1823
1824 net_dev_put:
1825         if (net_dev)
1826                 dev_put(net_dev);
1827
1828         return ret;
1829 }
1830
1831 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
1832 {
1833         if (addr->sa_family == AF_IB)
1834                 return ((struct sockaddr_ib *) addr)->sib_sid;
1835
1836         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
1837 }
1838 EXPORT_SYMBOL(rdma_get_service_id);
1839
1840 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1841 {
1842         struct rdma_id_private *id_priv = iw_id->context;
1843         struct rdma_cm_event event;
1844         int ret = 0;
1845         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1846         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1847
1848         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1849                 return 0;
1850
1851         memset(&event, 0, sizeof event);
1852         switch (iw_event->event) {
1853         case IW_CM_EVENT_CLOSE:
1854                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1855                 break;
1856         case IW_CM_EVENT_CONNECT_REPLY:
1857                 memcpy(cma_src_addr(id_priv), laddr,
1858                        rdma_addr_size(laddr));
1859                 memcpy(cma_dst_addr(id_priv), raddr,
1860                        rdma_addr_size(raddr));
1861                 switch (iw_event->status) {
1862                 case 0:
1863                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1864                         event.param.conn.initiator_depth = iw_event->ird;
1865                         event.param.conn.responder_resources = iw_event->ord;
1866                         break;
1867                 case -ECONNRESET:
1868                 case -ECONNREFUSED:
1869                         event.event = RDMA_CM_EVENT_REJECTED;
1870                         break;
1871                 case -ETIMEDOUT:
1872                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1873                         break;
1874                 default:
1875                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1876                         break;
1877                 }
1878                 break;
1879         case IW_CM_EVENT_ESTABLISHED:
1880                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1881                 event.param.conn.initiator_depth = iw_event->ird;
1882                 event.param.conn.responder_resources = iw_event->ord;
1883                 break;
1884         default:
1885                 BUG_ON(1);
1886         }
1887
1888         event.status = iw_event->status;
1889         event.param.conn.private_data = iw_event->private_data;
1890         event.param.conn.private_data_len = iw_event->private_data_len;
1891         ret = id_priv->id.event_handler(&id_priv->id, &event);
1892         if (ret) {
1893                 /* Destroy the CM ID by returning a non-zero value. */
1894                 id_priv->cm_id.iw = NULL;
1895                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1896                 mutex_unlock(&id_priv->handler_mutex);
1897                 rdma_destroy_id(&id_priv->id);
1898                 return ret;
1899         }
1900
1901         mutex_unlock(&id_priv->handler_mutex);
1902         return ret;
1903 }
1904
1905 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1906                                struct iw_cm_event *iw_event)
1907 {
1908         struct rdma_cm_id *new_cm_id;
1909         struct rdma_id_private *listen_id, *conn_id;
1910         struct rdma_cm_event event;
1911         int ret;
1912         struct ib_device_attr attr;
1913         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1914         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1915
1916         listen_id = cm_id->context;
1917         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1918                 return -ECONNABORTED;
1919
1920         /* Create a new RDMA id for the new IW CM ID */
1921         new_cm_id = rdma_create_id(listen_id->id.route.addr.dev_addr.net,
1922                                    listen_id->id.event_handler,
1923                                    listen_id->id.context,
1924                                    RDMA_PS_TCP, IB_QPT_RC);
1925         if (IS_ERR(new_cm_id)) {
1926                 ret = -ENOMEM;
1927                 goto out;
1928         }
1929         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1930         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1931         conn_id->state = RDMA_CM_CONNECT;
1932
1933         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL);
1934         if (ret) {
1935                 mutex_unlock(&conn_id->handler_mutex);
1936                 rdma_destroy_id(new_cm_id);
1937                 goto out;
1938         }
1939
1940         ret = cma_acquire_dev(conn_id, listen_id);
1941         if (ret) {
1942                 mutex_unlock(&conn_id->handler_mutex);
1943                 rdma_destroy_id(new_cm_id);
1944                 goto out;
1945         }
1946
1947         conn_id->cm_id.iw = cm_id;
1948         cm_id->context = conn_id;
1949         cm_id->cm_handler = cma_iw_handler;
1950
1951         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
1952         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
1953
1954         ret = ib_query_device(conn_id->id.device, &attr);
1955         if (ret) {
1956                 mutex_unlock(&conn_id->handler_mutex);
1957                 rdma_destroy_id(new_cm_id);
1958                 goto out;
1959         }
1960
1961         memset(&event, 0, sizeof event);
1962         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1963         event.param.conn.private_data = iw_event->private_data;
1964         event.param.conn.private_data_len = iw_event->private_data_len;
1965         event.param.conn.initiator_depth = iw_event->ird;
1966         event.param.conn.responder_resources = iw_event->ord;
1967
1968         /*
1969          * Protect against the user destroying conn_id from another thread
1970          * until we're done accessing it.
1971          */
1972         atomic_inc(&conn_id->refcount);
1973         ret = conn_id->id.event_handler(&conn_id->id, &event);
1974         if (ret) {
1975                 /* User wants to destroy the CM ID */
1976                 conn_id->cm_id.iw = NULL;
1977                 cma_exch(conn_id, RDMA_CM_DESTROYING);
1978                 mutex_unlock(&conn_id->handler_mutex);
1979                 mutex_unlock(&listen_id->handler_mutex);
1980                 cma_deref_id(conn_id);
1981                 rdma_destroy_id(&conn_id->id);
1982                 return ret;
1983         }
1984
1985         mutex_unlock(&conn_id->handler_mutex);
1986         cma_deref_id(conn_id);
1987
1988 out:
1989         mutex_unlock(&listen_id->handler_mutex);
1990         return ret;
1991 }
1992
1993 static int cma_ib_listen(struct rdma_id_private *id_priv)
1994 {
1995         struct sockaddr *addr;
1996         struct ib_cm_id *id;
1997         __be64 svc_id;
1998
1999         addr = cma_src_addr(id_priv);
2000         svc_id = rdma_get_service_id(&id_priv->id, addr);
2001         id = ib_cm_insert_listen(id_priv->id.device, cma_req_handler, svc_id);
2002         if (IS_ERR(id))
2003                 return PTR_ERR(id);
2004         id_priv->cm_id.ib = id;
2005
2006         return 0;
2007 }
2008
2009 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2010 {
2011         int ret;
2012         struct iw_cm_id *id;
2013
2014         id = iw_create_cm_id(id_priv->id.device,
2015                              iw_conn_req_handler,
2016                              id_priv);
2017         if (IS_ERR(id))
2018                 return PTR_ERR(id);
2019
2020         id->tos = id_priv->tos;
2021         id_priv->cm_id.iw = id;
2022
2023         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2024                rdma_addr_size(cma_src_addr(id_priv)));
2025
2026         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2027
2028         if (ret) {
2029                 iw_destroy_cm_id(id_priv->cm_id.iw);
2030                 id_priv->cm_id.iw = NULL;
2031         }
2032
2033         return ret;
2034 }
2035
2036 static int cma_listen_handler(struct rdma_cm_id *id,
2037                               struct rdma_cm_event *event)
2038 {
2039         struct rdma_id_private *id_priv = id->context;
2040
2041         id->context = id_priv->id.context;
2042         id->event_handler = id_priv->id.event_handler;
2043         return id_priv->id.event_handler(id, event);
2044 }
2045
2046 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2047                               struct cma_device *cma_dev)
2048 {
2049         struct rdma_id_private *dev_id_priv;
2050         struct rdma_cm_id *id;
2051         struct net *net = id_priv->id.route.addr.dev_addr.net;
2052         int ret;
2053
2054         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2055                 return;
2056
2057         id = rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2058                             id_priv->id.qp_type);
2059         if (IS_ERR(id))
2060                 return;
2061
2062         dev_id_priv = container_of(id, struct rdma_id_private, id);
2063
2064         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2065         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2066                rdma_addr_size(cma_src_addr(id_priv)));
2067
2068         cma_attach_to_dev(dev_id_priv, cma_dev);
2069         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2070         atomic_inc(&id_priv->refcount);
2071         dev_id_priv->internal_id = 1;
2072         dev_id_priv->afonly = id_priv->afonly;
2073
2074         ret = rdma_listen(id, id_priv->backlog);
2075         if (ret)
2076                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
2077                        "listening on device %s\n", ret, cma_dev->device->name);
2078 }
2079
2080 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2081 {
2082         struct cma_device *cma_dev;
2083
2084         mutex_lock(&lock);
2085         list_add_tail(&id_priv->list, &listen_any_list);
2086         list_for_each_entry(cma_dev, &dev_list, list)
2087                 cma_listen_on_dev(id_priv, cma_dev);
2088         mutex_unlock(&lock);
2089 }
2090
2091 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2092 {
2093         struct rdma_id_private *id_priv;
2094
2095         id_priv = container_of(id, struct rdma_id_private, id);
2096         id_priv->tos = (u8) tos;
2097 }
2098 EXPORT_SYMBOL(rdma_set_service_type);
2099
2100 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
2101                               void *context)
2102 {
2103         struct cma_work *work = context;
2104         struct rdma_route *route;
2105
2106         route = &work->id->id.route;
2107
2108         if (!status) {
2109                 route->num_paths = 1;
2110                 *route->path_rec = *path_rec;
2111         } else {
2112                 work->old_state = RDMA_CM_ROUTE_QUERY;
2113                 work->new_state = RDMA_CM_ADDR_RESOLVED;
2114                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2115                 work->event.status = status;
2116         }
2117
2118         queue_work(cma_wq, &work->work);
2119 }
2120
2121 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
2122                               struct cma_work *work)
2123 {
2124         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2125         struct ib_sa_path_rec path_rec;
2126         ib_sa_comp_mask comp_mask;
2127         struct sockaddr_in6 *sin6;
2128         struct sockaddr_ib *sib;
2129
2130         memset(&path_rec, 0, sizeof path_rec);
2131         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2132         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2133         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2134         path_rec.numb_path = 1;
2135         path_rec.reversible = 1;
2136         path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2137
2138         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2139                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2140                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2141
2142         switch (cma_family(id_priv)) {
2143         case AF_INET:
2144                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2145                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2146                 break;
2147         case AF_INET6:
2148                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2149                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2150                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2151                 break;
2152         case AF_IB:
2153                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2154                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2155                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2156                 break;
2157         }
2158
2159         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2160                                                id_priv->id.port_num, &path_rec,
2161                                                comp_mask, timeout_ms,
2162                                                GFP_KERNEL, cma_query_handler,
2163                                                work, &id_priv->query);
2164
2165         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2166 }
2167
2168 static void cma_work_handler(struct work_struct *_work)
2169 {
2170         struct cma_work *work = container_of(_work, struct cma_work, work);
2171         struct rdma_id_private *id_priv = work->id;
2172         int destroy = 0;
2173
2174         mutex_lock(&id_priv->handler_mutex);
2175         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2176                 goto out;
2177
2178         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2179                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2180                 destroy = 1;
2181         }
2182 out:
2183         mutex_unlock(&id_priv->handler_mutex);
2184         cma_deref_id(id_priv);
2185         if (destroy)
2186                 rdma_destroy_id(&id_priv->id);
2187         kfree(work);
2188 }
2189
2190 static void cma_ndev_work_handler(struct work_struct *_work)
2191 {
2192         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2193         struct rdma_id_private *id_priv = work->id;
2194         int destroy = 0;
2195
2196         mutex_lock(&id_priv->handler_mutex);
2197         if (id_priv->state == RDMA_CM_DESTROYING ||
2198             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2199                 goto out;
2200
2201         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2202                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2203                 destroy = 1;
2204         }
2205
2206 out:
2207         mutex_unlock(&id_priv->handler_mutex);
2208         cma_deref_id(id_priv);
2209         if (destroy)
2210                 rdma_destroy_id(&id_priv->id);
2211         kfree(work);
2212 }
2213
2214 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
2215 {
2216         struct rdma_route *route = &id_priv->id.route;
2217         struct cma_work *work;
2218         int ret;
2219
2220         work = kzalloc(sizeof *work, GFP_KERNEL);
2221         if (!work)
2222                 return -ENOMEM;
2223
2224         work->id = id_priv;
2225         INIT_WORK(&work->work, cma_work_handler);
2226         work->old_state = RDMA_CM_ROUTE_QUERY;
2227         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2228         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2229
2230         if (!route->path_rec)
2231                 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2232         if (!route->path_rec) {
2233                 ret = -ENOMEM;
2234                 goto err1;
2235         }
2236
2237         ret = cma_query_ib_route(id_priv, timeout_ms, work);
2238         if (ret)
2239                 goto err2;
2240
2241         return 0;
2242 err2:
2243         kfree(route->path_rec);
2244         route->path_rec = NULL;
2245 err1:
2246         kfree(work);
2247         return ret;
2248 }
2249
2250 int rdma_set_ib_paths(struct rdma_cm_id *id,
2251                       struct ib_sa_path_rec *path_rec, int num_paths)
2252 {
2253         struct rdma_id_private *id_priv;
2254         int ret;
2255
2256         id_priv = container_of(id, struct rdma_id_private, id);
2257         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2258                            RDMA_CM_ROUTE_RESOLVED))
2259                 return -EINVAL;
2260
2261         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
2262                                      GFP_KERNEL);
2263         if (!id->route.path_rec) {
2264                 ret = -ENOMEM;
2265                 goto err;
2266         }
2267
2268         id->route.num_paths = num_paths;
2269         return 0;
2270 err:
2271         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2272         return ret;
2273 }
2274 EXPORT_SYMBOL(rdma_set_ib_paths);
2275
2276 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2277 {
2278         struct cma_work *work;
2279
2280         work = kzalloc(sizeof *work, GFP_KERNEL);
2281         if (!work)
2282                 return -ENOMEM;
2283
2284         work->id = id_priv;
2285         INIT_WORK(&work->work, cma_work_handler);
2286         work->old_state = RDMA_CM_ROUTE_QUERY;
2287         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2288         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2289         queue_work(cma_wq, &work->work);
2290         return 0;
2291 }
2292
2293 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2294 {
2295         int prio;
2296         struct net_device *dev;
2297
2298         prio = rt_tos2priority(tos);
2299         dev = ndev->priv_flags & IFF_802_1Q_VLAN ?
2300                 vlan_dev_real_dev(ndev) : ndev;
2301
2302         if (dev->num_tc)
2303                 return netdev_get_prio_tc_map(dev, prio);
2304
2305 #if IS_ENABLED(CONFIG_VLAN_8021Q)
2306         if (ndev->priv_flags & IFF_802_1Q_VLAN)
2307                 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
2308                         VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2309 #endif
2310         return 0;
2311 }
2312
2313 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2314 {
2315         struct rdma_route *route = &id_priv->id.route;
2316         struct rdma_addr *addr = &route->addr;
2317         struct cma_work *work;
2318         int ret;
2319         struct net_device *ndev = NULL;
2320
2321
2322         work = kzalloc(sizeof *work, GFP_KERNEL);
2323         if (!work)
2324                 return -ENOMEM;
2325
2326         work->id = id_priv;
2327         INIT_WORK(&work->work, cma_work_handler);
2328
2329         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2330         if (!route->path_rec) {
2331                 ret = -ENOMEM;
2332                 goto err1;
2333         }
2334
2335         route->num_paths = 1;
2336
2337         if (addr->dev_addr.bound_dev_if) {
2338                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
2339                 route->path_rec->net = &init_net;
2340                 route->path_rec->ifindex = addr->dev_addr.bound_dev_if;
2341         }
2342         if (!ndev) {
2343                 ret = -ENODEV;
2344                 goto err2;
2345         }
2346
2347         memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN);
2348
2349         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2350                     &route->path_rec->sgid);
2351         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2352                     &route->path_rec->dgid);
2353
2354         route->path_rec->hop_limit = 1;
2355         route->path_rec->reversible = 1;
2356         route->path_rec->pkey = cpu_to_be16(0xffff);
2357         route->path_rec->mtu_selector = IB_SA_EQ;
2358         route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos);
2359         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2360         route->path_rec->rate_selector = IB_SA_EQ;
2361         route->path_rec->rate = iboe_get_rate(ndev);
2362         dev_put(ndev);
2363         route->path_rec->packet_life_time_selector = IB_SA_EQ;
2364         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2365         if (!route->path_rec->mtu) {
2366                 ret = -EINVAL;
2367                 goto err2;
2368         }
2369
2370         work->old_state = RDMA_CM_ROUTE_QUERY;
2371         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2372         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2373         work->event.status = 0;
2374
2375         queue_work(cma_wq, &work->work);
2376
2377         return 0;
2378
2379 err2:
2380         kfree(route->path_rec);
2381         route->path_rec = NULL;
2382         route->num_paths = 0;
2383 err1:
2384         kfree(work);
2385         return ret;
2386 }
2387
2388 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2389 {
2390         struct rdma_id_private *id_priv;
2391         int ret;
2392
2393         id_priv = container_of(id, struct rdma_id_private, id);
2394         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2395                 return -EINVAL;
2396
2397         atomic_inc(&id_priv->refcount);
2398         if (rdma_cap_ib_sa(id->device, id->port_num))
2399                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2400         else if (rdma_protocol_roce(id->device, id->port_num))
2401                 ret = cma_resolve_iboe_route(id_priv);
2402         else if (rdma_protocol_iwarp(id->device, id->port_num))
2403                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2404         else
2405                 ret = -ENOSYS;
2406
2407         if (ret)
2408                 goto err;
2409
2410         return 0;
2411 err:
2412         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2413         cma_deref_id(id_priv);
2414         return ret;
2415 }
2416 EXPORT_SYMBOL(rdma_resolve_route);
2417
2418 static void cma_set_loopback(struct sockaddr *addr)
2419 {
2420         switch (addr->sa_family) {
2421         case AF_INET:
2422                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2423                 break;
2424         case AF_INET6:
2425                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2426                               0, 0, 0, htonl(1));
2427                 break;
2428         default:
2429                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2430                             0, 0, 0, htonl(1));
2431                 break;
2432         }
2433 }
2434
2435 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2436 {
2437         struct cma_device *cma_dev, *cur_dev;
2438         struct ib_port_attr port_attr;
2439         union ib_gid gid;
2440         u16 pkey;
2441         int ret;
2442         u8 p;
2443
2444         cma_dev = NULL;
2445         mutex_lock(&lock);
2446         list_for_each_entry(cur_dev, &dev_list, list) {
2447                 if (cma_family(id_priv) == AF_IB &&
2448                     !rdma_cap_ib_cm(cur_dev->device, 1))
2449                         continue;
2450
2451                 if (!cma_dev)
2452                         cma_dev = cur_dev;
2453
2454                 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2455                         if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2456                             port_attr.state == IB_PORT_ACTIVE) {
2457                                 cma_dev = cur_dev;
2458                                 goto port_found;
2459                         }
2460                 }
2461         }
2462
2463         if (!cma_dev) {
2464                 ret = -ENODEV;
2465                 goto out;
2466         }
2467
2468         p = 1;
2469
2470 port_found:
2471         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid, NULL);
2472         if (ret)
2473                 goto out;
2474
2475         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2476         if (ret)
2477                 goto out;
2478
2479         id_priv->id.route.addr.dev_addr.dev_type =
2480                 (rdma_protocol_ib(cma_dev->device, p)) ?
2481                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2482
2483         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2484         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2485         id_priv->id.port_num = p;
2486         cma_attach_to_dev(id_priv, cma_dev);
2487         cma_set_loopback(cma_src_addr(id_priv));
2488 out:
2489         mutex_unlock(&lock);
2490         return ret;
2491 }
2492
2493 static void addr_handler(int status, struct sockaddr *src_addr,
2494                          struct rdma_dev_addr *dev_addr, void *context)
2495 {
2496         struct rdma_id_private *id_priv = context;
2497         struct rdma_cm_event event;
2498
2499         memset(&event, 0, sizeof event);
2500         mutex_lock(&id_priv->handler_mutex);
2501         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2502                            RDMA_CM_ADDR_RESOLVED))
2503                 goto out;
2504
2505         memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2506         if (!status && !id_priv->cma_dev)
2507                 status = cma_acquire_dev(id_priv, NULL);
2508
2509         if (status) {
2510                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2511                                    RDMA_CM_ADDR_BOUND))
2512                         goto out;
2513                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2514                 event.status = status;
2515         } else
2516                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2517
2518         if (id_priv->id.event_handler(&id_priv->id, &event)) {
2519                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2520                 mutex_unlock(&id_priv->handler_mutex);
2521                 cma_deref_id(id_priv);
2522                 rdma_destroy_id(&id_priv->id);
2523                 return;
2524         }
2525 out:
2526         mutex_unlock(&id_priv->handler_mutex);
2527         cma_deref_id(id_priv);
2528 }
2529
2530 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2531 {
2532         struct cma_work *work;
2533         union ib_gid gid;
2534         int ret;
2535
2536         work = kzalloc(sizeof *work, GFP_KERNEL);
2537         if (!work)
2538                 return -ENOMEM;
2539
2540         if (!id_priv->cma_dev) {
2541                 ret = cma_bind_loopback(id_priv);
2542                 if (ret)
2543                         goto err;
2544         }
2545
2546         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2547         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2548
2549         work->id = id_priv;
2550         INIT_WORK(&work->work, cma_work_handler);
2551         work->old_state = RDMA_CM_ADDR_QUERY;
2552         work->new_state = RDMA_CM_ADDR_RESOLVED;
2553         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2554         queue_work(cma_wq, &work->work);
2555         return 0;
2556 err:
2557         kfree(work);
2558         return ret;
2559 }
2560
2561 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2562 {
2563         struct cma_work *work;
2564         int ret;
2565
2566         work = kzalloc(sizeof *work, GFP_KERNEL);
2567         if (!work)
2568                 return -ENOMEM;
2569
2570         if (!id_priv->cma_dev) {
2571                 ret = cma_resolve_ib_dev(id_priv);
2572                 if (ret)
2573                         goto err;
2574         }
2575
2576         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2577                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2578
2579         work->id = id_priv;
2580         INIT_WORK(&work->work, cma_work_handler);
2581         work->old_state = RDMA_CM_ADDR_QUERY;
2582         work->new_state = RDMA_CM_ADDR_RESOLVED;
2583         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2584         queue_work(cma_wq, &work->work);
2585         return 0;
2586 err:
2587         kfree(work);
2588         return ret;
2589 }
2590
2591 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2592                          struct sockaddr *dst_addr)
2593 {
2594         if (!src_addr || !src_addr->sa_family) {
2595                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2596                 src_addr->sa_family = dst_addr->sa_family;
2597                 if (IS_ENABLED(CONFIG_IPV6) &&
2598                     dst_addr->sa_family == AF_INET6) {
2599                         struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
2600                         struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
2601                         src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
2602                         if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
2603                                 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
2604                 } else if (dst_addr->sa_family == AF_IB) {
2605                         ((struct sockaddr_ib *) src_addr)->sib_pkey =
2606                                 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2607                 }
2608         }
2609         return rdma_bind_addr(id, src_addr);
2610 }
2611
2612 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2613                       struct sockaddr *dst_addr, int timeout_ms)
2614 {
2615         struct rdma_id_private *id_priv;
2616         int ret;
2617
2618         id_priv = container_of(id, struct rdma_id_private, id);
2619         if (id_priv->state == RDMA_CM_IDLE) {
2620                 ret = cma_bind_addr(id, src_addr, dst_addr);
2621                 if (ret)
2622                         return ret;
2623         }
2624
2625         if (cma_family(id_priv) != dst_addr->sa_family)
2626                 return -EINVAL;
2627
2628         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2629                 return -EINVAL;
2630
2631         atomic_inc(&id_priv->refcount);
2632         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2633         if (cma_any_addr(dst_addr)) {
2634                 ret = cma_resolve_loopback(id_priv);
2635         } else {
2636                 if (dst_addr->sa_family == AF_IB) {
2637                         ret = cma_resolve_ib_addr(id_priv);
2638                 } else {
2639                         ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2640                                               dst_addr, &id->route.addr.dev_addr,
2641                                               timeout_ms, addr_handler, id_priv);
2642                 }
2643         }
2644         if (ret)
2645                 goto err;
2646
2647         return 0;
2648 err:
2649         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2650         cma_deref_id(id_priv);
2651         return ret;
2652 }
2653 EXPORT_SYMBOL(rdma_resolve_addr);
2654
2655 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2656 {
2657         struct rdma_id_private *id_priv;
2658         unsigned long flags;
2659         int ret;
2660
2661         id_priv = container_of(id, struct rdma_id_private, id);
2662         spin_lock_irqsave(&id_priv->lock, flags);
2663         if (reuse || id_priv->state == RDMA_CM_IDLE) {
2664                 id_priv->reuseaddr = reuse;
2665                 ret = 0;
2666         } else {
2667                 ret = -EINVAL;
2668         }
2669         spin_unlock_irqrestore(&id_priv->lock, flags);
2670         return ret;
2671 }
2672 EXPORT_SYMBOL(rdma_set_reuseaddr);
2673
2674 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2675 {
2676         struct rdma_id_private *id_priv;
2677         unsigned long flags;
2678         int ret;
2679
2680         id_priv = container_of(id, struct rdma_id_private, id);
2681         spin_lock_irqsave(&id_priv->lock, flags);
2682         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2683                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2684                 id_priv->afonly = afonly;
2685                 ret = 0;
2686         } else {
2687                 ret = -EINVAL;
2688         }
2689         spin_unlock_irqrestore(&id_priv->lock, flags);
2690         return ret;
2691 }
2692 EXPORT_SYMBOL(rdma_set_afonly);
2693
2694 static void cma_bind_port(struct rdma_bind_list *bind_list,
2695                           struct rdma_id_private *id_priv)
2696 {
2697         struct sockaddr *addr;
2698         struct sockaddr_ib *sib;
2699         u64 sid, mask;
2700         __be16 port;
2701
2702         addr = cma_src_addr(id_priv);
2703         port = htons(bind_list->port);
2704
2705         switch (addr->sa_family) {
2706         case AF_INET:
2707                 ((struct sockaddr_in *) addr)->sin_port = port;
2708                 break;
2709         case AF_INET6:
2710                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2711                 break;
2712         case AF_IB:
2713                 sib = (struct sockaddr_ib *) addr;
2714                 sid = be64_to_cpu(sib->sib_sid);
2715                 mask = be64_to_cpu(sib->sib_sid_mask);
2716                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2717                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2718                 break;
2719         }
2720         id_priv->bind_list = bind_list;
2721         hlist_add_head(&id_priv->node, &bind_list->owners);
2722 }
2723
2724 static int cma_alloc_port(enum rdma_port_space ps,
2725                           struct rdma_id_private *id_priv, unsigned short snum)
2726 {
2727         struct rdma_bind_list *bind_list;
2728         int ret;
2729
2730         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2731         if (!bind_list)
2732                 return -ENOMEM;
2733
2734         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
2735                            snum);
2736         if (ret < 0)
2737                 goto err;
2738
2739         bind_list->ps = ps;
2740         bind_list->port = (unsigned short)ret;
2741         cma_bind_port(bind_list, id_priv);
2742         return 0;
2743 err:
2744         kfree(bind_list);
2745         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2746 }
2747
2748 static int cma_alloc_any_port(enum rdma_port_space ps,
2749                               struct rdma_id_private *id_priv)
2750 {
2751         static unsigned int last_used_port;
2752         int low, high, remaining;
2753         unsigned int rover;
2754         struct net *net = id_priv->id.route.addr.dev_addr.net;
2755
2756         inet_get_local_port_range(net, &low, &high);
2757         remaining = (high - low) + 1;
2758         rover = prandom_u32() % remaining + low;
2759 retry:
2760         if (last_used_port != rover &&
2761             !cma_ps_find(net, ps, (unsigned short)rover)) {
2762                 int ret = cma_alloc_port(ps, id_priv, rover);
2763                 /*
2764                  * Remember previously used port number in order to avoid
2765                  * re-using same port immediately after it is closed.
2766                  */
2767                 if (!ret)
2768                         last_used_port = rover;
2769                 if (ret != -EADDRNOTAVAIL)
2770                         return ret;
2771         }
2772         if (--remaining) {
2773                 rover++;
2774                 if ((rover < low) || (rover > high))
2775                         rover = low;
2776                 goto retry;
2777         }
2778         return -EADDRNOTAVAIL;
2779 }
2780
2781 /*
2782  * Check that the requested port is available.  This is called when trying to
2783  * bind to a specific port, or when trying to listen on a bound port.  In
2784  * the latter case, the provided id_priv may already be on the bind_list, but
2785  * we still need to check that it's okay to start listening.
2786  */
2787 static int cma_check_port(struct rdma_bind_list *bind_list,
2788                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2789 {
2790         struct rdma_id_private *cur_id;
2791         struct sockaddr *addr, *cur_addr;
2792
2793         addr = cma_src_addr(id_priv);
2794         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2795                 if (id_priv == cur_id)
2796                         continue;
2797
2798                 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2799                     cur_id->reuseaddr)
2800                         continue;
2801
2802                 cur_addr = cma_src_addr(cur_id);
2803                 if (id_priv->afonly && cur_id->afonly &&
2804                     (addr->sa_family != cur_addr->sa_family))
2805                         continue;
2806
2807                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2808                         return -EADDRNOTAVAIL;
2809
2810                 if (!cma_addr_cmp(addr, cur_addr))
2811                         return -EADDRINUSE;
2812         }
2813         return 0;
2814 }
2815
2816 static int cma_use_port(enum rdma_port_space ps,
2817                         struct rdma_id_private *id_priv)
2818 {
2819         struct rdma_bind_list *bind_list;
2820         unsigned short snum;
2821         int ret;
2822
2823         snum = ntohs(cma_port(cma_src_addr(id_priv)));
2824         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2825                 return -EACCES;
2826
2827         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
2828         if (!bind_list) {
2829                 ret = cma_alloc_port(ps, id_priv, snum);
2830         } else {
2831                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2832                 if (!ret)
2833                         cma_bind_port(bind_list, id_priv);
2834         }
2835         return ret;
2836 }
2837
2838 static int cma_bind_listen(struct rdma_id_private *id_priv)
2839 {
2840         struct rdma_bind_list *bind_list = id_priv->bind_list;
2841         int ret = 0;
2842
2843         mutex_lock(&lock);
2844         if (bind_list->owners.first->next)
2845                 ret = cma_check_port(bind_list, id_priv, 0);
2846         mutex_unlock(&lock);
2847         return ret;
2848 }
2849
2850 static enum rdma_port_space cma_select_inet_ps(
2851                 struct rdma_id_private *id_priv)
2852 {
2853         switch (id_priv->id.ps) {
2854         case RDMA_PS_TCP:
2855         case RDMA_PS_UDP:
2856         case RDMA_PS_IPOIB:
2857         case RDMA_PS_IB:
2858                 return id_priv->id.ps;
2859         default:
2860
2861                 return 0;
2862         }
2863 }
2864
2865 static enum rdma_port_space cma_select_ib_ps(struct rdma_id_private *id_priv)
2866 {
2867         enum rdma_port_space ps = 0;
2868         struct sockaddr_ib *sib;
2869         u64 sid_ps, mask, sid;
2870
2871         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2872         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
2873         sid = be64_to_cpu(sib->sib_sid) & mask;
2874
2875         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
2876                 sid_ps = RDMA_IB_IP_PS_IB;
2877                 ps = RDMA_PS_IB;
2878         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
2879                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
2880                 sid_ps = RDMA_IB_IP_PS_TCP;
2881                 ps = RDMA_PS_TCP;
2882         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
2883                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
2884                 sid_ps = RDMA_IB_IP_PS_UDP;
2885                 ps = RDMA_PS_UDP;
2886         }
2887
2888         if (ps) {
2889                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
2890                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
2891                                                 be64_to_cpu(sib->sib_sid_mask));
2892         }
2893         return ps;
2894 }
2895
2896 static int cma_get_port(struct rdma_id_private *id_priv)
2897 {
2898         enum rdma_port_space ps;
2899         int ret;
2900
2901         if (cma_family(id_priv) != AF_IB)
2902                 ps = cma_select_inet_ps(id_priv);
2903         else
2904                 ps = cma_select_ib_ps(id_priv);
2905         if (!ps)
2906                 return -EPROTONOSUPPORT;
2907
2908         mutex_lock(&lock);
2909         if (cma_any_port(cma_src_addr(id_priv)))
2910                 ret = cma_alloc_any_port(ps, id_priv);
2911         else
2912                 ret = cma_use_port(ps, id_priv);
2913         mutex_unlock(&lock);
2914
2915         return ret;
2916 }
2917
2918 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2919                                struct sockaddr *addr)
2920 {
2921 #if IS_ENABLED(CONFIG_IPV6)
2922         struct sockaddr_in6 *sin6;
2923
2924         if (addr->sa_family != AF_INET6)
2925                 return 0;
2926
2927         sin6 = (struct sockaddr_in6 *) addr;
2928
2929         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
2930                 return 0;
2931
2932         if (!sin6->sin6_scope_id)
2933                         return -EINVAL;
2934
2935         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2936 #endif
2937         return 0;
2938 }
2939
2940 int rdma_listen(struct rdma_cm_id *id, int backlog)
2941 {
2942         struct rdma_id_private *id_priv;
2943         int ret;
2944
2945         id_priv = container_of(id, struct rdma_id_private, id);
2946         if (id_priv->state == RDMA_CM_IDLE) {
2947                 id->route.addr.src_addr.ss_family = AF_INET;
2948                 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
2949                 if (ret)
2950                         return ret;
2951         }
2952
2953         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2954                 return -EINVAL;
2955
2956         if (id_priv->reuseaddr) {
2957                 ret = cma_bind_listen(id_priv);
2958                 if (ret)
2959                         goto err;
2960         }
2961
2962         id_priv->backlog = backlog;
2963         if (id->device) {
2964                 if (rdma_cap_ib_cm(id->device, 1)) {
2965                         ret = cma_ib_listen(id_priv);
2966                         if (ret)
2967                                 goto err;
2968                 } else if (rdma_cap_iw_cm(id->device, 1)) {
2969                         ret = cma_iw_listen(id_priv, backlog);
2970                         if (ret)
2971                                 goto err;
2972                 } else {
2973                         ret = -ENOSYS;
2974                         goto err;
2975                 }
2976         } else
2977                 cma_listen_on_all(id_priv);
2978
2979         return 0;
2980 err:
2981         id_priv->backlog = 0;
2982         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2983         return ret;
2984 }
2985 EXPORT_SYMBOL(rdma_listen);
2986
2987 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2988 {
2989         struct rdma_id_private *id_priv;
2990         int ret;
2991
2992         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
2993             addr->sa_family != AF_IB)
2994                 return -EAFNOSUPPORT;
2995
2996         id_priv = container_of(id, struct rdma_id_private, id);
2997         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2998                 return -EINVAL;
2999
3000         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3001         if (ret)
3002                 goto err1;
3003
3004         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3005         if (!cma_any_addr(addr)) {
3006                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3007                 if (ret)
3008                         goto err1;
3009
3010                 ret = cma_acquire_dev(id_priv, NULL);
3011                 if (ret)
3012                         goto err1;
3013         }
3014
3015         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3016                 if (addr->sa_family == AF_INET)
3017                         id_priv->afonly = 1;
3018 #if IS_ENABLED(CONFIG_IPV6)
3019                 else if (addr->sa_family == AF_INET6) {
3020                         struct net *net = id_priv->id.route.addr.dev_addr.net;
3021
3022                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
3023                 }
3024 #endif
3025         }
3026         ret = cma_get_port(id_priv);
3027         if (ret)
3028                 goto err2;
3029
3030         return 0;
3031 err2:
3032         if (id_priv->cma_dev)
3033                 cma_release_dev(id_priv);
3034 err1:
3035         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3036         return ret;
3037 }
3038 EXPORT_SYMBOL(rdma_bind_addr);
3039
3040 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3041 {
3042         struct cma_hdr *cma_hdr;
3043
3044         cma_hdr = hdr;
3045         cma_hdr->cma_version = CMA_VERSION;
3046         if (cma_family(id_priv) == AF_INET) {
3047                 struct sockaddr_in *src4, *dst4;
3048
3049                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3050                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3051
3052                 cma_set_ip_ver(cma_hdr, 4);
3053                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3054                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3055                 cma_hdr->port = src4->sin_port;
3056         } else if (cma_family(id_priv) == AF_INET6) {
3057                 struct sockaddr_in6 *src6, *dst6;
3058
3059                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3060                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3061
3062                 cma_set_ip_ver(cma_hdr, 6);
3063                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3064                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3065                 cma_hdr->port = src6->sin6_port;
3066         }
3067         return 0;
3068 }
3069
3070 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3071                                 struct ib_cm_event *ib_event)
3072 {
3073         struct rdma_id_private *id_priv = cm_id->context;
3074         struct rdma_cm_event event;
3075         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
3076         int ret = 0;
3077
3078         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
3079                 return 0;
3080
3081         memset(&event, 0, sizeof event);
3082         switch (ib_event->event) {
3083         case IB_CM_SIDR_REQ_ERROR:
3084                 event.event = RDMA_CM_EVENT_UNREACHABLE;
3085                 event.status = -ETIMEDOUT;
3086                 break;
3087         case IB_CM_SIDR_REP_RECEIVED:
3088                 event.param.ud.private_data = ib_event->private_data;
3089                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3090                 if (rep->status != IB_SIDR_SUCCESS) {
3091                         event.event = RDMA_CM_EVENT_UNREACHABLE;
3092                         event.status = ib_event->param.sidr_rep_rcvd.status;
3093                         break;
3094                 }
3095                 ret = cma_set_qkey(id_priv, rep->qkey);
3096                 if (ret) {
3097                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
3098                         event.status = ret;
3099                         break;
3100                 }
3101                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
3102                                      id_priv->id.route.path_rec,
3103                                      &event.param.ud.ah_attr);
3104                 event.param.ud.qp_num = rep->qpn;
3105                 event.param.ud.qkey = rep->qkey;
3106                 event.event = RDMA_CM_EVENT_ESTABLISHED;
3107                 event.status = 0;
3108                 break;
3109         default:
3110                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
3111                        ib_event->event);
3112                 goto out;
3113         }
3114
3115         ret = id_priv->id.event_handler(&id_priv->id, &event);
3116         if (ret) {
3117                 /* Destroy the CM ID by returning a non-zero value. */
3118                 id_priv->cm_id.ib = NULL;
3119                 cma_exch(id_priv, RDMA_CM_DESTROYING);
3120                 mutex_unlock(&id_priv->handler_mutex);
3121                 rdma_destroy_id(&id_priv->id);
3122                 return ret;
3123         }
3124 out:
3125         mutex_unlock(&id_priv->handler_mutex);
3126         return ret;
3127 }
3128
3129 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3130                               struct rdma_conn_param *conn_param)
3131 {
3132         struct ib_cm_sidr_req_param req;
3133         struct ib_cm_id *id;
3134         void *private_data;
3135         u8 offset;
3136         int ret;
3137
3138         memset(&req, 0, sizeof req);
3139         offset = cma_user_data_offset(id_priv);
3140         req.private_data_len = offset + conn_param->private_data_len;
3141         if (req.private_data_len < conn_param->private_data_len)
3142                 return -EINVAL;
3143
3144         if (req.private_data_len) {
3145                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3146                 if (!private_data)
3147                         return -ENOMEM;
3148         } else {
3149                 private_data = NULL;
3150         }
3151
3152         if (conn_param->private_data && conn_param->private_data_len)
3153                 memcpy(private_data + offset, conn_param->private_data,
3154                        conn_param->private_data_len);
3155
3156         if (private_data) {
3157                 ret = cma_format_hdr(private_data, id_priv);
3158                 if (ret)
3159                         goto out;
3160                 req.private_data = private_data;
3161         }
3162
3163         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3164                              id_priv);
3165         if (IS_ERR(id)) {
3166                 ret = PTR_ERR(id);
3167                 goto out;
3168         }
3169         id_priv->cm_id.ib = id;
3170
3171         req.path = id_priv->id.route.path_rec;
3172         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3173         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3174         req.max_cm_retries = CMA_MAX_CM_RETRIES;
3175
3176         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3177         if (ret) {
3178                 ib_destroy_cm_id(id_priv->cm_id.ib);
3179                 id_priv->cm_id.ib = NULL;
3180         }
3181 out:
3182         kfree(private_data);
3183         return ret;
3184 }
3185
3186 static int cma_connect_ib(struct rdma_id_private *id_priv,
3187                           struct rdma_conn_param *conn_param)
3188 {
3189         struct ib_cm_req_param req;
3190         struct rdma_route *route;
3191         void *private_data;
3192         struct ib_cm_id *id;
3193         u8 offset;
3194         int ret;
3195
3196         memset(&req, 0, sizeof req);
3197         offset = cma_user_data_offset(id_priv);
3198         req.private_data_len = offset + conn_param->private_data_len;
3199         if (req.private_data_len < conn_param->private_data_len)
3200                 return -EINVAL;
3201
3202         if (req.private_data_len) {
3203                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3204                 if (!private_data)
3205                         return -ENOMEM;
3206         } else {
3207                 private_data = NULL;
3208         }
3209
3210         if (conn_param->private_data && conn_param->private_data_len)
3211                 memcpy(private_data + offset, conn_param->private_data,
3212                        conn_param->private_data_len);
3213
3214         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3215         if (IS_ERR(id)) {
3216                 ret = PTR_ERR(id);
3217                 goto out;
3218         }
3219         id_priv->cm_id.ib = id;
3220
3221         route = &id_priv->id.route;
3222         if (private_data) {
3223                 ret = cma_format_hdr(private_data, id_priv);
3224                 if (ret)
3225                         goto out;
3226                 req.private_data = private_data;
3227         }
3228
3229         req.primary_path = &route->path_rec[0];
3230         if (route->num_paths == 2)
3231                 req.alternate_path = &route->path_rec[1];
3232
3233         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3234         req.qp_num = id_priv->qp_num;
3235         req.qp_type = id_priv->id.qp_type;
3236         req.starting_psn = id_priv->seq_num;
3237         req.responder_resources = conn_param->responder_resources;
3238         req.initiator_depth = conn_param->initiator_depth;
3239         req.flow_control = conn_param->flow_control;
3240         req.retry_count = min_t(u8, 7, conn_param->retry_count);
3241         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3242         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3243         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3244         req.max_cm_retries = CMA_MAX_CM_RETRIES;
3245         req.srq = id_priv->srq ? 1 : 0;
3246
3247         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3248 out:
3249         if (ret && !IS_ERR(id)) {
3250                 ib_destroy_cm_id(id);
3251                 id_priv->cm_id.ib = NULL;
3252         }
3253
3254         kfree(private_data);
3255         return ret;
3256 }
3257
3258 static int cma_connect_iw(struct rdma_id_private *id_priv,
3259                           struct rdma_conn_param *conn_param)
3260 {
3261         struct iw_cm_id *cm_id;
3262         int ret;
3263         struct iw_cm_conn_param iw_param;
3264
3265         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3266         if (IS_ERR(cm_id))
3267                 return PTR_ERR(cm_id);
3268
3269         cm_id->tos = id_priv->tos;
3270         id_priv->cm_id.iw = cm_id;
3271
3272         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3273                rdma_addr_size(cma_src_addr(id_priv)));
3274         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3275                rdma_addr_size(cma_dst_addr(id_priv)));
3276
3277         ret = cma_modify_qp_rtr(id_priv, conn_param);
3278         if (ret)
3279                 goto out;
3280
3281         if (conn_param) {
3282                 iw_param.ord = conn_param->initiator_depth;
3283                 iw_param.ird = conn_param->responder_resources;
3284                 iw_param.private_data = conn_param->private_data;
3285                 iw_param.private_data_len = conn_param->private_data_len;
3286                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3287         } else {
3288                 memset(&iw_param, 0, sizeof iw_param);
3289                 iw_param.qpn = id_priv->qp_num;
3290         }
3291         ret = iw_cm_connect(cm_id, &iw_param);
3292 out:
3293         if (ret) {
3294                 iw_destroy_cm_id(cm_id);
3295                 id_priv->cm_id.iw = NULL;
3296         }
3297         return ret;
3298 }
3299
3300 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3301 {
3302         struct rdma_id_private *id_priv;
3303         int ret;
3304
3305         id_priv = container_of(id, struct rdma_id_private, id);
3306         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3307                 return -EINVAL;
3308
3309         if (!id->qp) {
3310                 id_priv->qp_num = conn_param->qp_num;
3311                 id_priv->srq = conn_param->srq;
3312         }
3313
3314         if (rdma_cap_ib_cm(id->device, id->port_num)) {
3315                 if (id->qp_type == IB_QPT_UD)
3316                         ret = cma_resolve_ib_udp(id_priv, conn_param);
3317                 else
3318                         ret = cma_connect_ib(id_priv, conn_param);
3319         } else if (rdma_cap_iw_cm(id->device, id->port_num))
3320                 ret = cma_connect_iw(id_priv, conn_param);
3321         else
3322                 ret = -ENOSYS;
3323         if (ret)
3324                 goto err;
3325
3326         return 0;
3327 err:
3328         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3329         return ret;
3330 }
3331 EXPORT_SYMBOL(rdma_connect);
3332
3333 static int cma_accept_ib(struct rdma_id_private *id_priv,
3334                          struct rdma_conn_param *conn_param)
3335 {
3336         struct ib_cm_rep_param rep;
3337         int ret;
3338
3339         ret = cma_modify_qp_rtr(id_priv, conn_param);
3340         if (ret)
3341                 goto out;
3342
3343         ret = cma_modify_qp_rts(id_priv, conn_param);
3344         if (ret)
3345                 goto out;
3346
3347         memset(&rep, 0, sizeof rep);
3348         rep.qp_num = id_priv->qp_num;
3349         rep.starting_psn = id_priv->seq_num;
3350         rep.private_data = conn_param->private_data;
3351         rep.private_data_len = conn_param->private_data_len;
3352         rep.responder_resources = conn_param->responder_resources;
3353         rep.initiator_depth = conn_param->initiator_depth;
3354         rep.failover_accepted = 0;
3355         rep.flow_control = conn_param->flow_control;
3356         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3357         rep.srq = id_priv->srq ? 1 : 0;
3358
3359         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3360 out:
3361         return ret;
3362 }
3363
3364 static int cma_accept_iw(struct rdma_id_private *id_priv,
3365                   struct rdma_conn_param *conn_param)
3366 {
3367         struct iw_cm_conn_param iw_param;
3368         int ret;
3369
3370         if (!conn_param)
3371                 return -EINVAL;
3372
3373         ret = cma_modify_qp_rtr(id_priv, conn_param);
3374         if (ret)
3375                 return ret;
3376
3377         iw_param.ord = conn_param->initiator_depth;
3378         iw_param.ird = conn_param->responder_resources;
3379         iw_param.private_data = conn_param->private_data;
3380         iw_param.private_data_len = conn_param->private_data_len;
3381         if (id_priv->id.qp) {
3382                 iw_param.qpn = id_priv->qp_num;
3383         } else
3384                 iw_param.qpn = conn_param->qp_num;
3385
3386         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3387 }
3388
3389 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3390                              enum ib_cm_sidr_status status, u32 qkey,
3391                              const void *private_data, int private_data_len)
3392 {
3393         struct ib_cm_sidr_rep_param rep;
3394         int ret;
3395
3396         memset(&rep, 0, sizeof rep);
3397         rep.status = status;
3398         if (status == IB_SIDR_SUCCESS) {
3399                 ret = cma_set_qkey(id_priv, qkey);
3400                 if (ret)
3401                         return ret;
3402                 rep.qp_num = id_priv->qp_num;
3403                 rep.qkey = id_priv->qkey;
3404         }
3405         rep.private_data = private_data;
3406         rep.private_data_len = private_data_len;
3407
3408         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3409 }
3410
3411 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3412 {
3413         struct rdma_id_private *id_priv;
3414         int ret;
3415
3416         id_priv = container_of(id, struct rdma_id_private, id);
3417
3418         id_priv->owner = task_pid_nr(current);
3419
3420         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3421                 return -EINVAL;
3422
3423         if (!id->qp && conn_param) {
3424                 id_priv->qp_num = conn_param->qp_num;
3425                 id_priv->srq = conn_param->srq;
3426         }
3427
3428         if (rdma_cap_ib_cm(id->device, id->port_num)) {
3429                 if (id->qp_type == IB_QPT_UD) {
3430                         if (conn_param)
3431                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3432                                                         conn_param->qkey,
3433                                                         conn_param->private_data,
3434                                                         conn_param->private_data_len);
3435                         else
3436                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3437                                                         0, NULL, 0);
3438                 } else {
3439                         if (conn_param)
3440                                 ret = cma_accept_ib(id_priv, conn_param);
3441                         else
3442                                 ret = cma_rep_recv(id_priv);
3443                 }
3444         } else if (rdma_cap_iw_cm(id->device, id->port_num))
3445                 ret = cma_accept_iw(id_priv, conn_param);
3446         else
3447                 ret = -ENOSYS;
3448
3449         if (ret)
3450                 goto reject;
3451
3452         return 0;
3453 reject:
3454         cma_modify_qp_err(id_priv);
3455         rdma_reject(id, NULL, 0);
3456         return ret;
3457 }
3458 EXPORT_SYMBOL(rdma_accept);
3459
3460 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3461 {
3462         struct rdma_id_private *id_priv;
3463         int ret;
3464
3465         id_priv = container_of(id, struct rdma_id_private, id);
3466         if (!id_priv->cm_id.ib)
3467                 return -EINVAL;
3468
3469         switch (id->device->node_type) {
3470         case RDMA_NODE_IB_CA:
3471                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3472                 break;
3473         default:
3474                 ret = 0;
3475                 break;
3476         }
3477         return ret;
3478 }
3479 EXPORT_SYMBOL(rdma_notify);
3480
3481 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3482                 u8 private_data_len)
3483 {
3484         struct rdma_id_private *id_priv;
3485         int ret;
3486
3487         id_priv = container_of(id, struct rdma_id_private, id);
3488         if (!id_priv->cm_id.ib)
3489                 return -EINVAL;
3490
3491         if (rdma_cap_ib_cm(id->device, id->port_num)) {
3492                 if (id->qp_type == IB_QPT_UD)
3493                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3494                                                 private_data, private_data_len);
3495                 else
3496                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
3497                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
3498                                              0, private_data, private_data_len);
3499         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3500                 ret = iw_cm_reject(id_priv->cm_id.iw,
3501                                    private_data, private_data_len);
3502         } else
3503                 ret = -ENOSYS;
3504
3505         return ret;
3506 }
3507 EXPORT_SYMBOL(rdma_reject);
3508
3509 int rdma_disconnect(struct rdma_cm_id *id)
3510 {
3511         struct rdma_id_private *id_priv;
3512         int ret;
3513
3514         id_priv = container_of(id, struct rdma_id_private, id);
3515         if (!id_priv->cm_id.ib)
3516                 return -EINVAL;
3517
3518         if (rdma_cap_ib_cm(id->device, id->port_num)) {
3519                 ret = cma_modify_qp_err(id_priv);
3520                 if (ret)
3521                         goto out;
3522                 /* Initiate or respond to a disconnect. */
3523                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3524                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3525         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3526                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3527         } else
3528                 ret = -EINVAL;
3529
3530 out:
3531         return ret;
3532 }
3533 EXPORT_SYMBOL(rdma_disconnect);
3534
3535 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3536 {
3537         struct rdma_id_private *id_priv;
3538         struct cma_multicast *mc = multicast->context;
3539         struct rdma_cm_event event;
3540         int ret;
3541
3542         id_priv = mc->id_priv;
3543         if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
3544             cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
3545                 return 0;
3546
3547         if (!status)
3548                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3549         mutex_lock(&id_priv->qp_mutex);
3550         if (!status && id_priv->id.qp)
3551                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3552                                          be16_to_cpu(multicast->rec.mlid));
3553         mutex_unlock(&id_priv->qp_mutex);
3554
3555         memset(&event, 0, sizeof event);
3556         event.status = status;
3557         event.param.ud.private_data = mc->context;
3558         if (!status) {
3559                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3560                 ib_init_ah_from_mcmember(id_priv->id.device,
3561                                          id_priv->id.port_num, &multicast->rec,
3562                                          &event.param.ud.ah_attr);
3563                 event.param.ud.qp_num = 0xFFFFFF;
3564                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3565         } else
3566                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3567
3568         ret = id_priv->id.event_handler(&id_priv->id, &event);
3569         if (ret) {
3570                 cma_exch(id_priv, RDMA_CM_DESTROYING);
3571                 mutex_unlock(&id_priv->handler_mutex);
3572                 rdma_destroy_id(&id_priv->id);
3573                 return 0;
3574         }
3575
3576         mutex_unlock(&id_priv->handler_mutex);
3577         return 0;
3578 }
3579
3580 static void cma_set_mgid(struct rdma_id_private *id_priv,
3581                          struct sockaddr *addr, union ib_gid *mgid)
3582 {
3583         unsigned char mc_map[MAX_ADDR_LEN];
3584         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3585         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3586         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3587
3588         if (cma_any_addr(addr)) {
3589                 memset(mgid, 0, sizeof *mgid);
3590         } else if ((addr->sa_family == AF_INET6) &&
3591                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3592                                                                  0xFF10A01B)) {
3593                 /* IPv6 address is an SA assigned MGID. */
3594                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3595         } else if (addr->sa_family == AF_IB) {
3596                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3597         } else if ((addr->sa_family == AF_INET6)) {
3598                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3599                 if (id_priv->id.ps == RDMA_PS_UDP)
3600                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3601                 *mgid = *(union ib_gid *) (mc_map + 4);
3602         } else {
3603                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3604                 if (id_priv->id.ps == RDMA_PS_UDP)
3605                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3606                 *mgid = *(union ib_gid *) (mc_map + 4);
3607         }
3608 }
3609
3610 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3611                                  struct cma_multicast *mc)
3612 {
3613         struct ib_sa_mcmember_rec rec;
3614         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3615         ib_sa_comp_mask comp_mask;
3616         int ret;
3617
3618         ib_addr_get_mgid(dev_addr, &rec.mgid);
3619         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3620                                      &rec.mgid, &rec);
3621         if (ret)
3622                 return ret;
3623
3624         ret = cma_set_qkey(id_priv, 0);
3625         if (ret)
3626                 return ret;
3627
3628         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3629         rec.qkey = cpu_to_be32(id_priv->qkey);
3630         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3631         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3632         rec.join_state = 1;
3633
3634         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3635                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3636                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3637                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
3638                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3639
3640         if (id_priv->id.ps == RDMA_PS_IPOIB)
3641                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3642                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3643                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3644                              IB_SA_MCMEMBER_REC_MTU |
3645                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
3646
3647         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3648                                                 id_priv->id.port_num, &rec,
3649                                                 comp_mask, GFP_KERNEL,
3650                                                 cma_ib_mc_handler, mc);
3651         return PTR_ERR_OR_ZERO(mc->multicast.ib);
3652 }
3653
3654 static void iboe_mcast_work_handler(struct work_struct *work)
3655 {
3656         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3657         struct cma_multicast *mc = mw->mc;
3658         struct ib_sa_multicast *m = mc->multicast.ib;
3659
3660         mc->multicast.ib->context = mc;
3661         cma_ib_mc_handler(0, m);
3662         kref_put(&mc->mcref, release_mc);
3663         kfree(mw);
3664 }
3665
3666 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3667 {
3668         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3669         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3670
3671         if (cma_any_addr(addr)) {
3672                 memset(mgid, 0, sizeof *mgid);
3673         } else if (addr->sa_family == AF_INET6) {
3674                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3675         } else {
3676                 mgid->raw[0] = 0xff;
3677                 mgid->raw[1] = 0x0e;
3678                 mgid->raw[2] = 0;
3679                 mgid->raw[3] = 0;
3680                 mgid->raw[4] = 0;
3681                 mgid->raw[5] = 0;
3682                 mgid->raw[6] = 0;
3683                 mgid->raw[7] = 0;
3684                 mgid->raw[8] = 0;
3685                 mgid->raw[9] = 0;
3686                 mgid->raw[10] = 0xff;
3687                 mgid->raw[11] = 0xff;
3688                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3689         }
3690 }
3691
3692 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3693                                    struct cma_multicast *mc)
3694 {
3695         struct iboe_mcast_work *work;
3696         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3697         int err;
3698         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3699         struct net_device *ndev = NULL;
3700
3701         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3702                 return -EINVAL;
3703
3704         work = kzalloc(sizeof *work, GFP_KERNEL);
3705         if (!work)
3706                 return -ENOMEM;
3707
3708         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3709         if (!mc->multicast.ib) {
3710                 err = -ENOMEM;
3711                 goto out1;
3712         }
3713
3714         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3715
3716         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3717         if (id_priv->id.ps == RDMA_PS_UDP)
3718                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3719
3720         if (dev_addr->bound_dev_if)
3721                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3722         if (!ndev) {
3723                 err = -ENODEV;
3724                 goto out2;
3725         }
3726         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3727         mc->multicast.ib->rec.hop_limit = 1;
3728         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3729         dev_put(ndev);
3730         if (!mc->multicast.ib->rec.mtu) {
3731                 err = -EINVAL;
3732                 goto out2;
3733         }
3734         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3735                     &mc->multicast.ib->rec.port_gid);
3736         work->id = id_priv;
3737         work->mc = mc;
3738         INIT_WORK(&work->work, iboe_mcast_work_handler);
3739         kref_get(&mc->mcref);
3740         queue_work(cma_wq, &work->work);
3741
3742         return 0;
3743
3744 out2:
3745         kfree(mc->multicast.ib);
3746 out1:
3747         kfree(work);
3748         return err;
3749 }
3750
3751 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3752                         void *context)
3753 {
3754         struct rdma_id_private *id_priv;
3755         struct cma_multicast *mc;
3756         int ret;
3757
3758         if (!id->device)
3759                 return -EINVAL;
3760
3761         id_priv = container_of(id, struct rdma_id_private, id);
3762         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3763             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3764                 return -EINVAL;
3765
3766         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3767         if (!mc)
3768                 return -ENOMEM;
3769
3770         memcpy(&mc->addr, addr, rdma_addr_size(addr));
3771         mc->context = context;
3772         mc->id_priv = id_priv;
3773
3774         spin_lock(&id_priv->lock);
3775         list_add(&mc->list, &id_priv->mc_list);
3776         spin_unlock(&id_priv->lock);
3777
3778         if (rdma_protocol_roce(id->device, id->port_num)) {
3779                 kref_init(&mc->mcref);
3780                 ret = cma_iboe_join_multicast(id_priv, mc);
3781         } else if (rdma_cap_ib_mcast(id->device, id->port_num))
3782                 ret = cma_join_ib_multicast(id_priv, mc);
3783         else
3784                 ret = -ENOSYS;
3785
3786         if (ret) {
3787                 spin_lock_irq(&id_priv->lock);
3788                 list_del(&mc->list);
3789                 spin_unlock_irq(&id_priv->lock);
3790                 kfree(mc);
3791         }
3792         return ret;
3793 }
3794 EXPORT_SYMBOL(rdma_join_multicast);
3795
3796 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3797 {
3798         struct rdma_id_private *id_priv;
3799         struct cma_multicast *mc;
3800
3801         id_priv = container_of(id, struct rdma_id_private, id);
3802         spin_lock_irq(&id_priv->lock);
3803         list_for_each_entry(mc, &id_priv->mc_list, list) {
3804                 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
3805                         list_del(&mc->list);
3806                         spin_unlock_irq(&id_priv->lock);
3807
3808                         if (id->qp)
3809                                 ib_detach_mcast(id->qp,
3810                                                 &mc->multicast.ib->rec.mgid,
3811                                                 be16_to_cpu(mc->multicast.ib->rec.mlid));
3812
3813                         BUG_ON(id_priv->cma_dev->device != id->device);
3814
3815                         if (rdma_cap_ib_mcast(id->device, id->port_num)) {
3816                                 ib_sa_free_multicast(mc->multicast.ib);
3817                                 kfree(mc);
3818                         } else if (rdma_protocol_roce(id->device, id->port_num))
3819                                 kref_put(&mc->mcref, release_mc);
3820
3821                         return;
3822                 }
3823         }
3824         spin_unlock_irq(&id_priv->lock);
3825 }
3826 EXPORT_SYMBOL(rdma_leave_multicast);
3827
3828 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3829 {
3830         struct rdma_dev_addr *dev_addr;
3831         struct cma_ndev_work *work;
3832
3833         dev_addr = &id_priv->id.route.addr.dev_addr;
3834
3835         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3836             (net_eq(dev_net(ndev), dev_addr->net)) &&
3837             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3838                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3839                        ndev->name, &id_priv->id);
3840                 work = kzalloc(sizeof *work, GFP_KERNEL);
3841                 if (!work)
3842                         return -ENOMEM;
3843
3844                 INIT_WORK(&work->work, cma_ndev_work_handler);
3845                 work->id = id_priv;
3846                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3847                 atomic_inc(&id_priv->refcount);
3848                 queue_work(cma_wq, &work->work);
3849         }
3850
3851         return 0;
3852 }
3853
3854 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3855                                void *ptr)
3856 {
3857         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
3858         struct cma_device *cma_dev;
3859         struct rdma_id_private *id_priv;
3860         int ret = NOTIFY_DONE;
3861
3862         if (event != NETDEV_BONDING_FAILOVER)
3863                 return NOTIFY_DONE;
3864
3865         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3866                 return NOTIFY_DONE;
3867
3868         mutex_lock(&lock);
3869         list_for_each_entry(cma_dev, &dev_list, list)
3870                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3871                         ret = cma_netdev_change(ndev, id_priv);
3872                         if (ret)
3873                                 goto out;
3874                 }
3875
3876 out:
3877         mutex_unlock(&lock);
3878         return ret;
3879 }
3880
3881 static struct notifier_block cma_nb = {
3882         .notifier_call = cma_netdev_callback
3883 };
3884
3885 static void cma_add_one(struct ib_device *device)
3886 {
3887         struct cma_device *cma_dev;
3888         struct rdma_id_private *id_priv;
3889
3890         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3891         if (!cma_dev)
3892                 return;
3893
3894         cma_dev->device = device;
3895
3896         init_completion(&cma_dev->comp);
3897         atomic_set(&cma_dev->refcount, 1);
3898         INIT_LIST_HEAD(&cma_dev->id_list);
3899         ib_set_client_data(device, &cma_client, cma_dev);
3900
3901         mutex_lock(&lock);
3902         list_add_tail(&cma_dev->list, &dev_list);
3903         list_for_each_entry(id_priv, &listen_any_list, list)
3904                 cma_listen_on_dev(id_priv, cma_dev);
3905         mutex_unlock(&lock);
3906 }
3907
3908 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3909 {
3910         struct rdma_cm_event event;
3911         enum rdma_cm_state state;
3912         int ret = 0;
3913
3914         /* Record that we want to remove the device */
3915         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3916         if (state == RDMA_CM_DESTROYING)
3917                 return 0;
3918
3919         cma_cancel_operation(id_priv, state);
3920         mutex_lock(&id_priv->handler_mutex);
3921
3922         /* Check for destruction from another callback. */
3923         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3924                 goto out;
3925
3926         memset(&event, 0, sizeof event);
3927         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3928         ret = id_priv->id.event_handler(&id_priv->id, &event);
3929 out:
3930         mutex_unlock(&id_priv->handler_mutex);
3931         return ret;
3932 }
3933
3934 static void cma_process_remove(struct cma_device *cma_dev)
3935 {
3936         struct rdma_id_private *id_priv;
3937         int ret;
3938
3939         mutex_lock(&lock);
3940         while (!list_empty(&cma_dev->id_list)) {
3941                 id_priv = list_entry(cma_dev->id_list.next,
3942                                      struct rdma_id_private, list);
3943
3944                 list_del(&id_priv->listen_list);
3945                 list_del_init(&id_priv->list);
3946                 atomic_inc(&id_priv->refcount);
3947                 mutex_unlock(&lock);
3948
3949                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3950                 cma_deref_id(id_priv);
3951                 if (ret)
3952                         rdma_destroy_id(&id_priv->id);
3953
3954                 mutex_lock(&lock);
3955         }
3956         mutex_unlock(&lock);
3957
3958         cma_deref_dev(cma_dev);
3959         wait_for_completion(&cma_dev->comp);
3960 }
3961
3962 static void cma_remove_one(struct ib_device *device, void *client_data)
3963 {
3964         struct cma_device *cma_dev = client_data;
3965
3966         if (!cma_dev)
3967                 return;
3968
3969         mutex_lock(&lock);
3970         list_del(&cma_dev->list);
3971         mutex_unlock(&lock);
3972
3973         cma_process_remove(cma_dev);
3974         kfree(cma_dev);
3975 }
3976
3977 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3978 {
3979         struct nlmsghdr *nlh;
3980         struct rdma_cm_id_stats *id_stats;
3981         struct rdma_id_private *id_priv;
3982         struct rdma_cm_id *id = NULL;
3983         struct cma_device *cma_dev;
3984         int i_dev = 0, i_id = 0;
3985
3986         /*
3987          * We export all of the IDs as a sequence of messages.  Each
3988          * ID gets its own netlink message.
3989          */
3990         mutex_lock(&lock);
3991
3992         list_for_each_entry(cma_dev, &dev_list, list) {
3993                 if (i_dev < cb->args[0]) {
3994                         i_dev++;
3995                         continue;
3996                 }
3997
3998                 i_id = 0;
3999                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4000                         if (i_id < cb->args[1]) {
4001                                 i_id++;
4002                                 continue;
4003                         }
4004
4005                         id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
4006                                                 sizeof *id_stats, RDMA_NL_RDMA_CM,
4007                                                 RDMA_NL_RDMA_CM_ID_STATS,
4008                                                 NLM_F_MULTI);
4009                         if (!id_stats)
4010                                 goto out;
4011
4012                         memset(id_stats, 0, sizeof *id_stats);
4013                         id = &id_priv->id;
4014                         id_stats->node_type = id->route.addr.dev_addr.dev_type;
4015                         id_stats->port_num = id->port_num;
4016                         id_stats->bound_dev_if =
4017                                 id->route.addr.dev_addr.bound_dev_if;
4018
4019                         if (ibnl_put_attr(skb, nlh,
4020                                           rdma_addr_size(cma_src_addr(id_priv)),
4021                                           cma_src_addr(id_priv),
4022                                           RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
4023                                 goto out;
4024                         if (ibnl_put_attr(skb, nlh,
4025                                           rdma_addr_size(cma_dst_addr(id_priv)),
4026                                           cma_dst_addr(id_priv),
4027                                           RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
4028                                 goto out;
4029
4030                         id_stats->pid           = id_priv->owner;
4031                         id_stats->port_space    = id->ps;
4032                         id_stats->cm_state      = id_priv->state;
4033                         id_stats->qp_num        = id_priv->qp_num;
4034                         id_stats->qp_type       = id->qp_type;
4035
4036                         i_id++;
4037                 }
4038
4039                 cb->args[1] = 0;
4040                 i_dev++;
4041         }
4042
4043 out:
4044         mutex_unlock(&lock);
4045         cb->args[0] = i_dev;
4046         cb->args[1] = i_id;
4047
4048         return skb->len;
4049 }
4050
4051 static const struct ibnl_client_cbs cma_cb_table[] = {
4052         [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
4053                                        .module = THIS_MODULE },
4054 };
4055
4056 static int cma_init_net(struct net *net)
4057 {
4058         struct cma_pernet *pernet = cma_pernet(net);
4059
4060         idr_init(&pernet->tcp_ps);
4061         idr_init(&pernet->udp_ps);
4062         idr_init(&pernet->ipoib_ps);
4063         idr_init(&pernet->ib_ps);
4064
4065         return 0;
4066 }
4067
4068 static void cma_exit_net(struct net *net)
4069 {
4070         struct cma_pernet *pernet = cma_pernet(net);
4071
4072         idr_destroy(&pernet->tcp_ps);
4073         idr_destroy(&pernet->udp_ps);
4074         idr_destroy(&pernet->ipoib_ps);
4075         idr_destroy(&pernet->ib_ps);
4076 }
4077
4078 static struct pernet_operations cma_pernet_operations = {
4079         .init = cma_init_net,
4080         .exit = cma_exit_net,
4081         .id = &cma_pernet_id,
4082         .size = sizeof(struct cma_pernet),
4083 };
4084
4085 static int __init cma_init(void)
4086 {
4087         int ret;
4088
4089         cma_wq = create_singlethread_workqueue("rdma_cm");
4090         if (!cma_wq)
4091                 return -ENOMEM;
4092
4093         ret = register_pernet_subsys(&cma_pernet_operations);
4094         if (ret)
4095                 goto err_wq;
4096
4097         ib_sa_register_client(&sa_client);
4098         rdma_addr_register_client(&addr_client);
4099         register_netdevice_notifier(&cma_nb);
4100
4101         ret = ib_register_client(&cma_client);
4102         if (ret)
4103                 goto err;
4104
4105         if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
4106                 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
4107
4108         return 0;
4109
4110 err:
4111         unregister_netdevice_notifier(&cma_nb);
4112         rdma_addr_unregister_client(&addr_client);
4113         ib_sa_unregister_client(&sa_client);
4114         unregister_pernet_subsys(&cma_pernet_operations);
4115 err_wq:
4116         destroy_workqueue(cma_wq);
4117         return ret;
4118 }
4119
4120 static void __exit cma_cleanup(void)
4121 {
4122         ibnl_remove_client(RDMA_NL_RDMA_CM);
4123         ib_unregister_client(&cma_client);
4124         unregister_netdevice_notifier(&cma_nb);
4125         rdma_addr_unregister_client(&addr_client);
4126         ib_sa_unregister_client(&sa_client);
4127         unregister_pernet_subsys(&cma_pernet_operations);
4128         destroy_workqueue(cma_wq);
4129 }
4130
4131 module_init(cma_init);
4132 module_exit(cma_cleanup);