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