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         [RDMA_NL_IWPM_HELLO] = {.dump = iwpm_hello_cb}
  92 };
  93
  94 static struct workqueue_struct *iwcm_wq;
  95 struct iwcm_work {
  96         struct work_struct work;
  97         struct iwcm_id_private *cm_id;
  98         struct list_head list;
  99         struct iw_cm_event event;
 100         struct list_head free_list;
 101 };
 102
 103 static unsigned int default_backlog = 256;
 104
 105 static struct ctl_table_header *iwcm_ctl_table_hdr;
 106 static struct ctl_table iwcm_ctl_table[] = {
 107         {
 108                 .procname       = "default_backlog",
 109                 .data           = &default_backlog,
 110                 .maxlen         = sizeof(default_backlog),
 111                 .mode           = 0644,
 112                 .proc_handler   = proc_dointvec,
 113         },
 114         { }
 115 };
 116
 117 /*
 118  * The following services provide a mechanism for pre-allocating iwcm_work
 119  * elements.  The design pre-allocates them  based on the cm_id type:
 120  *      LISTENING IDS:  Get enough elements preallocated to handle the
 121  *                      listen backlog.
 122  *      ACTIVE IDS:     4: CONNECT_REPLY, ESTABLISHED, DISCONNECT, CLOSE
 123  *      PASSIVE IDS:    3: ESTABLISHED, DISCONNECT, CLOSE
 124  *
 125  * Allocating them in connect and listen avoids having to deal
 126  * with allocation failures on the event upcall from the provider (which
 127  * is called in the interrupt context).
 128  *
 129  * One exception is when creating the cm_id for incoming connection requests.
 130  * There are two cases:
 131  * 1) in the event upcall, cm_event_handler(), for a listening cm_id.  If
 132  *    the backlog is exceeded, then no more connection request events will
 133  *    be processed.  cm_event_handler() returns -ENOMEM in this case.  Its up
 134  *    to the provider to reject the connection request.
 135  * 2) in the connection request workqueue handler, cm_conn_req_handler().
 136  *    If work elements cannot be allocated for the new connect request cm_id,
 137  *    then IWCM will call the provider reject method.  This is ok since
 138  *    cm_conn_req_handler() runs in the workqueue thread context.
 139  */
 140
 141 static struct iwcm_work *get_work(struct iwcm_id_private *cm_id_priv)
 142 {
 143         struct iwcm_work *work;
 144
 145         if (list_empty(&cm_id_priv->work_free_list))
 146                 return NULL;
 147         work = list_entry(cm_id_priv->work_free_list.next, struct iwcm_work,
 148                           free_list);
 149         list_del_init(&work->free_list);
 150         return work;
 151 }
 152
 153 static void put_work(struct iwcm_work *work)
 154 {
 155         list_add(&work->free_list, &work->cm_id->work_free_list);
 156 }
 157
 158 static void dealloc_work_entries(struct iwcm_id_private *cm_id_priv)
 159 {
 160         struct list_head *e, *tmp;
 161
 162         list_for_each_safe(e, tmp, &cm_id_priv->work_free_list)
 163                 kfree(list_entry(e, struct iwcm_work, free_list));
 164 }
 165
 166 static int alloc_work_entries(struct iwcm_id_private *cm_id_priv, int count)
 167 {
 168         struct iwcm_work *work;
 169
 170         BUG_ON(!list_empty(&cm_id_priv->work_free_list));
 171         while (count--) {
 172                 work = kmalloc(sizeof(struct iwcm_work), GFP_KERNEL);
 173                 if (!work) {
 174                         dealloc_work_entries(cm_id_priv);
 175                         return -ENOMEM;
 176                 }
 177                 work->cm_id = cm_id_priv;
 178                 INIT_LIST_HEAD(&work->list);
 179                 put_work(work);
 180         }
 181         return 0;
 182 }
 183
 184 /*
 185  * Save private data from incoming connection requests to
 186  * iw_cm_event, so the low level driver doesn't have to. Adjust
 187  * the event ptr to point to the local copy.
 188  */
 189 static int copy_private_data(struct iw_cm_event *event)
 190 {
 191         void *p;
 192
 193         p = kmemdup(event->private_data, event->private_data_len, GFP_ATOMIC);
 194         if (!p)
 195                 return -ENOMEM;
 196         event->private_data = p;
 197         return 0;
 198 }
 199
 200 static void free_cm_id(struct iwcm_id_private *cm_id_priv)
 201 {
 202         dealloc_work_entries(cm_id_priv);
 203         kfree(cm_id_priv);
 204 }
 205
 206 /*
 207  * Release a reference on cm_id. If the last reference is being
 208  * released, free the cm_id and return 1.
 209  */
 210 static int iwcm_deref_id(struct iwcm_id_private *cm_id_priv)
 211 {
 212         BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
 213         if (atomic_dec_and_test(&cm_id_priv->refcount)) {
 214                 BUG_ON(!list_empty(&cm_id_priv->work_list));
 215                 free_cm_id(cm_id_priv);
 216                 return 1;
 217         }
 218
 219         return 0;
 220 }
 221
 222 static void add_ref(struct iw_cm_id *cm_id)
 223 {
 224         struct iwcm_id_private *cm_id_priv;
 225         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 226         atomic_inc(&cm_id_priv->refcount);
 227 }
 228
 229 static void rem_ref(struct iw_cm_id *cm_id)
 230 {
 231         struct iwcm_id_private *cm_id_priv;
 232
 233         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 234
 235         (void)iwcm_deref_id(cm_id_priv);
 236 }
 237
 238 static int cm_event_handler(struct iw_cm_id *cm_id, struct iw_cm_event *event);
 239
 240 struct iw_cm_id *iw_create_cm_id(struct ib_device *device,
 241                                  iw_cm_handler cm_handler,
 242                                  void *context)
 243 {
 244         struct iwcm_id_private *cm_id_priv;
 245
 246         cm_id_priv = kzalloc(sizeof(*cm_id_priv), GFP_KERNEL);
 247         if (!cm_id_priv)
 248                 return ERR_PTR(-ENOMEM);
 249
 250         cm_id_priv->state = IW_CM_STATE_IDLE;
 251         cm_id_priv->id.device = device;
 252         cm_id_priv->id.cm_handler = cm_handler;
 253         cm_id_priv->id.context = context;
 254         cm_id_priv->id.event_handler = cm_event_handler;
 255         cm_id_priv->id.add_ref = add_ref;
 256         cm_id_priv->id.rem_ref = rem_ref;
 257         spin_lock_init(&cm_id_priv->lock);
 258         atomic_set(&cm_id_priv->refcount, 1);
 259         init_waitqueue_head(&cm_id_priv->connect_wait);
 260         init_completion(&cm_id_priv->destroy_comp);
 261         INIT_LIST_HEAD(&cm_id_priv->work_list);
 262         INIT_LIST_HEAD(&cm_id_priv->work_free_list);
 263
 264         return &cm_id_priv->id;
 265 }
 266 EXPORT_SYMBOL(iw_create_cm_id);
 267
 268
 269 static int iwcm_modify_qp_err(struct ib_qp *qp)
 270 {
 271         struct ib_qp_attr qp_attr;
 272
 273         if (!qp)
 274                 return -EINVAL;
 275
 276         qp_attr.qp_state = IB_QPS_ERR;
 277         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 278 }
 279
 280 /*
 281  * This is really the RDMAC CLOSING state. It is most similar to the
 282  * IB SQD QP state.
 283  */
 284 static int iwcm_modify_qp_sqd(struct ib_qp *qp)
 285 {
 286         struct ib_qp_attr qp_attr;
 287
 288         BUG_ON(qp == NULL);
 289         qp_attr.qp_state = IB_QPS_SQD;
 290         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 291 }
 292
 293 /*
 294  * CM_ID <-- CLOSING
 295  *
 296  * Block if a passive or active connection is currently being processed. Then
 297  * process the event as follows:
 298  * - If we are ESTABLISHED, move to CLOSING and modify the QP state
 299  *   based on the abrupt flag
 300  * - If the connection is already in the CLOSING or IDLE state, the peer is
 301  *   disconnecting concurrently with us and we've already seen the
 302  *   DISCONNECT event -- ignore the request and return 0
 303  * - Disconnect on a listening endpoint returns -EINVAL
 304  */
 305 int iw_cm_disconnect(struct iw_cm_id *cm_id, int abrupt)
 306 {
 307         struct iwcm_id_private *cm_id_priv;
 308         unsigned long flags;
 309         int ret = 0;
 310         struct ib_qp *qp = NULL;
 311
 312         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 313         /* Wait if we're currently in a connect or accept downcall */
 314         wait_event(cm_id_priv->connect_wait,
 315                    !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
 316
 317         spin_lock_irqsave(&cm_id_priv->lock, flags);
 318         switch (cm_id_priv->state) {
 319         case IW_CM_STATE_ESTABLISHED:
 320                 cm_id_priv->state = IW_CM_STATE_CLOSING;
 321
 322                 /* QP could be <nul> for user-mode client */
 323                 if (cm_id_priv->qp)
 324                         qp = cm_id_priv->qp;
 325                 else
 326                         ret = -EINVAL;
 327                 break;
 328         case IW_CM_STATE_LISTEN:
 329                 ret = -EINVAL;
 330                 break;
 331         case IW_CM_STATE_CLOSING:
 332                 /* remote peer closed first */
 333         case IW_CM_STATE_IDLE:
 334                 /* accept or connect returned !0 */
 335                 break;
 336         case IW_CM_STATE_CONN_RECV:
 337                 /*
 338                  * App called disconnect before/without calling accept after
 339                  * connect_request event delivered.
 340                  */
 341                 break;
 342         case IW_CM_STATE_CONN_SENT:
 343                 /* Can only get here if wait above fails */
 344         default:
 345                 BUG();
 346         }
 347         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 348
 349         if (qp) {
 350                 if (abrupt)
 351                         ret = iwcm_modify_qp_err(qp);
 352                 else
 353                         ret = iwcm_modify_qp_sqd(qp);
 354
 355                 /*
 356                  * If both sides are disconnecting the QP could
 357                  * already be in ERR or SQD states
 358                  */
 359                 ret = 0;
 360         }
 361
 362         return ret;
 363 }
 364 EXPORT_SYMBOL(iw_cm_disconnect);
 365
 366 /*
 367  * CM_ID <-- DESTROYING
 368  *
 369  * Clean up all resources associated with the connection and release
 370  * the initial reference taken by iw_create_cm_id.
 371  */
 372 static void destroy_cm_id(struct iw_cm_id *cm_id)
 373 {
 374         struct iwcm_id_private *cm_id_priv;
 375         unsigned long flags;
 376
 377         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 378         /*
 379          * Wait if we're currently in a connect or accept downcall. A
 380          * listening endpoint should never block here.
 381          */
 382         wait_event(cm_id_priv->connect_wait,
 383                    !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
 384
 385         /*
 386          * Since we're deleting the cm_id, drop any events that
 387          * might arrive before the last dereference.
 388          */
 389         set_bit(IWCM_F_DROP_EVENTS, &cm_id_priv->flags);
 390
 391         spin_lock_irqsave(&cm_id_priv->lock, flags);
 392         switch (cm_id_priv->state) {
 393         case IW_CM_STATE_LISTEN:
 394                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 395                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 396                 /* destroy the listening endpoint */
 397                 cm_id->device->ops.iw_destroy_listen(cm_id);
 398                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 399                 break;
 400         case IW_CM_STATE_ESTABLISHED:
 401                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 402                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 403                 /* Abrupt close of the connection */
 404                 (void)iwcm_modify_qp_err(cm_id_priv->qp);
 405                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 406                 break;
 407         case IW_CM_STATE_IDLE:
 408         case IW_CM_STATE_CLOSING:
 409                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 410                 break;
 411         case IW_CM_STATE_CONN_RECV:
 412                 /*
 413                  * App called destroy before/without calling accept after
 414                  * receiving connection request event notification or
 415                  * returned non zero from the event callback function.
 416                  * In either case, must tell the provider to reject.
 417                  */
 418                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 419                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 420                 cm_id->device->ops.iw_reject(cm_id, NULL, 0);
 421                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 422                 break;
 423         case IW_CM_STATE_CONN_SENT:
 424         case IW_CM_STATE_DESTROYING:
 425         default:
 426                 BUG();
 427                 break;
 428         }
 429         if (cm_id_priv->qp) {
 430                 cm_id_priv->id.device->ops.iw_rem_ref(cm_id_priv->qp);
 431                 cm_id_priv->qp = NULL;
 432         }
 433         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 434
 435         if (cm_id->mapped) {
 436                 iwpm_remove_mapinfo(&cm_id->local_addr, &cm_id->m_local_addr);
 437                 iwpm_remove_mapping(&cm_id->local_addr, RDMA_NL_IWCM);
 438         }
 439
 440         (void)iwcm_deref_id(cm_id_priv);
 441 }
 442
 443 /*
 444  * This function is only called by the application thread and cannot
 445  * be called by the event thread. The function will wait for all
 446  * references to be released on the cm_id and then kfree the cm_id
 447  * object.
 448  */
 449 void iw_destroy_cm_id(struct iw_cm_id *cm_id)
 450 {
 451         destroy_cm_id(cm_id);
 452 }
 453 EXPORT_SYMBOL(iw_destroy_cm_id);
 454
 455 /**
 456  * iw_cm_check_wildcard - If IP address is 0 then use original
 457  * @pm_addr: sockaddr containing the ip to check for wildcard
 458  * @cm_addr: sockaddr containing the actual IP address
 459  * @cm_outaddr: sockaddr to set IP addr which leaving port
 460  *
 461  *  Checks the pm_addr for wildcard and then sets cm_outaddr's
 462  *  IP to the actual (cm_addr).
 463  */
 464 static void iw_cm_check_wildcard(struct sockaddr_storage *pm_addr,
 465                                  struct sockaddr_storage *cm_addr,
 466                                  struct sockaddr_storage *cm_outaddr)
 467 {
 468         if (pm_addr->ss_family == AF_INET) {
 469                 struct sockaddr_in *pm4_addr = (struct sockaddr_in *)pm_addr;
 470
 471                 if (pm4_addr->sin_addr.s_addr == htonl(INADDR_ANY)) {
 472                         struct sockaddr_in *cm4_addr =
 473                                 (struct sockaddr_in *)cm_addr;
 474                         struct sockaddr_in *cm4_outaddr =
 475                                 (struct sockaddr_in *)cm_outaddr;
 476
 477                         cm4_outaddr->sin_addr = cm4_addr->sin_addr;
 478                 }
 479         } else {
 480                 struct sockaddr_in6 *pm6_addr = (struct sockaddr_in6 *)pm_addr;
 481
 482                 if (ipv6_addr_type(&pm6_addr->sin6_addr) == IPV6_ADDR_ANY) {
 483                         struct sockaddr_in6 *cm6_addr =
 484                                 (struct sockaddr_in6 *)cm_addr;
 485                         struct sockaddr_in6 *cm6_outaddr =
 486                                 (struct sockaddr_in6 *)cm_outaddr;
 487
 488                         cm6_outaddr->sin6_addr = cm6_addr->sin6_addr;
 489                 }
 490         }
 491 }
 492
 493 /**
 494  * iw_cm_map - Use portmapper to map the ports
 495  * @cm_id: connection manager pointer
 496  * @active: Indicates the active side when true
 497  * returns nonzero for error only if iwpm_create_mapinfo() fails
 498  *
 499  * Tries to add a mapping for a port using the Portmapper. If
 500  * successful in mapping the IP/Port it will check the remote
 501  * mapped IP address for a wildcard IP address and replace the
 502  * zero IP address with the remote_addr.
 503  */
 504 static int iw_cm_map(struct iw_cm_id *cm_id, bool active)
 505 {
 506         const char *devname = dev_name(&cm_id->device->dev);
 507         const char *ifname = cm_id->device->iw_ifname;
 508         struct iwpm_dev_data pm_reg_msg = {};
 509         struct iwpm_sa_data pm_msg;
 510         int status;
 511
 512         if (strlen(devname) >= sizeof(pm_reg_msg.dev_name) ||
 513             strlen(ifname) >= sizeof(pm_reg_msg.if_name))
 514                 return -EINVAL;
 515
 516         cm_id->m_local_addr = cm_id->local_addr;
 517         cm_id->m_remote_addr = cm_id->remote_addr;
 518
 519         strcpy(pm_reg_msg.dev_name, devname);
 520         strcpy(pm_reg_msg.if_name, ifname);
 521
 522         if (iwpm_register_pid(&pm_reg_msg, RDMA_NL_IWCM) ||
 523             !iwpm_valid_pid())
 524                 return 0;
 525
 526         cm_id->mapped = true;
 527         pm_msg.loc_addr = cm_id->local_addr;
 528         pm_msg.rem_addr = cm_id->remote_addr;
 529         pm_msg.flags = (cm_id->device->iw_driver_flags & IW_F_NO_PORT_MAP) ?
 530                        IWPM_FLAGS_NO_PORT_MAP : 0;
 531         if (active)
 532                 status = iwpm_add_and_query_mapping(&pm_msg,
 533                                                     RDMA_NL_IWCM);
 534         else
 535                 status = iwpm_add_mapping(&pm_msg, RDMA_NL_IWCM);
 536
 537         if (!status) {
 538                 cm_id->m_local_addr = pm_msg.mapped_loc_addr;
 539                 if (active) {
 540                         cm_id->m_remote_addr = pm_msg.mapped_rem_addr;
 541                         iw_cm_check_wildcard(&pm_msg.mapped_rem_addr,
 542                                              &cm_id->remote_addr,
 543                                              &cm_id->m_remote_addr);
 544                 }
 545         }
 546
 547         return iwpm_create_mapinfo(&cm_id->local_addr,
 548                                    &cm_id->m_local_addr,
 549                                    RDMA_NL_IWCM, pm_msg.flags);
 550 }
 551
 552 /*
 553  * CM_ID <-- LISTEN
 554  *
 555  * Start listening for connect requests. Generates one CONNECT_REQUEST
 556  * event for each inbound connect request.
 557  */
 558 int iw_cm_listen(struct iw_cm_id *cm_id, int backlog)
 559 {
 560         struct iwcm_id_private *cm_id_priv;
 561         unsigned long flags;
 562         int ret;
 563
 564         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 565
 566         if (!backlog)
 567                 backlog = default_backlog;
 568
 569         ret = alloc_work_entries(cm_id_priv, backlog);
 570         if (ret)
 571                 return ret;
 572
 573         spin_lock_irqsave(&cm_id_priv->lock, flags);
 574         switch (cm_id_priv->state) {
 575         case IW_CM_STATE_IDLE:
 576                 cm_id_priv->state = IW_CM_STATE_LISTEN;
 577                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 578                 ret = iw_cm_map(cm_id, false);
 579                 if (!ret)
 580                         ret = cm_id->device->ops.iw_create_listen(cm_id,
 581                                                                   backlog);
 582                 if (ret)
 583                         cm_id_priv->state = IW_CM_STATE_IDLE;
 584                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 585                 break;
 586         default:
 587                 ret = -EINVAL;
 588         }
 589         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 590
 591         return ret;
 592 }
 593 EXPORT_SYMBOL(iw_cm_listen);
 594
 595 /*
 596  * CM_ID <-- IDLE
 597  *
 598  * Rejects an inbound connection request. No events are generated.
 599  */
 600 int iw_cm_reject(struct iw_cm_id *cm_id,
 601                  const void *private_data,
 602                  u8 private_data_len)
 603 {
 604         struct iwcm_id_private *cm_id_priv;
 605         unsigned long flags;
 606         int ret;
 607
 608         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 609         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 610
 611         spin_lock_irqsave(&cm_id_priv->lock, flags);
 612         if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
 613                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 614                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 615                 wake_up_all(&cm_id_priv->connect_wait);
 616                 return -EINVAL;
 617         }
 618         cm_id_priv->state = IW_CM_STATE_IDLE;
 619         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 620
 621         ret = cm_id->device->ops.iw_reject(cm_id, private_data,
 622                                           private_data_len);
 623
 624         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 625         wake_up_all(&cm_id_priv->connect_wait);
 626
 627         return ret;
 628 }
 629 EXPORT_SYMBOL(iw_cm_reject);
 630
 631 /*
 632  * CM_ID <-- ESTABLISHED
 633  *
 634  * Accepts an inbound connection request and generates an ESTABLISHED
 635  * event. Callers of iw_cm_disconnect and iw_destroy_cm_id will block
 636  * until the ESTABLISHED event is received from the provider.
 637  */
 638 int iw_cm_accept(struct iw_cm_id *cm_id,
 639                  struct iw_cm_conn_param *iw_param)
 640 {
 641         struct iwcm_id_private *cm_id_priv;
 642         struct ib_qp *qp;
 643         unsigned long flags;
 644         int ret;
 645
 646         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 647         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 648
 649         spin_lock_irqsave(&cm_id_priv->lock, flags);
 650         if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
 651                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 652                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 653                 wake_up_all(&cm_id_priv->connect_wait);
 654                 return -EINVAL;
 655         }
 656         /* Get the ib_qp given the QPN */
 657         qp = cm_id->device->ops.iw_get_qp(cm_id->device, iw_param->qpn);
 658         if (!qp) {
 659                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 660                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 661                 wake_up_all(&cm_id_priv->connect_wait);
 662                 return -EINVAL;
 663         }
 664         cm_id->device->ops.iw_add_ref(qp);
 665         cm_id_priv->qp = qp;
 666         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 667
 668         ret = cm_id->device->ops.iw_accept(cm_id, iw_param);
 669         if (ret) {
 670                 /* An error on accept precludes provider events */
 671                 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
 672                 cm_id_priv->state = IW_CM_STATE_IDLE;
 673                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 674                 if (cm_id_priv->qp) {
 675                         cm_id->device->ops.iw_rem_ref(qp);
 676                         cm_id_priv->qp = NULL;
 677                 }
 678                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 679                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 680                 wake_up_all(&cm_id_priv->connect_wait);
 681         }
 682
 683         return ret;
 684 }
 685 EXPORT_SYMBOL(iw_cm_accept);
 686
 687 /*
 688  * Active Side: CM_ID <-- CONN_SENT
 689  *
 690  * If successful, results in the generation of a CONNECT_REPLY
 691  * event. iw_cm_disconnect and iw_cm_destroy will block until the
 692  * CONNECT_REPLY event is received from the provider.
 693  */
 694 int iw_cm_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *iw_param)
 695 {
 696         struct iwcm_id_private *cm_id_priv;
 697         int ret;
 698         unsigned long flags;
 699         struct ib_qp *qp;
 700
 701         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 702
 703         ret = alloc_work_entries(cm_id_priv, 4);
 704         if (ret)
 705                 return ret;
 706
 707         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 708         spin_lock_irqsave(&cm_id_priv->lock, flags);
 709
 710         if (cm_id_priv->state != IW_CM_STATE_IDLE) {
 711                 ret = -EINVAL;
 712                 goto err;
 713         }
 714
 715         /* Get the ib_qp given the QPN */
 716         qp = cm_id->device->ops.iw_get_qp(cm_id->device, iw_param->qpn);
 717         if (!qp) {
 718                 ret = -EINVAL;
 719                 goto err;
 720         }
 721         cm_id->device->ops.iw_add_ref(qp);
 722         cm_id_priv->qp = qp;
 723         cm_id_priv->state = IW_CM_STATE_CONN_SENT;
 724         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 725
 726         ret = iw_cm_map(cm_id, true);
 727         if (!ret)
 728                 ret = cm_id->device->ops.iw_connect(cm_id, iw_param);
 729         if (!ret)
 730                 return 0;       /* success */
 731
 732         spin_lock_irqsave(&cm_id_priv->lock, flags);
 733         if (cm_id_priv->qp) {
 734                 cm_id->device->ops.iw_rem_ref(qp);
 735                 cm_id_priv->qp = NULL;
 736         }
 737         cm_id_priv->state = IW_CM_STATE_IDLE;
 738 err:
 739         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 740         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 741         wake_up_all(&cm_id_priv->connect_wait);
 742         return ret;
 743 }
 744 EXPORT_SYMBOL(iw_cm_connect);
 745
 746 /*
 747  * Passive Side: new CM_ID <-- CONN_RECV
 748  *
 749  * Handles an inbound connect request. The function creates a new
 750  * iw_cm_id to represent the new connection and inherits the client
 751  * callback function and other attributes from the listening parent.
 752  *
 753  * The work item contains a pointer to the listen_cm_id and the event. The
 754  * listen_cm_id contains the client cm_handler, context and
 755  * device. These are copied when the device is cloned. The event
 756  * contains the new four tuple.
 757  *
 758  * An error on the child should not affect the parent, so this
 759  * function does not return a value.
 760  */
 761 static void cm_conn_req_handler(struct iwcm_id_private *listen_id_priv,
 762                                 struct iw_cm_event *iw_event)
 763 {
 764         unsigned long flags;
 765         struct iw_cm_id *cm_id;
 766         struct iwcm_id_private *cm_id_priv;
 767         int ret;
 768
 769         /*
 770          * The provider should never generate a connection request
 771          * event with a bad status.
 772          */
 773         BUG_ON(iw_event->status);
 774
 775         cm_id = iw_create_cm_id(listen_id_priv->id.device,
 776                                 listen_id_priv->id.cm_handler,
 777                                 listen_id_priv->id.context);
 778         /* If the cm_id could not be created, ignore the request */
 779         if (IS_ERR(cm_id))
 780                 goto out;
 781
 782         cm_id->provider_data = iw_event->provider_data;
 783         cm_id->m_local_addr = iw_event->local_addr;
 784         cm_id->m_remote_addr = iw_event->remote_addr;
 785         cm_id->local_addr = listen_id_priv->id.local_addr;
 786
 787         ret = iwpm_get_remote_info(&listen_id_priv->id.m_local_addr,
 788                                    &iw_event->remote_addr,
 789                                    &cm_id->remote_addr,
 790                                    RDMA_NL_IWCM);
 791         if (ret) {
 792                 cm_id->remote_addr = iw_event->remote_addr;
 793         } else {
 794                 iw_cm_check_wildcard(&listen_id_priv->id.m_local_addr,
 795                                      &iw_event->local_addr,
 796                                      &cm_id->local_addr);
 797                 iw_event->local_addr = cm_id->local_addr;
 798                 iw_event->remote_addr = cm_id->remote_addr;
 799         }
 800
 801         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 802         cm_id_priv->state = IW_CM_STATE_CONN_RECV;
 803
 804         /*
 805          * We could be destroying the listening id. If so, ignore this
 806          * upcall.
 807          */
 808         spin_lock_irqsave(&listen_id_priv->lock, flags);
 809         if (listen_id_priv->state != IW_CM_STATE_LISTEN) {
 810                 spin_unlock_irqrestore(&listen_id_priv->lock, flags);
 811                 iw_cm_reject(cm_id, NULL, 0);
 812                 iw_destroy_cm_id(cm_id);
 813                 goto out;
 814         }
 815         spin_unlock_irqrestore(&listen_id_priv->lock, flags);
 816
 817         ret = alloc_work_entries(cm_id_priv, 3);
 818         if (ret) {
 819                 iw_cm_reject(cm_id, NULL, 0);
 820                 iw_destroy_cm_id(cm_id);
 821                 goto out;
 822         }
 823
 824         /* Call the client CM handler */
 825         ret = cm_id->cm_handler(cm_id, iw_event);
 826         if (ret) {
 827                 iw_cm_reject(cm_id, NULL, 0);
 828                 iw_destroy_cm_id(cm_id);
 829         }
 830
 831 out:
 832         if (iw_event->private_data_len)
 833                 kfree(iw_event->private_data);
 834 }
 835
 836 /*
 837  * Passive Side: CM_ID <-- ESTABLISHED
 838  *
 839  * The provider generated an ESTABLISHED event which means that
 840  * the MPA negotion has completed successfully and we are now in MPA
 841  * FPDU mode.
 842  *
 843  * This event can only be received in the CONN_RECV state. If the
 844  * remote peer closed, the ESTABLISHED event would be received followed
 845  * by the CLOSE event. If the app closes, it will block until we wake
 846  * it up after processing this event.
 847  */
 848 static int cm_conn_est_handler(struct iwcm_id_private *cm_id_priv,
 849                                struct iw_cm_event *iw_event)
 850 {
 851         unsigned long flags;
 852         int ret;
 853
 854         spin_lock_irqsave(&cm_id_priv->lock, flags);
 855
 856         /*
 857          * We clear the CONNECT_WAIT bit here to allow the callback
 858          * function to call iw_cm_disconnect. Calling iw_destroy_cm_id
 859          * from a callback handler is not allowed.
 860          */
 861         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 862         BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
 863         cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
 864         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 865         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 866         wake_up_all(&cm_id_priv->connect_wait);
 867
 868         return ret;
 869 }
 870
 871 /*
 872  * Active Side: CM_ID <-- ESTABLISHED
 873  *
 874  * The app has called connect and is waiting for the established event to
 875  * post it's requests to the server. This event will wake up anyone
 876  * blocked in iw_cm_disconnect or iw_destroy_id.
 877  */
 878 static int cm_conn_rep_handler(struct iwcm_id_private *cm_id_priv,
 879                                struct iw_cm_event *iw_event)
 880 {
 881         unsigned long flags;
 882         int ret;
 883
 884         spin_lock_irqsave(&cm_id_priv->lock, flags);
 885         /*
 886          * Clear the connect wait bit so a callback function calling
 887          * iw_cm_disconnect will not wait and deadlock this thread
 888          */
 889         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 890         BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
 891         if (iw_event->status == 0) {
 892                 cm_id_priv->id.m_local_addr = iw_event->local_addr;
 893                 cm_id_priv->id.m_remote_addr = iw_event->remote_addr;
 894                 iw_event->local_addr = cm_id_priv->id.local_addr;
 895                 iw_event->remote_addr = cm_id_priv->id.remote_addr;
 896                 cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
 897         } else {
 898                 /* REJECTED or RESET */
 899                 cm_id_priv->id.device->ops.iw_rem_ref(cm_id_priv->qp);
 900                 cm_id_priv->qp = NULL;
 901                 cm_id_priv->state = IW_CM_STATE_IDLE;
 902         }
 903         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 904         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 905
 906         if (iw_event->private_data_len)
 907                 kfree(iw_event->private_data);
 908
 909         /* Wake up waiters on connect complete */
 910         wake_up_all(&cm_id_priv->connect_wait);
 911
 912         return ret;
 913 }
 914
 915 /*
 916  * CM_ID <-- CLOSING
 917  *
 918  * If in the ESTABLISHED state, move to CLOSING.
 919  */
 920 static void cm_disconnect_handler(struct iwcm_id_private *cm_id_priv,
 921                                   struct iw_cm_event *iw_event)
 922 {
 923         unsigned long flags;
 924
 925         spin_lock_irqsave(&cm_id_priv->lock, flags);
 926         if (cm_id_priv->state == IW_CM_STATE_ESTABLISHED)
 927                 cm_id_priv->state = IW_CM_STATE_CLOSING;
 928         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 929 }
 930
 931 /*
 932  * CM_ID <-- IDLE
 933  *
 934  * If in the ESTBLISHED or CLOSING states, the QP will have have been
 935  * moved by the provider to the ERR state. Disassociate the CM_ID from
 936  * the QP,  move to IDLE, and remove the 'connected' reference.
 937  *
 938  * If in some other state, the cm_id was destroyed asynchronously.
 939  * This is the last reference that will result in waking up
 940  * the app thread blocked in iw_destroy_cm_id.
 941  */
 942 static int cm_close_handler(struct iwcm_id_private *cm_id_priv,
 943                                   struct iw_cm_event *iw_event)
 944 {
 945         unsigned long flags;
 946         int ret = 0;
 947         spin_lock_irqsave(&cm_id_priv->lock, flags);
 948
 949         if (cm_id_priv->qp) {
 950                 cm_id_priv->id.device->ops.iw_rem_ref(cm_id_priv->qp);
 951                 cm_id_priv->qp = NULL;
 952         }
 953         switch (cm_id_priv->state) {
 954         case IW_CM_STATE_ESTABLISHED:
 955         case IW_CM_STATE_CLOSING:
 956                 cm_id_priv->state = IW_CM_STATE_IDLE;
 957                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 958                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 959                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 960                 break;
 961         case IW_CM_STATE_DESTROYING:
 962                 break;
 963         default:
 964                 BUG();
 965         }
 966         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 967
 968         return ret;
 969 }
 970
 971 static int process_event(struct iwcm_id_private *cm_id_priv,
 972                          struct iw_cm_event *iw_event)
 973 {
 974         int ret = 0;
 975
 976         switch (iw_event->event) {
 977         case IW_CM_EVENT_CONNECT_REQUEST:
 978                 cm_conn_req_handler(cm_id_priv, iw_event);
 979                 break;
 980         case IW_CM_EVENT_CONNECT_REPLY:
 981                 ret = cm_conn_rep_handler(cm_id_priv, iw_event);
 982                 break;
 983         case IW_CM_EVENT_ESTABLISHED:
 984                 ret = cm_conn_est_handler(cm_id_priv, iw_event);
 985                 break;
 986         case IW_CM_EVENT_DISCONNECT:
 987                 cm_disconnect_handler(cm_id_priv, iw_event);
 988                 break;
 989         case IW_CM_EVENT_CLOSE:
 990                 ret = cm_close_handler(cm_id_priv, iw_event);
 991                 break;
 992         default:
 993                 BUG();
 994         }
 995
 996         return ret;
 997 }
 998
 999 /*
1000  * Process events on the work_list for the cm_id. If the callback
1001  * function requests that the cm_id be deleted, a flag is set in the
1002  * cm_id flags to indicate that when the last reference is
1003  * removed, the cm_id is to be destroyed. This is necessary to
1004  * distinguish between an object that will be destroyed by the app
1005  * thread asleep on the destroy_comp list vs. an object destroyed
1006  * here synchronously when the last reference is removed.
1007  */
1008 static void cm_work_handler(struct work_struct *_work)
1009 {
1010         struct iwcm_work *work = container_of(_work, struct iwcm_work, work);
1011         struct iw_cm_event levent;
1012         struct iwcm_id_private *cm_id_priv = work->cm_id;
1013         unsigned long flags;
1014         int empty;
1015         int ret = 0;
1016
1017         spin_lock_irqsave(&cm_id_priv->lock, flags);
1018         empty = list_empty(&cm_id_priv->work_list);
1019         while (!empty) {
1020                 work = list_entry(cm_id_priv->work_list.next,
1021                                   struct iwcm_work, list);
1022                 list_del_init(&work->list);
1023                 empty = list_empty(&cm_id_priv->work_list);
1024                 levent = work->event;
1025                 put_work(work);
1026                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1027
1028                 if (!test_bit(IWCM_F_DROP_EVENTS, &cm_id_priv->flags)) {
1029                         ret = process_event(cm_id_priv, &levent);
1030                         if (ret)
1031                                 destroy_cm_id(&cm_id_priv->id);
1032                 } else
1033                         pr_debug("dropping event %d\n", levent.event);
1034                 if (iwcm_deref_id(cm_id_priv))
1035                         return;
1036                 if (empty)
1037                         return;
1038                 spin_lock_irqsave(&cm_id_priv->lock, flags);
1039         }
1040         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1041 }
1042
1043 /*
1044  * This function is called on interrupt context. Schedule events on
1045  * the iwcm_wq thread to allow callback functions to downcall into
1046  * the CM and/or block.  Events are queued to a per-CM_ID
1047  * work_list. If this is the first event on the work_list, the work
1048  * element is also queued on the iwcm_wq thread.
1049  *
1050  * Each event holds a reference on the cm_id. Until the last posted
1051  * event has been delivered and processed, the cm_id cannot be
1052  * deleted.
1053  *
1054  * Returns:
1055  *            0 - the event was handled.
1056  *      -ENOMEM - the event was not handled due to lack of resources.
1057  */
1058 static int cm_event_handler(struct iw_cm_id *cm_id,
1059                              struct iw_cm_event *iw_event)
1060 {
1061         struct iwcm_work *work;
1062         struct iwcm_id_private *cm_id_priv;
1063         unsigned long flags;
1064         int ret = 0;
1065
1066         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1067
1068         spin_lock_irqsave(&cm_id_priv->lock, flags);
1069         work = get_work(cm_id_priv);
1070         if (!work) {
1071                 ret = -ENOMEM;
1072                 goto out;
1073         }
1074
1075         INIT_WORK(&work->work, cm_work_handler);
1076         work->cm_id = cm_id_priv;
1077         work->event = *iw_event;
1078
1079         if ((work->event.event == IW_CM_EVENT_CONNECT_REQUEST ||
1080              work->event.event == IW_CM_EVENT_CONNECT_REPLY) &&
1081             work->event.private_data_len) {
1082                 ret = copy_private_data(&work->event);
1083                 if (ret) {
1084                         put_work(work);
1085                         goto out;
1086                 }
1087         }
1088
1089         atomic_inc(&cm_id_priv->refcount);
1090         if (list_empty(&cm_id_priv->work_list)) {
1091                 list_add_tail(&work->list, &cm_id_priv->work_list);
1092                 queue_work(iwcm_wq, &work->work);
1093         } else
1094                 list_add_tail(&work->list, &cm_id_priv->work_list);
1095 out:
1096         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1097         return ret;
1098 }
1099
1100 static int iwcm_init_qp_init_attr(struct iwcm_id_private *cm_id_priv,
1101                                   struct ib_qp_attr *qp_attr,
1102                                   int *qp_attr_mask)
1103 {
1104         unsigned long flags;
1105         int ret;
1106
1107         spin_lock_irqsave(&cm_id_priv->lock, flags);
1108         switch (cm_id_priv->state) {
1109         case IW_CM_STATE_IDLE:
1110         case IW_CM_STATE_CONN_SENT:
1111         case IW_CM_STATE_CONN_RECV:
1112         case IW_CM_STATE_ESTABLISHED:
1113                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1114                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE|
1115                                            IB_ACCESS_REMOTE_READ;
1116                 ret = 0;
1117                 break;
1118         default:
1119                 ret = -EINVAL;
1120                 break;
1121         }
1122         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1123         return ret;
1124 }
1125
1126 static int iwcm_init_qp_rts_attr(struct iwcm_id_private *cm_id_priv,
1127                                   struct ib_qp_attr *qp_attr,
1128                                   int *qp_attr_mask)
1129 {
1130         unsigned long flags;
1131         int ret;
1132
1133         spin_lock_irqsave(&cm_id_priv->lock, flags);
1134         switch (cm_id_priv->state) {
1135         case IW_CM_STATE_IDLE:
1136         case IW_CM_STATE_CONN_SENT:
1137         case IW_CM_STATE_CONN_RECV:
1138         case IW_CM_STATE_ESTABLISHED:
1139                 *qp_attr_mask = 0;
1140                 ret = 0;
1141                 break;
1142         default:
1143                 ret = -EINVAL;
1144                 break;
1145         }
1146         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1147         return ret;
1148 }
1149
1150 int iw_cm_init_qp_attr(struct iw_cm_id *cm_id,
1151                        struct ib_qp_attr *qp_attr,
1152                        int *qp_attr_mask)
1153 {
1154         struct iwcm_id_private *cm_id_priv;
1155         int ret;
1156
1157         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1158         switch (qp_attr->qp_state) {
1159         case IB_QPS_INIT:
1160         case IB_QPS_RTR:
1161                 ret = iwcm_init_qp_init_attr(cm_id_priv,
1162                                              qp_attr, qp_attr_mask);
1163                 break;
1164         case IB_QPS_RTS:
1165                 ret = iwcm_init_qp_rts_attr(cm_id_priv,
1166                                             qp_attr, qp_attr_mask);
1167                 break;
1168         default:
1169                 ret = -EINVAL;
1170                 break;
1171         }
1172         return ret;
1173 }
1174 EXPORT_SYMBOL(iw_cm_init_qp_attr);
1175
1176 static int __init iw_cm_init(void)
1177 {
1178         int ret;
1179
1180         ret = iwpm_init(RDMA_NL_IWCM);
1181         if (ret)
1182                 pr_err("iw_cm: couldn't init iwpm\n");
1183         else
1184                 rdma_nl_register(RDMA_NL_IWCM, iwcm_nl_cb_table);
1185         iwcm_wq = alloc_ordered_workqueue("iw_cm_wq", 0);
1186         if (!iwcm_wq)
1187                 return -ENOMEM;
1188
1189         iwcm_ctl_table_hdr = register_net_sysctl(&init_net, "net/iw_cm",
1190                                                  iwcm_ctl_table);
1191         if (!iwcm_ctl_table_hdr) {
1192                 pr_err("iw_cm: couldn't register sysctl paths\n");
1193                 destroy_workqueue(iwcm_wq);
1194                 return -ENOMEM;
1195         }
1196
1197         return 0;
1198 }
1199
1200 static void __exit iw_cm_cleanup(void)
1201 {
1202         unregister_net_sysctl_table(iwcm_ctl_table_hdr);
1203         destroy_workqueue(iwcm_wq);
1204         rdma_nl_unregister(RDMA_NL_IWCM);
1205         iwpm_exit(RDMA_NL_IWCM);
1206 }
1207
1208 MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_IWCM, 2);
1209
1210 module_init(iw_cm_init);
1211 module_exit(iw_cm_cleanup);