drivers/infiniband/core/iwcm.c

   1 /*
   2  * Copyright (c) 2004, 2005 Intel Corporation.  All rights reserved.
   3  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
   4  * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
   5  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   6  * Copyright (c) 2005 Open Grid Computing, Inc. All rights reserved.
   7  * Copyright (c) 2005 Network Appliance, Inc. All rights reserved.
   8  *
   9  * This software is available to you under a choice of one of two
  10  * licenses.  You may choose to be licensed under the terms of the GNU
  11  * General Public License (GPL) Version 2, available from the file
  12  * COPYING in the main directory of this source tree, or the
  13  * OpenIB.org BSD license below:
  14  *
  15  *     Redistribution and use in source and binary forms, with or
  16  *     without modification, are permitted provided that the following
  17  *     conditions are met:
  18  *
  19  *      - Redistributions of source code must retain the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer.
  22  *
  23  *      - Redistributions in binary form must reproduce the above
  24  *        copyright notice, this list of conditions and the following
  25  *        disclaimer in the documentation and/or other materials
  26  *        provided with the distribution.
  27  *
  28  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  29  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  30  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  31  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  32  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  33  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  34  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  35  * SOFTWARE.
  36  *
  37  */
  38 #include <linux/dma-mapping.h>
  39 #include <linux/err.h>
  40 #include <linux/idr.h>
  41 #include <linux/interrupt.h>
  42 #include <linux/rbtree.h>
  43 #include <linux/sched.h>
  44 #include <linux/spinlock.h>
  45 #include <linux/workqueue.h>
  46 #include <linux/completion.h>
  47 #include <linux/slab.h>
  48 #include <linux/module.h>
  49 #include <linux/sysctl.h>
  50
  51 #include <rdma/iw_cm.h>
  52 #include <rdma/ib_addr.h>
  53 #include <rdma/iw_portmap.h>
  54 #include <rdma/rdma_netlink.h>
  55
  56 #include "iwcm.h"
  57
  58 MODULE_AUTHOR("Tom Tucker");
  59 MODULE_DESCRIPTION("iWARP CM");
  60 MODULE_LICENSE("Dual BSD/GPL");
  61
  62 static const char * const iwcm_rej_reason_strs[] = {
  63         [ECONNRESET]                    = "reset by remote host",
  64         [ECONNREFUSED]                  = "refused by remote application",
  65         [ETIMEDOUT]                     = "setup timeout",
  66 };
  67
  68 const char *__attribute_const__ iwcm_reject_msg(int reason)
  69 {
  70         size_t index;
  71
  72         /* iWARP uses negative errnos */
  73         index = -reason;
  74
  75         if (index < ARRAY_SIZE(iwcm_rej_reason_strs) &&
  76             iwcm_rej_reason_strs[index])
  77                 return iwcm_rej_reason_strs[index];
  78         else
  79                 return "unrecognized reason";
  80 }
  81 EXPORT_SYMBOL(iwcm_reject_msg);
  82
  83 static struct rdma_nl_cbs iwcm_nl_cb_table[RDMA_NL_IWPM_NUM_OPS] = {
  84         [RDMA_NL_IWPM_REG_PID] = {.dump = iwpm_register_pid_cb},
  85         [RDMA_NL_IWPM_ADD_MAPPING] = {.dump = iwpm_add_mapping_cb},
  86         [RDMA_NL_IWPM_QUERY_MAPPING] = {.dump = iwpm_add_and_query_mapping_cb},
  87         [RDMA_NL_IWPM_REMOTE_INFO] = {.dump = iwpm_remote_info_cb},
  88         [RDMA_NL_IWPM_HANDLE_ERR] = {.dump = iwpm_mapping_error_cb},
  89         [RDMA_NL_IWPM_MAPINFO] = {.dump = iwpm_mapping_info_cb},
  90         [RDMA_NL_IWPM_MAPINFO_NUM] = {.dump = iwpm_ack_mapping_info_cb}
  91 };
  92
  93 static struct workqueue_struct *iwcm_wq;
  94 struct iwcm_work {
  95         struct work_struct work;
  96         struct iwcm_id_private *cm_id;
  97         struct list_head list;
  98         struct iw_cm_event event;
  99         struct list_head free_list;
 100 };
 101
 102 static unsigned int default_backlog = 256;
 103
 104 static struct ctl_table_header *iwcm_ctl_table_hdr;
 105 static struct ctl_table iwcm_ctl_table[] = {
 106         {
 107                 .procname       = "default_backlog",
 108                 .data           = &default_backlog,
 109                 .maxlen         = sizeof(default_backlog),
 110                 .mode           = 0644,
 111                 .proc_handler   = proc_dointvec,
 112         },
 113         { }
 114 };
 115
 116 /*
 117  * The following services provide a mechanism for pre-allocating iwcm_work
 118  * elements.  The design pre-allocates them  based on the cm_id type:
 119  *      LISTENING IDS:  Get enough elements preallocated to handle the
 120  *                      listen backlog.
 121  *      ACTIVE IDS:     4: CONNECT_REPLY, ESTABLISHED, DISCONNECT, CLOSE
 122  *      PASSIVE IDS:    3: ESTABLISHED, DISCONNECT, CLOSE
 123  *
 124  * Allocating them in connect and listen avoids having to deal
 125  * with allocation failures on the event upcall from the provider (which
 126  * is called in the interrupt context).
 127  *
 128  * One exception is when creating the cm_id for incoming connection requests.
 129  * There are two cases:
 130  * 1) in the event upcall, cm_event_handler(), for a listening cm_id.  If
 131  *    the backlog is exceeded, then no more connection request events will
 132  *    be processed.  cm_event_handler() returns -ENOMEM in this case.  Its up
 133  *    to the provider to reject the connection request.
 134  * 2) in the connection request workqueue handler, cm_conn_req_handler().
 135  *    If work elements cannot be allocated for the new connect request cm_id,
 136  *    then IWCM will call the provider reject method.  This is ok since
 137  *    cm_conn_req_handler() runs in the workqueue thread context.
 138  */
 139
 140 static struct iwcm_work *get_work(struct iwcm_id_private *cm_id_priv)
 141 {
 142         struct iwcm_work *work;
 143
 144         if (list_empty(&cm_id_priv->work_free_list))
 145                 return NULL;
 146         work = list_entry(cm_id_priv->work_free_list.next, struct iwcm_work,
 147                           free_list);
 148         list_del_init(&work->free_list);
 149         return work;
 150 }
 151
 152 static void put_work(struct iwcm_work *work)
 153 {
 154         list_add(&work->free_list, &work->cm_id->work_free_list);
 155 }
 156
 157 static void dealloc_work_entries(struct iwcm_id_private *cm_id_priv)
 158 {
 159         struct list_head *e, *tmp;
 160
 161         list_for_each_safe(e, tmp, &cm_id_priv->work_free_list) {
 162                 list_del(e);
 163                 kfree(list_entry(e, struct iwcm_work, free_list));
 164         }
 165 }
 166
 167 static int alloc_work_entries(struct iwcm_id_private *cm_id_priv, int count)
 168 {
 169         struct iwcm_work *work;
 170
 171         BUG_ON(!list_empty(&cm_id_priv->work_free_list));
 172         while (count--) {
 173                 work = kmalloc(sizeof(struct iwcm_work), GFP_KERNEL);
 174                 if (!work) {
 175                         dealloc_work_entries(cm_id_priv);
 176                         return -ENOMEM;
 177                 }
 178                 work->cm_id = cm_id_priv;
 179                 INIT_LIST_HEAD(&work->list);
 180                 put_work(work);
 181         }
 182         return 0;
 183 }
 184
 185 /*
 186  * Save private data from incoming connection requests to
 187  * iw_cm_event, so the low level driver doesn't have to. Adjust
 188  * the event ptr to point to the local copy.
 189  */
 190 static int copy_private_data(struct iw_cm_event *event)
 191 {
 192         void *p;
 193
 194         p = kmemdup(event->private_data, event->private_data_len, GFP_ATOMIC);
 195         if (!p)
 196                 return -ENOMEM;
 197         event->private_data = p;
 198         return 0;
 199 }
 200
 201 static void free_cm_id(struct iwcm_id_private *cm_id_priv)
 202 {
 203         dealloc_work_entries(cm_id_priv);
 204         kfree(cm_id_priv);
 205 }
 206
 207 /*
 208  * Release a reference on cm_id. If the last reference is being
 209  * released, free the cm_id and return 1.
 210  */
 211 static int iwcm_deref_id(struct iwcm_id_private *cm_id_priv)
 212 {
 213         BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
 214         if (atomic_dec_and_test(&cm_id_priv->refcount)) {
 215                 BUG_ON(!list_empty(&cm_id_priv->work_list));
 216                 free_cm_id(cm_id_priv);
 217                 return 1;
 218         }
 219
 220         return 0;
 221 }
 222
 223 static void add_ref(struct iw_cm_id *cm_id)
 224 {
 225         struct iwcm_id_private *cm_id_priv;
 226         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 227         atomic_inc(&cm_id_priv->refcount);
 228 }
 229
 230 static void rem_ref(struct iw_cm_id *cm_id)
 231 {
 232         struct iwcm_id_private *cm_id_priv;
 233
 234         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 235
 236         (void)iwcm_deref_id(cm_id_priv);
 237 }
 238
 239 static int cm_event_handler(struct iw_cm_id *cm_id, struct iw_cm_event *event);
 240
 241 struct iw_cm_id *iw_create_cm_id(struct ib_device *device,
 242                                  iw_cm_handler cm_handler,
 243                                  void *context)
 244 {
 245         struct iwcm_id_private *cm_id_priv;
 246
 247         cm_id_priv = kzalloc(sizeof(*cm_id_priv), GFP_KERNEL);
 248         if (!cm_id_priv)
 249                 return ERR_PTR(-ENOMEM);
 250
 251         cm_id_priv->state = IW_CM_STATE_IDLE;
 252         cm_id_priv->id.device = device;
 253         cm_id_priv->id.cm_handler = cm_handler;
 254         cm_id_priv->id.context = context;
 255         cm_id_priv->id.event_handler = cm_event_handler;
 256         cm_id_priv->id.add_ref = add_ref;
 257         cm_id_priv->id.rem_ref = rem_ref;
 258         spin_lock_init(&cm_id_priv->lock);
 259         atomic_set(&cm_id_priv->refcount, 1);
 260         init_waitqueue_head(&cm_id_priv->connect_wait);
 261         init_completion(&cm_id_priv->destroy_comp);
 262         INIT_LIST_HEAD(&cm_id_priv->work_list);
 263         INIT_LIST_HEAD(&cm_id_priv->work_free_list);
 264
 265         return &cm_id_priv->id;
 266 }
 267 EXPORT_SYMBOL(iw_create_cm_id);
 268
 269
 270 static int iwcm_modify_qp_err(struct ib_qp *qp)
 271 {
 272         struct ib_qp_attr qp_attr;
 273
 274         if (!qp)
 275                 return -EINVAL;
 276
 277         qp_attr.qp_state = IB_QPS_ERR;
 278         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 279 }
 280
 281 /*
 282  * This is really the RDMAC CLOSING state. It is most similar to the
 283  * IB SQD QP state.
 284  */
 285 static int iwcm_modify_qp_sqd(struct ib_qp *qp)
 286 {
 287         struct ib_qp_attr qp_attr;
 288
 289         BUG_ON(qp == NULL);
 290         qp_attr.qp_state = IB_QPS_SQD;
 291         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 292 }
 293
 294 /*
 295  * CM_ID <-- CLOSING
 296  *
 297  * Block if a passive or active connection is currently being processed. Then
 298  * process the event as follows:
 299  * - If we are ESTABLISHED, move to CLOSING and modify the QP state
 300  *   based on the abrupt flag
 301  * - If the connection is already in the CLOSING or IDLE state, the peer is
 302  *   disconnecting concurrently with us and we've already seen the
 303  *   DISCONNECT event -- ignore the request and return 0
 304  * - Disconnect on a listening endpoint returns -EINVAL
 305  */
 306 int iw_cm_disconnect(struct iw_cm_id *cm_id, int abrupt)
 307 {
 308         struct iwcm_id_private *cm_id_priv;
 309         unsigned long flags;
 310         int ret = 0;
 311         struct ib_qp *qp = NULL;
 312
 313         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 314         /* Wait if we're currently in a connect or accept downcall */
 315         wait_event(cm_id_priv->connect_wait,
 316                    !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
 317
 318         spin_lock_irqsave(&cm_id_priv->lock, flags);
 319         switch (cm_id_priv->state) {
 320         case IW_CM_STATE_ESTABLISHED:
 321                 cm_id_priv->state = IW_CM_STATE_CLOSING;
 322
 323                 /* QP could be <nul> for user-mode client */
 324                 if (cm_id_priv->qp)
 325                         qp = cm_id_priv->qp;
 326                 else
 327                         ret = -EINVAL;
 328                 break;
 329         case IW_CM_STATE_LISTEN:
 330                 ret = -EINVAL;
 331                 break;
 332         case IW_CM_STATE_CLOSING:
 333                 /* remote peer closed first */
 334         case IW_CM_STATE_IDLE:
 335                 /* accept or connect returned !0 */
 336                 break;
 337         case IW_CM_STATE_CONN_RECV:
 338                 /*
 339                  * App called disconnect before/without calling accept after
 340                  * connect_request event delivered.
 341                  */
 342                 break;
 343         case IW_CM_STATE_CONN_SENT:
 344                 /* Can only get here if wait above fails */
 345         default:
 346                 BUG();
 347         }
 348         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 349
 350         if (qp) {
 351                 if (abrupt)
 352                         ret = iwcm_modify_qp_err(qp);
 353                 else
 354                         ret = iwcm_modify_qp_sqd(qp);
 355
 356                 /*
 357                  * If both sides are disconnecting the QP could
 358                  * already be in ERR or SQD states
 359                  */
 360                 ret = 0;
 361         }
 362
 363         return ret;
 364 }
 365 EXPORT_SYMBOL(iw_cm_disconnect);
 366
 367 /*
 368  * CM_ID <-- DESTROYING
 369  *
 370  * Clean up all resources associated with the connection and release
 371  * the initial reference taken by iw_create_cm_id.
 372  */
 373 static void destroy_cm_id(struct iw_cm_id *cm_id)
 374 {
 375         struct iwcm_id_private *cm_id_priv;
 376         unsigned long flags;
 377
 378         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 379         /*
 380          * Wait if we're currently in a connect or accept downcall. A
 381          * listening endpoint should never block here.
 382          */
 383         wait_event(cm_id_priv->connect_wait,
 384                    !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
 385
 386         /*
 387          * Since we're deleting the cm_id, drop any events that
 388          * might arrive before the last dereference.
 389          */
 390         set_bit(IWCM_F_DROP_EVENTS, &cm_id_priv->flags);
 391
 392         spin_lock_irqsave(&cm_id_priv->lock, flags);
 393         switch (cm_id_priv->state) {
 394         case IW_CM_STATE_LISTEN:
 395                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 396                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 397                 /* destroy the listening endpoint */
 398                 cm_id->device->iwcm->destroy_listen(cm_id);
 399                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 400                 break;
 401         case IW_CM_STATE_ESTABLISHED:
 402                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 403                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 404                 /* Abrupt close of the connection */
 405                 (void)iwcm_modify_qp_err(cm_id_priv->qp);
 406                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 407                 break;
 408         case IW_CM_STATE_IDLE:
 409         case IW_CM_STATE_CLOSING:
 410                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 411                 break;
 412         case IW_CM_STATE_CONN_RECV:
 413                 /*
 414                  * App called destroy before/without calling accept after
 415                  * receiving connection request event notification or
 416                  * returned non zero from the event callback function.
 417                  * In either case, must tell the provider to reject.
 418                  */
 419                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 420                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 421                 cm_id->device->iwcm->reject(cm_id, NULL, 0);
 422                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 423                 break;
 424         case IW_CM_STATE_CONN_SENT:
 425         case IW_CM_STATE_DESTROYING:
 426         default:
 427                 BUG();
 428                 break;
 429         }
 430         if (cm_id_priv->qp) {
 431                 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
 432                 cm_id_priv->qp = NULL;
 433         }
 434         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 435
 436         if (cm_id->mapped) {
 437                 iwpm_remove_mapinfo(&cm_id->local_addr, &cm_id->m_local_addr);
 438                 iwpm_remove_mapping(&cm_id->local_addr, RDMA_NL_IWCM);
 439         }
 440
 441         (void)iwcm_deref_id(cm_id_priv);
 442 }
 443
 444 /*
 445  * This function is only called by the application thread and cannot
 446  * be called by the event thread. The function will wait for all
 447  * references to be released on the cm_id and then kfree the cm_id
 448  * object.
 449  */
 450 void iw_destroy_cm_id(struct iw_cm_id *cm_id)
 451 {
 452         destroy_cm_id(cm_id);
 453 }
 454 EXPORT_SYMBOL(iw_destroy_cm_id);
 455
 456 /**
 457  * iw_cm_check_wildcard - If IP address is 0 then use original
 458  * @pm_addr: sockaddr containing the ip to check for wildcard
 459  * @cm_addr: sockaddr containing the actual IP address
 460  * @cm_outaddr: sockaddr to set IP addr which leaving port
 461  *
 462  *  Checks the pm_addr for wildcard and then sets cm_outaddr's
 463  *  IP to the actual (cm_addr).
 464  */
 465 static void iw_cm_check_wildcard(struct sockaddr_storage *pm_addr,
 466                                  struct sockaddr_storage *cm_addr,
 467                                  struct sockaddr_storage *cm_outaddr)
 468 {
 469         if (pm_addr->ss_family == AF_INET) {
 470                 struct sockaddr_in *pm4_addr = (struct sockaddr_in *)pm_addr;
 471
 472                 if (pm4_addr->sin_addr.s_addr == htonl(INADDR_ANY)) {
 473                         struct sockaddr_in *cm4_addr =
 474                                 (struct sockaddr_in *)cm_addr;
 475                         struct sockaddr_in *cm4_outaddr =
 476                                 (struct sockaddr_in *)cm_outaddr;
 477
 478                         cm4_outaddr->sin_addr = cm4_addr->sin_addr;
 479                 }
 480         } else {
 481                 struct sockaddr_in6 *pm6_addr = (struct sockaddr_in6 *)pm_addr;
 482
 483                 if (ipv6_addr_type(&pm6_addr->sin6_addr) == IPV6_ADDR_ANY) {
 484                         struct sockaddr_in6 *cm6_addr =
 485                                 (struct sockaddr_in6 *)cm_addr;
 486                         struct sockaddr_in6 *cm6_outaddr =
 487                                 (struct sockaddr_in6 *)cm_outaddr;
 488
 489                         cm6_outaddr->sin6_addr = cm6_addr->sin6_addr;
 490                 }
 491         }
 492 }
 493
 494 /**
 495  * iw_cm_map - Use portmapper to map the ports
 496  * @cm_id: connection manager pointer
 497  * @active: Indicates the active side when true
 498  * returns nonzero for error only if iwpm_create_mapinfo() fails
 499  *
 500  * Tries to add a mapping for a port using the Portmapper. If
 501  * successful in mapping the IP/Port it will check the remote
 502  * mapped IP address for a wildcard IP address and replace the
 503  * zero IP address with the remote_addr.
 504  */
 505 static int iw_cm_map(struct iw_cm_id *cm_id, bool active)
 506 {
 507         struct iwpm_dev_data pm_reg_msg;
 508         struct iwpm_sa_data pm_msg;
 509         int status;
 510
 511         cm_id->m_local_addr = cm_id->local_addr;
 512         cm_id->m_remote_addr = cm_id->remote_addr;
 513
 514         memcpy(pm_reg_msg.dev_name, cm_id->device->name,
 515                sizeof(pm_reg_msg.dev_name));
 516         memcpy(pm_reg_msg.if_name, cm_id->device->iwcm->ifname,
 517                sizeof(pm_reg_msg.if_name));
 518
 519         if (iwpm_register_pid(&pm_reg_msg, RDMA_NL_IWCM) ||
 520             !iwpm_valid_pid())
 521                 return 0;
 522
 523         cm_id->mapped = true;
 524         pm_msg.loc_addr = cm_id->local_addr;
 525         pm_msg.rem_addr = cm_id->remote_addr;
 526         if (active)
 527                 status = iwpm_add_and_query_mapping(&pm_msg,
 528                                                     RDMA_NL_IWCM);
 529         else
 530                 status = iwpm_add_mapping(&pm_msg, RDMA_NL_IWCM);
 531
 532         if (!status) {
 533                 cm_id->m_local_addr = pm_msg.mapped_loc_addr;
 534                 if (active) {
 535                         cm_id->m_remote_addr = pm_msg.mapped_rem_addr;
 536                         iw_cm_check_wildcard(&pm_msg.mapped_rem_addr,
 537                                              &cm_id->remote_addr,
 538                                              &cm_id->m_remote_addr);
 539                 }
 540         }
 541
 542         return iwpm_create_mapinfo(&cm_id->local_addr,
 543                                    &cm_id->m_local_addr,
 544                                    RDMA_NL_IWCM);
 545 }
 546
 547 /*
 548  * CM_ID <-- LISTEN
 549  *
 550  * Start listening for connect requests. Generates one CONNECT_REQUEST
 551  * event for each inbound connect request.
 552  */
 553 int iw_cm_listen(struct iw_cm_id *cm_id, int backlog)
 554 {
 555         struct iwcm_id_private *cm_id_priv;
 556         unsigned long flags;
 557         int ret;
 558
 559         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 560
 561         if (!backlog)
 562                 backlog = default_backlog;
 563
 564         ret = alloc_work_entries(cm_id_priv, backlog);
 565         if (ret)
 566                 return ret;
 567
 568         spin_lock_irqsave(&cm_id_priv->lock, flags);
 569         switch (cm_id_priv->state) {
 570         case IW_CM_STATE_IDLE:
 571                 cm_id_priv->state = IW_CM_STATE_LISTEN;
 572                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 573                 ret = iw_cm_map(cm_id, false);
 574                 if (!ret)
 575                         ret = cm_id->device->iwcm->create_listen(cm_id, backlog);
 576                 if (ret)
 577                         cm_id_priv->state = IW_CM_STATE_IDLE;
 578                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 579                 break;
 580         default:
 581                 ret = -EINVAL;
 582         }
 583         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 584
 585         return ret;
 586 }
 587 EXPORT_SYMBOL(iw_cm_listen);
 588
 589 /*
 590  * CM_ID <-- IDLE
 591  *
 592  * Rejects an inbound connection request. No events are generated.
 593  */
 594 int iw_cm_reject(struct iw_cm_id *cm_id,
 595                  const void *private_data,
 596                  u8 private_data_len)
 597 {
 598         struct iwcm_id_private *cm_id_priv;
 599         unsigned long flags;
 600         int ret;
 601
 602         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 603         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 604
 605         spin_lock_irqsave(&cm_id_priv->lock, flags);
 606         if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
 607                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 608                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 609                 wake_up_all(&cm_id_priv->connect_wait);
 610                 return -EINVAL;
 611         }
 612         cm_id_priv->state = IW_CM_STATE_IDLE;
 613         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 614
 615         ret = cm_id->device->iwcm->reject(cm_id, private_data,
 616                                           private_data_len);
 617
 618         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 619         wake_up_all(&cm_id_priv->connect_wait);
 620
 621         return ret;
 622 }
 623 EXPORT_SYMBOL(iw_cm_reject);
 624
 625 /*
 626  * CM_ID <-- ESTABLISHED
 627  *
 628  * Accepts an inbound connection request and generates an ESTABLISHED
 629  * event. Callers of iw_cm_disconnect and iw_destroy_cm_id will block
 630  * until the ESTABLISHED event is received from the provider.
 631  */
 632 int iw_cm_accept(struct iw_cm_id *cm_id,
 633                  struct iw_cm_conn_param *iw_param)
 634 {
 635         struct iwcm_id_private *cm_id_priv;
 636         struct ib_qp *qp;
 637         unsigned long flags;
 638         int ret;
 639
 640         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 641         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 642
 643         spin_lock_irqsave(&cm_id_priv->lock, flags);
 644         if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
 645                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 646                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 647                 wake_up_all(&cm_id_priv->connect_wait);
 648                 return -EINVAL;
 649         }
 650         /* Get the ib_qp given the QPN */
 651         qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
 652         if (!qp) {
 653                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 654                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 655                 wake_up_all(&cm_id_priv->connect_wait);
 656                 return -EINVAL;
 657         }
 658         cm_id->device->iwcm->add_ref(qp);
 659         cm_id_priv->qp = qp;
 660         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 661
 662         ret = cm_id->device->iwcm->accept(cm_id, iw_param);
 663         if (ret) {
 664                 /* An error on accept precludes provider events */
 665                 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
 666                 cm_id_priv->state = IW_CM_STATE_IDLE;
 667                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 668                 if (cm_id_priv->qp) {
 669                         cm_id->device->iwcm->rem_ref(qp);
 670                         cm_id_priv->qp = NULL;
 671                 }
 672                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 673                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 674                 wake_up_all(&cm_id_priv->connect_wait);
 675         }
 676
 677         return ret;
 678 }
 679 EXPORT_SYMBOL(iw_cm_accept);
 680
 681 /*
 682  * Active Side: CM_ID <-- CONN_SENT
 683  *
 684  * If successful, results in the generation of a CONNECT_REPLY
 685  * event. iw_cm_disconnect and iw_cm_destroy will block until the
 686  * CONNECT_REPLY event is received from the provider.
 687  */
 688 int iw_cm_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *iw_param)
 689 {
 690         struct iwcm_id_private *cm_id_priv;
 691         int ret;
 692         unsigned long flags;
 693         struct ib_qp *qp;
 694
 695         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 696
 697         ret = alloc_work_entries(cm_id_priv, 4);
 698         if (ret)
 699                 return ret;
 700
 701         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 702         spin_lock_irqsave(&cm_id_priv->lock, flags);
 703
 704         if (cm_id_priv->state != IW_CM_STATE_IDLE) {
 705                 ret = -EINVAL;
 706                 goto err;
 707         }
 708
 709         /* Get the ib_qp given the QPN */
 710         qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
 711         if (!qp) {
 712                 ret = -EINVAL;
 713                 goto err;
 714         }
 715         cm_id->device->iwcm->add_ref(qp);
 716         cm_id_priv->qp = qp;
 717         cm_id_priv->state = IW_CM_STATE_CONN_SENT;
 718         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 719
 720         ret = iw_cm_map(cm_id, true);
 721         if (!ret)
 722                 ret = cm_id->device->iwcm->connect(cm_id, iw_param);
 723         if (!ret)
 724                 return 0;       /* success */
 725
 726         spin_lock_irqsave(&cm_id_priv->lock, flags);
 727         if (cm_id_priv->qp) {
 728                 cm_id->device->iwcm->rem_ref(qp);
 729                 cm_id_priv->qp = NULL;
 730         }
 731         cm_id_priv->state = IW_CM_STATE_IDLE;
 732 err:
 733         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 734         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 735         wake_up_all(&cm_id_priv->connect_wait);
 736         return ret;
 737 }
 738 EXPORT_SYMBOL(iw_cm_connect);
 739
 740 /*
 741  * Passive Side: new CM_ID <-- CONN_RECV
 742  *
 743  * Handles an inbound connect request. The function creates a new
 744  * iw_cm_id to represent the new connection and inherits the client
 745  * callback function and other attributes from the listening parent.
 746  *
 747  * The work item contains a pointer to the listen_cm_id and the event. The
 748  * listen_cm_id contains the client cm_handler, context and
 749  * device. These are copied when the device is cloned. The event
 750  * contains the new four tuple.
 751  *
 752  * An error on the child should not affect the parent, so this
 753  * function does not return a value.
 754  */
 755 static void cm_conn_req_handler(struct iwcm_id_private *listen_id_priv,
 756                                 struct iw_cm_event *iw_event)
 757 {
 758         unsigned long flags;
 759         struct iw_cm_id *cm_id;
 760         struct iwcm_id_private *cm_id_priv;
 761         int ret;
 762
 763         /*
 764          * The provider should never generate a connection request
 765          * event with a bad status.
 766          */
 767         BUG_ON(iw_event->status);
 768
 769         cm_id = iw_create_cm_id(listen_id_priv->id.device,
 770                                 listen_id_priv->id.cm_handler,
 771                                 listen_id_priv->id.context);
 772         /* If the cm_id could not be created, ignore the request */
 773         if (IS_ERR(cm_id))
 774                 goto out;
 775
 776         cm_id->provider_data = iw_event->provider_data;
 777         cm_id->m_local_addr = iw_event->local_addr;
 778         cm_id->m_remote_addr = iw_event->remote_addr;
 779         cm_id->local_addr = listen_id_priv->id.local_addr;
 780
 781         ret = iwpm_get_remote_info(&listen_id_priv->id.m_local_addr,
 782                                    &iw_event->remote_addr,
 783                                    &cm_id->remote_addr,
 784                                    RDMA_NL_IWCM);
 785         if (ret) {
 786                 cm_id->remote_addr = iw_event->remote_addr;
 787         } else {
 788                 iw_cm_check_wildcard(&listen_id_priv->id.m_local_addr,
 789                                      &iw_event->local_addr,
 790                                      &cm_id->local_addr);
 791                 iw_event->local_addr = cm_id->local_addr;
 792                 iw_event->remote_addr = cm_id->remote_addr;
 793         }
 794
 795         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 796         cm_id_priv->state = IW_CM_STATE_CONN_RECV;
 797
 798         /*
 799          * We could be destroying the listening id. If so, ignore this
 800          * upcall.
 801          */
 802         spin_lock_irqsave(&listen_id_priv->lock, flags);
 803         if (listen_id_priv->state != IW_CM_STATE_LISTEN) {
 804                 spin_unlock_irqrestore(&listen_id_priv->lock, flags);
 805                 iw_cm_reject(cm_id, NULL, 0);
 806                 iw_destroy_cm_id(cm_id);
 807                 goto out;
 808         }
 809         spin_unlock_irqrestore(&listen_id_priv->lock, flags);
 810
 811         ret = alloc_work_entries(cm_id_priv, 3);
 812         if (ret) {
 813                 iw_cm_reject(cm_id, NULL, 0);
 814                 iw_destroy_cm_id(cm_id);
 815                 goto out;
 816         }
 817
 818         /* Call the client CM handler */
 819         ret = cm_id->cm_handler(cm_id, iw_event);
 820         if (ret) {
 821                 iw_cm_reject(cm_id, NULL, 0);
 822                 iw_destroy_cm_id(cm_id);
 823         }
 824
 825 out:
 826         if (iw_event->private_data_len)
 827                 kfree(iw_event->private_data);
 828 }
 829
 830 /*
 831  * Passive Side: CM_ID <-- ESTABLISHED
 832  *
 833  * The provider generated an ESTABLISHED event which means that
 834  * the MPA negotion has completed successfully and we are now in MPA
 835  * FPDU mode.
 836  *
 837  * This event can only be received in the CONN_RECV state. If the
 838  * remote peer closed, the ESTABLISHED event would be received followed
 839  * by the CLOSE event. If the app closes, it will block until we wake
 840  * it up after processing this event.
 841  */
 842 static int cm_conn_est_handler(struct iwcm_id_private *cm_id_priv,
 843                                struct iw_cm_event *iw_event)
 844 {
 845         unsigned long flags;
 846         int ret;
 847
 848         spin_lock_irqsave(&cm_id_priv->lock, flags);
 849
 850         /*
 851          * We clear the CONNECT_WAIT bit here to allow the callback
 852          * function to call iw_cm_disconnect. Calling iw_destroy_cm_id
 853          * from a callback handler is not allowed.
 854          */
 855         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 856         BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
 857         cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
 858         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 859         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 860         wake_up_all(&cm_id_priv->connect_wait);
 861
 862         return ret;
 863 }
 864
 865 /*
 866  * Active Side: CM_ID <-- ESTABLISHED
 867  *
 868  * The app has called connect and is waiting for the established event to
 869  * post it's requests to the server. This event will wake up anyone
 870  * blocked in iw_cm_disconnect or iw_destroy_id.
 871  */
 872 static int cm_conn_rep_handler(struct iwcm_id_private *cm_id_priv,
 873                                struct iw_cm_event *iw_event)
 874 {
 875         unsigned long flags;
 876         int ret;
 877
 878         spin_lock_irqsave(&cm_id_priv->lock, flags);
 879         /*
 880          * Clear the connect wait bit so a callback function calling
 881          * iw_cm_disconnect will not wait and deadlock this thread
 882          */
 883         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 884         BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
 885         if (iw_event->status == 0) {
 886                 cm_id_priv->id.m_local_addr = iw_event->local_addr;
 887                 cm_id_priv->id.m_remote_addr = iw_event->remote_addr;
 888                 iw_event->local_addr = cm_id_priv->id.local_addr;
 889                 iw_event->remote_addr = cm_id_priv->id.remote_addr;
 890                 cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
 891         } else {
 892                 /* REJECTED or RESET */
 893                 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
 894                 cm_id_priv->qp = NULL;
 895                 cm_id_priv->state = IW_CM_STATE_IDLE;
 896         }
 897         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 898         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 899
 900         if (iw_event->private_data_len)
 901                 kfree(iw_event->private_data);
 902
 903         /* Wake up waiters on connect complete */
 904         wake_up_all(&cm_id_priv->connect_wait);
 905
 906         return ret;
 907 }
 908
 909 /*
 910  * CM_ID <-- CLOSING
 911  *
 912  * If in the ESTABLISHED state, move to CLOSING.
 913  */
 914 static void cm_disconnect_handler(struct iwcm_id_private *cm_id_priv,
 915                                   struct iw_cm_event *iw_event)
 916 {
 917         unsigned long flags;
 918
 919         spin_lock_irqsave(&cm_id_priv->lock, flags);
 920         if (cm_id_priv->state == IW_CM_STATE_ESTABLISHED)
 921                 cm_id_priv->state = IW_CM_STATE_CLOSING;
 922         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 923 }
 924
 925 /*
 926  * CM_ID <-- IDLE
 927  *
 928  * If in the ESTBLISHED or CLOSING states, the QP will have have been
 929  * moved by the provider to the ERR state. Disassociate the CM_ID from
 930  * the QP,  move to IDLE, and remove the 'connected' reference.
 931  *
 932  * If in some other state, the cm_id was destroyed asynchronously.
 933  * This is the last reference that will result in waking up
 934  * the app thread blocked in iw_destroy_cm_id.
 935  */
 936 static int cm_close_handler(struct iwcm_id_private *cm_id_priv,
 937                                   struct iw_cm_event *iw_event)
 938 {
 939         unsigned long flags;
 940         int ret = 0;
 941         spin_lock_irqsave(&cm_id_priv->lock, flags);
 942
 943         if (cm_id_priv->qp) {
 944                 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
 945                 cm_id_priv->qp = NULL;
 946         }
 947         switch (cm_id_priv->state) {
 948         case IW_CM_STATE_ESTABLISHED:
 949         case IW_CM_STATE_CLOSING:
 950                 cm_id_priv->state = IW_CM_STATE_IDLE;
 951                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 952                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 953                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 954                 break;
 955         case IW_CM_STATE_DESTROYING:
 956                 break;
 957         default:
 958                 BUG();
 959         }
 960         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 961
 962         return ret;
 963 }
 964
 965 static int process_event(struct iwcm_id_private *cm_id_priv,
 966                          struct iw_cm_event *iw_event)
 967 {
 968         int ret = 0;
 969
 970         switch (iw_event->event) {
 971         case IW_CM_EVENT_CONNECT_REQUEST:
 972                 cm_conn_req_handler(cm_id_priv, iw_event);
 973                 break;
 974         case IW_CM_EVENT_CONNECT_REPLY:
 975                 ret = cm_conn_rep_handler(cm_id_priv, iw_event);
 976                 break;
 977         case IW_CM_EVENT_ESTABLISHED:
 978                 ret = cm_conn_est_handler(cm_id_priv, iw_event);
 979                 break;
 980         case IW_CM_EVENT_DISCONNECT:
 981                 cm_disconnect_handler(cm_id_priv, iw_event);
 982                 break;
 983         case IW_CM_EVENT_CLOSE:
 984                 ret = cm_close_handler(cm_id_priv, iw_event);
 985                 break;
 986         default:
 987                 BUG();
 988         }
 989
 990         return ret;
 991 }
 992
 993 /*
 994  * Process events on the work_list for the cm_id. If the callback
 995  * function requests that the cm_id be deleted, a flag is set in the
 996  * cm_id flags to indicate that when the last reference is
 997  * removed, the cm_id is to be destroyed. This is necessary to
 998  * distinguish between an object that will be destroyed by the app
 999  * thread asleep on the destroy_comp list vs. an object destroyed
1000  * here synchronously when the last reference is removed.
1001  */
1002 static void cm_work_handler(struct work_struct *_work)
1003 {
1004         struct iwcm_work *work = container_of(_work, struct iwcm_work, work);
1005         struct iw_cm_event levent;
1006         struct iwcm_id_private *cm_id_priv = work->cm_id;
1007         unsigned long flags;
1008         int empty;
1009         int ret = 0;
1010
1011         spin_lock_irqsave(&cm_id_priv->lock, flags);
1012         empty = list_empty(&cm_id_priv->work_list);
1013         while (!empty) {
1014                 work = list_entry(cm_id_priv->work_list.next,
1015                                   struct iwcm_work, list);
1016                 list_del_init(&work->list);
1017                 empty = list_empty(&cm_id_priv->work_list);
1018                 levent = work->event;
1019                 put_work(work);
1020                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1021
1022                 if (!test_bit(IWCM_F_DROP_EVENTS, &cm_id_priv->flags)) {
1023                         ret = process_event(cm_id_priv, &levent);
1024                         if (ret)
1025                                 destroy_cm_id(&cm_id_priv->id);
1026                 } else
1027                         pr_debug("dropping event %d\n", levent.event);
1028                 if (iwcm_deref_id(cm_id_priv))
1029                         return;
1030                 if (empty)
1031                         return;
1032                 spin_lock_irqsave(&cm_id_priv->lock, flags);
1033         }
1034         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1035 }
1036
1037 /*
1038  * This function is called on interrupt context. Schedule events on
1039  * the iwcm_wq thread to allow callback functions to downcall into
1040  * the CM and/or block.  Events are queued to a per-CM_ID
1041  * work_list. If this is the first event on the work_list, the work
1042  * element is also queued on the iwcm_wq thread.
1043  *
1044  * Each event holds a reference on the cm_id. Until the last posted
1045  * event has been delivered and processed, the cm_id cannot be
1046  * deleted.
1047  *
1048  * Returns:
1049  *            0 - the event was handled.
1050  *      -ENOMEM - the event was not handled due to lack of resources.
1051  */
1052 static int cm_event_handler(struct iw_cm_id *cm_id,
1053                              struct iw_cm_event *iw_event)
1054 {
1055         struct iwcm_work *work;
1056         struct iwcm_id_private *cm_id_priv;
1057         unsigned long flags;
1058         int ret = 0;
1059
1060         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1061
1062         spin_lock_irqsave(&cm_id_priv->lock, flags);
1063         work = get_work(cm_id_priv);
1064         if (!work) {
1065                 ret = -ENOMEM;
1066                 goto out;
1067         }
1068
1069         INIT_WORK(&work->work, cm_work_handler);
1070         work->cm_id = cm_id_priv;
1071         work->event = *iw_event;
1072
1073         if ((work->event.event == IW_CM_EVENT_CONNECT_REQUEST ||
1074              work->event.event == IW_CM_EVENT_CONNECT_REPLY) &&
1075             work->event.private_data_len) {
1076                 ret = copy_private_data(&work->event);
1077                 if (ret) {
1078                         put_work(work);
1079                         goto out;
1080                 }
1081         }
1082
1083         atomic_inc(&cm_id_priv->refcount);
1084         if (list_empty(&cm_id_priv->work_list)) {
1085                 list_add_tail(&work->list, &cm_id_priv->work_list);
1086                 queue_work(iwcm_wq, &work->work);
1087         } else
1088                 list_add_tail(&work->list, &cm_id_priv->work_list);
1089 out:
1090         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1091         return ret;
1092 }
1093
1094 static int iwcm_init_qp_init_attr(struct iwcm_id_private *cm_id_priv,
1095                                   struct ib_qp_attr *qp_attr,
1096                                   int *qp_attr_mask)
1097 {
1098         unsigned long flags;
1099         int ret;
1100
1101         spin_lock_irqsave(&cm_id_priv->lock, flags);
1102         switch (cm_id_priv->state) {
1103         case IW_CM_STATE_IDLE:
1104         case IW_CM_STATE_CONN_SENT:
1105         case IW_CM_STATE_CONN_RECV:
1106         case IW_CM_STATE_ESTABLISHED:
1107                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1108                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE|
1109                                            IB_ACCESS_REMOTE_READ;
1110                 ret = 0;
1111                 break;
1112         default:
1113                 ret = -EINVAL;
1114                 break;
1115         }
1116         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1117         return ret;
1118 }
1119
1120 static int iwcm_init_qp_rts_attr(struct iwcm_id_private *cm_id_priv,
1121                                   struct ib_qp_attr *qp_attr,
1122                                   int *qp_attr_mask)
1123 {
1124         unsigned long flags;
1125         int ret;
1126
1127         spin_lock_irqsave(&cm_id_priv->lock, flags);
1128         switch (cm_id_priv->state) {
1129         case IW_CM_STATE_IDLE:
1130         case IW_CM_STATE_CONN_SENT:
1131         case IW_CM_STATE_CONN_RECV:
1132         case IW_CM_STATE_ESTABLISHED:
1133                 *qp_attr_mask = 0;
1134                 ret = 0;
1135                 break;
1136         default:
1137                 ret = -EINVAL;
1138                 break;
1139         }
1140         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1141         return ret;
1142 }
1143
1144 int iw_cm_init_qp_attr(struct iw_cm_id *cm_id,
1145                        struct ib_qp_attr *qp_attr,
1146                        int *qp_attr_mask)
1147 {
1148         struct iwcm_id_private *cm_id_priv;
1149         int ret;
1150
1151         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1152         switch (qp_attr->qp_state) {
1153         case IB_QPS_INIT:
1154         case IB_QPS_RTR:
1155                 ret = iwcm_init_qp_init_attr(cm_id_priv,
1156                                              qp_attr, qp_attr_mask);
1157                 break;
1158         case IB_QPS_RTS:
1159                 ret = iwcm_init_qp_rts_attr(cm_id_priv,
1160                                             qp_attr, qp_attr_mask);
1161                 break;
1162         default:
1163                 ret = -EINVAL;
1164                 break;
1165         }
1166         return ret;
1167 }
1168 EXPORT_SYMBOL(iw_cm_init_qp_attr);
1169
1170 static int __init iw_cm_init(void)
1171 {
1172         int ret;
1173
1174         ret = iwpm_init(RDMA_NL_IWCM);
1175         if (ret)
1176                 return ret;
1177
1178         iwcm_wq = alloc_ordered_workqueue("iw_cm_wq", 0);
1179         if (!iwcm_wq)
1180                 goto err_alloc;
1181
1182         iwcm_ctl_table_hdr = register_net_sysctl(&init_net, "net/iw_cm",
1183                                                  iwcm_ctl_table);
1184         if (!iwcm_ctl_table_hdr) {
1185                 pr_err("iw_cm: couldn't register sysctl paths\n");
1186                 goto err_sysctl;
1187         }
1188
1189         rdma_nl_register(RDMA_NL_IWCM, iwcm_nl_cb_table);
1190         return 0;
1191
1192 err_sysctl:
1193         destroy_workqueue(iwcm_wq);
1194 err_alloc:
1195         iwpm_exit(RDMA_NL_IWCM);
1196         return -ENOMEM;
1197 }
1198
1199 static void __exit iw_cm_cleanup(void)
1200 {
1201         rdma_nl_unregister(RDMA_NL_IWCM);
1202         unregister_net_sysctl_table(iwcm_ctl_table_hdr);
1203         destroy_workqueue(iwcm_wq);
1204         iwpm_exit(RDMA_NL_IWCM);
1205 }
1206
1207 MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_IWCM, 2);
1208
1209 module_init(iw_cm_init);
1210 module_exit(iw_cm_cleanup);