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/tcp.h>
  48 #include <net/ipv6.h>
  49
  50 #include <rdma/rdma_cm.h>
  51 #include <rdma/rdma_cm_ib.h>
  52 #include <rdma/rdma_netlink.h>
  53 #include <rdma/ib.h>
  54 #include <rdma/ib_cache.h>
  55 #include <rdma/ib_cm.h>
  56 #include <rdma/ib_sa.h>
  57 #include <rdma/iw_cm.h>
  58
  59 MODULE_AUTHOR("Sean Hefty");
  60 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  61 MODULE_LICENSE("Dual BSD/GPL");
  62
  63 #define CMA_CM_RESPONSE_TIMEOUT 20
  64 #define CMA_MAX_CM_RETRIES 15
  65 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  66 #define CMA_IBOE_PACKET_LIFETIME 18
  67
  68 static void cma_add_one(struct ib_device *device);
  69 static void cma_remove_one(struct ib_device *device);
  70
  71 static struct ib_client cma_client = {
  72         .name   = "cma",
  73         .add    = cma_add_one,
  74         .remove = cma_remove_one
  75 };
  76
  77 static struct ib_sa_client sa_client;
  78 static struct rdma_addr_client addr_client;
  79 static LIST_HEAD(dev_list);
  80 static LIST_HEAD(listen_any_list);
  81 static DEFINE_MUTEX(lock);
  82 static struct workqueue_struct *cma_wq;
  83 static DEFINE_IDR(tcp_ps);
  84 static DEFINE_IDR(udp_ps);
  85 static DEFINE_IDR(ipoib_ps);
  86 static DEFINE_IDR(ib_ps);
  87
  88 struct cma_device {
  89         struct list_head        list;
  90         struct ib_device        *device;
  91         struct completion       comp;
  92         atomic_t                refcount;
  93         struct list_head        id_list;
  94 };
  95
  96 struct rdma_bind_list {
  97         struct idr              *ps;
  98         struct hlist_head       owners;
  99         unsigned short          port;
 100 };
 101
 102 enum {
 103         CMA_OPTION_AFONLY,
 104 };
 105
 106 /*
 107  * Device removal can occur at anytime, so we need extra handling to
 108  * serialize notifying the user of device removal with other callbacks.
 109  * We do this by disabling removal notification while a callback is in process,
 110  * and reporting it after the callback completes.
 111  */
 112 struct rdma_id_private {
 113         struct rdma_cm_id       id;
 114
 115         struct rdma_bind_list   *bind_list;
 116         struct hlist_node       node;
 117         struct list_head        list; /* listen_any_list or cma_device.list */
 118         struct list_head        listen_list; /* per device listens */
 119         struct cma_device       *cma_dev;
 120         struct list_head        mc_list;
 121
 122         int                     internal_id;
 123         enum rdma_cm_state      state;
 124         spinlock_t              lock;
 125         struct mutex            qp_mutex;
 126
 127         struct completion       comp;
 128         atomic_t                refcount;
 129         struct mutex            handler_mutex;
 130
 131         int                     backlog;
 132         int                     timeout_ms;
 133         struct ib_sa_query      *query;
 134         int                     query_id;
 135         union {
 136                 struct ib_cm_id *ib;
 137                 struct iw_cm_id *iw;
 138         } cm_id;
 139
 140         u32                     seq_num;
 141         u32                     qkey;
 142         u32                     qp_num;
 143         pid_t                   owner;
 144         u32                     options;
 145         u8                      srq;
 146         u8                      tos;
 147         u8                      reuseaddr;
 148         u8                      afonly;
 149 };
 150
 151 struct cma_multicast {
 152         struct rdma_id_private *id_priv;
 153         union {
 154                 struct ib_sa_multicast *ib;
 155         } multicast;
 156         struct list_head        list;
 157         void                    *context;
 158         struct sockaddr_storage addr;
 159         struct kref             mcref;
 160 };
 161
 162 struct cma_work {
 163         struct work_struct      work;
 164         struct rdma_id_private  *id;
 165         enum rdma_cm_state      old_state;
 166         enum rdma_cm_state      new_state;
 167         struct rdma_cm_event    event;
 168 };
 169
 170 struct cma_ndev_work {
 171         struct work_struct      work;
 172         struct rdma_id_private  *id;
 173         struct rdma_cm_event    event;
 174 };
 175
 176 struct iboe_mcast_work {
 177         struct work_struct       work;
 178         struct rdma_id_private  *id;
 179         struct cma_multicast    *mc;
 180 };
 181
 182 union cma_ip_addr {
 183         struct in6_addr ip6;
 184         struct {
 185                 __be32 pad[3];
 186                 __be32 addr;
 187         } ip4;
 188 };
 189
 190 struct cma_hdr {
 191         u8 cma_version;
 192         u8 ip_version;  /* IP version: 7:4 */
 193         __be16 port;
 194         union cma_ip_addr src_addr;
 195         union cma_ip_addr dst_addr;
 196 };
 197
 198 #define CMA_VERSION 0x00
 199
 200 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 201 {
 202         unsigned long flags;
 203         int ret;
 204
 205         spin_lock_irqsave(&id_priv->lock, flags);
 206         ret = (id_priv->state == comp);
 207         spin_unlock_irqrestore(&id_priv->lock, flags);
 208         return ret;
 209 }
 210
 211 static int cma_comp_exch(struct rdma_id_private *id_priv,
 212                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 213 {
 214         unsigned long flags;
 215         int ret;
 216
 217         spin_lock_irqsave(&id_priv->lock, flags);
 218         if ((ret = (id_priv->state == comp)))
 219                 id_priv->state = exch;
 220         spin_unlock_irqrestore(&id_priv->lock, flags);
 221         return ret;
 222 }
 223
 224 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 225                                    enum rdma_cm_state exch)
 226 {
 227         unsigned long flags;
 228         enum rdma_cm_state old;
 229
 230         spin_lock_irqsave(&id_priv->lock, flags);
 231         old = id_priv->state;
 232         id_priv->state = exch;
 233         spin_unlock_irqrestore(&id_priv->lock, flags);
 234         return old;
 235 }
 236
 237 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 238 {
 239         return hdr->ip_version >> 4;
 240 }
 241
 242 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 243 {
 244         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 245 }
 246
 247 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 248                               struct cma_device *cma_dev)
 249 {
 250         atomic_inc(&cma_dev->refcount);
 251         id_priv->cma_dev = cma_dev;
 252         id_priv->id.device = cma_dev->device;
 253         id_priv->id.route.addr.dev_addr.transport =
 254                 rdma_node_get_transport(cma_dev->device->node_type);
 255         list_add_tail(&id_priv->list, &cma_dev->id_list);
 256 }
 257
 258 static inline void cma_deref_dev(struct cma_device *cma_dev)
 259 {
 260         if (atomic_dec_and_test(&cma_dev->refcount))
 261                 complete(&cma_dev->comp);
 262 }
 263
 264 static inline void release_mc(struct kref *kref)
 265 {
 266         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 267
 268         kfree(mc->multicast.ib);
 269         kfree(mc);
 270 }
 271
 272 static void cma_release_dev(struct rdma_id_private *id_priv)
 273 {
 274         mutex_lock(&lock);
 275         list_del(&id_priv->list);
 276         cma_deref_dev(id_priv->cma_dev);
 277         id_priv->cma_dev = NULL;
 278         mutex_unlock(&lock);
 279 }
 280
 281 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 282 {
 283         return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
 284 }
 285
 286 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 287 {
 288         return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
 289 }
 290
 291 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 292 {
 293         return id_priv->id.route.addr.src_addr.ss_family;
 294 }
 295
 296 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 297 {
 298         struct ib_sa_mcmember_rec rec;
 299         int ret = 0;
 300
 301         if (id_priv->qkey) {
 302                 if (qkey && id_priv->qkey != qkey)
 303                         return -EINVAL;
 304                 return 0;
 305         }
 306
 307         if (qkey) {
 308                 id_priv->qkey = qkey;
 309                 return 0;
 310         }
 311
 312         switch (id_priv->id.ps) {
 313         case RDMA_PS_UDP:
 314         case RDMA_PS_IB:
 315                 id_priv->qkey = RDMA_UDP_QKEY;
 316                 break;
 317         case RDMA_PS_IPOIB:
 318                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 319                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 320                                              id_priv->id.port_num, &rec.mgid,
 321                                              &rec);
 322                 if (!ret)
 323                         id_priv->qkey = be32_to_cpu(rec.qkey);
 324                 break;
 325         default:
 326                 break;
 327         }
 328         return ret;
 329 }
 330
 331 static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
 332 {
 333         int i;
 334         int err;
 335         struct ib_port_attr props;
 336         union ib_gid tmp;
 337
 338         err = ib_query_port(device, port_num, &props);
 339         if (err)
 340                 return err;
 341
 342         for (i = 0; i < props.gid_tbl_len; ++i) {
 343                 err = ib_query_gid(device, port_num, i, &tmp);
 344                 if (err)
 345                         return err;
 346                 if (!memcmp(&tmp, gid, sizeof tmp))
 347                         return 0;
 348         }
 349
 350         return -EADDRNOTAVAIL;
 351 }
 352
 353 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 354 {
 355         dev_addr->dev_type = ARPHRD_INFINIBAND;
 356         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 357         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 358 }
 359
 360 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 361 {
 362         int ret;
 363
 364         if (addr->sa_family != AF_IB) {
 365                 ret = rdma_translate_ip(addr, dev_addr);
 366         } else {
 367                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 368                 ret = 0;
 369         }
 370
 371         return ret;
 372 }
 373
 374 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 375 {
 376         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 377         struct cma_device *cma_dev;
 378         union ib_gid gid, iboe_gid;
 379         int ret = -ENODEV;
 380         u8 port;
 381         enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
 382                 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
 383
 384         if (dev_ll != IB_LINK_LAYER_INFINIBAND &&
 385             id_priv->id.ps == RDMA_PS_IPOIB)
 386                 return -EINVAL;
 387
 388         mutex_lock(&lock);
 389         iboe_addr_get_sgid(dev_addr, &iboe_gid);
 390         memcpy(&gid, dev_addr->src_dev_addr +
 391                rdma_addr_gid_offset(dev_addr), sizeof gid);
 392         list_for_each_entry(cma_dev, &dev_list, list) {
 393                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
 394                         if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
 395                                 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
 396                                     rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
 397                                         ret = find_gid_port(cma_dev->device, &iboe_gid, port);
 398                                 else
 399                                         ret = find_gid_port(cma_dev->device, &gid, port);
 400
 401                                 if (!ret) {
 402                                         id_priv->id.port_num = port;
 403                                         goto out;
 404                                 }
 405                         }
 406                 }
 407         }
 408
 409 out:
 410         if (!ret)
 411                 cma_attach_to_dev(id_priv, cma_dev);
 412
 413         mutex_unlock(&lock);
 414         return ret;
 415 }
 416
 417 /*
 418  * Select the source IB device and address to reach the destination IB address.
 419  */
 420 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 421 {
 422         struct cma_device *cma_dev, *cur_dev;
 423         struct sockaddr_ib *addr;
 424         union ib_gid gid, sgid, *dgid;
 425         u16 pkey, index;
 426         u8 p;
 427         int i;
 428
 429         cma_dev = NULL;
 430         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 431         dgid = (union ib_gid *) &addr->sib_addr;
 432         pkey = ntohs(addr->sib_pkey);
 433
 434         list_for_each_entry(cur_dev, &dev_list, list) {
 435                 if (rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
 436                         continue;
 437
 438                 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
 439                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 440                                 continue;
 441
 442                         for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i, &gid); i++) {
 443                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 444                                         cma_dev = cur_dev;
 445                                         sgid = gid;
 446                                         id_priv->id.port_num = p;
 447                                         goto found;
 448                                 }
 449
 450                                 if (!cma_dev && (gid.global.subnet_prefix ==
 451                                                  dgid->global.subnet_prefix)) {
 452                                         cma_dev = cur_dev;
 453                                         sgid = gid;
 454                                         id_priv->id.port_num = p;
 455                                 }
 456                         }
 457                 }
 458         }
 459
 460         if (!cma_dev)
 461                 return -ENODEV;
 462
 463 found:
 464         cma_attach_to_dev(id_priv, cma_dev);
 465         addr = (struct sockaddr_ib *) cma_src_addr(id_priv);
 466         memcpy(&addr->sib_addr, &sgid, sizeof sgid);
 467         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 468         return 0;
 469 }
 470
 471 static void cma_deref_id(struct rdma_id_private *id_priv)
 472 {
 473         if (atomic_dec_and_test(&id_priv->refcount))
 474                 complete(&id_priv->comp);
 475 }
 476
 477 static int cma_disable_callback(struct rdma_id_private *id_priv,
 478                                 enum rdma_cm_state state)
 479 {
 480         mutex_lock(&id_priv->handler_mutex);
 481         if (id_priv->state != state) {
 482                 mutex_unlock(&id_priv->handler_mutex);
 483                 return -EINVAL;
 484         }
 485         return 0;
 486 }
 487
 488 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 489                                   void *context, enum rdma_port_space ps,
 490                                   enum ib_qp_type qp_type)
 491 {
 492         struct rdma_id_private *id_priv;
 493
 494         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 495         if (!id_priv)
 496                 return ERR_PTR(-ENOMEM);
 497
 498         id_priv->owner = task_pid_nr(current);
 499         id_priv->state = RDMA_CM_IDLE;
 500         id_priv->id.context = context;
 501         id_priv->id.event_handler = event_handler;
 502         id_priv->id.ps = ps;
 503         id_priv->id.qp_type = qp_type;
 504         spin_lock_init(&id_priv->lock);
 505         mutex_init(&id_priv->qp_mutex);
 506         init_completion(&id_priv->comp);
 507         atomic_set(&id_priv->refcount, 1);
 508         mutex_init(&id_priv->handler_mutex);
 509         INIT_LIST_HEAD(&id_priv->listen_list);
 510         INIT_LIST_HEAD(&id_priv->mc_list);
 511         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 512
 513         return &id_priv->id;
 514 }
 515 EXPORT_SYMBOL(rdma_create_id);
 516
 517 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 518 {
 519         struct ib_qp_attr qp_attr;
 520         int qp_attr_mask, ret;
 521
 522         qp_attr.qp_state = IB_QPS_INIT;
 523         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 524         if (ret)
 525                 return ret;
 526
 527         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 528         if (ret)
 529                 return ret;
 530
 531         qp_attr.qp_state = IB_QPS_RTR;
 532         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 533         if (ret)
 534                 return ret;
 535
 536         qp_attr.qp_state = IB_QPS_RTS;
 537         qp_attr.sq_psn = 0;
 538         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 539
 540         return ret;
 541 }
 542
 543 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 544 {
 545         struct ib_qp_attr qp_attr;
 546         int qp_attr_mask, ret;
 547
 548         qp_attr.qp_state = IB_QPS_INIT;
 549         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 550         if (ret)
 551                 return ret;
 552
 553         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 554 }
 555
 556 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 557                    struct ib_qp_init_attr *qp_init_attr)
 558 {
 559         struct rdma_id_private *id_priv;
 560         struct ib_qp *qp;
 561         int ret;
 562
 563         id_priv = container_of(id, struct rdma_id_private, id);
 564         if (id->device != pd->device)
 565                 return -EINVAL;
 566
 567         qp = ib_create_qp(pd, qp_init_attr);
 568         if (IS_ERR(qp))
 569                 return PTR_ERR(qp);
 570
 571         if (id->qp_type == IB_QPT_UD)
 572                 ret = cma_init_ud_qp(id_priv, qp);
 573         else
 574                 ret = cma_init_conn_qp(id_priv, qp);
 575         if (ret)
 576                 goto err;
 577
 578         id->qp = qp;
 579         id_priv->qp_num = qp->qp_num;
 580         id_priv->srq = (qp->srq != NULL);
 581         return 0;
 582 err:
 583         ib_destroy_qp(qp);
 584         return ret;
 585 }
 586 EXPORT_SYMBOL(rdma_create_qp);
 587
 588 void rdma_destroy_qp(struct rdma_cm_id *id)
 589 {
 590         struct rdma_id_private *id_priv;
 591
 592         id_priv = container_of(id, struct rdma_id_private, id);
 593         mutex_lock(&id_priv->qp_mutex);
 594         ib_destroy_qp(id_priv->id.qp);
 595         id_priv->id.qp = NULL;
 596         mutex_unlock(&id_priv->qp_mutex);
 597 }
 598 EXPORT_SYMBOL(rdma_destroy_qp);
 599
 600 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 601                              struct rdma_conn_param *conn_param)
 602 {
 603         struct ib_qp_attr qp_attr;
 604         int qp_attr_mask, ret;
 605
 606         mutex_lock(&id_priv->qp_mutex);
 607         if (!id_priv->id.qp) {
 608                 ret = 0;
 609                 goto out;
 610         }
 611
 612         /* Need to update QP attributes from default values. */
 613         qp_attr.qp_state = IB_QPS_INIT;
 614         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 615         if (ret)
 616                 goto out;
 617
 618         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 619         if (ret)
 620                 goto out;
 621
 622         qp_attr.qp_state = IB_QPS_RTR;
 623         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 624         if (ret)
 625                 goto out;
 626
 627         if (conn_param)
 628                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 629         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 630 out:
 631         mutex_unlock(&id_priv->qp_mutex);
 632         return ret;
 633 }
 634
 635 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 636                              struct rdma_conn_param *conn_param)
 637 {
 638         struct ib_qp_attr qp_attr;
 639         int qp_attr_mask, ret;
 640
 641         mutex_lock(&id_priv->qp_mutex);
 642         if (!id_priv->id.qp) {
 643                 ret = 0;
 644                 goto out;
 645         }
 646
 647         qp_attr.qp_state = IB_QPS_RTS;
 648         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 649         if (ret)
 650                 goto out;
 651
 652         if (conn_param)
 653                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 654         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 655 out:
 656         mutex_unlock(&id_priv->qp_mutex);
 657         return ret;
 658 }
 659
 660 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 661 {
 662         struct ib_qp_attr qp_attr;
 663         int ret;
 664
 665         mutex_lock(&id_priv->qp_mutex);
 666         if (!id_priv->id.qp) {
 667                 ret = 0;
 668                 goto out;
 669         }
 670
 671         qp_attr.qp_state = IB_QPS_ERR;
 672         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 673 out:
 674         mutex_unlock(&id_priv->qp_mutex);
 675         return ret;
 676 }
 677
 678 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 679                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 680 {
 681         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 682         int ret;
 683         u16 pkey;
 684
 685         if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
 686             IB_LINK_LAYER_INFINIBAND)
 687                 pkey = ib_addr_get_pkey(dev_addr);
 688         else
 689                 pkey = 0xffff;
 690
 691         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 692                                   pkey, &qp_attr->pkey_index);
 693         if (ret)
 694                 return ret;
 695
 696         qp_attr->port_num = id_priv->id.port_num;
 697         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 698
 699         if (id_priv->id.qp_type == IB_QPT_UD) {
 700                 ret = cma_set_qkey(id_priv, 0);
 701                 if (ret)
 702                         return ret;
 703
 704                 qp_attr->qkey = id_priv->qkey;
 705                 *qp_attr_mask |= IB_QP_QKEY;
 706         } else {
 707                 qp_attr->qp_access_flags = 0;
 708                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 709         }
 710         return 0;
 711 }
 712
 713 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 714                        int *qp_attr_mask)
 715 {
 716         struct rdma_id_private *id_priv;
 717         int ret = 0;
 718
 719         id_priv = container_of(id, struct rdma_id_private, id);
 720         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 721         case RDMA_TRANSPORT_IB:
 722                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
 723                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 724                 else
 725                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 726                                                  qp_attr_mask);
 727                 if (qp_attr->qp_state == IB_QPS_RTR)
 728                         qp_attr->rq_psn = id_priv->seq_num;
 729                 break;
 730         case RDMA_TRANSPORT_IWARP:
 731                 if (!id_priv->cm_id.iw) {
 732                         qp_attr->qp_access_flags = 0;
 733                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 734                 } else
 735                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 736                                                  qp_attr_mask);
 737                 break;
 738         default:
 739                 ret = -ENOSYS;
 740                 break;
 741         }
 742
 743         return ret;
 744 }
 745 EXPORT_SYMBOL(rdma_init_qp_attr);
 746
 747 static inline int cma_zero_addr(struct sockaddr *addr)
 748 {
 749         switch (addr->sa_family) {
 750         case AF_INET:
 751                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
 752         case AF_INET6:
 753                 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
 754         case AF_IB:
 755                 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
 756         default:
 757                 return 0;
 758         }
 759 }
 760
 761 static inline int cma_loopback_addr(struct sockaddr *addr)
 762 {
 763         switch (addr->sa_family) {
 764         case AF_INET:
 765                 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
 766         case AF_INET6:
 767                 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
 768         case AF_IB:
 769                 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
 770         default:
 771                 return 0;
 772         }
 773 }
 774
 775 static inline int cma_any_addr(struct sockaddr *addr)
 776 {
 777         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 778 }
 779
 780 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
 781 {
 782         if (src->sa_family != dst->sa_family)
 783                 return -1;
 784
 785         switch (src->sa_family) {
 786         case AF_INET:
 787                 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
 788                        ((struct sockaddr_in *) dst)->sin_addr.s_addr;
 789         case AF_INET6:
 790                 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
 791                                      &((struct sockaddr_in6 *) dst)->sin6_addr);
 792         default:
 793                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
 794                                    &((struct sockaddr_ib *) dst)->sib_addr);
 795         }
 796 }
 797
 798 static __be16 cma_port(struct sockaddr *addr)
 799 {
 800         struct sockaddr_ib *sib;
 801
 802         switch (addr->sa_family) {
 803         case AF_INET:
 804                 return ((struct sockaddr_in *) addr)->sin_port;
 805         case AF_INET6:
 806                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 807         case AF_IB:
 808                 sib = (struct sockaddr_ib *) addr;
 809                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
 810                                     be64_to_cpu(sib->sib_sid_mask)));
 811         default:
 812                 return 0;
 813         }
 814 }
 815
 816 static inline int cma_any_port(struct sockaddr *addr)
 817 {
 818         return !cma_port(addr);
 819 }
 820
 821 static void cma_save_ib_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
 822                              struct ib_sa_path_rec *path)
 823 {
 824         struct sockaddr_ib *listen_ib, *ib;
 825
 826         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
 827         ib = (struct sockaddr_ib *) &id->route.addr.src_addr;
 828         ib->sib_family = listen_ib->sib_family;
 829         ib->sib_pkey = path->pkey;
 830         ib->sib_flowinfo = path->flow_label;
 831         memcpy(&ib->sib_addr, &path->sgid, 16);
 832         ib->sib_sid = listen_ib->sib_sid;
 833         ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
 834         ib->sib_scope_id = listen_ib->sib_scope_id;
 835
 836         ib = (struct sockaddr_ib *) &id->route.addr.dst_addr;
 837         ib->sib_family = listen_ib->sib_family;
 838         ib->sib_pkey = path->pkey;
 839         ib->sib_flowinfo = path->flow_label;
 840         memcpy(&ib->sib_addr, &path->dgid, 16);
 841 }
 842
 843 static void cma_save_ip4_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
 844                               struct cma_hdr *hdr)
 845 {
 846         struct sockaddr_in *listen4, *ip4;
 847
 848         listen4 = (struct sockaddr_in *) &listen_id->route.addr.src_addr;
 849         ip4 = (struct sockaddr_in *) &id->route.addr.src_addr;
 850         ip4->sin_family = listen4->sin_family;
 851         ip4->sin_addr.s_addr = hdr->dst_addr.ip4.addr;
 852         ip4->sin_port = listen4->sin_port;
 853
 854         ip4 = (struct sockaddr_in *) &id->route.addr.dst_addr;
 855         ip4->sin_family = listen4->sin_family;
 856         ip4->sin_addr.s_addr = hdr->src_addr.ip4.addr;
 857         ip4->sin_port = hdr->port;
 858 }
 859
 860 static void cma_save_ip6_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
 861                               struct cma_hdr *hdr)
 862 {
 863         struct sockaddr_in6 *listen6, *ip6;
 864
 865         listen6 = (struct sockaddr_in6 *) &listen_id->route.addr.src_addr;
 866         ip6 = (struct sockaddr_in6 *) &id->route.addr.src_addr;
 867         ip6->sin6_family = listen6->sin6_family;
 868         ip6->sin6_addr = hdr->dst_addr.ip6;
 869         ip6->sin6_port = listen6->sin6_port;
 870
 871         ip6 = (struct sockaddr_in6 *) &id->route.addr.dst_addr;
 872         ip6->sin6_family = listen6->sin6_family;
 873         ip6->sin6_addr = hdr->src_addr.ip6;
 874         ip6->sin6_port = hdr->port;
 875 }
 876
 877 static int cma_save_net_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
 878                              struct ib_cm_event *ib_event)
 879 {
 880         struct cma_hdr *hdr;
 881
 882         if (listen_id->route.addr.src_addr.ss_family == AF_IB) {
 883                 cma_save_ib_info(id, listen_id, ib_event->param.req_rcvd.primary_path);
 884                 return 0;
 885         }
 886
 887         hdr = ib_event->private_data;
 888         if (hdr->cma_version != CMA_VERSION)
 889                 return -EINVAL;
 890
 891         switch (cma_get_ip_ver(hdr)) {
 892         case 4:
 893                 cma_save_ip4_info(id, listen_id, hdr);
 894                 break;
 895         case 6:
 896                 cma_save_ip6_info(id, listen_id, hdr);
 897                 break;
 898         default:
 899                 return -EINVAL;
 900         }
 901         return 0;
 902 }
 903
 904 static inline int cma_user_data_offset(struct rdma_id_private *id_priv)
 905 {
 906         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
 907 }
 908
 909 static void cma_cancel_route(struct rdma_id_private *id_priv)
 910 {
 911         switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
 912         case IB_LINK_LAYER_INFINIBAND:
 913                 if (id_priv->query)
 914                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 915                 break;
 916         default:
 917                 break;
 918         }
 919 }
 920
 921 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 922 {
 923         struct rdma_id_private *dev_id_priv;
 924
 925         /*
 926          * Remove from listen_any_list to prevent added devices from spawning
 927          * additional listen requests.
 928          */
 929         mutex_lock(&lock);
 930         list_del(&id_priv->list);
 931
 932         while (!list_empty(&id_priv->listen_list)) {
 933                 dev_id_priv = list_entry(id_priv->listen_list.next,
 934                                          struct rdma_id_private, listen_list);
 935                 /* sync with device removal to avoid duplicate destruction */
 936                 list_del_init(&dev_id_priv->list);
 937                 list_del(&dev_id_priv->listen_list);
 938                 mutex_unlock(&lock);
 939
 940                 rdma_destroy_id(&dev_id_priv->id);
 941                 mutex_lock(&lock);
 942         }
 943         mutex_unlock(&lock);
 944 }
 945
 946 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 947                                  enum rdma_cm_state state)
 948 {
 949         switch (state) {
 950         case RDMA_CM_ADDR_QUERY:
 951                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 952                 break;
 953         case RDMA_CM_ROUTE_QUERY:
 954                 cma_cancel_route(id_priv);
 955                 break;
 956         case RDMA_CM_LISTEN:
 957                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
 958                         cma_cancel_listens(id_priv);
 959                 break;
 960         default:
 961                 break;
 962         }
 963 }
 964
 965 static void cma_release_port(struct rdma_id_private *id_priv)
 966 {
 967         struct rdma_bind_list *bind_list = id_priv->bind_list;
 968
 969         if (!bind_list)
 970                 return;
 971
 972         mutex_lock(&lock);
 973         hlist_del(&id_priv->node);
 974         if (hlist_empty(&bind_list->owners)) {
 975                 idr_remove(bind_list->ps, bind_list->port);
 976                 kfree(bind_list);
 977         }
 978         mutex_unlock(&lock);
 979 }
 980
 981 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 982 {
 983         struct cma_multicast *mc;
 984
 985         while (!list_empty(&id_priv->mc_list)) {
 986                 mc = container_of(id_priv->mc_list.next,
 987                                   struct cma_multicast, list);
 988                 list_del(&mc->list);
 989                 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
 990                 case IB_LINK_LAYER_INFINIBAND:
 991                         ib_sa_free_multicast(mc->multicast.ib);
 992                         kfree(mc);
 993                         break;
 994                 case IB_LINK_LAYER_ETHERNET:
 995                         kref_put(&mc->mcref, release_mc);
 996                         break;
 997                 default:
 998                         break;
 999                 }
1000         }
1001 }
1002
1003 void rdma_destroy_id(struct rdma_cm_id *id)
1004 {
1005         struct rdma_id_private *id_priv;
1006         enum rdma_cm_state state;
1007
1008         id_priv = container_of(id, struct rdma_id_private, id);
1009         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1010         cma_cancel_operation(id_priv, state);
1011
1012         /*
1013          * Wait for any active callback to finish.  New callbacks will find
1014          * the id_priv state set to destroying and abort.
1015          */
1016         mutex_lock(&id_priv->handler_mutex);
1017         mutex_unlock(&id_priv->handler_mutex);
1018
1019         if (id_priv->cma_dev) {
1020                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
1021                 case RDMA_TRANSPORT_IB:
1022                         if (id_priv->cm_id.ib)
1023                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1024                         break;
1025                 case RDMA_TRANSPORT_IWARP:
1026                         if (id_priv->cm_id.iw)
1027                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1028                         break;
1029                 default:
1030                         break;
1031                 }
1032                 cma_leave_mc_groups(id_priv);
1033                 cma_release_dev(id_priv);
1034         }
1035
1036         cma_release_port(id_priv);
1037         cma_deref_id(id_priv);
1038         wait_for_completion(&id_priv->comp);
1039
1040         if (id_priv->internal_id)
1041                 cma_deref_id(id_priv->id.context);
1042
1043         kfree(id_priv->id.route.path_rec);
1044         kfree(id_priv);
1045 }
1046 EXPORT_SYMBOL(rdma_destroy_id);
1047
1048 static int cma_rep_recv(struct rdma_id_private *id_priv)
1049 {
1050         int ret;
1051
1052         ret = cma_modify_qp_rtr(id_priv, NULL);
1053         if (ret)
1054                 goto reject;
1055
1056         ret = cma_modify_qp_rts(id_priv, NULL);
1057         if (ret)
1058                 goto reject;
1059
1060         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1061         if (ret)
1062                 goto reject;
1063
1064         return 0;
1065 reject:
1066         cma_modify_qp_err(id_priv);
1067         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1068                        NULL, 0, NULL, 0);
1069         return ret;
1070 }
1071
1072 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1073                                    struct ib_cm_rep_event_param *rep_data,
1074                                    void *private_data)
1075 {
1076         event->param.conn.private_data = private_data;
1077         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1078         event->param.conn.responder_resources = rep_data->responder_resources;
1079         event->param.conn.initiator_depth = rep_data->initiator_depth;
1080         event->param.conn.flow_control = rep_data->flow_control;
1081         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1082         event->param.conn.srq = rep_data->srq;
1083         event->param.conn.qp_num = rep_data->remote_qpn;
1084 }
1085
1086 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1087 {
1088         struct rdma_id_private *id_priv = cm_id->context;
1089         struct rdma_cm_event event;
1090         int ret = 0;
1091
1092         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1093                 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1094             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1095                 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1096                 return 0;
1097
1098         memset(&event, 0, sizeof event);
1099         switch (ib_event->event) {
1100         case IB_CM_REQ_ERROR:
1101         case IB_CM_REP_ERROR:
1102                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1103                 event.status = -ETIMEDOUT;
1104                 break;
1105         case IB_CM_REP_RECEIVED:
1106                 if (id_priv->id.qp) {
1107                         event.status = cma_rep_recv(id_priv);
1108                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1109                                                      RDMA_CM_EVENT_ESTABLISHED;
1110                 } else {
1111                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1112                 }
1113                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1114                                        ib_event->private_data);
1115                 break;
1116         case IB_CM_RTU_RECEIVED:
1117         case IB_CM_USER_ESTABLISHED:
1118                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1119                 break;
1120         case IB_CM_DREQ_ERROR:
1121                 event.status = -ETIMEDOUT; /* fall through */
1122         case IB_CM_DREQ_RECEIVED:
1123         case IB_CM_DREP_RECEIVED:
1124                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1125                                    RDMA_CM_DISCONNECT))
1126                         goto out;
1127                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1128                 break;
1129         case IB_CM_TIMEWAIT_EXIT:
1130                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1131                 break;
1132         case IB_CM_MRA_RECEIVED:
1133                 /* ignore event */
1134                 goto out;
1135         case IB_CM_REJ_RECEIVED:
1136                 cma_modify_qp_err(id_priv);
1137                 event.status = ib_event->param.rej_rcvd.reason;
1138                 event.event = RDMA_CM_EVENT_REJECTED;
1139                 event.param.conn.private_data = ib_event->private_data;
1140                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1141                 break;
1142         default:
1143                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1144                        ib_event->event);
1145                 goto out;
1146         }
1147
1148         ret = id_priv->id.event_handler(&id_priv->id, &event);
1149         if (ret) {
1150                 /* Destroy the CM ID by returning a non-zero value. */
1151                 id_priv->cm_id.ib = NULL;
1152                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1153                 mutex_unlock(&id_priv->handler_mutex);
1154                 rdma_destroy_id(&id_priv->id);
1155                 return ret;
1156         }
1157 out:
1158         mutex_unlock(&id_priv->handler_mutex);
1159         return ret;
1160 }
1161
1162 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1163                                                struct ib_cm_event *ib_event)
1164 {
1165         struct rdma_id_private *id_priv;
1166         struct rdma_cm_id *id;
1167         struct rdma_route *rt;
1168         int ret;
1169
1170         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1171                             listen_id->ps, ib_event->param.req_rcvd.qp_type);
1172         if (IS_ERR(id))
1173                 return NULL;
1174
1175         id_priv = container_of(id, struct rdma_id_private, id);
1176         if (cma_save_net_info(id, listen_id, ib_event))
1177                 goto err;
1178
1179         rt = &id->route;
1180         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1181         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1182                                GFP_KERNEL);
1183         if (!rt->path_rec)
1184                 goto err;
1185
1186         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1187         if (rt->num_paths == 2)
1188                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1189
1190         if (cma_any_addr(cma_src_addr(id_priv))) {
1191                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1192                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1193                 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1194         } else {
1195                 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1196                 if (ret)
1197                         goto err;
1198         }
1199         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1200
1201         id_priv->state = RDMA_CM_CONNECT;
1202         return id_priv;
1203
1204 err:
1205         rdma_destroy_id(id);
1206         return NULL;
1207 }
1208
1209 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1210                                               struct ib_cm_event *ib_event)
1211 {
1212         struct rdma_id_private *id_priv;
1213         struct rdma_cm_id *id;
1214         int ret;
1215
1216         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1217                             listen_id->ps, IB_QPT_UD);
1218         if (IS_ERR(id))
1219                 return NULL;
1220
1221         id_priv = container_of(id, struct rdma_id_private, id);
1222         if (cma_save_net_info(id, listen_id, ib_event))
1223                 goto err;
1224
1225         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1226                 ret = cma_translate_addr(cma_src_addr(id_priv), &id->route.addr.dev_addr);
1227                 if (ret)
1228                         goto err;
1229         }
1230
1231         id_priv->state = RDMA_CM_CONNECT;
1232         return id_priv;
1233 err:
1234         rdma_destroy_id(id);
1235         return NULL;
1236 }
1237
1238 static void cma_set_req_event_data(struct rdma_cm_event *event,
1239                                    struct ib_cm_req_event_param *req_data,
1240                                    void *private_data, int offset)
1241 {
1242         event->param.conn.private_data = private_data + offset;
1243         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1244         event->param.conn.responder_resources = req_data->responder_resources;
1245         event->param.conn.initiator_depth = req_data->initiator_depth;
1246         event->param.conn.flow_control = req_data->flow_control;
1247         event->param.conn.retry_count = req_data->retry_count;
1248         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1249         event->param.conn.srq = req_data->srq;
1250         event->param.conn.qp_num = req_data->remote_qpn;
1251 }
1252
1253 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1254 {
1255         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1256                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1257                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1258                  (id->qp_type == IB_QPT_UD)) ||
1259                 (!id->qp_type));
1260 }
1261
1262 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1263 {
1264         struct rdma_id_private *listen_id, *conn_id;
1265         struct rdma_cm_event event;
1266         int offset, ret;
1267
1268         listen_id = cm_id->context;
1269         if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1270                 return -EINVAL;
1271
1272         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1273                 return -ECONNABORTED;
1274
1275         memset(&event, 0, sizeof event);
1276         offset = cma_user_data_offset(listen_id);
1277         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1278         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1279                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1280                 event.param.ud.private_data = ib_event->private_data + offset;
1281                 event.param.ud.private_data_len =
1282                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1283         } else {
1284                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1285                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1286                                        ib_event->private_data, offset);
1287         }
1288         if (!conn_id) {
1289                 ret = -ENOMEM;
1290                 goto err1;
1291         }
1292
1293         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1294         ret = cma_acquire_dev(conn_id);
1295         if (ret)
1296                 goto err2;
1297
1298         conn_id->cm_id.ib = cm_id;
1299         cm_id->context = conn_id;
1300         cm_id->cm_handler = cma_ib_handler;
1301
1302         /*
1303          * Protect against the user destroying conn_id from another thread
1304          * until we're done accessing it.
1305          */
1306         atomic_inc(&conn_id->refcount);
1307         ret = conn_id->id.event_handler(&conn_id->id, &event);
1308         if (ret)
1309                 goto err3;
1310
1311         /*
1312          * Acquire mutex to prevent user executing rdma_destroy_id()
1313          * while we're accessing the cm_id.
1314          */
1315         mutex_lock(&lock);
1316         if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD))
1317                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1318         mutex_unlock(&lock);
1319         mutex_unlock(&conn_id->handler_mutex);
1320         mutex_unlock(&listen_id->handler_mutex);
1321         cma_deref_id(conn_id);
1322         return 0;
1323
1324 err3:
1325         cma_deref_id(conn_id);
1326         /* Destroy the CM ID by returning a non-zero value. */
1327         conn_id->cm_id.ib = NULL;
1328 err2:
1329         cma_exch(conn_id, RDMA_CM_DESTROYING);
1330         mutex_unlock(&conn_id->handler_mutex);
1331 err1:
1332         mutex_unlock(&listen_id->handler_mutex);
1333         if (conn_id)
1334                 rdma_destroy_id(&conn_id->id);
1335         return ret;
1336 }
1337
1338 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
1339 {
1340         if (addr->sa_family == AF_IB)
1341                 return ((struct sockaddr_ib *) addr)->sib_sid;
1342
1343         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
1344 }
1345 EXPORT_SYMBOL(rdma_get_service_id);
1346
1347 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1348                                  struct ib_cm_compare_data *compare)
1349 {
1350         struct cma_hdr *cma_data, *cma_mask;
1351         __be32 ip4_addr;
1352         struct in6_addr ip6_addr;
1353
1354         memset(compare, 0, sizeof *compare);
1355         cma_data = (void *) compare->data;
1356         cma_mask = (void *) compare->mask;
1357
1358         switch (addr->sa_family) {
1359         case AF_INET:
1360                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1361                 cma_set_ip_ver(cma_data, 4);
1362                 cma_set_ip_ver(cma_mask, 0xF);
1363                 if (!cma_any_addr(addr)) {
1364                         cma_data->dst_addr.ip4.addr = ip4_addr;
1365                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1366                 }
1367                 break;
1368         case AF_INET6:
1369                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1370                 cma_set_ip_ver(cma_data, 6);
1371                 cma_set_ip_ver(cma_mask, 0xF);
1372                 if (!cma_any_addr(addr)) {
1373                         cma_data->dst_addr.ip6 = ip6_addr;
1374                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1375                                sizeof cma_mask->dst_addr.ip6);
1376                 }
1377                 break;
1378         default:
1379                 break;
1380         }
1381 }
1382
1383 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1384 {
1385         struct rdma_id_private *id_priv = iw_id->context;
1386         struct rdma_cm_event event;
1387         struct sockaddr_in *sin;
1388         int ret = 0;
1389
1390         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1391                 return 0;
1392
1393         memset(&event, 0, sizeof event);
1394         switch (iw_event->event) {
1395         case IW_CM_EVENT_CLOSE:
1396                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1397                 break;
1398         case IW_CM_EVENT_CONNECT_REPLY:
1399                 sin = (struct sockaddr_in *) cma_src_addr(id_priv);
1400                 *sin = iw_event->local_addr;
1401                 sin = (struct sockaddr_in *) cma_dst_addr(id_priv);
1402                 *sin = iw_event->remote_addr;
1403                 switch (iw_event->status) {
1404                 case 0:
1405                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1406                         event.param.conn.initiator_depth = iw_event->ird;
1407                         event.param.conn.responder_resources = iw_event->ord;
1408                         break;
1409                 case -ECONNRESET:
1410                 case -ECONNREFUSED:
1411                         event.event = RDMA_CM_EVENT_REJECTED;
1412                         break;
1413                 case -ETIMEDOUT:
1414                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1415                         break;
1416                 default:
1417                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1418                         break;
1419                 }
1420                 break;
1421         case IW_CM_EVENT_ESTABLISHED:
1422                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1423                 event.param.conn.initiator_depth = iw_event->ird;
1424                 event.param.conn.responder_resources = iw_event->ord;
1425                 break;
1426         default:
1427                 BUG_ON(1);
1428         }
1429
1430         event.status = iw_event->status;
1431         event.param.conn.private_data = iw_event->private_data;
1432         event.param.conn.private_data_len = iw_event->private_data_len;
1433         ret = id_priv->id.event_handler(&id_priv->id, &event);
1434         if (ret) {
1435                 /* Destroy the CM ID by returning a non-zero value. */
1436                 id_priv->cm_id.iw = NULL;
1437                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1438                 mutex_unlock(&id_priv->handler_mutex);
1439                 rdma_destroy_id(&id_priv->id);
1440                 return ret;
1441         }
1442
1443         mutex_unlock(&id_priv->handler_mutex);
1444         return ret;
1445 }
1446
1447 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1448                                struct iw_cm_event *iw_event)
1449 {
1450         struct rdma_cm_id *new_cm_id;
1451         struct rdma_id_private *listen_id, *conn_id;
1452         struct sockaddr_in *sin;
1453         struct net_device *dev = NULL;
1454         struct rdma_cm_event event;
1455         int ret;
1456         struct ib_device_attr attr;
1457
1458         listen_id = cm_id->context;
1459         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1460                 return -ECONNABORTED;
1461
1462         /* Create a new RDMA id for the new IW CM ID */
1463         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1464                                    listen_id->id.context,
1465                                    RDMA_PS_TCP, IB_QPT_RC);
1466         if (IS_ERR(new_cm_id)) {
1467                 ret = -ENOMEM;
1468                 goto out;
1469         }
1470         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1471         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1472         conn_id->state = RDMA_CM_CONNECT;
1473
1474         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1475         if (!dev) {
1476                 ret = -EADDRNOTAVAIL;
1477                 mutex_unlock(&conn_id->handler_mutex);
1478                 rdma_destroy_id(new_cm_id);
1479                 goto out;
1480         }
1481         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1482         if (ret) {
1483                 mutex_unlock(&conn_id->handler_mutex);
1484                 rdma_destroy_id(new_cm_id);
1485                 goto out;
1486         }
1487
1488         ret = cma_acquire_dev(conn_id);
1489         if (ret) {
1490                 mutex_unlock(&conn_id->handler_mutex);
1491                 rdma_destroy_id(new_cm_id);
1492                 goto out;
1493         }
1494
1495         conn_id->cm_id.iw = cm_id;
1496         cm_id->context = conn_id;
1497         cm_id->cm_handler = cma_iw_handler;
1498
1499         sin = (struct sockaddr_in *) cma_src_addr(conn_id);
1500         *sin = iw_event->local_addr;
1501         sin = (struct sockaddr_in *) cma_dst_addr(conn_id);
1502         *sin = iw_event->remote_addr;
1503
1504         ret = ib_query_device(conn_id->id.device, &attr);
1505         if (ret) {
1506                 mutex_unlock(&conn_id->handler_mutex);
1507                 rdma_destroy_id(new_cm_id);
1508                 goto out;
1509         }
1510
1511         memset(&event, 0, sizeof event);
1512         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1513         event.param.conn.private_data = iw_event->private_data;
1514         event.param.conn.private_data_len = iw_event->private_data_len;
1515         event.param.conn.initiator_depth = iw_event->ird;
1516         event.param.conn.responder_resources = iw_event->ord;
1517
1518         /*
1519          * Protect against the user destroying conn_id from another thread
1520          * until we're done accessing it.
1521          */
1522         atomic_inc(&conn_id->refcount);
1523         ret = conn_id->id.event_handler(&conn_id->id, &event);
1524         if (ret) {
1525                 /* User wants to destroy the CM ID */
1526                 conn_id->cm_id.iw = NULL;
1527                 cma_exch(conn_id, RDMA_CM_DESTROYING);
1528                 mutex_unlock(&conn_id->handler_mutex);
1529                 cma_deref_id(conn_id);
1530                 rdma_destroy_id(&conn_id->id);
1531                 goto out;
1532         }
1533
1534         mutex_unlock(&conn_id->handler_mutex);
1535         cma_deref_id(conn_id);
1536
1537 out:
1538         if (dev)
1539                 dev_put(dev);
1540         mutex_unlock(&listen_id->handler_mutex);
1541         return ret;
1542 }
1543
1544 static int cma_ib_listen(struct rdma_id_private *id_priv)
1545 {
1546         struct ib_cm_compare_data compare_data;
1547         struct sockaddr *addr;
1548         struct ib_cm_id *id;
1549         __be64 svc_id;
1550         int ret;
1551
1552         id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1553         if (IS_ERR(id))
1554                 return PTR_ERR(id);
1555
1556         id_priv->cm_id.ib = id;
1557
1558         addr = cma_src_addr(id_priv);
1559         svc_id = rdma_get_service_id(&id_priv->id, addr);
1560         if (cma_any_addr(addr) && !id_priv->afonly)
1561                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1562         else {
1563                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1564                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1565         }
1566
1567         if (ret) {
1568                 ib_destroy_cm_id(id_priv->cm_id.ib);
1569                 id_priv->cm_id.ib = NULL;
1570         }
1571
1572         return ret;
1573 }
1574
1575 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1576 {
1577         int ret;
1578         struct sockaddr_in *sin;
1579         struct iw_cm_id *id;
1580
1581         id = iw_create_cm_id(id_priv->id.device,
1582                              iw_conn_req_handler,
1583                              id_priv);
1584         if (IS_ERR(id))
1585                 return PTR_ERR(id);
1586
1587         id_priv->cm_id.iw = id;
1588
1589         sin = (struct sockaddr_in *) cma_src_addr(id_priv);
1590         id_priv->cm_id.iw->local_addr = *sin;
1591
1592         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1593
1594         if (ret) {
1595                 iw_destroy_cm_id(id_priv->cm_id.iw);
1596                 id_priv->cm_id.iw = NULL;
1597         }
1598
1599         return ret;
1600 }
1601
1602 static int cma_listen_handler(struct rdma_cm_id *id,
1603                               struct rdma_cm_event *event)
1604 {
1605         struct rdma_id_private *id_priv = id->context;
1606
1607         id->context = id_priv->id.context;
1608         id->event_handler = id_priv->id.event_handler;
1609         return id_priv->id.event_handler(id, event);
1610 }
1611
1612 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1613                               struct cma_device *cma_dev)
1614 {
1615         struct rdma_id_private *dev_id_priv;
1616         struct rdma_cm_id *id;
1617         int ret;
1618
1619         if (cma_family(id_priv) == AF_IB &&
1620             rdma_node_get_transport(cma_dev->device->node_type) != RDMA_TRANSPORT_IB)
1621                 return;
1622
1623         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1624                             id_priv->id.qp_type);
1625         if (IS_ERR(id))
1626                 return;
1627
1628         dev_id_priv = container_of(id, struct rdma_id_private, id);
1629
1630         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1631         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
1632                rdma_addr_size(cma_src_addr(id_priv)));
1633
1634         cma_attach_to_dev(dev_id_priv, cma_dev);
1635         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1636         atomic_inc(&id_priv->refcount);
1637         dev_id_priv->internal_id = 1;
1638         dev_id_priv->afonly = id_priv->afonly;
1639
1640         ret = rdma_listen(id, id_priv->backlog);
1641         if (ret)
1642                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1643                        "listening on device %s\n", ret, cma_dev->device->name);
1644 }
1645
1646 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1647 {
1648         struct cma_device *cma_dev;
1649
1650         mutex_lock(&lock);
1651         list_add_tail(&id_priv->list, &listen_any_list);
1652         list_for_each_entry(cma_dev, &dev_list, list)
1653                 cma_listen_on_dev(id_priv, cma_dev);
1654         mutex_unlock(&lock);
1655 }
1656
1657 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1658 {
1659         struct rdma_id_private *id_priv;
1660
1661         id_priv = container_of(id, struct rdma_id_private, id);
1662         id_priv->tos = (u8) tos;
1663 }
1664 EXPORT_SYMBOL(rdma_set_service_type);
1665
1666 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1667                               void *context)
1668 {
1669         struct cma_work *work = context;
1670         struct rdma_route *route;
1671
1672         route = &work->id->id.route;
1673
1674         if (!status) {
1675                 route->num_paths = 1;
1676                 *route->path_rec = *path_rec;
1677         } else {
1678                 work->old_state = RDMA_CM_ROUTE_QUERY;
1679                 work->new_state = RDMA_CM_ADDR_RESOLVED;
1680                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1681                 work->event.status = status;
1682         }
1683
1684         queue_work(cma_wq, &work->work);
1685 }
1686
1687 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1688                               struct cma_work *work)
1689 {
1690         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1691         struct ib_sa_path_rec path_rec;
1692         ib_sa_comp_mask comp_mask;
1693         struct sockaddr_in6 *sin6;
1694         struct sockaddr_ib *sib;
1695
1696         memset(&path_rec, 0, sizeof path_rec);
1697         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
1698         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
1699         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
1700         path_rec.numb_path = 1;
1701         path_rec.reversible = 1;
1702         path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
1703
1704         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1705                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1706                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1707
1708         switch (cma_family(id_priv)) {
1709         case AF_INET:
1710                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1711                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1712                 break;
1713         case AF_INET6:
1714                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
1715                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1716                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1717                 break;
1718         case AF_IB:
1719                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
1720                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
1721                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1722                 break;
1723         }
1724
1725         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1726                                                id_priv->id.port_num, &path_rec,
1727                                                comp_mask, timeout_ms,
1728                                                GFP_KERNEL, cma_query_handler,
1729                                                work, &id_priv->query);
1730
1731         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1732 }
1733
1734 static void cma_work_handler(struct work_struct *_work)
1735 {
1736         struct cma_work *work = container_of(_work, struct cma_work, work);
1737         struct rdma_id_private *id_priv = work->id;
1738         int destroy = 0;
1739
1740         mutex_lock(&id_priv->handler_mutex);
1741         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1742                 goto out;
1743
1744         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1745                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1746                 destroy = 1;
1747         }
1748 out:
1749         mutex_unlock(&id_priv->handler_mutex);
1750         cma_deref_id(id_priv);
1751         if (destroy)
1752                 rdma_destroy_id(&id_priv->id);
1753         kfree(work);
1754 }
1755
1756 static void cma_ndev_work_handler(struct work_struct *_work)
1757 {
1758         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1759         struct rdma_id_private *id_priv = work->id;
1760         int destroy = 0;
1761
1762         mutex_lock(&id_priv->handler_mutex);
1763         if (id_priv->state == RDMA_CM_DESTROYING ||
1764             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1765                 goto out;
1766
1767         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1768                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1769                 destroy = 1;
1770         }
1771
1772 out:
1773         mutex_unlock(&id_priv->handler_mutex);
1774         cma_deref_id(id_priv);
1775         if (destroy)
1776                 rdma_destroy_id(&id_priv->id);
1777         kfree(work);
1778 }
1779
1780 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1781 {
1782         struct rdma_route *route = &id_priv->id.route;
1783         struct cma_work *work;
1784         int ret;
1785
1786         work = kzalloc(sizeof *work, GFP_KERNEL);
1787         if (!work)
1788                 return -ENOMEM;
1789
1790         work->id = id_priv;
1791         INIT_WORK(&work->work, cma_work_handler);
1792         work->old_state = RDMA_CM_ROUTE_QUERY;
1793         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1794         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1795
1796         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1797         if (!route->path_rec) {
1798                 ret = -ENOMEM;
1799                 goto err1;
1800         }
1801
1802         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1803         if (ret)
1804                 goto err2;
1805
1806         return 0;
1807 err2:
1808         kfree(route->path_rec);
1809         route->path_rec = NULL;
1810 err1:
1811         kfree(work);
1812         return ret;
1813 }
1814
1815 int rdma_set_ib_paths(struct rdma_cm_id *id,
1816                       struct ib_sa_path_rec *path_rec, int num_paths)
1817 {
1818         struct rdma_id_private *id_priv;
1819         int ret;
1820
1821         id_priv = container_of(id, struct rdma_id_private, id);
1822         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1823                            RDMA_CM_ROUTE_RESOLVED))
1824                 return -EINVAL;
1825
1826         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1827                                      GFP_KERNEL);
1828         if (!id->route.path_rec) {
1829                 ret = -ENOMEM;
1830                 goto err;
1831         }
1832
1833         id->route.num_paths = num_paths;
1834         return 0;
1835 err:
1836         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1837         return ret;
1838 }
1839 EXPORT_SYMBOL(rdma_set_ib_paths);
1840
1841 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1842 {
1843         struct cma_work *work;
1844
1845         work = kzalloc(sizeof *work, GFP_KERNEL);
1846         if (!work)
1847                 return -ENOMEM;
1848
1849         work->id = id_priv;
1850         INIT_WORK(&work->work, cma_work_handler);
1851         work->old_state = RDMA_CM_ROUTE_QUERY;
1852         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1853         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1854         queue_work(cma_wq, &work->work);
1855         return 0;
1856 }
1857
1858 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1859 {
1860         struct rdma_route *route = &id_priv->id.route;
1861         struct rdma_addr *addr = &route->addr;
1862         struct cma_work *work;
1863         int ret;
1864         struct net_device *ndev = NULL;
1865         u16 vid;
1866
1867         work = kzalloc(sizeof *work, GFP_KERNEL);
1868         if (!work)
1869                 return -ENOMEM;
1870
1871         work->id = id_priv;
1872         INIT_WORK(&work->work, cma_work_handler);
1873
1874         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1875         if (!route->path_rec) {
1876                 ret = -ENOMEM;
1877                 goto err1;
1878         }
1879
1880         route->num_paths = 1;
1881
1882         if (addr->dev_addr.bound_dev_if)
1883                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1884         if (!ndev) {
1885                 ret = -ENODEV;
1886                 goto err2;
1887         }
1888
1889         vid = rdma_vlan_dev_vlan_id(ndev);
1890
1891         iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1892         iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1893
1894         route->path_rec->hop_limit = 1;
1895         route->path_rec->reversible = 1;
1896         route->path_rec->pkey = cpu_to_be16(0xffff);
1897         route->path_rec->mtu_selector = IB_SA_EQ;
1898         route->path_rec->sl = netdev_get_prio_tc_map(
1899                         ndev->priv_flags & IFF_802_1Q_VLAN ?
1900                                 vlan_dev_real_dev(ndev) : ndev,
1901                         rt_tos2priority(id_priv->tos));
1902
1903         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1904         route->path_rec->rate_selector = IB_SA_EQ;
1905         route->path_rec->rate = iboe_get_rate(ndev);
1906         dev_put(ndev);
1907         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1908         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1909         if (!route->path_rec->mtu) {
1910                 ret = -EINVAL;
1911                 goto err2;
1912         }
1913
1914         work->old_state = RDMA_CM_ROUTE_QUERY;
1915         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1916         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1917         work->event.status = 0;
1918
1919         queue_work(cma_wq, &work->work);
1920
1921         return 0;
1922
1923 err2:
1924         kfree(route->path_rec);
1925         route->path_rec = NULL;
1926 err1:
1927         kfree(work);
1928         return ret;
1929 }
1930
1931 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1932 {
1933         struct rdma_id_private *id_priv;
1934         int ret;
1935
1936         id_priv = container_of(id, struct rdma_id_private, id);
1937         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1938                 return -EINVAL;
1939
1940         atomic_inc(&id_priv->refcount);
1941         switch (rdma_node_get_transport(id->device->node_type)) {
1942         case RDMA_TRANSPORT_IB:
1943                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1944                 case IB_LINK_LAYER_INFINIBAND:
1945                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1946                         break;
1947                 case IB_LINK_LAYER_ETHERNET:
1948                         ret = cma_resolve_iboe_route(id_priv);
1949                         break;
1950                 default:
1951                         ret = -ENOSYS;
1952                 }
1953                 break;
1954         case RDMA_TRANSPORT_IWARP:
1955                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1956                 break;
1957         default:
1958                 ret = -ENOSYS;
1959                 break;
1960         }
1961         if (ret)
1962                 goto err;
1963
1964         return 0;
1965 err:
1966         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1967         cma_deref_id(id_priv);
1968         return ret;
1969 }
1970 EXPORT_SYMBOL(rdma_resolve_route);
1971
1972 static void cma_set_loopback(struct sockaddr *addr)
1973 {
1974         switch (addr->sa_family) {
1975         case AF_INET:
1976                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
1977                 break;
1978         case AF_INET6:
1979                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
1980                               0, 0, 0, htonl(1));
1981                 break;
1982         default:
1983                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
1984                             0, 0, 0, htonl(1));
1985                 break;
1986         }
1987 }
1988
1989 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1990 {
1991         struct cma_device *cma_dev, *cur_dev;
1992         struct ib_port_attr port_attr;
1993         union ib_gid gid;
1994         u16 pkey;
1995         int ret;
1996         u8 p;
1997
1998         cma_dev = NULL;
1999         mutex_lock(&lock);
2000         list_for_each_entry(cur_dev, &dev_list, list) {
2001                 if (cma_family(id_priv) == AF_IB &&
2002                     rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
2003                         continue;
2004
2005                 if (!cma_dev)
2006                         cma_dev = cur_dev;
2007
2008                 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2009                         if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2010                             port_attr.state == IB_PORT_ACTIVE) {
2011                                 cma_dev = cur_dev;
2012                                 goto port_found;
2013                         }
2014                 }
2015         }
2016
2017         if (!cma_dev) {
2018                 ret = -ENODEV;
2019                 goto out;
2020         }
2021
2022         p = 1;
2023
2024 port_found:
2025         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
2026         if (ret)
2027                 goto out;
2028
2029         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2030         if (ret)
2031                 goto out;
2032
2033         id_priv->id.route.addr.dev_addr.dev_type =
2034                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
2035                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2036
2037         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2038         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2039         id_priv->id.port_num = p;
2040         cma_attach_to_dev(id_priv, cma_dev);
2041         cma_set_loopback(cma_src_addr(id_priv));
2042 out:
2043         mutex_unlock(&lock);
2044         return ret;
2045 }
2046
2047 static void addr_handler(int status, struct sockaddr *src_addr,
2048                          struct rdma_dev_addr *dev_addr, void *context)
2049 {
2050         struct rdma_id_private *id_priv = context;
2051         struct rdma_cm_event event;
2052
2053         memset(&event, 0, sizeof event);
2054         mutex_lock(&id_priv->handler_mutex);
2055         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2056                            RDMA_CM_ADDR_RESOLVED))
2057                 goto out;
2058
2059         if (!status && !id_priv->cma_dev)
2060                 status = cma_acquire_dev(id_priv);
2061
2062         if (status) {
2063                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2064                                    RDMA_CM_ADDR_BOUND))
2065                         goto out;
2066                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2067                 event.status = status;
2068         } else {
2069                 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2070                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2071         }
2072
2073         if (id_priv->id.event_handler(&id_priv->id, &event)) {
2074                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2075                 mutex_unlock(&id_priv->handler_mutex);
2076                 cma_deref_id(id_priv);
2077                 rdma_destroy_id(&id_priv->id);
2078                 return;
2079         }
2080 out:
2081         mutex_unlock(&id_priv->handler_mutex);
2082         cma_deref_id(id_priv);
2083 }
2084
2085 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2086 {
2087         struct cma_work *work;
2088         union ib_gid gid;
2089         int ret;
2090
2091         work = kzalloc(sizeof *work, GFP_KERNEL);
2092         if (!work)
2093                 return -ENOMEM;
2094
2095         if (!id_priv->cma_dev) {
2096                 ret = cma_bind_loopback(id_priv);
2097                 if (ret)
2098                         goto err;
2099         }
2100
2101         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2102         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2103
2104         work->id = id_priv;
2105         INIT_WORK(&work->work, cma_work_handler);
2106         work->old_state = RDMA_CM_ADDR_QUERY;
2107         work->new_state = RDMA_CM_ADDR_RESOLVED;
2108         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2109         queue_work(cma_wq, &work->work);
2110         return 0;
2111 err:
2112         kfree(work);
2113         return ret;
2114 }
2115
2116 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2117 {
2118         struct cma_work *work;
2119         int ret;
2120
2121         work = kzalloc(sizeof *work, GFP_KERNEL);
2122         if (!work)
2123                 return -ENOMEM;
2124
2125         if (!id_priv->cma_dev) {
2126                 ret = cma_resolve_ib_dev(id_priv);
2127                 if (ret)
2128                         goto err;
2129         }
2130
2131         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2132                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2133
2134         work->id = id_priv;
2135         INIT_WORK(&work->work, cma_work_handler);
2136         work->old_state = RDMA_CM_ADDR_QUERY;
2137         work->new_state = RDMA_CM_ADDR_RESOLVED;
2138         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2139         queue_work(cma_wq, &work->work);
2140         return 0;
2141 err:
2142         kfree(work);
2143         return ret;
2144 }
2145
2146 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2147                          struct sockaddr *dst_addr)
2148 {
2149         if (!src_addr || !src_addr->sa_family) {
2150                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2151                 src_addr->sa_family = dst_addr->sa_family;
2152                 if (dst_addr->sa_family == AF_INET6) {
2153                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2154                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2155                 } else if (dst_addr->sa_family == AF_IB) {
2156                         ((struct sockaddr_ib *) src_addr)->sib_pkey =
2157                                 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2158                 }
2159         }
2160         return rdma_bind_addr(id, src_addr);
2161 }
2162
2163 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2164                       struct sockaddr *dst_addr, int timeout_ms)
2165 {
2166         struct rdma_id_private *id_priv;
2167         int ret;
2168
2169         id_priv = container_of(id, struct rdma_id_private, id);
2170         if (id_priv->state == RDMA_CM_IDLE) {
2171                 ret = cma_bind_addr(id, src_addr, dst_addr);
2172                 if (ret)
2173                         return ret;
2174         }
2175
2176         if (cma_family(id_priv) != dst_addr->sa_family)
2177                 return -EINVAL;
2178
2179         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2180                 return -EINVAL;
2181
2182         atomic_inc(&id_priv->refcount);
2183         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2184         if (cma_any_addr(dst_addr)) {
2185                 ret = cma_resolve_loopback(id_priv);
2186         } else {
2187                 if (dst_addr->sa_family == AF_IB) {
2188                         ret = cma_resolve_ib_addr(id_priv);
2189                 } else {
2190                         ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2191                                               dst_addr, &id->route.addr.dev_addr,
2192                                               timeout_ms, addr_handler, id_priv);
2193                 }
2194         }
2195         if (ret)
2196                 goto err;
2197
2198         return 0;
2199 err:
2200         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2201         cma_deref_id(id_priv);
2202         return ret;
2203 }
2204 EXPORT_SYMBOL(rdma_resolve_addr);
2205
2206 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2207 {
2208         struct rdma_id_private *id_priv;
2209         unsigned long flags;
2210         int ret;
2211
2212         id_priv = container_of(id, struct rdma_id_private, id);
2213         spin_lock_irqsave(&id_priv->lock, flags);
2214         if (reuse || id_priv->state == RDMA_CM_IDLE) {
2215                 id_priv->reuseaddr = reuse;
2216                 ret = 0;
2217         } else {
2218                 ret = -EINVAL;
2219         }
2220         spin_unlock_irqrestore(&id_priv->lock, flags);
2221         return ret;
2222 }
2223 EXPORT_SYMBOL(rdma_set_reuseaddr);
2224
2225 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2226 {
2227         struct rdma_id_private *id_priv;
2228         unsigned long flags;
2229         int ret;
2230
2231         id_priv = container_of(id, struct rdma_id_private, id);
2232         spin_lock_irqsave(&id_priv->lock, flags);
2233         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2234                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2235                 id_priv->afonly = afonly;
2236                 ret = 0;
2237         } else {
2238                 ret = -EINVAL;
2239         }
2240         spin_unlock_irqrestore(&id_priv->lock, flags);
2241         return ret;
2242 }
2243 EXPORT_SYMBOL(rdma_set_afonly);
2244
2245 static void cma_bind_port(struct rdma_bind_list *bind_list,
2246                           struct rdma_id_private *id_priv)
2247 {
2248         struct sockaddr *addr;
2249         struct sockaddr_ib *sib;
2250         u64 sid, mask;
2251         __be16 port;
2252
2253         addr = cma_src_addr(id_priv);
2254         port = htons(bind_list->port);
2255
2256         switch (addr->sa_family) {
2257         case AF_INET:
2258                 ((struct sockaddr_in *) addr)->sin_port = port;
2259                 break;
2260         case AF_INET6:
2261                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2262                 break;
2263         case AF_IB:
2264                 sib = (struct sockaddr_ib *) addr;
2265                 sid = be64_to_cpu(sib->sib_sid);
2266                 mask = be64_to_cpu(sib->sib_sid_mask);
2267                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2268                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2269                 break;
2270         }
2271         id_priv->bind_list = bind_list;
2272         hlist_add_head(&id_priv->node, &bind_list->owners);
2273 }
2274
2275 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2276                           unsigned short snum)
2277 {
2278         struct rdma_bind_list *bind_list;
2279         int ret;
2280
2281         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2282         if (!bind_list)
2283                 return -ENOMEM;
2284
2285         ret = idr_alloc(ps, bind_list, snum, snum + 1, GFP_KERNEL);
2286         if (ret < 0)
2287                 goto err;
2288
2289         bind_list->ps = ps;
2290         bind_list->port = (unsigned short)ret;
2291         cma_bind_port(bind_list, id_priv);
2292         return 0;
2293 err:
2294         kfree(bind_list);
2295         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2296 }
2297
2298 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2299 {
2300         static unsigned int last_used_port;
2301         int low, high, remaining;
2302         unsigned int rover;
2303
2304         inet_get_local_port_range(&low, &high);
2305         remaining = (high - low) + 1;
2306         rover = net_random() % remaining + low;
2307 retry:
2308         if (last_used_port != rover &&
2309             !idr_find(ps, (unsigned short) rover)) {
2310                 int ret = cma_alloc_port(ps, id_priv, rover);
2311                 /*
2312                  * Remember previously used port number in order to avoid
2313                  * re-using same port immediately after it is closed.
2314                  */
2315                 if (!ret)
2316                         last_used_port = rover;
2317                 if (ret != -EADDRNOTAVAIL)
2318                         return ret;
2319         }
2320         if (--remaining) {
2321                 rover++;
2322                 if ((rover < low) || (rover > high))
2323                         rover = low;
2324                 goto retry;
2325         }
2326         return -EADDRNOTAVAIL;
2327 }
2328
2329 /*
2330  * Check that the requested port is available.  This is called when trying to
2331  * bind to a specific port, or when trying to listen on a bound port.  In
2332  * the latter case, the provided id_priv may already be on the bind_list, but
2333  * we still need to check that it's okay to start listening.
2334  */
2335 static int cma_check_port(struct rdma_bind_list *bind_list,
2336                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2337 {
2338         struct rdma_id_private *cur_id;
2339         struct sockaddr *addr, *cur_addr;
2340
2341         addr = cma_src_addr(id_priv);
2342         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2343                 if (id_priv == cur_id)
2344                         continue;
2345
2346                 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2347                     cur_id->reuseaddr)
2348                         continue;
2349
2350                 cur_addr = cma_src_addr(cur_id);
2351                 if (id_priv->afonly && cur_id->afonly &&
2352                     (addr->sa_family != cur_addr->sa_family))
2353                         continue;
2354
2355                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2356                         return -EADDRNOTAVAIL;
2357
2358                 if (!cma_addr_cmp(addr, cur_addr))
2359                         return -EADDRINUSE;
2360         }
2361         return 0;
2362 }
2363
2364 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2365 {
2366         struct rdma_bind_list *bind_list;
2367         unsigned short snum;
2368         int ret;
2369
2370         snum = ntohs(cma_port(cma_src_addr(id_priv)));
2371         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2372                 return -EACCES;
2373
2374         bind_list = idr_find(ps, snum);
2375         if (!bind_list) {
2376                 ret = cma_alloc_port(ps, id_priv, snum);
2377         } else {
2378                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2379                 if (!ret)
2380                         cma_bind_port(bind_list, id_priv);
2381         }
2382         return ret;
2383 }
2384
2385 static int cma_bind_listen(struct rdma_id_private *id_priv)
2386 {
2387         struct rdma_bind_list *bind_list = id_priv->bind_list;
2388         int ret = 0;
2389
2390         mutex_lock(&lock);
2391         if (bind_list->owners.first->next)
2392                 ret = cma_check_port(bind_list, id_priv, 0);
2393         mutex_unlock(&lock);
2394         return ret;
2395 }
2396
2397 static struct idr *cma_select_inet_ps(struct rdma_id_private *id_priv)
2398 {
2399         switch (id_priv->id.ps) {
2400         case RDMA_PS_TCP:
2401                 return &tcp_ps;
2402         case RDMA_PS_UDP:
2403                 return &udp_ps;
2404         case RDMA_PS_IPOIB:
2405                 return &ipoib_ps;
2406         case RDMA_PS_IB:
2407                 return &ib_ps;
2408         default:
2409                 return NULL;
2410         }
2411 }
2412
2413 static struct idr *cma_select_ib_ps(struct rdma_id_private *id_priv)
2414 {
2415         struct idr *ps = NULL;
2416         struct sockaddr_ib *sib;
2417         u64 sid_ps, mask, sid;
2418
2419         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2420         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
2421         sid = be64_to_cpu(sib->sib_sid) & mask;
2422
2423         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
2424                 sid_ps = RDMA_IB_IP_PS_IB;
2425                 ps = &ib_ps;
2426         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
2427                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
2428                 sid_ps = RDMA_IB_IP_PS_TCP;
2429                 ps = &tcp_ps;
2430         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
2431                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
2432                 sid_ps = RDMA_IB_IP_PS_UDP;
2433                 ps = &udp_ps;
2434         }
2435
2436         if (ps) {
2437                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
2438                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
2439                                                 be64_to_cpu(sib->sib_sid_mask));
2440         }
2441         return ps;
2442 }
2443
2444 static int cma_get_port(struct rdma_id_private *id_priv)
2445 {
2446         struct idr *ps;
2447         int ret;
2448
2449         if (cma_family(id_priv) != AF_IB)
2450                 ps = cma_select_inet_ps(id_priv);
2451         else
2452                 ps = cma_select_ib_ps(id_priv);
2453         if (!ps)
2454                 return -EPROTONOSUPPORT;
2455
2456         mutex_lock(&lock);
2457         if (cma_any_port(cma_src_addr(id_priv)))
2458                 ret = cma_alloc_any_port(ps, id_priv);
2459         else
2460                 ret = cma_use_port(ps, id_priv);
2461         mutex_unlock(&lock);
2462
2463         return ret;
2464 }
2465
2466 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2467                                struct sockaddr *addr)
2468 {
2469 #if IS_ENABLED(CONFIG_IPV6)
2470         struct sockaddr_in6 *sin6;
2471
2472         if (addr->sa_family != AF_INET6)
2473                 return 0;
2474
2475         sin6 = (struct sockaddr_in6 *) addr;
2476         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2477             !sin6->sin6_scope_id)
2478                         return -EINVAL;
2479
2480         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2481 #endif
2482         return 0;
2483 }
2484
2485 int rdma_listen(struct rdma_cm_id *id, int backlog)
2486 {
2487         struct rdma_id_private *id_priv;
2488         int ret;
2489
2490         id_priv = container_of(id, struct rdma_id_private, id);
2491         if (id_priv->state == RDMA_CM_IDLE) {
2492                 id->route.addr.src_addr.ss_family = AF_INET;
2493                 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
2494                 if (ret)
2495                         return ret;
2496         }
2497
2498         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2499                 return -EINVAL;
2500
2501         if (id_priv->reuseaddr) {
2502                 ret = cma_bind_listen(id_priv);
2503                 if (ret)
2504                         goto err;
2505         }
2506
2507         id_priv->backlog = backlog;
2508         if (id->device) {
2509                 switch (rdma_node_get_transport(id->device->node_type)) {
2510                 case RDMA_TRANSPORT_IB:
2511                         ret = cma_ib_listen(id_priv);
2512                         if (ret)
2513                                 goto err;
2514                         break;
2515                 case RDMA_TRANSPORT_IWARP:
2516                         ret = cma_iw_listen(id_priv, backlog);
2517                         if (ret)
2518                                 goto err;
2519                         break;
2520                 default:
2521                         ret = -ENOSYS;
2522                         goto err;
2523                 }
2524         } else
2525                 cma_listen_on_all(id_priv);
2526
2527         return 0;
2528 err:
2529         id_priv->backlog = 0;
2530         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2531         return ret;
2532 }
2533 EXPORT_SYMBOL(rdma_listen);
2534
2535 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2536 {
2537         struct rdma_id_private *id_priv;
2538         int ret;
2539
2540         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
2541             addr->sa_family != AF_IB)
2542                 return -EAFNOSUPPORT;
2543
2544         id_priv = container_of(id, struct rdma_id_private, id);
2545         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2546                 return -EINVAL;
2547
2548         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2549         if (ret)
2550                 goto err1;
2551
2552         if (!cma_any_addr(addr)) {
2553                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
2554                 if (ret)
2555                         goto err1;
2556
2557                 ret = cma_acquire_dev(id_priv);
2558                 if (ret)
2559                         goto err1;
2560         }
2561
2562         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
2563         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
2564                 if (addr->sa_family == AF_INET)
2565                         id_priv->afonly = 1;
2566 #if IS_ENABLED(CONFIG_IPV6)
2567                 else if (addr->sa_family == AF_INET6)
2568                         id_priv->afonly = init_net.ipv6.sysctl.bindv6only;
2569 #endif
2570         }
2571         ret = cma_get_port(id_priv);
2572         if (ret)
2573                 goto err2;
2574
2575         return 0;
2576 err2:
2577         if (id_priv->cma_dev)
2578                 cma_release_dev(id_priv);
2579 err1:
2580         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2581         return ret;
2582 }
2583 EXPORT_SYMBOL(rdma_bind_addr);
2584
2585 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
2586 {
2587         struct cma_hdr *cma_hdr;
2588
2589         cma_hdr = hdr;
2590         cma_hdr->cma_version = CMA_VERSION;
2591         if (cma_family(id_priv) == AF_INET) {
2592                 struct sockaddr_in *src4, *dst4;
2593
2594                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
2595                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
2596
2597                 cma_set_ip_ver(cma_hdr, 4);
2598                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2599                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2600                 cma_hdr->port = src4->sin_port;
2601         } else if (cma_family(id_priv) == AF_INET6) {
2602                 struct sockaddr_in6 *src6, *dst6;
2603
2604                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2605                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
2606
2607                 cma_set_ip_ver(cma_hdr, 6);
2608                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
2609                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2610                 cma_hdr->port = src6->sin6_port;
2611         }
2612         return 0;
2613 }
2614
2615 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2616                                 struct ib_cm_event *ib_event)
2617 {
2618         struct rdma_id_private *id_priv = cm_id->context;
2619         struct rdma_cm_event event;
2620         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2621         int ret = 0;
2622
2623         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2624                 return 0;
2625
2626         memset(&event, 0, sizeof event);
2627         switch (ib_event->event) {
2628         case IB_CM_SIDR_REQ_ERROR:
2629                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2630                 event.status = -ETIMEDOUT;
2631                 break;
2632         case IB_CM_SIDR_REP_RECEIVED:
2633                 event.param.ud.private_data = ib_event->private_data;
2634                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2635                 if (rep->status != IB_SIDR_SUCCESS) {
2636                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2637                         event.status = ib_event->param.sidr_rep_rcvd.status;
2638                         break;
2639                 }
2640                 ret = cma_set_qkey(id_priv, rep->qkey);
2641                 if (ret) {
2642                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2643                         event.status = ret;
2644                         break;
2645                 }
2646                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2647                                      id_priv->id.route.path_rec,
2648                                      &event.param.ud.ah_attr);
2649                 event.param.ud.qp_num = rep->qpn;
2650                 event.param.ud.qkey = rep->qkey;
2651                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2652                 event.status = 0;
2653                 break;
2654         default:
2655                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2656                        ib_event->event);
2657                 goto out;
2658         }
2659
2660         ret = id_priv->id.event_handler(&id_priv->id, &event);
2661         if (ret) {
2662                 /* Destroy the CM ID by returning a non-zero value. */
2663                 id_priv->cm_id.ib = NULL;
2664                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2665                 mutex_unlock(&id_priv->handler_mutex);
2666                 rdma_destroy_id(&id_priv->id);
2667                 return ret;
2668         }
2669 out:
2670         mutex_unlock(&id_priv->handler_mutex);
2671         return ret;
2672 }
2673
2674 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2675                               struct rdma_conn_param *conn_param)
2676 {
2677         struct ib_cm_sidr_req_param req;
2678         struct ib_cm_id *id;
2679         void *private_data;
2680         int offset, ret;
2681
2682         memset(&req, 0, sizeof req);
2683         offset = cma_user_data_offset(id_priv);
2684         req.private_data_len = offset + conn_param->private_data_len;
2685         if (req.private_data_len < conn_param->private_data_len)
2686                 return -EINVAL;
2687
2688         if (req.private_data_len) {
2689                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2690                 if (!private_data)
2691                         return -ENOMEM;
2692         } else {
2693                 private_data = NULL;
2694         }
2695
2696         if (conn_param->private_data && conn_param->private_data_len)
2697                 memcpy(private_data + offset, conn_param->private_data,
2698                        conn_param->private_data_len);
2699
2700         if (private_data) {
2701                 ret = cma_format_hdr(private_data, id_priv);
2702                 if (ret)
2703                         goto out;
2704                 req.private_data = private_data;
2705         }
2706
2707         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2708                              id_priv);
2709         if (IS_ERR(id)) {
2710                 ret = PTR_ERR(id);
2711                 goto out;
2712         }
2713         id_priv->cm_id.ib = id;
2714
2715         req.path = id_priv->id.route.path_rec;
2716         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2717         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2718         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2719
2720         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2721         if (ret) {
2722                 ib_destroy_cm_id(id_priv->cm_id.ib);
2723                 id_priv->cm_id.ib = NULL;
2724         }
2725 out:
2726         kfree(private_data);
2727         return ret;
2728 }
2729
2730 static int cma_connect_ib(struct rdma_id_private *id_priv,
2731                           struct rdma_conn_param *conn_param)
2732 {
2733         struct ib_cm_req_param req;
2734         struct rdma_route *route;
2735         void *private_data;
2736         struct ib_cm_id *id;
2737         int offset, ret;
2738
2739         memset(&req, 0, sizeof req);
2740         offset = cma_user_data_offset(id_priv);
2741         req.private_data_len = offset + conn_param->private_data_len;
2742         if (req.private_data_len < conn_param->private_data_len)
2743                 return -EINVAL;
2744
2745         if (req.private_data_len) {
2746                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2747                 if (!private_data)
2748                         return -ENOMEM;
2749         } else {
2750                 private_data = NULL;
2751         }
2752
2753         if (conn_param->private_data && conn_param->private_data_len)
2754                 memcpy(private_data + offset, conn_param->private_data,
2755                        conn_param->private_data_len);
2756
2757         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2758         if (IS_ERR(id)) {
2759                 ret = PTR_ERR(id);
2760                 goto out;
2761         }
2762         id_priv->cm_id.ib = id;
2763
2764         route = &id_priv->id.route;
2765         if (private_data) {
2766                 ret = cma_format_hdr(private_data, id_priv);
2767                 if (ret)
2768                         goto out;
2769                 req.private_data = private_data;
2770         }
2771
2772         req.primary_path = &route->path_rec[0];
2773         if (route->num_paths == 2)
2774                 req.alternate_path = &route->path_rec[1];
2775
2776         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2777         req.qp_num = id_priv->qp_num;
2778         req.qp_type = id_priv->id.qp_type;
2779         req.starting_psn = id_priv->seq_num;
2780         req.responder_resources = conn_param->responder_resources;
2781         req.initiator_depth = conn_param->initiator_depth;
2782         req.flow_control = conn_param->flow_control;
2783         req.retry_count = min_t(u8, 7, conn_param->retry_count);
2784         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2785         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2786         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2787         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2788         req.srq = id_priv->srq ? 1 : 0;
2789
2790         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2791 out:
2792         if (ret && !IS_ERR(id)) {
2793                 ib_destroy_cm_id(id);
2794                 id_priv->cm_id.ib = NULL;
2795         }
2796
2797         kfree(private_data);
2798         return ret;
2799 }
2800
2801 static int cma_connect_iw(struct rdma_id_private *id_priv,
2802                           struct rdma_conn_param *conn_param)
2803 {
2804         struct iw_cm_id *cm_id;
2805         struct sockaddr_in* sin;
2806         int ret;
2807         struct iw_cm_conn_param iw_param;
2808
2809         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2810         if (IS_ERR(cm_id))
2811                 return PTR_ERR(cm_id);
2812
2813         id_priv->cm_id.iw = cm_id;
2814
2815         sin = (struct sockaddr_in *) cma_src_addr(id_priv);
2816         cm_id->local_addr = *sin;
2817
2818         sin = (struct sockaddr_in *) cma_dst_addr(id_priv);
2819         cm_id->remote_addr = *sin;
2820
2821         ret = cma_modify_qp_rtr(id_priv, conn_param);
2822         if (ret)
2823                 goto out;
2824
2825         if (conn_param) {
2826                 iw_param.ord = conn_param->initiator_depth;
2827                 iw_param.ird = conn_param->responder_resources;
2828                 iw_param.private_data = conn_param->private_data;
2829                 iw_param.private_data_len = conn_param->private_data_len;
2830                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2831         } else {
2832                 memset(&iw_param, 0, sizeof iw_param);
2833                 iw_param.qpn = id_priv->qp_num;
2834         }
2835         ret = iw_cm_connect(cm_id, &iw_param);
2836 out:
2837         if (ret) {
2838                 iw_destroy_cm_id(cm_id);
2839                 id_priv->cm_id.iw = NULL;
2840         }
2841         return ret;
2842 }
2843
2844 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2845 {
2846         struct rdma_id_private *id_priv;
2847         int ret;
2848
2849         id_priv = container_of(id, struct rdma_id_private, id);
2850         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2851                 return -EINVAL;
2852
2853         if (!id->qp) {
2854                 id_priv->qp_num = conn_param->qp_num;
2855                 id_priv->srq = conn_param->srq;
2856         }
2857
2858         switch (rdma_node_get_transport(id->device->node_type)) {
2859         case RDMA_TRANSPORT_IB:
2860                 if (id->qp_type == IB_QPT_UD)
2861                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2862                 else
2863                         ret = cma_connect_ib(id_priv, conn_param);
2864                 break;
2865         case RDMA_TRANSPORT_IWARP:
2866                 ret = cma_connect_iw(id_priv, conn_param);
2867                 break;
2868         default:
2869                 ret = -ENOSYS;
2870                 break;
2871         }
2872         if (ret)
2873                 goto err;
2874
2875         return 0;
2876 err:
2877         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2878         return ret;
2879 }
2880 EXPORT_SYMBOL(rdma_connect);
2881
2882 static int cma_accept_ib(struct rdma_id_private *id_priv,
2883                          struct rdma_conn_param *conn_param)
2884 {
2885         struct ib_cm_rep_param rep;
2886         int ret;
2887
2888         ret = cma_modify_qp_rtr(id_priv, conn_param);
2889         if (ret)
2890                 goto out;
2891
2892         ret = cma_modify_qp_rts(id_priv, conn_param);
2893         if (ret)
2894                 goto out;
2895
2896         memset(&rep, 0, sizeof rep);
2897         rep.qp_num = id_priv->qp_num;
2898         rep.starting_psn = id_priv->seq_num;
2899         rep.private_data = conn_param->private_data;
2900         rep.private_data_len = conn_param->private_data_len;
2901         rep.responder_resources = conn_param->responder_resources;
2902         rep.initiator_depth = conn_param->initiator_depth;
2903         rep.failover_accepted = 0;
2904         rep.flow_control = conn_param->flow_control;
2905         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2906         rep.srq = id_priv->srq ? 1 : 0;
2907
2908         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2909 out:
2910         return ret;
2911 }
2912
2913 static int cma_accept_iw(struct rdma_id_private *id_priv,
2914                   struct rdma_conn_param *conn_param)
2915 {
2916         struct iw_cm_conn_param iw_param;
2917         int ret;
2918
2919         ret = cma_modify_qp_rtr(id_priv, conn_param);
2920         if (ret)
2921                 return ret;
2922
2923         iw_param.ord = conn_param->initiator_depth;
2924         iw_param.ird = conn_param->responder_resources;
2925         iw_param.private_data = conn_param->private_data;
2926         iw_param.private_data_len = conn_param->private_data_len;
2927         if (id_priv->id.qp) {
2928                 iw_param.qpn = id_priv->qp_num;
2929         } else
2930                 iw_param.qpn = conn_param->qp_num;
2931
2932         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2933 }
2934
2935 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2936                              enum ib_cm_sidr_status status, u32 qkey,
2937                              const void *private_data, int private_data_len)
2938 {
2939         struct ib_cm_sidr_rep_param rep;
2940         int ret;
2941
2942         memset(&rep, 0, sizeof rep);
2943         rep.status = status;
2944         if (status == IB_SIDR_SUCCESS) {
2945                 ret = cma_set_qkey(id_priv, qkey);
2946                 if (ret)
2947                         return ret;
2948                 rep.qp_num = id_priv->qp_num;
2949                 rep.qkey = id_priv->qkey;
2950         }
2951         rep.private_data = private_data;
2952         rep.private_data_len = private_data_len;
2953
2954         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2955 }
2956
2957 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2958 {
2959         struct rdma_id_private *id_priv;
2960         int ret;
2961
2962         id_priv = container_of(id, struct rdma_id_private, id);
2963
2964         id_priv->owner = task_pid_nr(current);
2965
2966         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2967                 return -EINVAL;
2968
2969         if (!id->qp && conn_param) {
2970                 id_priv->qp_num = conn_param->qp_num;
2971                 id_priv->srq = conn_param->srq;
2972         }
2973
2974         switch (rdma_node_get_transport(id->device->node_type)) {
2975         case RDMA_TRANSPORT_IB:
2976                 if (id->qp_type == IB_QPT_UD) {
2977                         if (conn_param)
2978                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2979                                                         conn_param->qkey,
2980                                                         conn_param->private_data,
2981                                                         conn_param->private_data_len);
2982                         else
2983                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2984                                                         0, NULL, 0);
2985                 } else {
2986                         if (conn_param)
2987                                 ret = cma_accept_ib(id_priv, conn_param);
2988                         else
2989                                 ret = cma_rep_recv(id_priv);
2990                 }
2991                 break;
2992         case RDMA_TRANSPORT_IWARP:
2993                 ret = cma_accept_iw(id_priv, conn_param);
2994                 break;
2995         default:
2996                 ret = -ENOSYS;
2997                 break;
2998         }
2999
3000         if (ret)
3001                 goto reject;
3002
3003         return 0;
3004 reject:
3005         cma_modify_qp_err(id_priv);
3006         rdma_reject(id, NULL, 0);
3007         return ret;
3008 }
3009 EXPORT_SYMBOL(rdma_accept);
3010
3011 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3012 {
3013         struct rdma_id_private *id_priv;
3014         int ret;
3015
3016         id_priv = container_of(id, struct rdma_id_private, id);
3017         if (!id_priv->cm_id.ib)
3018                 return -EINVAL;
3019
3020         switch (id->device->node_type) {
3021         case RDMA_NODE_IB_CA:
3022                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3023                 break;
3024         default:
3025                 ret = 0;
3026                 break;
3027         }
3028         return ret;
3029 }
3030 EXPORT_SYMBOL(rdma_notify);
3031
3032 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3033                 u8 private_data_len)
3034 {
3035         struct rdma_id_private *id_priv;
3036         int ret;
3037
3038         id_priv = container_of(id, struct rdma_id_private, id);
3039         if (!id_priv->cm_id.ib)
3040                 return -EINVAL;
3041
3042         switch (rdma_node_get_transport(id->device->node_type)) {
3043         case RDMA_TRANSPORT_IB:
3044                 if (id->qp_type == IB_QPT_UD)
3045                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3046                                                 private_data, private_data_len);
3047                 else
3048                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
3049                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
3050                                              0, private_data, private_data_len);
3051                 break;
3052         case RDMA_TRANSPORT_IWARP:
3053                 ret = iw_cm_reject(id_priv->cm_id.iw,
3054                                    private_data, private_data_len);
3055                 break;
3056         default:
3057                 ret = -ENOSYS;
3058                 break;
3059         }
3060         return ret;
3061 }
3062 EXPORT_SYMBOL(rdma_reject);
3063
3064 int rdma_disconnect(struct rdma_cm_id *id)
3065 {
3066         struct rdma_id_private *id_priv;
3067         int ret;
3068
3069         id_priv = container_of(id, struct rdma_id_private, id);
3070         if (!id_priv->cm_id.ib)
3071                 return -EINVAL;
3072
3073         switch (rdma_node_get_transport(id->device->node_type)) {
3074         case RDMA_TRANSPORT_IB:
3075                 ret = cma_modify_qp_err(id_priv);
3076                 if (ret)
3077                         goto out;
3078                 /* Initiate or respond to a disconnect. */
3079                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3080                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3081                 break;
3082         case RDMA_TRANSPORT_IWARP:
3083                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3084                 break;
3085         default:
3086                 ret = -EINVAL;
3087                 break;
3088         }
3089 out:
3090         return ret;
3091 }
3092 EXPORT_SYMBOL(rdma_disconnect);
3093
3094 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3095 {
3096         struct rdma_id_private *id_priv;
3097         struct cma_multicast *mc = multicast->context;
3098         struct rdma_cm_event event;
3099         int ret;
3100
3101         id_priv = mc->id_priv;
3102         if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
3103             cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
3104                 return 0;
3105
3106         if (!status)
3107                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3108         mutex_lock(&id_priv->qp_mutex);
3109         if (!status && id_priv->id.qp)
3110                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3111                                          be16_to_cpu(multicast->rec.mlid));
3112         mutex_unlock(&id_priv->qp_mutex);
3113
3114         memset(&event, 0, sizeof event);
3115         event.status = status;
3116         event.param.ud.private_data = mc->context;
3117         if (!status) {
3118                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3119                 ib_init_ah_from_mcmember(id_priv->id.device,
3120                                          id_priv->id.port_num, &multicast->rec,
3121                                          &event.param.ud.ah_attr);
3122                 event.param.ud.qp_num = 0xFFFFFF;
3123                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3124         } else
3125                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3126
3127         ret = id_priv->id.event_handler(&id_priv->id, &event);
3128         if (ret) {
3129                 cma_exch(id_priv, RDMA_CM_DESTROYING);
3130                 mutex_unlock(&id_priv->handler_mutex);
3131                 rdma_destroy_id(&id_priv->id);
3132                 return 0;
3133         }
3134
3135         mutex_unlock(&id_priv->handler_mutex);
3136         return 0;
3137 }
3138
3139 static void cma_set_mgid(struct rdma_id_private *id_priv,
3140                          struct sockaddr *addr, union ib_gid *mgid)
3141 {
3142         unsigned char mc_map[MAX_ADDR_LEN];
3143         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3144         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3145         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3146
3147         if (cma_any_addr(addr)) {
3148                 memset(mgid, 0, sizeof *mgid);
3149         } else if ((addr->sa_family == AF_INET6) &&
3150                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3151                                                                  0xFF10A01B)) {
3152                 /* IPv6 address is an SA assigned MGID. */
3153                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3154         } else if (addr->sa_family == AF_IB) {
3155                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3156         } else if ((addr->sa_family == AF_INET6)) {
3157                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3158                 if (id_priv->id.ps == RDMA_PS_UDP)
3159                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3160                 *mgid = *(union ib_gid *) (mc_map + 4);
3161         } else {
3162                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3163                 if (id_priv->id.ps == RDMA_PS_UDP)
3164                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3165                 *mgid = *(union ib_gid *) (mc_map + 4);
3166         }
3167 }
3168
3169 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3170                                  struct cma_multicast *mc)
3171 {
3172         struct ib_sa_mcmember_rec rec;
3173         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3174         ib_sa_comp_mask comp_mask;
3175         int ret;
3176
3177         ib_addr_get_mgid(dev_addr, &rec.mgid);
3178         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3179                                      &rec.mgid, &rec);
3180         if (ret)
3181                 return ret;
3182
3183         ret = cma_set_qkey(id_priv, 0);
3184         if (ret)
3185                 return ret;
3186
3187         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3188         rec.qkey = cpu_to_be32(id_priv->qkey);
3189         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3190         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3191         rec.join_state = 1;
3192
3193         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3194                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3195                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3196                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
3197                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3198
3199         if (id_priv->id.ps == RDMA_PS_IPOIB)
3200                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3201                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3202                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3203                              IB_SA_MCMEMBER_REC_MTU |
3204                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
3205
3206         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3207                                                 id_priv->id.port_num, &rec,
3208                                                 comp_mask, GFP_KERNEL,
3209                                                 cma_ib_mc_handler, mc);
3210         return PTR_RET(mc->multicast.ib);
3211 }
3212
3213 static void iboe_mcast_work_handler(struct work_struct *work)
3214 {
3215         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3216         struct cma_multicast *mc = mw->mc;
3217         struct ib_sa_multicast *m = mc->multicast.ib;
3218
3219         mc->multicast.ib->context = mc;
3220         cma_ib_mc_handler(0, m);
3221         kref_put(&mc->mcref, release_mc);
3222         kfree(mw);
3223 }
3224
3225 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3226 {
3227         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3228         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3229
3230         if (cma_any_addr(addr)) {
3231                 memset(mgid, 0, sizeof *mgid);
3232         } else if (addr->sa_family == AF_INET6) {
3233                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3234         } else {
3235                 mgid->raw[0] = 0xff;
3236                 mgid->raw[1] = 0x0e;
3237                 mgid->raw[2] = 0;
3238                 mgid->raw[3] = 0;
3239                 mgid->raw[4] = 0;
3240                 mgid->raw[5] = 0;
3241                 mgid->raw[6] = 0;
3242                 mgid->raw[7] = 0;
3243                 mgid->raw[8] = 0;
3244                 mgid->raw[9] = 0;
3245                 mgid->raw[10] = 0xff;
3246                 mgid->raw[11] = 0xff;
3247                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3248         }
3249 }
3250
3251 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3252                                    struct cma_multicast *mc)
3253 {
3254         struct iboe_mcast_work *work;
3255         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3256         int err;
3257         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3258         struct net_device *ndev = NULL;
3259
3260         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3261                 return -EINVAL;
3262
3263         work = kzalloc(sizeof *work, GFP_KERNEL);
3264         if (!work)
3265                 return -ENOMEM;
3266
3267         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3268         if (!mc->multicast.ib) {
3269                 err = -ENOMEM;
3270                 goto out1;
3271         }
3272
3273         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3274
3275         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3276         if (id_priv->id.ps == RDMA_PS_UDP)
3277                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3278
3279         if (dev_addr->bound_dev_if)
3280                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3281         if (!ndev) {
3282                 err = -ENODEV;
3283                 goto out2;
3284         }
3285         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3286         mc->multicast.ib->rec.hop_limit = 1;
3287         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3288         dev_put(ndev);
3289         if (!mc->multicast.ib->rec.mtu) {
3290                 err = -EINVAL;
3291                 goto out2;
3292         }
3293         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3294         work->id = id_priv;
3295         work->mc = mc;
3296         INIT_WORK(&work->work, iboe_mcast_work_handler);
3297         kref_get(&mc->mcref);
3298         queue_work(cma_wq, &work->work);
3299
3300         return 0;
3301
3302 out2:
3303         kfree(mc->multicast.ib);
3304 out1:
3305         kfree(work);
3306         return err;
3307 }
3308
3309 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3310                         void *context)
3311 {
3312         struct rdma_id_private *id_priv;
3313         struct cma_multicast *mc;
3314         int ret;
3315
3316         id_priv = container_of(id, struct rdma_id_private, id);
3317         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3318             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3319                 return -EINVAL;
3320
3321         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3322         if (!mc)
3323                 return -ENOMEM;
3324
3325         memcpy(&mc->addr, addr, rdma_addr_size(addr));
3326         mc->context = context;
3327         mc->id_priv = id_priv;
3328
3329         spin_lock(&id_priv->lock);
3330         list_add(&mc->list, &id_priv->mc_list);
3331         spin_unlock(&id_priv->lock);
3332
3333         switch (rdma_node_get_transport(id->device->node_type)) {
3334         case RDMA_TRANSPORT_IB:
3335                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3336                 case IB_LINK_LAYER_INFINIBAND:
3337                         ret = cma_join_ib_multicast(id_priv, mc);
3338                         break;
3339                 case IB_LINK_LAYER_ETHERNET:
3340                         kref_init(&mc->mcref);
3341                         ret = cma_iboe_join_multicast(id_priv, mc);
3342                         break;
3343                 default:
3344                         ret = -EINVAL;
3345                 }
3346                 break;
3347         default:
3348                 ret = -ENOSYS;
3349                 break;
3350         }
3351
3352         if (ret) {
3353                 spin_lock_irq(&id_priv->lock);
3354                 list_del(&mc->list);
3355                 spin_unlock_irq(&id_priv->lock);
3356                 kfree(mc);
3357         }
3358         return ret;
3359 }
3360 EXPORT_SYMBOL(rdma_join_multicast);
3361
3362 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3363 {
3364         struct rdma_id_private *id_priv;
3365         struct cma_multicast *mc;
3366
3367         id_priv = container_of(id, struct rdma_id_private, id);
3368         spin_lock_irq(&id_priv->lock);
3369         list_for_each_entry(mc, &id_priv->mc_list, list) {
3370                 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
3371                         list_del(&mc->list);
3372                         spin_unlock_irq(&id_priv->lock);
3373
3374                         if (id->qp)
3375                                 ib_detach_mcast(id->qp,
3376                                                 &mc->multicast.ib->rec.mgid,
3377                                                 be16_to_cpu(mc->multicast.ib->rec.mlid));
3378                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3379                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3380                                 case IB_LINK_LAYER_INFINIBAND:
3381                                         ib_sa_free_multicast(mc->multicast.ib);
3382                                         kfree(mc);
3383                                         break;
3384                                 case IB_LINK_LAYER_ETHERNET:
3385                                         kref_put(&mc->mcref, release_mc);
3386                                         break;
3387                                 default:
3388                                         break;
3389                                 }
3390                         }
3391                         return;
3392                 }
3393         }
3394         spin_unlock_irq(&id_priv->lock);
3395 }
3396 EXPORT_SYMBOL(rdma_leave_multicast);
3397
3398 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3399 {
3400         struct rdma_dev_addr *dev_addr;
3401         struct cma_ndev_work *work;
3402
3403         dev_addr = &id_priv->id.route.addr.dev_addr;
3404
3405         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3406             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3407                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3408                        ndev->name, &id_priv->id);
3409                 work = kzalloc(sizeof *work, GFP_KERNEL);
3410                 if (!work)
3411                         return -ENOMEM;
3412
3413                 INIT_WORK(&work->work, cma_ndev_work_handler);
3414                 work->id = id_priv;
3415                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3416                 atomic_inc(&id_priv->refcount);
3417                 queue_work(cma_wq, &work->work);
3418         }
3419
3420         return 0;
3421 }
3422
3423 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3424                                void *ptr)
3425 {
3426         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
3427         struct cma_device *cma_dev;
3428         struct rdma_id_private *id_priv;
3429         int ret = NOTIFY_DONE;
3430
3431         if (dev_net(ndev) != &init_net)
3432                 return NOTIFY_DONE;
3433
3434         if (event != NETDEV_BONDING_FAILOVER)
3435                 return NOTIFY_DONE;
3436
3437         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3438                 return NOTIFY_DONE;
3439
3440         mutex_lock(&lock);
3441         list_for_each_entry(cma_dev, &dev_list, list)
3442                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3443                         ret = cma_netdev_change(ndev, id_priv);
3444                         if (ret)
3445                                 goto out;
3446                 }
3447
3448 out:
3449         mutex_unlock(&lock);
3450         return ret;
3451 }
3452
3453 static struct notifier_block cma_nb = {
3454         .notifier_call = cma_netdev_callback
3455 };
3456
3457 static void cma_add_one(struct ib_device *device)
3458 {
3459         struct cma_device *cma_dev;
3460         struct rdma_id_private *id_priv;
3461
3462         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3463         if (!cma_dev)
3464                 return;
3465
3466         cma_dev->device = device;
3467
3468         init_completion(&cma_dev->comp);
3469         atomic_set(&cma_dev->refcount, 1);
3470         INIT_LIST_HEAD(&cma_dev->id_list);
3471         ib_set_client_data(device, &cma_client, cma_dev);
3472
3473         mutex_lock(&lock);
3474         list_add_tail(&cma_dev->list, &dev_list);
3475         list_for_each_entry(id_priv, &listen_any_list, list)
3476                 cma_listen_on_dev(id_priv, cma_dev);
3477         mutex_unlock(&lock);
3478 }
3479
3480 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3481 {
3482         struct rdma_cm_event event;
3483         enum rdma_cm_state state;
3484         int ret = 0;
3485
3486         /* Record that we want to remove the device */
3487         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3488         if (state == RDMA_CM_DESTROYING)
3489                 return 0;
3490
3491         cma_cancel_operation(id_priv, state);
3492         mutex_lock(&id_priv->handler_mutex);
3493
3494         /* Check for destruction from another callback. */
3495         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3496                 goto out;
3497
3498         memset(&event, 0, sizeof event);
3499         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3500         ret = id_priv->id.event_handler(&id_priv->id, &event);
3501 out:
3502         mutex_unlock(&id_priv->handler_mutex);
3503         return ret;
3504 }
3505
3506 static void cma_process_remove(struct cma_device *cma_dev)
3507 {
3508         struct rdma_id_private *id_priv;
3509         int ret;
3510
3511         mutex_lock(&lock);
3512         while (!list_empty(&cma_dev->id_list)) {
3513                 id_priv = list_entry(cma_dev->id_list.next,
3514                                      struct rdma_id_private, list);
3515
3516                 list_del(&id_priv->listen_list);
3517                 list_del_init(&id_priv->list);
3518                 atomic_inc(&id_priv->refcount);
3519                 mutex_unlock(&lock);
3520
3521                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3522                 cma_deref_id(id_priv);
3523                 if (ret)
3524                         rdma_destroy_id(&id_priv->id);
3525
3526                 mutex_lock(&lock);
3527         }
3528         mutex_unlock(&lock);
3529
3530         cma_deref_dev(cma_dev);
3531         wait_for_completion(&cma_dev->comp);
3532 }
3533
3534 static void cma_remove_one(struct ib_device *device)
3535 {
3536         struct cma_device *cma_dev;
3537
3538         cma_dev = ib_get_client_data(device, &cma_client);
3539         if (!cma_dev)
3540                 return;
3541
3542         mutex_lock(&lock);
3543         list_del(&cma_dev->list);
3544         mutex_unlock(&lock);
3545
3546         cma_process_remove(cma_dev);
3547         kfree(cma_dev);
3548 }
3549
3550 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3551 {
3552         struct nlmsghdr *nlh;
3553         struct rdma_cm_id_stats *id_stats;
3554         struct rdma_id_private *id_priv;
3555         struct rdma_cm_id *id = NULL;
3556         struct cma_device *cma_dev;
3557         int i_dev = 0, i_id = 0;
3558
3559         /*
3560          * We export all of the IDs as a sequence of messages.  Each
3561          * ID gets its own netlink message.
3562          */
3563         mutex_lock(&lock);
3564
3565         list_for_each_entry(cma_dev, &dev_list, list) {
3566                 if (i_dev < cb->args[0]) {
3567                         i_dev++;
3568                         continue;
3569                 }
3570
3571                 i_id = 0;
3572                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3573                         if (i_id < cb->args[1]) {
3574                                 i_id++;
3575                                 continue;
3576                         }
3577
3578                         id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3579                                                 sizeof *id_stats, RDMA_NL_RDMA_CM,
3580                                                 RDMA_NL_RDMA_CM_ID_STATS);
3581                         if (!id_stats)
3582                                 goto out;
3583
3584                         memset(id_stats, 0, sizeof *id_stats);
3585                         id = &id_priv->id;
3586                         id_stats->node_type = id->route.addr.dev_addr.dev_type;
3587                         id_stats->port_num = id->port_num;
3588                         id_stats->bound_dev_if =
3589                                 id->route.addr.dev_addr.bound_dev_if;
3590
3591                         if (ibnl_put_attr(skb, nlh,
3592                                           rdma_addr_size(cma_src_addr(id_priv)),
3593                                           cma_src_addr(id_priv),
3594                                           RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
3595                                 goto out;
3596                         if (ibnl_put_attr(skb, nlh,
3597                                           rdma_addr_size(cma_src_addr(id_priv)),
3598                                           cma_dst_addr(id_priv),
3599                                           RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
3600                                 goto out;
3601
3602                         id_stats->pid           = id_priv->owner;
3603                         id_stats->port_space    = id->ps;
3604                         id_stats->cm_state      = id_priv->state;
3605                         id_stats->qp_num        = id_priv->qp_num;
3606                         id_stats->qp_type       = id->qp_type;
3607
3608                         i_id++;
3609                 }
3610
3611                 cb->args[1] = 0;
3612                 i_dev++;
3613         }
3614
3615 out:
3616         mutex_unlock(&lock);
3617         cb->args[0] = i_dev;
3618         cb->args[1] = i_id;
3619
3620         return skb->len;
3621 }
3622
3623 static const struct ibnl_client_cbs cma_cb_table[] = {
3624         [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
3625                                        .module = THIS_MODULE },
3626 };
3627
3628 static int __init cma_init(void)
3629 {
3630         int ret;
3631
3632         cma_wq = create_singlethread_workqueue("rdma_cm");
3633         if (!cma_wq)
3634                 return -ENOMEM;
3635
3636         ib_sa_register_client(&sa_client);
3637         rdma_addr_register_client(&addr_client);
3638         register_netdevice_notifier(&cma_nb);
3639
3640         ret = ib_register_client(&cma_client);
3641         if (ret)
3642                 goto err;
3643
3644         if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3645                 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3646
3647         return 0;
3648
3649 err:
3650         unregister_netdevice_notifier(&cma_nb);
3651         rdma_addr_unregister_client(&addr_client);
3652         ib_sa_unregister_client(&sa_client);
3653         destroy_workqueue(cma_wq);
3654         return ret;
3655 }
3656
3657 static void __exit cma_cleanup(void)
3658 {
3659         ibnl_remove_client(RDMA_NL_RDMA_CM);
3660         ib_unregister_client(&cma_client);
3661         unregister_netdevice_notifier(&cma_nb);
3662         rdma_addr_unregister_client(&addr_client);
3663         ib_sa_unregister_client(&sa_client);
3664         destroy_workqueue(cma_wq);
3665         idr_destroy(&tcp_ps);
3666         idr_destroy(&udp_ps);
3667         idr_destroy(&ipoib_ps);
3668         idr_destroy(&ib_ps);
3669 }
3670
3671 module_init(cma_init);
3672 module_exit(cma_cleanup);