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             (ib_event->event == IB_CM_REQ_RECEIVED)) {
 884                 cma_save_ib_info(id, listen_id, ib_event->param.req_rcvd.primary_path);
 885                 return 0;
 886         }
 887
 888         hdr = ib_event->private_data;
 889         if (hdr->cma_version != CMA_VERSION)
 890                 return -EINVAL;
 891
 892         switch (cma_get_ip_ver(hdr)) {
 893         case 4:
 894                 cma_save_ip4_info(id, listen_id, hdr);
 895                 break;
 896         case 6:
 897                 cma_save_ip6_info(id, listen_id, hdr);
 898                 break;
 899         default:
 900                 return -EINVAL;
 901         }
 902         return 0;
 903 }
 904
 905 static inline int cma_user_data_offset(struct rdma_id_private *id_priv)
 906 {
 907         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
 908 }
 909
 910 static void cma_cancel_route(struct rdma_id_private *id_priv)
 911 {
 912         switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
 913         case IB_LINK_LAYER_INFINIBAND:
 914                 if (id_priv->query)
 915                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 916                 break;
 917         default:
 918                 break;
 919         }
 920 }
 921
 922 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 923 {
 924         struct rdma_id_private *dev_id_priv;
 925
 926         /*
 927          * Remove from listen_any_list to prevent added devices from spawning
 928          * additional listen requests.
 929          */
 930         mutex_lock(&lock);
 931         list_del(&id_priv->list);
 932
 933         while (!list_empty(&id_priv->listen_list)) {
 934                 dev_id_priv = list_entry(id_priv->listen_list.next,
 935                                          struct rdma_id_private, listen_list);
 936                 /* sync with device removal to avoid duplicate destruction */
 937                 list_del_init(&dev_id_priv->list);
 938                 list_del(&dev_id_priv->listen_list);
 939                 mutex_unlock(&lock);
 940
 941                 rdma_destroy_id(&dev_id_priv->id);
 942                 mutex_lock(&lock);
 943         }
 944         mutex_unlock(&lock);
 945 }
 946
 947 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 948                                  enum rdma_cm_state state)
 949 {
 950         switch (state) {
 951         case RDMA_CM_ADDR_QUERY:
 952                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 953                 break;
 954         case RDMA_CM_ROUTE_QUERY:
 955                 cma_cancel_route(id_priv);
 956                 break;
 957         case RDMA_CM_LISTEN:
 958                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
 959                         cma_cancel_listens(id_priv);
 960                 break;
 961         default:
 962                 break;
 963         }
 964 }
 965
 966 static void cma_release_port(struct rdma_id_private *id_priv)
 967 {
 968         struct rdma_bind_list *bind_list = id_priv->bind_list;
 969
 970         if (!bind_list)
 971                 return;
 972
 973         mutex_lock(&lock);
 974         hlist_del(&id_priv->node);
 975         if (hlist_empty(&bind_list->owners)) {
 976                 idr_remove(bind_list->ps, bind_list->port);
 977                 kfree(bind_list);
 978         }
 979         mutex_unlock(&lock);
 980 }
 981
 982 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 983 {
 984         struct cma_multicast *mc;
 985
 986         while (!list_empty(&id_priv->mc_list)) {
 987                 mc = container_of(id_priv->mc_list.next,
 988                                   struct cma_multicast, list);
 989                 list_del(&mc->list);
 990                 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
 991                 case IB_LINK_LAYER_INFINIBAND:
 992                         ib_sa_free_multicast(mc->multicast.ib);
 993                         kfree(mc);
 994                         break;
 995                 case IB_LINK_LAYER_ETHERNET:
 996                         kref_put(&mc->mcref, release_mc);
 997                         break;
 998                 default:
 999                         break;
1000                 }
1001         }
1002 }
1003
1004 void rdma_destroy_id(struct rdma_cm_id *id)
1005 {
1006         struct rdma_id_private *id_priv;
1007         enum rdma_cm_state state;
1008
1009         id_priv = container_of(id, struct rdma_id_private, id);
1010         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1011         cma_cancel_operation(id_priv, state);
1012
1013         /*
1014          * Wait for any active callback to finish.  New callbacks will find
1015          * the id_priv state set to destroying and abort.
1016          */
1017         mutex_lock(&id_priv->handler_mutex);
1018         mutex_unlock(&id_priv->handler_mutex);
1019
1020         if (id_priv->cma_dev) {
1021                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
1022                 case RDMA_TRANSPORT_IB:
1023                         if (id_priv->cm_id.ib)
1024                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1025                         break;
1026                 case RDMA_TRANSPORT_IWARP:
1027                         if (id_priv->cm_id.iw)
1028                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1029                         break;
1030                 default:
1031                         break;
1032                 }
1033                 cma_leave_mc_groups(id_priv);
1034                 cma_release_dev(id_priv);
1035         }
1036
1037         cma_release_port(id_priv);
1038         cma_deref_id(id_priv);
1039         wait_for_completion(&id_priv->comp);
1040
1041         if (id_priv->internal_id)
1042                 cma_deref_id(id_priv->id.context);
1043
1044         kfree(id_priv->id.route.path_rec);
1045         kfree(id_priv);
1046 }
1047 EXPORT_SYMBOL(rdma_destroy_id);
1048
1049 static int cma_rep_recv(struct rdma_id_private *id_priv)
1050 {
1051         int ret;
1052
1053         ret = cma_modify_qp_rtr(id_priv, NULL);
1054         if (ret)
1055                 goto reject;
1056
1057         ret = cma_modify_qp_rts(id_priv, NULL);
1058         if (ret)
1059                 goto reject;
1060
1061         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1062         if (ret)
1063                 goto reject;
1064
1065         return 0;
1066 reject:
1067         cma_modify_qp_err(id_priv);
1068         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1069                        NULL, 0, NULL, 0);
1070         return ret;
1071 }
1072
1073 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1074                                    struct ib_cm_rep_event_param *rep_data,
1075                                    void *private_data)
1076 {
1077         event->param.conn.private_data = private_data;
1078         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1079         event->param.conn.responder_resources = rep_data->responder_resources;
1080         event->param.conn.initiator_depth = rep_data->initiator_depth;
1081         event->param.conn.flow_control = rep_data->flow_control;
1082         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1083         event->param.conn.srq = rep_data->srq;
1084         event->param.conn.qp_num = rep_data->remote_qpn;
1085 }
1086
1087 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1088 {
1089         struct rdma_id_private *id_priv = cm_id->context;
1090         struct rdma_cm_event event;
1091         int ret = 0;
1092
1093         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1094                 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1095             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1096                 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1097                 return 0;
1098
1099         memset(&event, 0, sizeof event);
1100         switch (ib_event->event) {
1101         case IB_CM_REQ_ERROR:
1102         case IB_CM_REP_ERROR:
1103                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1104                 event.status = -ETIMEDOUT;
1105                 break;
1106         case IB_CM_REP_RECEIVED:
1107                 if (id_priv->id.qp) {
1108                         event.status = cma_rep_recv(id_priv);
1109                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1110                                                      RDMA_CM_EVENT_ESTABLISHED;
1111                 } else {
1112                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1113                 }
1114                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1115                                        ib_event->private_data);
1116                 break;
1117         case IB_CM_RTU_RECEIVED:
1118         case IB_CM_USER_ESTABLISHED:
1119                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1120                 break;
1121         case IB_CM_DREQ_ERROR:
1122                 event.status = -ETIMEDOUT; /* fall through */
1123         case IB_CM_DREQ_RECEIVED:
1124         case IB_CM_DREP_RECEIVED:
1125                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1126                                    RDMA_CM_DISCONNECT))
1127                         goto out;
1128                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1129                 break;
1130         case IB_CM_TIMEWAIT_EXIT:
1131                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1132                 break;
1133         case IB_CM_MRA_RECEIVED:
1134                 /* ignore event */
1135                 goto out;
1136         case IB_CM_REJ_RECEIVED:
1137                 cma_modify_qp_err(id_priv);
1138                 event.status = ib_event->param.rej_rcvd.reason;
1139                 event.event = RDMA_CM_EVENT_REJECTED;
1140                 event.param.conn.private_data = ib_event->private_data;
1141                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1142                 break;
1143         default:
1144                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1145                        ib_event->event);
1146                 goto out;
1147         }
1148
1149         ret = id_priv->id.event_handler(&id_priv->id, &event);
1150         if (ret) {
1151                 /* Destroy the CM ID by returning a non-zero value. */
1152                 id_priv->cm_id.ib = NULL;
1153                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1154                 mutex_unlock(&id_priv->handler_mutex);
1155                 rdma_destroy_id(&id_priv->id);
1156                 return ret;
1157         }
1158 out:
1159         mutex_unlock(&id_priv->handler_mutex);
1160         return ret;
1161 }
1162
1163 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1164                                                struct ib_cm_event *ib_event)
1165 {
1166         struct rdma_id_private *id_priv;
1167         struct rdma_cm_id *id;
1168         struct rdma_route *rt;
1169         int ret;
1170
1171         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1172                             listen_id->ps, ib_event->param.req_rcvd.qp_type);
1173         if (IS_ERR(id))
1174                 return NULL;
1175
1176         id_priv = container_of(id, struct rdma_id_private, id);
1177         if (cma_save_net_info(id, listen_id, ib_event))
1178                 goto err;
1179
1180         rt = &id->route;
1181         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1182         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1183                                GFP_KERNEL);
1184         if (!rt->path_rec)
1185                 goto err;
1186
1187         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1188         if (rt->num_paths == 2)
1189                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1190
1191         if (cma_any_addr(cma_src_addr(id_priv))) {
1192                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1193                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1194                 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1195         } else {
1196                 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1197                 if (ret)
1198                         goto err;
1199         }
1200         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1201
1202         id_priv->state = RDMA_CM_CONNECT;
1203         return id_priv;
1204
1205 err:
1206         rdma_destroy_id(id);
1207         return NULL;
1208 }
1209
1210 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1211                                               struct ib_cm_event *ib_event)
1212 {
1213         struct rdma_id_private *id_priv;
1214         struct rdma_cm_id *id;
1215         int ret;
1216
1217         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1218                             listen_id->ps, IB_QPT_UD);
1219         if (IS_ERR(id))
1220                 return NULL;
1221
1222         id_priv = container_of(id, struct rdma_id_private, id);
1223         if (cma_save_net_info(id, listen_id, ib_event))
1224                 goto err;
1225
1226         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1227                 ret = cma_translate_addr(cma_src_addr(id_priv), &id->route.addr.dev_addr);
1228                 if (ret)
1229                         goto err;
1230         }
1231
1232         id_priv->state = RDMA_CM_CONNECT;
1233         return id_priv;
1234 err:
1235         rdma_destroy_id(id);
1236         return NULL;
1237 }
1238
1239 static void cma_set_req_event_data(struct rdma_cm_event *event,
1240                                    struct ib_cm_req_event_param *req_data,
1241                                    void *private_data, int offset)
1242 {
1243         event->param.conn.private_data = private_data + offset;
1244         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1245         event->param.conn.responder_resources = req_data->responder_resources;
1246         event->param.conn.initiator_depth = req_data->initiator_depth;
1247         event->param.conn.flow_control = req_data->flow_control;
1248         event->param.conn.retry_count = req_data->retry_count;
1249         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1250         event->param.conn.srq = req_data->srq;
1251         event->param.conn.qp_num = req_data->remote_qpn;
1252 }
1253
1254 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1255 {
1256         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1257                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1258                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1259                  (id->qp_type == IB_QPT_UD)) ||
1260                 (!id->qp_type));
1261 }
1262
1263 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1264 {
1265         struct rdma_id_private *listen_id, *conn_id;
1266         struct rdma_cm_event event;
1267         int offset, ret;
1268
1269         listen_id = cm_id->context;
1270         if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1271                 return -EINVAL;
1272
1273         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1274                 return -ECONNABORTED;
1275
1276         memset(&event, 0, sizeof event);
1277         offset = cma_user_data_offset(listen_id);
1278         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1279         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1280                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1281                 event.param.ud.private_data = ib_event->private_data + offset;
1282                 event.param.ud.private_data_len =
1283                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1284         } else {
1285                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1286                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1287                                        ib_event->private_data, offset);
1288         }
1289         if (!conn_id) {
1290                 ret = -ENOMEM;
1291                 goto err1;
1292         }
1293
1294         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1295         ret = cma_acquire_dev(conn_id);
1296         if (ret)
1297                 goto err2;
1298
1299         conn_id->cm_id.ib = cm_id;
1300         cm_id->context = conn_id;
1301         cm_id->cm_handler = cma_ib_handler;
1302
1303         /*
1304          * Protect against the user destroying conn_id from another thread
1305          * until we're done accessing it.
1306          */
1307         atomic_inc(&conn_id->refcount);
1308         ret = conn_id->id.event_handler(&conn_id->id, &event);
1309         if (ret)
1310                 goto err3;
1311
1312         /*
1313          * Acquire mutex to prevent user executing rdma_destroy_id()
1314          * while we're accessing the cm_id.
1315          */
1316         mutex_lock(&lock);
1317         if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD))
1318                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1319         mutex_unlock(&lock);
1320         mutex_unlock(&conn_id->handler_mutex);
1321         mutex_unlock(&listen_id->handler_mutex);
1322         cma_deref_id(conn_id);
1323         return 0;
1324
1325 err3:
1326         cma_deref_id(conn_id);
1327         /* Destroy the CM ID by returning a non-zero value. */
1328         conn_id->cm_id.ib = NULL;
1329 err2:
1330         cma_exch(conn_id, RDMA_CM_DESTROYING);
1331         mutex_unlock(&conn_id->handler_mutex);
1332 err1:
1333         mutex_unlock(&listen_id->handler_mutex);
1334         if (conn_id)
1335                 rdma_destroy_id(&conn_id->id);
1336         return ret;
1337 }
1338
1339 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
1340 {
1341         if (addr->sa_family == AF_IB)
1342                 return ((struct sockaddr_ib *) addr)->sib_sid;
1343
1344         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
1345 }
1346 EXPORT_SYMBOL(rdma_get_service_id);
1347
1348 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1349                                  struct ib_cm_compare_data *compare)
1350 {
1351         struct cma_hdr *cma_data, *cma_mask;
1352         __be32 ip4_addr;
1353         struct in6_addr ip6_addr;
1354
1355         memset(compare, 0, sizeof *compare);
1356         cma_data = (void *) compare->data;
1357         cma_mask = (void *) compare->mask;
1358
1359         switch (addr->sa_family) {
1360         case AF_INET:
1361                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1362                 cma_set_ip_ver(cma_data, 4);
1363                 cma_set_ip_ver(cma_mask, 0xF);
1364                 if (!cma_any_addr(addr)) {
1365                         cma_data->dst_addr.ip4.addr = ip4_addr;
1366                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1367                 }
1368                 break;
1369         case AF_INET6:
1370                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1371                 cma_set_ip_ver(cma_data, 6);
1372                 cma_set_ip_ver(cma_mask, 0xF);
1373                 if (!cma_any_addr(addr)) {
1374                         cma_data->dst_addr.ip6 = ip6_addr;
1375                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1376                                sizeof cma_mask->dst_addr.ip6);
1377                 }
1378                 break;
1379         default:
1380                 break;
1381         }
1382 }
1383
1384 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1385 {
1386         struct rdma_id_private *id_priv = iw_id->context;
1387         struct rdma_cm_event event;
1388         struct sockaddr_in *sin;
1389         int ret = 0;
1390
1391         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1392                 return 0;
1393
1394         memset(&event, 0, sizeof event);
1395         switch (iw_event->event) {
1396         case IW_CM_EVENT_CLOSE:
1397                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1398                 break;
1399         case IW_CM_EVENT_CONNECT_REPLY:
1400                 sin = (struct sockaddr_in *) cma_src_addr(id_priv);
1401                 *sin = iw_event->local_addr;
1402                 sin = (struct sockaddr_in *) cma_dst_addr(id_priv);
1403                 *sin = iw_event->remote_addr;
1404                 switch (iw_event->status) {
1405                 case 0:
1406                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1407                         event.param.conn.initiator_depth = iw_event->ird;
1408                         event.param.conn.responder_resources = iw_event->ord;
1409                         break;
1410                 case -ECONNRESET:
1411                 case -ECONNREFUSED:
1412                         event.event = RDMA_CM_EVENT_REJECTED;
1413                         break;
1414                 case -ETIMEDOUT:
1415                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1416                         break;
1417                 default:
1418                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1419                         break;
1420                 }
1421                 break;
1422         case IW_CM_EVENT_ESTABLISHED:
1423                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1424                 event.param.conn.initiator_depth = iw_event->ird;
1425                 event.param.conn.responder_resources = iw_event->ord;
1426                 break;
1427         default:
1428                 BUG_ON(1);
1429         }
1430
1431         event.status = iw_event->status;
1432         event.param.conn.private_data = iw_event->private_data;
1433         event.param.conn.private_data_len = iw_event->private_data_len;
1434         ret = id_priv->id.event_handler(&id_priv->id, &event);
1435         if (ret) {
1436                 /* Destroy the CM ID by returning a non-zero value. */
1437                 id_priv->cm_id.iw = NULL;
1438                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1439                 mutex_unlock(&id_priv->handler_mutex);
1440                 rdma_destroy_id(&id_priv->id);
1441                 return ret;
1442         }
1443
1444         mutex_unlock(&id_priv->handler_mutex);
1445         return ret;
1446 }
1447
1448 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1449                                struct iw_cm_event *iw_event)
1450 {
1451         struct rdma_cm_id *new_cm_id;
1452         struct rdma_id_private *listen_id, *conn_id;
1453         struct sockaddr_in *sin;
1454         struct net_device *dev = NULL;
1455         struct rdma_cm_event event;
1456         int ret;
1457         struct ib_device_attr attr;
1458
1459         listen_id = cm_id->context;
1460         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1461                 return -ECONNABORTED;
1462
1463         /* Create a new RDMA id for the new IW CM ID */
1464         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1465                                    listen_id->id.context,
1466                                    RDMA_PS_TCP, IB_QPT_RC);
1467         if (IS_ERR(new_cm_id)) {
1468                 ret = -ENOMEM;
1469                 goto out;
1470         }
1471         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1472         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1473         conn_id->state = RDMA_CM_CONNECT;
1474
1475         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1476         if (!dev) {
1477                 ret = -EADDRNOTAVAIL;
1478                 mutex_unlock(&conn_id->handler_mutex);
1479                 rdma_destroy_id(new_cm_id);
1480                 goto out;
1481         }
1482         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1483         if (ret) {
1484                 mutex_unlock(&conn_id->handler_mutex);
1485                 rdma_destroy_id(new_cm_id);
1486                 goto out;
1487         }
1488
1489         ret = cma_acquire_dev(conn_id);
1490         if (ret) {
1491                 mutex_unlock(&conn_id->handler_mutex);
1492                 rdma_destroy_id(new_cm_id);
1493                 goto out;
1494         }
1495
1496         conn_id->cm_id.iw = cm_id;
1497         cm_id->context = conn_id;
1498         cm_id->cm_handler = cma_iw_handler;
1499
1500         sin = (struct sockaddr_in *) cma_src_addr(conn_id);
1501         *sin = iw_event->local_addr;
1502         sin = (struct sockaddr_in *) cma_dst_addr(conn_id);
1503         *sin = iw_event->remote_addr;
1504
1505         ret = ib_query_device(conn_id->id.device, &attr);
1506         if (ret) {
1507                 mutex_unlock(&conn_id->handler_mutex);
1508                 rdma_destroy_id(new_cm_id);
1509                 goto out;
1510         }
1511
1512         memset(&event, 0, sizeof event);
1513         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1514         event.param.conn.private_data = iw_event->private_data;
1515         event.param.conn.private_data_len = iw_event->private_data_len;
1516         event.param.conn.initiator_depth = iw_event->ird;
1517         event.param.conn.responder_resources = iw_event->ord;
1518
1519         /*
1520          * Protect against the user destroying conn_id from another thread
1521          * until we're done accessing it.
1522          */
1523         atomic_inc(&conn_id->refcount);
1524         ret = conn_id->id.event_handler(&conn_id->id, &event);
1525         if (ret) {
1526                 /* User wants to destroy the CM ID */
1527                 conn_id->cm_id.iw = NULL;
1528                 cma_exch(conn_id, RDMA_CM_DESTROYING);
1529                 mutex_unlock(&conn_id->handler_mutex);
1530                 cma_deref_id(conn_id);
1531                 rdma_destroy_id(&conn_id->id);
1532                 goto out;
1533         }
1534
1535         mutex_unlock(&conn_id->handler_mutex);
1536         cma_deref_id(conn_id);
1537
1538 out:
1539         if (dev)
1540                 dev_put(dev);
1541         mutex_unlock(&listen_id->handler_mutex);
1542         return ret;
1543 }
1544
1545 static int cma_ib_listen(struct rdma_id_private *id_priv)
1546 {
1547         struct ib_cm_compare_data compare_data;
1548         struct sockaddr *addr;
1549         struct ib_cm_id *id;
1550         __be64 svc_id;
1551         int ret;
1552
1553         id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1554         if (IS_ERR(id))
1555                 return PTR_ERR(id);
1556
1557         id_priv->cm_id.ib = id;
1558
1559         addr = cma_src_addr(id_priv);
1560         svc_id = rdma_get_service_id(&id_priv->id, addr);
1561         if (cma_any_addr(addr) && !id_priv->afonly)
1562                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1563         else {
1564                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1565                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1566         }
1567
1568         if (ret) {
1569                 ib_destroy_cm_id(id_priv->cm_id.ib);
1570                 id_priv->cm_id.ib = NULL;
1571         }
1572
1573         return ret;
1574 }
1575
1576 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1577 {
1578         int ret;
1579         struct sockaddr_in *sin;
1580         struct iw_cm_id *id;
1581
1582         id = iw_create_cm_id(id_priv->id.device,
1583                              iw_conn_req_handler,
1584                              id_priv);
1585         if (IS_ERR(id))
1586                 return PTR_ERR(id);
1587
1588         id_priv->cm_id.iw = id;
1589
1590         sin = (struct sockaddr_in *) cma_src_addr(id_priv);
1591         id_priv->cm_id.iw->local_addr = *sin;
1592
1593         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1594
1595         if (ret) {
1596                 iw_destroy_cm_id(id_priv->cm_id.iw);
1597                 id_priv->cm_id.iw = NULL;
1598         }
1599
1600         return ret;
1601 }
1602
1603 static int cma_listen_handler(struct rdma_cm_id *id,
1604                               struct rdma_cm_event *event)
1605 {
1606         struct rdma_id_private *id_priv = id->context;
1607
1608         id->context = id_priv->id.context;
1609         id->event_handler = id_priv->id.event_handler;
1610         return id_priv->id.event_handler(id, event);
1611 }
1612
1613 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1614                               struct cma_device *cma_dev)
1615 {
1616         struct rdma_id_private *dev_id_priv;
1617         struct rdma_cm_id *id;
1618         int ret;
1619
1620         if (cma_family(id_priv) == AF_IB &&
1621             rdma_node_get_transport(cma_dev->device->node_type) != RDMA_TRANSPORT_IB)
1622                 return;
1623
1624         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1625                             id_priv->id.qp_type);
1626         if (IS_ERR(id))
1627                 return;
1628
1629         dev_id_priv = container_of(id, struct rdma_id_private, id);
1630
1631         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1632         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
1633                rdma_addr_size(cma_src_addr(id_priv)));
1634
1635         cma_attach_to_dev(dev_id_priv, cma_dev);
1636         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1637         atomic_inc(&id_priv->refcount);
1638         dev_id_priv->internal_id = 1;
1639         dev_id_priv->afonly = id_priv->afonly;
1640
1641         ret = rdma_listen(id, id_priv->backlog);
1642         if (ret)
1643                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1644                        "listening on device %s\n", ret, cma_dev->device->name);
1645 }
1646
1647 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1648 {
1649         struct cma_device *cma_dev;
1650
1651         mutex_lock(&lock);
1652         list_add_tail(&id_priv->list, &listen_any_list);
1653         list_for_each_entry(cma_dev, &dev_list, list)
1654                 cma_listen_on_dev(id_priv, cma_dev);
1655         mutex_unlock(&lock);
1656 }
1657
1658 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1659 {
1660         struct rdma_id_private *id_priv;
1661
1662         id_priv = container_of(id, struct rdma_id_private, id);
1663         id_priv->tos = (u8) tos;
1664 }
1665 EXPORT_SYMBOL(rdma_set_service_type);
1666
1667 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1668                               void *context)
1669 {
1670         struct cma_work *work = context;
1671         struct rdma_route *route;
1672
1673         route = &work->id->id.route;
1674
1675         if (!status) {
1676                 route->num_paths = 1;
1677                 *route->path_rec = *path_rec;
1678         } else {
1679                 work->old_state = RDMA_CM_ROUTE_QUERY;
1680                 work->new_state = RDMA_CM_ADDR_RESOLVED;
1681                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1682                 work->event.status = status;
1683         }
1684
1685         queue_work(cma_wq, &work->work);
1686 }
1687
1688 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1689                               struct cma_work *work)
1690 {
1691         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1692         struct ib_sa_path_rec path_rec;
1693         ib_sa_comp_mask comp_mask;
1694         struct sockaddr_in6 *sin6;
1695         struct sockaddr_ib *sib;
1696
1697         memset(&path_rec, 0, sizeof path_rec);
1698         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
1699         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
1700         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
1701         path_rec.numb_path = 1;
1702         path_rec.reversible = 1;
1703         path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
1704
1705         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1706                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1707                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1708
1709         switch (cma_family(id_priv)) {
1710         case AF_INET:
1711                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1712                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1713                 break;
1714         case AF_INET6:
1715                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
1716                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1717                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1718                 break;
1719         case AF_IB:
1720                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
1721                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
1722                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1723                 break;
1724         }
1725
1726         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1727                                                id_priv->id.port_num, &path_rec,
1728                                                comp_mask, timeout_ms,
1729                                                GFP_KERNEL, cma_query_handler,
1730                                                work, &id_priv->query);
1731
1732         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1733 }
1734
1735 static void cma_work_handler(struct work_struct *_work)
1736 {
1737         struct cma_work *work = container_of(_work, struct cma_work, work);
1738         struct rdma_id_private *id_priv = work->id;
1739         int destroy = 0;
1740
1741         mutex_lock(&id_priv->handler_mutex);
1742         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1743                 goto out;
1744
1745         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1746                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1747                 destroy = 1;
1748         }
1749 out:
1750         mutex_unlock(&id_priv->handler_mutex);
1751         cma_deref_id(id_priv);
1752         if (destroy)
1753                 rdma_destroy_id(&id_priv->id);
1754         kfree(work);
1755 }
1756
1757 static void cma_ndev_work_handler(struct work_struct *_work)
1758 {
1759         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1760         struct rdma_id_private *id_priv = work->id;
1761         int destroy = 0;
1762
1763         mutex_lock(&id_priv->handler_mutex);
1764         if (id_priv->state == RDMA_CM_DESTROYING ||
1765             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1766                 goto out;
1767
1768         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1769                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1770                 destroy = 1;
1771         }
1772
1773 out:
1774         mutex_unlock(&id_priv->handler_mutex);
1775         cma_deref_id(id_priv);
1776         if (destroy)
1777                 rdma_destroy_id(&id_priv->id);
1778         kfree(work);
1779 }
1780
1781 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1782 {
1783         struct rdma_route *route = &id_priv->id.route;
1784         struct cma_work *work;
1785         int ret;
1786
1787         work = kzalloc(sizeof *work, GFP_KERNEL);
1788         if (!work)
1789                 return -ENOMEM;
1790
1791         work->id = id_priv;
1792         INIT_WORK(&work->work, cma_work_handler);
1793         work->old_state = RDMA_CM_ROUTE_QUERY;
1794         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1795         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1796
1797         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1798         if (!route->path_rec) {
1799                 ret = -ENOMEM;
1800                 goto err1;
1801         }
1802
1803         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1804         if (ret)
1805                 goto err2;
1806
1807         return 0;
1808 err2:
1809         kfree(route->path_rec);
1810         route->path_rec = NULL;
1811 err1:
1812         kfree(work);
1813         return ret;
1814 }
1815
1816 int rdma_set_ib_paths(struct rdma_cm_id *id,
1817                       struct ib_sa_path_rec *path_rec, int num_paths)
1818 {
1819         struct rdma_id_private *id_priv;
1820         int ret;
1821
1822         id_priv = container_of(id, struct rdma_id_private, id);
1823         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1824                            RDMA_CM_ROUTE_RESOLVED))
1825                 return -EINVAL;
1826
1827         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1828                                      GFP_KERNEL);
1829         if (!id->route.path_rec) {
1830                 ret = -ENOMEM;
1831                 goto err;
1832         }
1833
1834         id->route.num_paths = num_paths;
1835         return 0;
1836 err:
1837         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1838         return ret;
1839 }
1840 EXPORT_SYMBOL(rdma_set_ib_paths);
1841
1842 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1843 {
1844         struct cma_work *work;
1845
1846         work = kzalloc(sizeof *work, GFP_KERNEL);
1847         if (!work)
1848                 return -ENOMEM;
1849
1850         work->id = id_priv;
1851         INIT_WORK(&work->work, cma_work_handler);
1852         work->old_state = RDMA_CM_ROUTE_QUERY;
1853         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1854         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1855         queue_work(cma_wq, &work->work);
1856         return 0;
1857 }
1858
1859 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1860 {
1861         struct rdma_route *route = &id_priv->id.route;
1862         struct rdma_addr *addr = &route->addr;
1863         struct cma_work *work;
1864         int ret;
1865         struct net_device *ndev = NULL;
1866         u16 vid;
1867
1868         work = kzalloc(sizeof *work, GFP_KERNEL);
1869         if (!work)
1870                 return -ENOMEM;
1871
1872         work->id = id_priv;
1873         INIT_WORK(&work->work, cma_work_handler);
1874
1875         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1876         if (!route->path_rec) {
1877                 ret = -ENOMEM;
1878                 goto err1;
1879         }
1880
1881         route->num_paths = 1;
1882
1883         if (addr->dev_addr.bound_dev_if)
1884                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1885         if (!ndev) {
1886                 ret = -ENODEV;
1887                 goto err2;
1888         }
1889
1890         vid = rdma_vlan_dev_vlan_id(ndev);
1891
1892         iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1893         iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1894
1895         route->path_rec->hop_limit = 1;
1896         route->path_rec->reversible = 1;
1897         route->path_rec->pkey = cpu_to_be16(0xffff);
1898         route->path_rec->mtu_selector = IB_SA_EQ;
1899         route->path_rec->sl = netdev_get_prio_tc_map(
1900                         ndev->priv_flags & IFF_802_1Q_VLAN ?
1901                                 vlan_dev_real_dev(ndev) : ndev,
1902                         rt_tos2priority(id_priv->tos));
1903
1904         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1905         route->path_rec->rate_selector = IB_SA_EQ;
1906         route->path_rec->rate = iboe_get_rate(ndev);
1907         dev_put(ndev);
1908         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1909         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1910         if (!route->path_rec->mtu) {
1911                 ret = -EINVAL;
1912                 goto err2;
1913         }
1914
1915         work->old_state = RDMA_CM_ROUTE_QUERY;
1916         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1917         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1918         work->event.status = 0;
1919
1920         queue_work(cma_wq, &work->work);
1921
1922         return 0;
1923
1924 err2:
1925         kfree(route->path_rec);
1926         route->path_rec = NULL;
1927 err1:
1928         kfree(work);
1929         return ret;
1930 }
1931
1932 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1933 {
1934         struct rdma_id_private *id_priv;
1935         int ret;
1936
1937         id_priv = container_of(id, struct rdma_id_private, id);
1938         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1939                 return -EINVAL;
1940
1941         atomic_inc(&id_priv->refcount);
1942         switch (rdma_node_get_transport(id->device->node_type)) {
1943         case RDMA_TRANSPORT_IB:
1944                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1945                 case IB_LINK_LAYER_INFINIBAND:
1946                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1947                         break;
1948                 case IB_LINK_LAYER_ETHERNET:
1949                         ret = cma_resolve_iboe_route(id_priv);
1950                         break;
1951                 default:
1952                         ret = -ENOSYS;
1953                 }
1954                 break;
1955         case RDMA_TRANSPORT_IWARP:
1956                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1957                 break;
1958         default:
1959                 ret = -ENOSYS;
1960                 break;
1961         }
1962         if (ret)
1963                 goto err;
1964
1965         return 0;
1966 err:
1967         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1968         cma_deref_id(id_priv);
1969         return ret;
1970 }
1971 EXPORT_SYMBOL(rdma_resolve_route);
1972
1973 static void cma_set_loopback(struct sockaddr *addr)
1974 {
1975         switch (addr->sa_family) {
1976         case AF_INET:
1977                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
1978                 break;
1979         case AF_INET6:
1980                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
1981                               0, 0, 0, htonl(1));
1982                 break;
1983         default:
1984                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
1985                             0, 0, 0, htonl(1));
1986                 break;
1987         }
1988 }
1989
1990 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1991 {
1992         struct cma_device *cma_dev, *cur_dev;
1993         struct ib_port_attr port_attr;
1994         union ib_gid gid;
1995         u16 pkey;
1996         int ret;
1997         u8 p;
1998
1999         cma_dev = NULL;
2000         mutex_lock(&lock);
2001         list_for_each_entry(cur_dev, &dev_list, list) {
2002                 if (cma_family(id_priv) == AF_IB &&
2003                     rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
2004                         continue;
2005
2006                 if (!cma_dev)
2007                         cma_dev = cur_dev;
2008
2009                 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2010                         if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2011                             port_attr.state == IB_PORT_ACTIVE) {
2012                                 cma_dev = cur_dev;
2013                                 goto port_found;
2014                         }
2015                 }
2016         }
2017
2018         if (!cma_dev) {
2019                 ret = -ENODEV;
2020                 goto out;
2021         }
2022
2023         p = 1;
2024
2025 port_found:
2026         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
2027         if (ret)
2028                 goto out;
2029
2030         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2031         if (ret)
2032                 goto out;
2033
2034         id_priv->id.route.addr.dev_addr.dev_type =
2035                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
2036                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2037
2038         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2039         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2040         id_priv->id.port_num = p;
2041         cma_attach_to_dev(id_priv, cma_dev);
2042         cma_set_loopback(cma_src_addr(id_priv));
2043 out:
2044         mutex_unlock(&lock);
2045         return ret;
2046 }
2047
2048 static void addr_handler(int status, struct sockaddr *src_addr,
2049                          struct rdma_dev_addr *dev_addr, void *context)
2050 {
2051         struct rdma_id_private *id_priv = context;
2052         struct rdma_cm_event event;
2053
2054         memset(&event, 0, sizeof event);
2055         mutex_lock(&id_priv->handler_mutex);
2056         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2057                            RDMA_CM_ADDR_RESOLVED))
2058                 goto out;
2059
2060         if (!status && !id_priv->cma_dev)
2061                 status = cma_acquire_dev(id_priv);
2062
2063         if (status) {
2064                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2065                                    RDMA_CM_ADDR_BOUND))
2066                         goto out;
2067                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2068                 event.status = status;
2069         } else {
2070                 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2071                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2072         }
2073
2074         if (id_priv->id.event_handler(&id_priv->id, &event)) {
2075                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2076                 mutex_unlock(&id_priv->handler_mutex);
2077                 cma_deref_id(id_priv);
2078                 rdma_destroy_id(&id_priv->id);
2079                 return;
2080         }
2081 out:
2082         mutex_unlock(&id_priv->handler_mutex);
2083         cma_deref_id(id_priv);
2084 }
2085
2086 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2087 {
2088         struct cma_work *work;
2089         union ib_gid gid;
2090         int ret;
2091
2092         work = kzalloc(sizeof *work, GFP_KERNEL);
2093         if (!work)
2094                 return -ENOMEM;
2095
2096         if (!id_priv->cma_dev) {
2097                 ret = cma_bind_loopback(id_priv);
2098                 if (ret)
2099                         goto err;
2100         }
2101
2102         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2103         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2104
2105         work->id = id_priv;
2106         INIT_WORK(&work->work, cma_work_handler);
2107         work->old_state = RDMA_CM_ADDR_QUERY;
2108         work->new_state = RDMA_CM_ADDR_RESOLVED;
2109         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2110         queue_work(cma_wq, &work->work);
2111         return 0;
2112 err:
2113         kfree(work);
2114         return ret;
2115 }
2116
2117 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2118 {
2119         struct cma_work *work;
2120         int ret;
2121
2122         work = kzalloc(sizeof *work, GFP_KERNEL);
2123         if (!work)
2124                 return -ENOMEM;
2125
2126         if (!id_priv->cma_dev) {
2127                 ret = cma_resolve_ib_dev(id_priv);
2128                 if (ret)
2129                         goto err;
2130         }
2131
2132         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2133                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2134
2135         work->id = id_priv;
2136         INIT_WORK(&work->work, cma_work_handler);
2137         work->old_state = RDMA_CM_ADDR_QUERY;
2138         work->new_state = RDMA_CM_ADDR_RESOLVED;
2139         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2140         queue_work(cma_wq, &work->work);
2141         return 0;
2142 err:
2143         kfree(work);
2144         return ret;
2145 }
2146
2147 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2148                          struct sockaddr *dst_addr)
2149 {
2150         if (!src_addr || !src_addr->sa_family) {
2151                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2152                 src_addr->sa_family = dst_addr->sa_family;
2153                 if (dst_addr->sa_family == AF_INET6) {
2154                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2155                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2156                 } else if (dst_addr->sa_family == AF_IB) {
2157                         ((struct sockaddr_ib *) src_addr)->sib_pkey =
2158                                 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2159                 }
2160         }
2161         return rdma_bind_addr(id, src_addr);
2162 }
2163
2164 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2165                       struct sockaddr *dst_addr, int timeout_ms)
2166 {
2167         struct rdma_id_private *id_priv;
2168         int ret;
2169
2170         id_priv = container_of(id, struct rdma_id_private, id);
2171         if (id_priv->state == RDMA_CM_IDLE) {
2172                 ret = cma_bind_addr(id, src_addr, dst_addr);
2173                 if (ret)
2174                         return ret;
2175         }
2176
2177         if (cma_family(id_priv) != dst_addr->sa_family)
2178                 return -EINVAL;
2179
2180         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2181                 return -EINVAL;
2182
2183         atomic_inc(&id_priv->refcount);
2184         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2185         if (cma_any_addr(dst_addr)) {
2186                 ret = cma_resolve_loopback(id_priv);
2187         } else {
2188                 if (dst_addr->sa_family == AF_IB) {
2189                         ret = cma_resolve_ib_addr(id_priv);
2190                 } else {
2191                         ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2192                                               dst_addr, &id->route.addr.dev_addr,
2193                                               timeout_ms, addr_handler, id_priv);
2194                 }
2195         }
2196         if (ret)
2197                 goto err;
2198
2199         return 0;
2200 err:
2201         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2202         cma_deref_id(id_priv);
2203         return ret;
2204 }
2205 EXPORT_SYMBOL(rdma_resolve_addr);
2206
2207 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2208 {
2209         struct rdma_id_private *id_priv;
2210         unsigned long flags;
2211         int ret;
2212
2213         id_priv = container_of(id, struct rdma_id_private, id);
2214         spin_lock_irqsave(&id_priv->lock, flags);
2215         if (reuse || id_priv->state == RDMA_CM_IDLE) {
2216                 id_priv->reuseaddr = reuse;
2217                 ret = 0;
2218         } else {
2219                 ret = -EINVAL;
2220         }
2221         spin_unlock_irqrestore(&id_priv->lock, flags);
2222         return ret;
2223 }
2224 EXPORT_SYMBOL(rdma_set_reuseaddr);
2225
2226 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2227 {
2228         struct rdma_id_private *id_priv;
2229         unsigned long flags;
2230         int ret;
2231
2232         id_priv = container_of(id, struct rdma_id_private, id);
2233         spin_lock_irqsave(&id_priv->lock, flags);
2234         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2235                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2236                 id_priv->afonly = afonly;
2237                 ret = 0;
2238         } else {
2239                 ret = -EINVAL;
2240         }
2241         spin_unlock_irqrestore(&id_priv->lock, flags);
2242         return ret;
2243 }
2244 EXPORT_SYMBOL(rdma_set_afonly);
2245
2246 static void cma_bind_port(struct rdma_bind_list *bind_list,
2247                           struct rdma_id_private *id_priv)
2248 {
2249         struct sockaddr *addr;
2250         struct sockaddr_ib *sib;
2251         u64 sid, mask;
2252         __be16 port;
2253
2254         addr = cma_src_addr(id_priv);
2255         port = htons(bind_list->port);
2256
2257         switch (addr->sa_family) {
2258         case AF_INET:
2259                 ((struct sockaddr_in *) addr)->sin_port = port;
2260                 break;
2261         case AF_INET6:
2262                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2263                 break;
2264         case AF_IB:
2265                 sib = (struct sockaddr_ib *) addr;
2266                 sid = be64_to_cpu(sib->sib_sid);
2267                 mask = be64_to_cpu(sib->sib_sid_mask);
2268                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2269                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2270                 break;
2271         }
2272         id_priv->bind_list = bind_list;
2273         hlist_add_head(&id_priv->node, &bind_list->owners);
2274 }
2275
2276 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2277                           unsigned short snum)
2278 {
2279         struct rdma_bind_list *bind_list;
2280         int ret;
2281
2282         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2283         if (!bind_list)
2284                 return -ENOMEM;
2285
2286         ret = idr_alloc(ps, bind_list, snum, snum + 1, GFP_KERNEL);
2287         if (ret < 0)
2288                 goto err;
2289
2290         bind_list->ps = ps;
2291         bind_list->port = (unsigned short)ret;
2292         cma_bind_port(bind_list, id_priv);
2293         return 0;
2294 err:
2295         kfree(bind_list);
2296         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2297 }
2298
2299 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2300 {
2301         static unsigned int last_used_port;
2302         int low, high, remaining;
2303         unsigned int rover;
2304
2305         inet_get_local_port_range(&low, &high);
2306         remaining = (high - low) + 1;
2307         rover = net_random() % remaining + low;
2308 retry:
2309         if (last_used_port != rover &&
2310             !idr_find(ps, (unsigned short) rover)) {
2311                 int ret = cma_alloc_port(ps, id_priv, rover);
2312                 /*
2313                  * Remember previously used port number in order to avoid
2314                  * re-using same port immediately after it is closed.
2315                  */
2316                 if (!ret)
2317                         last_used_port = rover;
2318                 if (ret != -EADDRNOTAVAIL)
2319                         return ret;
2320         }
2321         if (--remaining) {
2322                 rover++;
2323                 if ((rover < low) || (rover > high))
2324                         rover = low;
2325                 goto retry;
2326         }
2327         return -EADDRNOTAVAIL;
2328 }
2329
2330 /*
2331  * Check that the requested port is available.  This is called when trying to
2332  * bind to a specific port, or when trying to listen on a bound port.  In
2333  * the latter case, the provided id_priv may already be on the bind_list, but
2334  * we still need to check that it's okay to start listening.
2335  */
2336 static int cma_check_port(struct rdma_bind_list *bind_list,
2337                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2338 {
2339         struct rdma_id_private *cur_id;
2340         struct sockaddr *addr, *cur_addr;
2341
2342         addr = cma_src_addr(id_priv);
2343         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2344                 if (id_priv == cur_id)
2345                         continue;
2346
2347                 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2348                     cur_id->reuseaddr)
2349                         continue;
2350
2351                 cur_addr = cma_src_addr(cur_id);
2352                 if (id_priv->afonly && cur_id->afonly &&
2353                     (addr->sa_family != cur_addr->sa_family))
2354                         continue;
2355
2356                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2357                         return -EADDRNOTAVAIL;
2358
2359                 if (!cma_addr_cmp(addr, cur_addr))
2360                         return -EADDRINUSE;
2361         }
2362         return 0;
2363 }
2364
2365 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2366 {
2367         struct rdma_bind_list *bind_list;
2368         unsigned short snum;
2369         int ret;
2370
2371         snum = ntohs(cma_port(cma_src_addr(id_priv)));
2372         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2373                 return -EACCES;
2374
2375         bind_list = idr_find(ps, snum);
2376         if (!bind_list) {
2377                 ret = cma_alloc_port(ps, id_priv, snum);
2378         } else {
2379                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2380                 if (!ret)
2381                         cma_bind_port(bind_list, id_priv);
2382         }
2383         return ret;
2384 }
2385
2386 static int cma_bind_listen(struct rdma_id_private *id_priv)
2387 {
2388         struct rdma_bind_list *bind_list = id_priv->bind_list;
2389         int ret = 0;
2390
2391         mutex_lock(&lock);
2392         if (bind_list->owners.first->next)
2393                 ret = cma_check_port(bind_list, id_priv, 0);
2394         mutex_unlock(&lock);
2395         return ret;
2396 }
2397
2398 static struct idr *cma_select_inet_ps(struct rdma_id_private *id_priv)
2399 {
2400         switch (id_priv->id.ps) {
2401         case RDMA_PS_TCP:
2402                 return &tcp_ps;
2403         case RDMA_PS_UDP:
2404                 return &udp_ps;
2405         case RDMA_PS_IPOIB:
2406                 return &ipoib_ps;
2407         case RDMA_PS_IB:
2408                 return &ib_ps;
2409         default:
2410                 return NULL;
2411         }
2412 }
2413
2414 static struct idr *cma_select_ib_ps(struct rdma_id_private *id_priv)
2415 {
2416         struct idr *ps = NULL;
2417         struct sockaddr_ib *sib;
2418         u64 sid_ps, mask, sid;
2419
2420         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2421         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
2422         sid = be64_to_cpu(sib->sib_sid) & mask;
2423
2424         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
2425                 sid_ps = RDMA_IB_IP_PS_IB;
2426                 ps = &ib_ps;
2427         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
2428                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
2429                 sid_ps = RDMA_IB_IP_PS_TCP;
2430                 ps = &tcp_ps;
2431         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
2432                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
2433                 sid_ps = RDMA_IB_IP_PS_UDP;
2434                 ps = &udp_ps;
2435         }
2436
2437         if (ps) {
2438                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
2439                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
2440                                                 be64_to_cpu(sib->sib_sid_mask));
2441         }
2442         return ps;
2443 }
2444
2445 static int cma_get_port(struct rdma_id_private *id_priv)
2446 {
2447         struct idr *ps;
2448         int ret;
2449
2450         if (cma_family(id_priv) != AF_IB)
2451                 ps = cma_select_inet_ps(id_priv);
2452         else
2453                 ps = cma_select_ib_ps(id_priv);
2454         if (!ps)
2455                 return -EPROTONOSUPPORT;
2456
2457         mutex_lock(&lock);
2458         if (cma_any_port(cma_src_addr(id_priv)))
2459                 ret = cma_alloc_any_port(ps, id_priv);
2460         else
2461                 ret = cma_use_port(ps, id_priv);
2462         mutex_unlock(&lock);
2463
2464         return ret;
2465 }
2466
2467 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2468                                struct sockaddr *addr)
2469 {
2470 #if IS_ENABLED(CONFIG_IPV6)
2471         struct sockaddr_in6 *sin6;
2472
2473         if (addr->sa_family != AF_INET6)
2474                 return 0;
2475
2476         sin6 = (struct sockaddr_in6 *) addr;
2477         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2478             !sin6->sin6_scope_id)
2479                         return -EINVAL;
2480
2481         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2482 #endif
2483         return 0;
2484 }
2485
2486 int rdma_listen(struct rdma_cm_id *id, int backlog)
2487 {
2488         struct rdma_id_private *id_priv;
2489         int ret;
2490
2491         id_priv = container_of(id, struct rdma_id_private, id);
2492         if (id_priv->state == RDMA_CM_IDLE) {
2493                 id->route.addr.src_addr.ss_family = AF_INET;
2494                 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
2495                 if (ret)
2496                         return ret;
2497         }
2498
2499         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2500                 return -EINVAL;
2501
2502         if (id_priv->reuseaddr) {
2503                 ret = cma_bind_listen(id_priv);
2504                 if (ret)
2505                         goto err;
2506         }
2507
2508         id_priv->backlog = backlog;
2509         if (id->device) {
2510                 switch (rdma_node_get_transport(id->device->node_type)) {
2511                 case RDMA_TRANSPORT_IB:
2512                         ret = cma_ib_listen(id_priv);
2513                         if (ret)
2514                                 goto err;
2515                         break;
2516                 case RDMA_TRANSPORT_IWARP:
2517                         ret = cma_iw_listen(id_priv, backlog);
2518                         if (ret)
2519                                 goto err;
2520                         break;
2521                 default:
2522                         ret = -ENOSYS;
2523                         goto err;
2524                 }
2525         } else
2526                 cma_listen_on_all(id_priv);
2527
2528         return 0;
2529 err:
2530         id_priv->backlog = 0;
2531         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2532         return ret;
2533 }
2534 EXPORT_SYMBOL(rdma_listen);
2535
2536 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2537 {
2538         struct rdma_id_private *id_priv;
2539         int ret;
2540
2541         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
2542             addr->sa_family != AF_IB)
2543                 return -EAFNOSUPPORT;
2544
2545         id_priv = container_of(id, struct rdma_id_private, id);
2546         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2547                 return -EINVAL;
2548
2549         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2550         if (ret)
2551                 goto err1;
2552
2553         if (!cma_any_addr(addr)) {
2554                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
2555                 if (ret)
2556                         goto err1;
2557
2558                 ret = cma_acquire_dev(id_priv);
2559                 if (ret)
2560                         goto err1;
2561         }
2562
2563         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
2564         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
2565                 if (addr->sa_family == AF_INET)
2566                         id_priv->afonly = 1;
2567 #if IS_ENABLED(CONFIG_IPV6)
2568                 else if (addr->sa_family == AF_INET6)
2569                         id_priv->afonly = init_net.ipv6.sysctl.bindv6only;
2570 #endif
2571         }
2572         ret = cma_get_port(id_priv);
2573         if (ret)
2574                 goto err2;
2575
2576         return 0;
2577 err2:
2578         if (id_priv->cma_dev)
2579                 cma_release_dev(id_priv);
2580 err1:
2581         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2582         return ret;
2583 }
2584 EXPORT_SYMBOL(rdma_bind_addr);
2585
2586 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
2587 {
2588         struct cma_hdr *cma_hdr;
2589
2590         cma_hdr = hdr;
2591         cma_hdr->cma_version = CMA_VERSION;
2592         if (cma_family(id_priv) == AF_INET) {
2593                 struct sockaddr_in *src4, *dst4;
2594
2595                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
2596                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
2597
2598                 cma_set_ip_ver(cma_hdr, 4);
2599                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2600                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2601                 cma_hdr->port = src4->sin_port;
2602         } else if (cma_family(id_priv) == AF_INET6) {
2603                 struct sockaddr_in6 *src6, *dst6;
2604
2605                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2606                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
2607
2608                 cma_set_ip_ver(cma_hdr, 6);
2609                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
2610                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2611                 cma_hdr->port = src6->sin6_port;
2612         }
2613         return 0;
2614 }
2615
2616 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2617                                 struct ib_cm_event *ib_event)
2618 {
2619         struct rdma_id_private *id_priv = cm_id->context;
2620         struct rdma_cm_event event;
2621         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2622         int ret = 0;
2623
2624         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2625                 return 0;
2626
2627         memset(&event, 0, sizeof event);
2628         switch (ib_event->event) {
2629         case IB_CM_SIDR_REQ_ERROR:
2630                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2631                 event.status = -ETIMEDOUT;
2632                 break;
2633         case IB_CM_SIDR_REP_RECEIVED:
2634                 event.param.ud.private_data = ib_event->private_data;
2635                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2636                 if (rep->status != IB_SIDR_SUCCESS) {
2637                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2638                         event.status = ib_event->param.sidr_rep_rcvd.status;
2639                         break;
2640                 }
2641                 ret = cma_set_qkey(id_priv, rep->qkey);
2642                 if (ret) {
2643                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2644                         event.status = ret;
2645                         break;
2646                 }
2647                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2648                                      id_priv->id.route.path_rec,
2649                                      &event.param.ud.ah_attr);
2650                 event.param.ud.qp_num = rep->qpn;
2651                 event.param.ud.qkey = rep->qkey;
2652                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2653                 event.status = 0;
2654                 break;
2655         default:
2656                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2657                        ib_event->event);
2658                 goto out;
2659         }
2660
2661         ret = id_priv->id.event_handler(&id_priv->id, &event);
2662         if (ret) {
2663                 /* Destroy the CM ID by returning a non-zero value. */
2664                 id_priv->cm_id.ib = NULL;
2665                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2666                 mutex_unlock(&id_priv->handler_mutex);
2667                 rdma_destroy_id(&id_priv->id);
2668                 return ret;
2669         }
2670 out:
2671         mutex_unlock(&id_priv->handler_mutex);
2672         return ret;
2673 }
2674
2675 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2676                               struct rdma_conn_param *conn_param)
2677 {
2678         struct ib_cm_sidr_req_param req;
2679         struct ib_cm_id *id;
2680         void *private_data;
2681         int offset, ret;
2682
2683         memset(&req, 0, sizeof req);
2684         offset = cma_user_data_offset(id_priv);
2685         req.private_data_len = offset + conn_param->private_data_len;
2686         if (req.private_data_len < conn_param->private_data_len)
2687                 return -EINVAL;
2688
2689         if (req.private_data_len) {
2690                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2691                 if (!private_data)
2692                         return -ENOMEM;
2693         } else {
2694                 private_data = NULL;
2695         }
2696
2697         if (conn_param->private_data && conn_param->private_data_len)
2698                 memcpy(private_data + offset, conn_param->private_data,
2699                        conn_param->private_data_len);
2700
2701         if (private_data) {
2702                 ret = cma_format_hdr(private_data, id_priv);
2703                 if (ret)
2704                         goto out;
2705                 req.private_data = private_data;
2706         }
2707
2708         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2709                              id_priv);
2710         if (IS_ERR(id)) {
2711                 ret = PTR_ERR(id);
2712                 goto out;
2713         }
2714         id_priv->cm_id.ib = id;
2715
2716         req.path = id_priv->id.route.path_rec;
2717         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2718         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2719         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2720
2721         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2722         if (ret) {
2723                 ib_destroy_cm_id(id_priv->cm_id.ib);
2724                 id_priv->cm_id.ib = NULL;
2725         }
2726 out:
2727         kfree(private_data);
2728         return ret;
2729 }
2730
2731 static int cma_connect_ib(struct rdma_id_private *id_priv,
2732                           struct rdma_conn_param *conn_param)
2733 {
2734         struct ib_cm_req_param req;
2735         struct rdma_route *route;
2736         void *private_data;
2737         struct ib_cm_id *id;
2738         int offset, ret;
2739
2740         memset(&req, 0, sizeof req);
2741         offset = cma_user_data_offset(id_priv);
2742         req.private_data_len = offset + conn_param->private_data_len;
2743         if (req.private_data_len < conn_param->private_data_len)
2744                 return -EINVAL;
2745
2746         if (req.private_data_len) {
2747                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2748                 if (!private_data)
2749                         return -ENOMEM;
2750         } else {
2751                 private_data = NULL;
2752         }
2753
2754         if (conn_param->private_data && conn_param->private_data_len)
2755                 memcpy(private_data + offset, conn_param->private_data,
2756                        conn_param->private_data_len);
2757
2758         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2759         if (IS_ERR(id)) {
2760                 ret = PTR_ERR(id);
2761                 goto out;
2762         }
2763         id_priv->cm_id.ib = id;
2764
2765         route = &id_priv->id.route;
2766         if (private_data) {
2767                 ret = cma_format_hdr(private_data, id_priv);
2768                 if (ret)
2769                         goto out;
2770                 req.private_data = private_data;
2771         }
2772
2773         req.primary_path = &route->path_rec[0];
2774         if (route->num_paths == 2)
2775                 req.alternate_path = &route->path_rec[1];
2776
2777         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2778         req.qp_num = id_priv->qp_num;
2779         req.qp_type = id_priv->id.qp_type;
2780         req.starting_psn = id_priv->seq_num;
2781         req.responder_resources = conn_param->responder_resources;
2782         req.initiator_depth = conn_param->initiator_depth;
2783         req.flow_control = conn_param->flow_control;
2784         req.retry_count = min_t(u8, 7, conn_param->retry_count);
2785         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2786         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2787         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2788         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2789         req.srq = id_priv->srq ? 1 : 0;
2790
2791         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2792 out:
2793         if (ret && !IS_ERR(id)) {
2794                 ib_destroy_cm_id(id);
2795                 id_priv->cm_id.ib = NULL;
2796         }
2797
2798         kfree(private_data);
2799         return ret;
2800 }
2801
2802 static int cma_connect_iw(struct rdma_id_private *id_priv,
2803                           struct rdma_conn_param *conn_param)
2804 {
2805         struct iw_cm_id *cm_id;
2806         struct sockaddr_in* sin;
2807         int ret;
2808         struct iw_cm_conn_param iw_param;
2809
2810         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2811         if (IS_ERR(cm_id))
2812                 return PTR_ERR(cm_id);
2813
2814         id_priv->cm_id.iw = cm_id;
2815
2816         sin = (struct sockaddr_in *) cma_src_addr(id_priv);
2817         cm_id->local_addr = *sin;
2818
2819         sin = (struct sockaddr_in *) cma_dst_addr(id_priv);
2820         cm_id->remote_addr = *sin;
2821
2822         ret = cma_modify_qp_rtr(id_priv, conn_param);
2823         if (ret)
2824                 goto out;
2825
2826         if (conn_param) {
2827                 iw_param.ord = conn_param->initiator_depth;
2828                 iw_param.ird = conn_param->responder_resources;
2829                 iw_param.private_data = conn_param->private_data;
2830                 iw_param.private_data_len = conn_param->private_data_len;
2831                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2832         } else {
2833                 memset(&iw_param, 0, sizeof iw_param);
2834                 iw_param.qpn = id_priv->qp_num;
2835         }
2836         ret = iw_cm_connect(cm_id, &iw_param);
2837 out:
2838         if (ret) {
2839                 iw_destroy_cm_id(cm_id);
2840                 id_priv->cm_id.iw = NULL;
2841         }
2842         return ret;
2843 }
2844
2845 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2846 {
2847         struct rdma_id_private *id_priv;
2848         int ret;
2849
2850         id_priv = container_of(id, struct rdma_id_private, id);
2851         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2852                 return -EINVAL;
2853
2854         if (!id->qp) {
2855                 id_priv->qp_num = conn_param->qp_num;
2856                 id_priv->srq = conn_param->srq;
2857         }
2858
2859         switch (rdma_node_get_transport(id->device->node_type)) {
2860         case RDMA_TRANSPORT_IB:
2861                 if (id->qp_type == IB_QPT_UD)
2862                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2863                 else
2864                         ret = cma_connect_ib(id_priv, conn_param);
2865                 break;
2866         case RDMA_TRANSPORT_IWARP:
2867                 ret = cma_connect_iw(id_priv, conn_param);
2868                 break;
2869         default:
2870                 ret = -ENOSYS;
2871                 break;
2872         }
2873         if (ret)
2874                 goto err;
2875
2876         return 0;
2877 err:
2878         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2879         return ret;
2880 }
2881 EXPORT_SYMBOL(rdma_connect);
2882
2883 static int cma_accept_ib(struct rdma_id_private *id_priv,
2884                          struct rdma_conn_param *conn_param)
2885 {
2886         struct ib_cm_rep_param rep;
2887         int ret;
2888
2889         ret = cma_modify_qp_rtr(id_priv, conn_param);
2890         if (ret)
2891                 goto out;
2892
2893         ret = cma_modify_qp_rts(id_priv, conn_param);
2894         if (ret)
2895                 goto out;
2896
2897         memset(&rep, 0, sizeof rep);
2898         rep.qp_num = id_priv->qp_num;
2899         rep.starting_psn = id_priv->seq_num;
2900         rep.private_data = conn_param->private_data;
2901         rep.private_data_len = conn_param->private_data_len;
2902         rep.responder_resources = conn_param->responder_resources;
2903         rep.initiator_depth = conn_param->initiator_depth;
2904         rep.failover_accepted = 0;
2905         rep.flow_control = conn_param->flow_control;
2906         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2907         rep.srq = id_priv->srq ? 1 : 0;
2908
2909         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2910 out:
2911         return ret;
2912 }
2913
2914 static int cma_accept_iw(struct rdma_id_private *id_priv,
2915                   struct rdma_conn_param *conn_param)
2916 {
2917         struct iw_cm_conn_param iw_param;
2918         int ret;
2919
2920         ret = cma_modify_qp_rtr(id_priv, conn_param);
2921         if (ret)
2922                 return ret;
2923
2924         iw_param.ord = conn_param->initiator_depth;
2925         iw_param.ird = conn_param->responder_resources;
2926         iw_param.private_data = conn_param->private_data;
2927         iw_param.private_data_len = conn_param->private_data_len;
2928         if (id_priv->id.qp) {
2929                 iw_param.qpn = id_priv->qp_num;
2930         } else
2931                 iw_param.qpn = conn_param->qp_num;
2932
2933         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2934 }
2935
2936 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2937                              enum ib_cm_sidr_status status, u32 qkey,
2938                              const void *private_data, int private_data_len)
2939 {
2940         struct ib_cm_sidr_rep_param rep;
2941         int ret;
2942
2943         memset(&rep, 0, sizeof rep);
2944         rep.status = status;
2945         if (status == IB_SIDR_SUCCESS) {
2946                 ret = cma_set_qkey(id_priv, qkey);
2947                 if (ret)
2948                         return ret;
2949                 rep.qp_num = id_priv->qp_num;
2950                 rep.qkey = id_priv->qkey;
2951         }
2952         rep.private_data = private_data;
2953         rep.private_data_len = private_data_len;
2954
2955         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2956 }
2957
2958 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2959 {
2960         struct rdma_id_private *id_priv;
2961         int ret;
2962
2963         id_priv = container_of(id, struct rdma_id_private, id);
2964
2965         id_priv->owner = task_pid_nr(current);
2966
2967         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2968                 return -EINVAL;
2969
2970         if (!id->qp && conn_param) {
2971                 id_priv->qp_num = conn_param->qp_num;
2972                 id_priv->srq = conn_param->srq;
2973         }
2974
2975         switch (rdma_node_get_transport(id->device->node_type)) {
2976         case RDMA_TRANSPORT_IB:
2977                 if (id->qp_type == IB_QPT_UD) {
2978                         if (conn_param)
2979                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2980                                                         conn_param->qkey,
2981                                                         conn_param->private_data,
2982                                                         conn_param->private_data_len);
2983                         else
2984                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2985                                                         0, NULL, 0);
2986                 } else {
2987                         if (conn_param)
2988                                 ret = cma_accept_ib(id_priv, conn_param);
2989                         else
2990                                 ret = cma_rep_recv(id_priv);
2991                 }
2992                 break;
2993         case RDMA_TRANSPORT_IWARP:
2994                 ret = cma_accept_iw(id_priv, conn_param);
2995                 break;
2996         default:
2997                 ret = -ENOSYS;
2998                 break;
2999         }
3000
3001         if (ret)
3002                 goto reject;
3003
3004         return 0;
3005 reject:
3006         cma_modify_qp_err(id_priv);
3007         rdma_reject(id, NULL, 0);
3008         return ret;
3009 }
3010 EXPORT_SYMBOL(rdma_accept);
3011
3012 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3013 {
3014         struct rdma_id_private *id_priv;
3015         int ret;
3016
3017         id_priv = container_of(id, struct rdma_id_private, id);
3018         if (!id_priv->cm_id.ib)
3019                 return -EINVAL;
3020
3021         switch (id->device->node_type) {
3022         case RDMA_NODE_IB_CA:
3023                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3024                 break;
3025         default:
3026                 ret = 0;
3027                 break;
3028         }
3029         return ret;
3030 }
3031 EXPORT_SYMBOL(rdma_notify);
3032
3033 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3034                 u8 private_data_len)
3035 {
3036         struct rdma_id_private *id_priv;
3037         int ret;
3038
3039         id_priv = container_of(id, struct rdma_id_private, id);
3040         if (!id_priv->cm_id.ib)
3041                 return -EINVAL;
3042
3043         switch (rdma_node_get_transport(id->device->node_type)) {
3044         case RDMA_TRANSPORT_IB:
3045                 if (id->qp_type == IB_QPT_UD)
3046                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3047                                                 private_data, private_data_len);
3048                 else
3049                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
3050                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
3051                                              0, private_data, private_data_len);
3052                 break;
3053         case RDMA_TRANSPORT_IWARP:
3054                 ret = iw_cm_reject(id_priv->cm_id.iw,
3055                                    private_data, private_data_len);
3056                 break;
3057         default:
3058                 ret = -ENOSYS;
3059                 break;
3060         }
3061         return ret;
3062 }
3063 EXPORT_SYMBOL(rdma_reject);
3064
3065 int rdma_disconnect(struct rdma_cm_id *id)
3066 {
3067         struct rdma_id_private *id_priv;
3068         int ret;
3069
3070         id_priv = container_of(id, struct rdma_id_private, id);
3071         if (!id_priv->cm_id.ib)
3072                 return -EINVAL;
3073
3074         switch (rdma_node_get_transport(id->device->node_type)) {
3075         case RDMA_TRANSPORT_IB:
3076                 ret = cma_modify_qp_err(id_priv);
3077                 if (ret)
3078                         goto out;
3079                 /* Initiate or respond to a disconnect. */
3080                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3081                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3082                 break;
3083         case RDMA_TRANSPORT_IWARP:
3084                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3085                 break;
3086         default:
3087                 ret = -EINVAL;
3088                 break;
3089         }
3090 out:
3091         return ret;
3092 }
3093 EXPORT_SYMBOL(rdma_disconnect);
3094
3095 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3096 {
3097         struct rdma_id_private *id_priv;
3098         struct cma_multicast *mc = multicast->context;
3099         struct rdma_cm_event event;
3100         int ret;
3101
3102         id_priv = mc->id_priv;
3103         if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
3104             cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
3105                 return 0;
3106
3107         if (!status)
3108                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3109         mutex_lock(&id_priv->qp_mutex);
3110         if (!status && id_priv->id.qp)
3111                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3112                                          be16_to_cpu(multicast->rec.mlid));
3113         mutex_unlock(&id_priv->qp_mutex);
3114
3115         memset(&event, 0, sizeof event);
3116         event.status = status;
3117         event.param.ud.private_data = mc->context;
3118         if (!status) {
3119                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3120                 ib_init_ah_from_mcmember(id_priv->id.device,
3121                                          id_priv->id.port_num, &multicast->rec,
3122                                          &event.param.ud.ah_attr);
3123                 event.param.ud.qp_num = 0xFFFFFF;
3124                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3125         } else
3126                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3127
3128         ret = id_priv->id.event_handler(&id_priv->id, &event);
3129         if (ret) {
3130                 cma_exch(id_priv, RDMA_CM_DESTROYING);
3131                 mutex_unlock(&id_priv->handler_mutex);
3132                 rdma_destroy_id(&id_priv->id);
3133                 return 0;
3134         }
3135
3136         mutex_unlock(&id_priv->handler_mutex);
3137         return 0;
3138 }
3139
3140 static void cma_set_mgid(struct rdma_id_private *id_priv,
3141                          struct sockaddr *addr, union ib_gid *mgid)
3142 {
3143         unsigned char mc_map[MAX_ADDR_LEN];
3144         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3145         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3146         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3147
3148         if (cma_any_addr(addr)) {
3149                 memset(mgid, 0, sizeof *mgid);
3150         } else if ((addr->sa_family == AF_INET6) &&
3151                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3152                                                                  0xFF10A01B)) {
3153                 /* IPv6 address is an SA assigned MGID. */
3154                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3155         } else if (addr->sa_family == AF_IB) {
3156                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3157         } else if ((addr->sa_family == AF_INET6)) {
3158                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3159                 if (id_priv->id.ps == RDMA_PS_UDP)
3160                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3161                 *mgid = *(union ib_gid *) (mc_map + 4);
3162         } else {
3163                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3164                 if (id_priv->id.ps == RDMA_PS_UDP)
3165                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3166                 *mgid = *(union ib_gid *) (mc_map + 4);
3167         }
3168 }
3169
3170 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3171                                  struct cma_multicast *mc)
3172 {
3173         struct ib_sa_mcmember_rec rec;
3174         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3175         ib_sa_comp_mask comp_mask;
3176         int ret;
3177
3178         ib_addr_get_mgid(dev_addr, &rec.mgid);
3179         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3180                                      &rec.mgid, &rec);
3181         if (ret)
3182                 return ret;
3183
3184         ret = cma_set_qkey(id_priv, 0);
3185         if (ret)
3186                 return ret;
3187
3188         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3189         rec.qkey = cpu_to_be32(id_priv->qkey);
3190         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3191         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3192         rec.join_state = 1;
3193
3194         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3195                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3196                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3197                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
3198                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3199
3200         if (id_priv->id.ps == RDMA_PS_IPOIB)
3201                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3202                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3203                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3204                              IB_SA_MCMEMBER_REC_MTU |
3205                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
3206
3207         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3208                                                 id_priv->id.port_num, &rec,
3209                                                 comp_mask, GFP_KERNEL,
3210                                                 cma_ib_mc_handler, mc);
3211         return PTR_RET(mc->multicast.ib);
3212 }
3213
3214 static void iboe_mcast_work_handler(struct work_struct *work)
3215 {
3216         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3217         struct cma_multicast *mc = mw->mc;
3218         struct ib_sa_multicast *m = mc->multicast.ib;
3219
3220         mc->multicast.ib->context = mc;
3221         cma_ib_mc_handler(0, m);
3222         kref_put(&mc->mcref, release_mc);
3223         kfree(mw);
3224 }
3225
3226 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3227 {
3228         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3229         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3230
3231         if (cma_any_addr(addr)) {
3232                 memset(mgid, 0, sizeof *mgid);
3233         } else if (addr->sa_family == AF_INET6) {
3234                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3235         } else {
3236                 mgid->raw[0] = 0xff;
3237                 mgid->raw[1] = 0x0e;
3238                 mgid->raw[2] = 0;
3239                 mgid->raw[3] = 0;
3240                 mgid->raw[4] = 0;
3241                 mgid->raw[5] = 0;
3242                 mgid->raw[6] = 0;
3243                 mgid->raw[7] = 0;
3244                 mgid->raw[8] = 0;
3245                 mgid->raw[9] = 0;
3246                 mgid->raw[10] = 0xff;
3247                 mgid->raw[11] = 0xff;
3248                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3249         }
3250 }
3251
3252 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3253                                    struct cma_multicast *mc)
3254 {
3255         struct iboe_mcast_work *work;
3256         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3257         int err;
3258         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3259         struct net_device *ndev = NULL;
3260
3261         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3262                 return -EINVAL;
3263
3264         work = kzalloc(sizeof *work, GFP_KERNEL);
3265         if (!work)
3266                 return -ENOMEM;
3267
3268         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3269         if (!mc->multicast.ib) {
3270                 err = -ENOMEM;
3271                 goto out1;
3272         }
3273
3274         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3275
3276         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3277         if (id_priv->id.ps == RDMA_PS_UDP)
3278                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3279
3280         if (dev_addr->bound_dev_if)
3281                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3282         if (!ndev) {
3283                 err = -ENODEV;
3284                 goto out2;
3285         }
3286         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3287         mc->multicast.ib->rec.hop_limit = 1;
3288         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3289         dev_put(ndev);
3290         if (!mc->multicast.ib->rec.mtu) {
3291                 err = -EINVAL;
3292                 goto out2;
3293         }
3294         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3295         work->id = id_priv;
3296         work->mc = mc;
3297         INIT_WORK(&work->work, iboe_mcast_work_handler);
3298         kref_get(&mc->mcref);
3299         queue_work(cma_wq, &work->work);
3300
3301         return 0;
3302
3303 out2:
3304         kfree(mc->multicast.ib);
3305 out1:
3306         kfree(work);
3307         return err;
3308 }
3309
3310 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3311                         void *context)
3312 {
3313         struct rdma_id_private *id_priv;
3314         struct cma_multicast *mc;
3315         int ret;
3316
3317         id_priv = container_of(id, struct rdma_id_private, id);
3318         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3319             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3320                 return -EINVAL;
3321
3322         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3323         if (!mc)
3324                 return -ENOMEM;
3325
3326         memcpy(&mc->addr, addr, rdma_addr_size(addr));
3327         mc->context = context;
3328         mc->id_priv = id_priv;
3329
3330         spin_lock(&id_priv->lock);
3331         list_add(&mc->list, &id_priv->mc_list);
3332         spin_unlock(&id_priv->lock);
3333
3334         switch (rdma_node_get_transport(id->device->node_type)) {
3335         case RDMA_TRANSPORT_IB:
3336                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3337                 case IB_LINK_LAYER_INFINIBAND:
3338                         ret = cma_join_ib_multicast(id_priv, mc);
3339                         break;
3340                 case IB_LINK_LAYER_ETHERNET:
3341                         kref_init(&mc->mcref);
3342                         ret = cma_iboe_join_multicast(id_priv, mc);
3343                         break;
3344                 default:
3345                         ret = -EINVAL;
3346                 }
3347                 break;
3348         default:
3349                 ret = -ENOSYS;
3350                 break;
3351         }
3352
3353         if (ret) {
3354                 spin_lock_irq(&id_priv->lock);
3355                 list_del(&mc->list);
3356                 spin_unlock_irq(&id_priv->lock);
3357                 kfree(mc);
3358         }
3359         return ret;
3360 }
3361 EXPORT_SYMBOL(rdma_join_multicast);
3362
3363 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3364 {
3365         struct rdma_id_private *id_priv;
3366         struct cma_multicast *mc;
3367
3368         id_priv = container_of(id, struct rdma_id_private, id);
3369         spin_lock_irq(&id_priv->lock);
3370         list_for_each_entry(mc, &id_priv->mc_list, list) {
3371                 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
3372                         list_del(&mc->list);
3373                         spin_unlock_irq(&id_priv->lock);
3374
3375                         if (id->qp)
3376                                 ib_detach_mcast(id->qp,
3377                                                 &mc->multicast.ib->rec.mgid,
3378                                                 be16_to_cpu(mc->multicast.ib->rec.mlid));
3379                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3380                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3381                                 case IB_LINK_LAYER_INFINIBAND:
3382                                         ib_sa_free_multicast(mc->multicast.ib);
3383                                         kfree(mc);
3384                                         break;
3385                                 case IB_LINK_LAYER_ETHERNET:
3386                                         kref_put(&mc->mcref, release_mc);
3387                                         break;
3388                                 default:
3389                                         break;
3390                                 }
3391                         }
3392                         return;
3393                 }
3394         }
3395         spin_unlock_irq(&id_priv->lock);
3396 }
3397 EXPORT_SYMBOL(rdma_leave_multicast);
3398
3399 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3400 {
3401         struct rdma_dev_addr *dev_addr;
3402         struct cma_ndev_work *work;
3403
3404         dev_addr = &id_priv->id.route.addr.dev_addr;
3405
3406         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3407             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3408                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3409                        ndev->name, &id_priv->id);
3410                 work = kzalloc(sizeof *work, GFP_KERNEL);
3411                 if (!work)
3412                         return -ENOMEM;
3413
3414                 INIT_WORK(&work->work, cma_ndev_work_handler);
3415                 work->id = id_priv;
3416                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3417                 atomic_inc(&id_priv->refcount);
3418                 queue_work(cma_wq, &work->work);
3419         }
3420
3421         return 0;
3422 }
3423
3424 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3425                                void *ptr)
3426 {
3427         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
3428         struct cma_device *cma_dev;
3429         struct rdma_id_private *id_priv;
3430         int ret = NOTIFY_DONE;
3431
3432         if (dev_net(ndev) != &init_net)
3433                 return NOTIFY_DONE;
3434
3435         if (event != NETDEV_BONDING_FAILOVER)
3436                 return NOTIFY_DONE;
3437
3438         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3439                 return NOTIFY_DONE;
3440
3441         mutex_lock(&lock);
3442         list_for_each_entry(cma_dev, &dev_list, list)
3443                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3444                         ret = cma_netdev_change(ndev, id_priv);
3445                         if (ret)
3446                                 goto out;
3447                 }
3448
3449 out:
3450         mutex_unlock(&lock);
3451         return ret;
3452 }
3453
3454 static struct notifier_block cma_nb = {
3455         .notifier_call = cma_netdev_callback
3456 };
3457
3458 static void cma_add_one(struct ib_device *device)
3459 {
3460         struct cma_device *cma_dev;
3461         struct rdma_id_private *id_priv;
3462
3463         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3464         if (!cma_dev)
3465                 return;
3466
3467         cma_dev->device = device;
3468
3469         init_completion(&cma_dev->comp);
3470         atomic_set(&cma_dev->refcount, 1);
3471         INIT_LIST_HEAD(&cma_dev->id_list);
3472         ib_set_client_data(device, &cma_client, cma_dev);
3473
3474         mutex_lock(&lock);
3475         list_add_tail(&cma_dev->list, &dev_list);
3476         list_for_each_entry(id_priv, &listen_any_list, list)
3477                 cma_listen_on_dev(id_priv, cma_dev);
3478         mutex_unlock(&lock);
3479 }
3480
3481 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3482 {
3483         struct rdma_cm_event event;
3484         enum rdma_cm_state state;
3485         int ret = 0;
3486
3487         /* Record that we want to remove the device */
3488         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3489         if (state == RDMA_CM_DESTROYING)
3490                 return 0;
3491
3492         cma_cancel_operation(id_priv, state);
3493         mutex_lock(&id_priv->handler_mutex);
3494
3495         /* Check for destruction from another callback. */
3496         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3497                 goto out;
3498
3499         memset(&event, 0, sizeof event);
3500         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3501         ret = id_priv->id.event_handler(&id_priv->id, &event);
3502 out:
3503         mutex_unlock(&id_priv->handler_mutex);
3504         return ret;
3505 }
3506
3507 static void cma_process_remove(struct cma_device *cma_dev)
3508 {
3509         struct rdma_id_private *id_priv;
3510         int ret;
3511
3512         mutex_lock(&lock);
3513         while (!list_empty(&cma_dev->id_list)) {
3514                 id_priv = list_entry(cma_dev->id_list.next,
3515                                      struct rdma_id_private, list);
3516
3517                 list_del(&id_priv->listen_list);
3518                 list_del_init(&id_priv->list);
3519                 atomic_inc(&id_priv->refcount);
3520                 mutex_unlock(&lock);
3521
3522                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3523                 cma_deref_id(id_priv);
3524                 if (ret)
3525                         rdma_destroy_id(&id_priv->id);
3526
3527                 mutex_lock(&lock);
3528         }
3529         mutex_unlock(&lock);
3530
3531         cma_deref_dev(cma_dev);
3532         wait_for_completion(&cma_dev->comp);
3533 }
3534
3535 static void cma_remove_one(struct ib_device *device)
3536 {
3537         struct cma_device *cma_dev;
3538
3539         cma_dev = ib_get_client_data(device, &cma_client);
3540         if (!cma_dev)
3541                 return;
3542
3543         mutex_lock(&lock);
3544         list_del(&cma_dev->list);
3545         mutex_unlock(&lock);
3546
3547         cma_process_remove(cma_dev);
3548         kfree(cma_dev);
3549 }
3550
3551 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3552 {
3553         struct nlmsghdr *nlh;
3554         struct rdma_cm_id_stats *id_stats;
3555         struct rdma_id_private *id_priv;
3556         struct rdma_cm_id *id = NULL;
3557         struct cma_device *cma_dev;
3558         int i_dev = 0, i_id = 0;
3559
3560         /*
3561          * We export all of the IDs as a sequence of messages.  Each
3562          * ID gets its own netlink message.
3563          */
3564         mutex_lock(&lock);
3565
3566         list_for_each_entry(cma_dev, &dev_list, list) {
3567                 if (i_dev < cb->args[0]) {
3568                         i_dev++;
3569                         continue;
3570                 }
3571
3572                 i_id = 0;
3573                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3574                         if (i_id < cb->args[1]) {
3575                                 i_id++;
3576                                 continue;
3577                         }
3578
3579                         id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3580                                                 sizeof *id_stats, RDMA_NL_RDMA_CM,
3581                                                 RDMA_NL_RDMA_CM_ID_STATS);
3582                         if (!id_stats)
3583                                 goto out;
3584
3585                         memset(id_stats, 0, sizeof *id_stats);
3586                         id = &id_priv->id;
3587                         id_stats->node_type = id->route.addr.dev_addr.dev_type;
3588                         id_stats->port_num = id->port_num;
3589                         id_stats->bound_dev_if =
3590                                 id->route.addr.dev_addr.bound_dev_if;
3591
3592                         if (ibnl_put_attr(skb, nlh,
3593                                           rdma_addr_size(cma_src_addr(id_priv)),
3594                                           cma_src_addr(id_priv),
3595                                           RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
3596                                 goto out;
3597                         if (ibnl_put_attr(skb, nlh,
3598                                           rdma_addr_size(cma_src_addr(id_priv)),
3599                                           cma_dst_addr(id_priv),
3600                                           RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
3601                                 goto out;
3602
3603                         id_stats->pid           = id_priv->owner;
3604                         id_stats->port_space    = id->ps;
3605                         id_stats->cm_state      = id_priv->state;
3606                         id_stats->qp_num        = id_priv->qp_num;
3607                         id_stats->qp_type       = id->qp_type;
3608
3609                         i_id++;
3610                 }
3611
3612                 cb->args[1] = 0;
3613                 i_dev++;
3614         }
3615
3616 out:
3617         mutex_unlock(&lock);
3618         cb->args[0] = i_dev;
3619         cb->args[1] = i_id;
3620
3621         return skb->len;
3622 }
3623
3624 static const struct ibnl_client_cbs cma_cb_table[] = {
3625         [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
3626                                        .module = THIS_MODULE },
3627 };
3628
3629 static int __init cma_init(void)
3630 {
3631         int ret;
3632
3633         cma_wq = create_singlethread_workqueue("rdma_cm");
3634         if (!cma_wq)
3635                 return -ENOMEM;
3636
3637         ib_sa_register_client(&sa_client);
3638         rdma_addr_register_client(&addr_client);
3639         register_netdevice_notifier(&cma_nb);
3640
3641         ret = ib_register_client(&cma_client);
3642         if (ret)
3643                 goto err;
3644
3645         if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3646                 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3647
3648         return 0;
3649
3650 err:
3651         unregister_netdevice_notifier(&cma_nb);
3652         rdma_addr_unregister_client(&addr_client);
3653         ib_sa_unregister_client(&sa_client);
3654         destroy_workqueue(cma_wq);
3655         return ret;
3656 }
3657
3658 static void __exit cma_cleanup(void)
3659 {
3660         ibnl_remove_client(RDMA_NL_RDMA_CM);
3661         ib_unregister_client(&cma_client);
3662         unregister_netdevice_notifier(&cma_nb);
3663         rdma_addr_unregister_client(&addr_client);
3664         ib_sa_unregister_client(&sa_client);
3665         destroy_workqueue(cma_wq);
3666         idr_destroy(&tcp_ps);
3667         idr_destroy(&udp_ps);
3668         idr_destroy(&ipoib_ps);
3669         idr_destroy(&ib_ps);
3670 }
3671
3672 module_init(cma_init);
3673 module_exit(cma_cleanup);