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                 list_del(e);
 164                 kfree(list_entry(e, struct iwcm_work, free_list));
 165         }
 166 }
 167
 168 static int alloc_work_entries(struct iwcm_id_private *cm_id_priv, int count)
 169 {
 170         struct iwcm_work *work;
 171
 172         BUG_ON(!list_empty(&cm_id_priv->work_free_list));
 173         while (count--) {
 174                 work = kmalloc(sizeof(struct iwcm_work), GFP_KERNEL);
 175                 if (!work) {
 176                         dealloc_work_entries(cm_id_priv);
 177                         return -ENOMEM;
 178                 }
 179                 work->cm_id = cm_id_priv;
 180                 INIT_LIST_HEAD(&work->list);
 181                 put_work(work);
 182         }
 183         return 0;
 184 }
 185
 186 /*
 187  * Save private data from incoming connection requests to
 188  * iw_cm_event, so the low level driver doesn't have to. Adjust
 189  * the event ptr to point to the local copy.
 190  */
 191 static int copy_private_data(struct iw_cm_event *event)
 192 {
 193         void *p;
 194
 195         p = kmemdup(event->private_data, event->private_data_len, GFP_ATOMIC);
 196         if (!p)
 197                 return -ENOMEM;
 198         event->private_data = p;
 199         return 0;
 200 }
 201
 202 static void free_cm_id(struct iwcm_id_private *cm_id_priv)
 203 {
 204         dealloc_work_entries(cm_id_priv);
 205         kfree(cm_id_priv);
 206 }
 207
 208 /*
 209  * Release a reference on cm_id. If the last reference is being
 210  * released, free the cm_id and return 1.
 211  */
 212 static int iwcm_deref_id(struct iwcm_id_private *cm_id_priv)
 213 {
 214         BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
 215         if (atomic_dec_and_test(&cm_id_priv->refcount)) {
 216                 BUG_ON(!list_empty(&cm_id_priv->work_list));
 217                 free_cm_id(cm_id_priv);
 218                 return 1;
 219         }
 220
 221         return 0;
 222 }
 223
 224 static void add_ref(struct iw_cm_id *cm_id)
 225 {
 226         struct iwcm_id_private *cm_id_priv;
 227         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 228         atomic_inc(&cm_id_priv->refcount);
 229 }
 230
 231 static void rem_ref(struct iw_cm_id *cm_id)
 232 {
 233         struct iwcm_id_private *cm_id_priv;
 234
 235         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 236
 237         (void)iwcm_deref_id(cm_id_priv);
 238 }
 239
 240 static int cm_event_handler(struct iw_cm_id *cm_id, struct iw_cm_event *event);
 241
 242 struct iw_cm_id *iw_create_cm_id(struct ib_device *device,
 243                                  iw_cm_handler cm_handler,
 244                                  void *context)
 245 {
 246         struct iwcm_id_private *cm_id_priv;
 247
 248         cm_id_priv = kzalloc(sizeof(*cm_id_priv), GFP_KERNEL);
 249         if (!cm_id_priv)
 250                 return ERR_PTR(-ENOMEM);
 251
 252         cm_id_priv->state = IW_CM_STATE_IDLE;
 253         cm_id_priv->id.device = device;
 254         cm_id_priv->id.cm_handler = cm_handler;
 255         cm_id_priv->id.context = context;
 256         cm_id_priv->id.event_handler = cm_event_handler;
 257         cm_id_priv->id.add_ref = add_ref;
 258         cm_id_priv->id.rem_ref = rem_ref;
 259         spin_lock_init(&cm_id_priv->lock);
 260         atomic_set(&cm_id_priv->refcount, 1);
 261         init_waitqueue_head(&cm_id_priv->connect_wait);
 262         init_completion(&cm_id_priv->destroy_comp);
 263         INIT_LIST_HEAD(&cm_id_priv->work_list);
 264         INIT_LIST_HEAD(&cm_id_priv->work_free_list);
 265
 266         return &cm_id_priv->id;
 267 }
 268 EXPORT_SYMBOL(iw_create_cm_id);
 269
 270
 271 static int iwcm_modify_qp_err(struct ib_qp *qp)
 272 {
 273         struct ib_qp_attr qp_attr;
 274
 275         if (!qp)
 276                 return -EINVAL;
 277
 278         qp_attr.qp_state = IB_QPS_ERR;
 279         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 280 }
 281
 282 /*
 283  * This is really the RDMAC CLOSING state. It is most similar to the
 284  * IB SQD QP state.
 285  */
 286 static int iwcm_modify_qp_sqd(struct ib_qp *qp)
 287 {
 288         struct ib_qp_attr qp_attr;
 289
 290         BUG_ON(qp == NULL);
 291         qp_attr.qp_state = IB_QPS_SQD;
 292         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 293 }
 294
 295 /*
 296  * CM_ID <-- CLOSING
 297  *
 298  * Block if a passive or active connection is currently being processed. Then
 299  * process the event as follows:
 300  * - If we are ESTABLISHED, move to CLOSING and modify the QP state
 301  *   based on the abrupt flag
 302  * - If the connection is already in the CLOSING or IDLE state, the peer is
 303  *   disconnecting concurrently with us and we've already seen the
 304  *   DISCONNECT event -- ignore the request and return 0
 305  * - Disconnect on a listening endpoint returns -EINVAL
 306  */
 307 int iw_cm_disconnect(struct iw_cm_id *cm_id, int abrupt)
 308 {
 309         struct iwcm_id_private *cm_id_priv;
 310         unsigned long flags;
 311         int ret = 0;
 312         struct ib_qp *qp = NULL;
 313
 314         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 315         /* Wait if we're currently in a connect or accept downcall */
 316         wait_event(cm_id_priv->connect_wait,
 317                    !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
 318
 319         spin_lock_irqsave(&cm_id_priv->lock, flags);
 320         switch (cm_id_priv->state) {
 321         case IW_CM_STATE_ESTABLISHED:
 322                 cm_id_priv->state = IW_CM_STATE_CLOSING;
 323
 324                 /* QP could be <nul> for user-mode client */
 325                 if (cm_id_priv->qp)
 326                         qp = cm_id_priv->qp;
 327                 else
 328                         ret = -EINVAL;
 329                 break;
 330         case IW_CM_STATE_LISTEN:
 331                 ret = -EINVAL;
 332                 break;
 333         case IW_CM_STATE_CLOSING:
 334                 /* remote peer closed first */
 335         case IW_CM_STATE_IDLE:
 336                 /* accept or connect returned !0 */
 337                 break;
 338         case IW_CM_STATE_CONN_RECV:
 339                 /*
 340                  * App called disconnect before/without calling accept after
 341                  * connect_request event delivered.
 342                  */
 343                 break;
 344         case IW_CM_STATE_CONN_SENT:
 345                 /* Can only get here if wait above fails */
 346         default:
 347                 BUG();
 348         }
 349         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 350
 351         if (qp) {
 352                 if (abrupt)
 353                         ret = iwcm_modify_qp_err(qp);
 354                 else
 355                         ret = iwcm_modify_qp_sqd(qp);
 356
 357                 /*
 358                  * If both sides are disconnecting the QP could
 359                  * already be in ERR or SQD states
 360                  */
 361                 ret = 0;
 362         }
 363
 364         return ret;
 365 }
 366 EXPORT_SYMBOL(iw_cm_disconnect);
 367
 368 /*
 369  * CM_ID <-- DESTROYING
 370  *
 371  * Clean up all resources associated with the connection and release
 372  * the initial reference taken by iw_create_cm_id.
 373  */
 374 static void destroy_cm_id(struct iw_cm_id *cm_id)
 375 {
 376         struct iwcm_id_private *cm_id_priv;
 377         struct ib_qp *qp;
 378         unsigned long flags;
 379
 380         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 381         /*
 382          * Wait if we're currently in a connect or accept downcall. A
 383          * listening endpoint should never block here.
 384          */
 385         wait_event(cm_id_priv->connect_wait,
 386                    !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
 387
 388         /*
 389          * Since we're deleting the cm_id, drop any events that
 390          * might arrive before the last dereference.
 391          */
 392         set_bit(IWCM_F_DROP_EVENTS, &cm_id_priv->flags);
 393
 394         spin_lock_irqsave(&cm_id_priv->lock, flags);
 395         qp = cm_id_priv->qp;
 396         cm_id_priv->qp = NULL;
 397
 398         switch (cm_id_priv->state) {
 399         case IW_CM_STATE_LISTEN:
 400                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 401                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 402                 /* destroy the listening endpoint */
 403                 cm_id->device->ops.iw_destroy_listen(cm_id);
 404                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 405                 break;
 406         case IW_CM_STATE_ESTABLISHED:
 407                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 408                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 409                 /* Abrupt close of the connection */
 410                 (void)iwcm_modify_qp_err(qp);
 411                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 412                 break;
 413         case IW_CM_STATE_IDLE:
 414         case IW_CM_STATE_CLOSING:
 415                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 416                 break;
 417         case IW_CM_STATE_CONN_RECV:
 418                 /*
 419                  * App called destroy before/without calling accept after
 420                  * receiving connection request event notification or
 421                  * returned non zero from the event callback function.
 422                  * In either case, must tell the provider to reject.
 423                  */
 424                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 425                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 426                 cm_id->device->ops.iw_reject(cm_id, NULL, 0);
 427                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 428                 break;
 429         case IW_CM_STATE_CONN_SENT:
 430         case IW_CM_STATE_DESTROYING:
 431         default:
 432                 BUG();
 433                 break;
 434         }
 435         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 436         if (qp)
 437                 cm_id_priv->id.device->ops.iw_rem_ref(qp);
 438
 439         if (cm_id->mapped) {
 440                 iwpm_remove_mapinfo(&cm_id->local_addr, &cm_id->m_local_addr);
 441                 iwpm_remove_mapping(&cm_id->local_addr, RDMA_NL_IWCM);
 442         }
 443
 444         (void)iwcm_deref_id(cm_id_priv);
 445 }
 446
 447 /*
 448  * This function is only called by the application thread and cannot
 449  * be called by the event thread. The function will wait for all
 450  * references to be released on the cm_id and then kfree the cm_id
 451  * object.
 452  */
 453 void iw_destroy_cm_id(struct iw_cm_id *cm_id)
 454 {
 455         destroy_cm_id(cm_id);
 456 }
 457 EXPORT_SYMBOL(iw_destroy_cm_id);
 458
 459 /**
 460  * iw_cm_check_wildcard - If IP address is 0 then use original
 461  * @pm_addr: sockaddr containing the ip to check for wildcard
 462  * @cm_addr: sockaddr containing the actual IP address
 463  * @cm_outaddr: sockaddr to set IP addr which leaving port
 464  *
 465  *  Checks the pm_addr for wildcard and then sets cm_outaddr's
 466  *  IP to the actual (cm_addr).
 467  */
 468 static void iw_cm_check_wildcard(struct sockaddr_storage *pm_addr,
 469                                  struct sockaddr_storage *cm_addr,
 470                                  struct sockaddr_storage *cm_outaddr)
 471 {
 472         if (pm_addr->ss_family == AF_INET) {
 473                 struct sockaddr_in *pm4_addr = (struct sockaddr_in *)pm_addr;
 474
 475                 if (pm4_addr->sin_addr.s_addr == htonl(INADDR_ANY)) {
 476                         struct sockaddr_in *cm4_addr =
 477                                 (struct sockaddr_in *)cm_addr;
 478                         struct sockaddr_in *cm4_outaddr =
 479                                 (struct sockaddr_in *)cm_outaddr;
 480
 481                         cm4_outaddr->sin_addr = cm4_addr->sin_addr;
 482                 }
 483         } else {
 484                 struct sockaddr_in6 *pm6_addr = (struct sockaddr_in6 *)pm_addr;
 485
 486                 if (ipv6_addr_type(&pm6_addr->sin6_addr) == IPV6_ADDR_ANY) {
 487                         struct sockaddr_in6 *cm6_addr =
 488                                 (struct sockaddr_in6 *)cm_addr;
 489                         struct sockaddr_in6 *cm6_outaddr =
 490                                 (struct sockaddr_in6 *)cm_outaddr;
 491
 492                         cm6_outaddr->sin6_addr = cm6_addr->sin6_addr;
 493                 }
 494         }
 495 }
 496
 497 /**
 498  * iw_cm_map - Use portmapper to map the ports
 499  * @cm_id: connection manager pointer
 500  * @active: Indicates the active side when true
 501  * returns nonzero for error only if iwpm_create_mapinfo() fails
 502  *
 503  * Tries to add a mapping for a port using the Portmapper. If
 504  * successful in mapping the IP/Port it will check the remote
 505  * mapped IP address for a wildcard IP address and replace the
 506  * zero IP address with the remote_addr.
 507  */
 508 static int iw_cm_map(struct iw_cm_id *cm_id, bool active)
 509 {
 510         const char *devname = dev_name(&cm_id->device->dev);
 511         const char *ifname = cm_id->device->iw_ifname;
 512         struct iwpm_dev_data pm_reg_msg = {};
 513         struct iwpm_sa_data pm_msg;
 514         int status;
 515
 516         if (strlen(devname) >= sizeof(pm_reg_msg.dev_name) ||
 517             strlen(ifname) >= sizeof(pm_reg_msg.if_name))
 518                 return -EINVAL;
 519
 520         cm_id->m_local_addr = cm_id->local_addr;
 521         cm_id->m_remote_addr = cm_id->remote_addr;
 522
 523         strcpy(pm_reg_msg.dev_name, devname);
 524         strcpy(pm_reg_msg.if_name, ifname);
 525
 526         if (iwpm_register_pid(&pm_reg_msg, RDMA_NL_IWCM) ||
 527             !iwpm_valid_pid())
 528                 return 0;
 529
 530         cm_id->mapped = true;
 531         pm_msg.loc_addr = cm_id->local_addr;
 532         pm_msg.rem_addr = cm_id->remote_addr;
 533         pm_msg.flags = (cm_id->device->iw_driver_flags & IW_F_NO_PORT_MAP) ?
 534                        IWPM_FLAGS_NO_PORT_MAP : 0;
 535         if (active)
 536                 status = iwpm_add_and_query_mapping(&pm_msg,
 537                                                     RDMA_NL_IWCM);
 538         else
 539                 status = iwpm_add_mapping(&pm_msg, RDMA_NL_IWCM);
 540
 541         if (!status) {
 542                 cm_id->m_local_addr = pm_msg.mapped_loc_addr;
 543                 if (active) {
 544                         cm_id->m_remote_addr = pm_msg.mapped_rem_addr;
 545                         iw_cm_check_wildcard(&pm_msg.mapped_rem_addr,
 546                                              &cm_id->remote_addr,
 547                                              &cm_id->m_remote_addr);
 548                 }
 549         }
 550
 551         return iwpm_create_mapinfo(&cm_id->local_addr,
 552                                    &cm_id->m_local_addr,
 553                                    RDMA_NL_IWCM, pm_msg.flags);
 554 }
 555
 556 /*
 557  * CM_ID <-- LISTEN
 558  *
 559  * Start listening for connect requests. Generates one CONNECT_REQUEST
 560  * event for each inbound connect request.
 561  */
 562 int iw_cm_listen(struct iw_cm_id *cm_id, int backlog)
 563 {
 564         struct iwcm_id_private *cm_id_priv;
 565         unsigned long flags;
 566         int ret;
 567
 568         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 569
 570         if (!backlog)
 571                 backlog = default_backlog;
 572
 573         ret = alloc_work_entries(cm_id_priv, backlog);
 574         if (ret)
 575                 return ret;
 576
 577         spin_lock_irqsave(&cm_id_priv->lock, flags);
 578         switch (cm_id_priv->state) {
 579         case IW_CM_STATE_IDLE:
 580                 cm_id_priv->state = IW_CM_STATE_LISTEN;
 581                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 582                 ret = iw_cm_map(cm_id, false);
 583                 if (!ret)
 584                         ret = cm_id->device->ops.iw_create_listen(cm_id,
 585                                                                   backlog);
 586                 if (ret)
 587                         cm_id_priv->state = IW_CM_STATE_IDLE;
 588                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 589                 break;
 590         default:
 591                 ret = -EINVAL;
 592         }
 593         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 594
 595         return ret;
 596 }
 597 EXPORT_SYMBOL(iw_cm_listen);
 598
 599 /*
 600  * CM_ID <-- IDLE
 601  *
 602  * Rejects an inbound connection request. No events are generated.
 603  */
 604 int iw_cm_reject(struct iw_cm_id *cm_id,
 605                  const void *private_data,
 606                  u8 private_data_len)
 607 {
 608         struct iwcm_id_private *cm_id_priv;
 609         unsigned long flags;
 610         int ret;
 611
 612         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 613         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 614
 615         spin_lock_irqsave(&cm_id_priv->lock, flags);
 616         if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
 617                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 618                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 619                 wake_up_all(&cm_id_priv->connect_wait);
 620                 return -EINVAL;
 621         }
 622         cm_id_priv->state = IW_CM_STATE_IDLE;
 623         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 624
 625         ret = cm_id->device->ops.iw_reject(cm_id, private_data,
 626                                           private_data_len);
 627
 628         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 629         wake_up_all(&cm_id_priv->connect_wait);
 630
 631         return ret;
 632 }
 633 EXPORT_SYMBOL(iw_cm_reject);
 634
 635 /*
 636  * CM_ID <-- ESTABLISHED
 637  *
 638  * Accepts an inbound connection request and generates an ESTABLISHED
 639  * event. Callers of iw_cm_disconnect and iw_destroy_cm_id will block
 640  * until the ESTABLISHED event is received from the provider.
 641  */
 642 int iw_cm_accept(struct iw_cm_id *cm_id,
 643                  struct iw_cm_conn_param *iw_param)
 644 {
 645         struct iwcm_id_private *cm_id_priv;
 646         struct ib_qp *qp;
 647         unsigned long flags;
 648         int ret;
 649
 650         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 651         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 652
 653         spin_lock_irqsave(&cm_id_priv->lock, flags);
 654         if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
 655                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 656                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 657                 wake_up_all(&cm_id_priv->connect_wait);
 658                 return -EINVAL;
 659         }
 660         /* Get the ib_qp given the QPN */
 661         qp = cm_id->device->ops.iw_get_qp(cm_id->device, iw_param->qpn);
 662         if (!qp) {
 663                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 664                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 665                 wake_up_all(&cm_id_priv->connect_wait);
 666                 return -EINVAL;
 667         }
 668         cm_id->device->ops.iw_add_ref(qp);
 669         cm_id_priv->qp = qp;
 670         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 671
 672         ret = cm_id->device->ops.iw_accept(cm_id, iw_param);
 673         if (ret) {
 674                 /* An error on accept precludes provider events */
 675                 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
 676                 cm_id_priv->state = IW_CM_STATE_IDLE;
 677                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 678                 qp = cm_id_priv->qp;
 679                 cm_id_priv->qp = NULL;
 680                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 681                 if (qp)
 682                         cm_id->device->ops.iw_rem_ref(qp);
 683                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 684                 wake_up_all(&cm_id_priv->connect_wait);
 685         }
 686
 687         return ret;
 688 }
 689 EXPORT_SYMBOL(iw_cm_accept);
 690
 691 /*
 692  * Active Side: CM_ID <-- CONN_SENT
 693  *
 694  * If successful, results in the generation of a CONNECT_REPLY
 695  * event. iw_cm_disconnect and iw_cm_destroy will block until the
 696  * CONNECT_REPLY event is received from the provider.
 697  */
 698 int iw_cm_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *iw_param)
 699 {
 700         struct iwcm_id_private *cm_id_priv;
 701         int ret;
 702         unsigned long flags;
 703         struct ib_qp *qp = NULL;
 704
 705         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 706
 707         ret = alloc_work_entries(cm_id_priv, 4);
 708         if (ret)
 709                 return ret;
 710
 711         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 712         spin_lock_irqsave(&cm_id_priv->lock, flags);
 713
 714         if (cm_id_priv->state != IW_CM_STATE_IDLE) {
 715                 ret = -EINVAL;
 716                 goto err;
 717         }
 718
 719         /* Get the ib_qp given the QPN */
 720         qp = cm_id->device->ops.iw_get_qp(cm_id->device, iw_param->qpn);
 721         if (!qp) {
 722                 ret = -EINVAL;
 723                 goto err;
 724         }
 725         cm_id->device->ops.iw_add_ref(qp);
 726         cm_id_priv->qp = qp;
 727         cm_id_priv->state = IW_CM_STATE_CONN_SENT;
 728         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 729
 730         ret = iw_cm_map(cm_id, true);
 731         if (!ret)
 732                 ret = cm_id->device->ops.iw_connect(cm_id, iw_param);
 733         if (!ret)
 734                 return 0;       /* success */
 735
 736         spin_lock_irqsave(&cm_id_priv->lock, flags);
 737         qp = cm_id_priv->qp;
 738         cm_id_priv->qp = NULL;
 739         cm_id_priv->state = IW_CM_STATE_IDLE;
 740 err:
 741         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 742         if (qp)
 743                 cm_id->device->ops.iw_rem_ref(qp);
 744         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 745         wake_up_all(&cm_id_priv->connect_wait);
 746         return ret;
 747 }
 748 EXPORT_SYMBOL(iw_cm_connect);
 749
 750 /*
 751  * Passive Side: new CM_ID <-- CONN_RECV
 752  *
 753  * Handles an inbound connect request. The function creates a new
 754  * iw_cm_id to represent the new connection and inherits the client
 755  * callback function and other attributes from the listening parent.
 756  *
 757  * The work item contains a pointer to the listen_cm_id and the event. The
 758  * listen_cm_id contains the client cm_handler, context and
 759  * device. These are copied when the device is cloned. The event
 760  * contains the new four tuple.
 761  *
 762  * An error on the child should not affect the parent, so this
 763  * function does not return a value.
 764  */
 765 static void cm_conn_req_handler(struct iwcm_id_private *listen_id_priv,
 766                                 struct iw_cm_event *iw_event)
 767 {
 768         unsigned long flags;
 769         struct iw_cm_id *cm_id;
 770         struct iwcm_id_private *cm_id_priv;
 771         int ret;
 772
 773         /*
 774          * The provider should never generate a connection request
 775          * event with a bad status.
 776          */
 777         BUG_ON(iw_event->status);
 778
 779         cm_id = iw_create_cm_id(listen_id_priv->id.device,
 780                                 listen_id_priv->id.cm_handler,
 781                                 listen_id_priv->id.context);
 782         /* If the cm_id could not be created, ignore the request */
 783         if (IS_ERR(cm_id))
 784                 goto out;
 785
 786         cm_id->provider_data = iw_event->provider_data;
 787         cm_id->m_local_addr = iw_event->local_addr;
 788         cm_id->m_remote_addr = iw_event->remote_addr;
 789         cm_id->local_addr = listen_id_priv->id.local_addr;
 790
 791         ret = iwpm_get_remote_info(&listen_id_priv->id.m_local_addr,
 792                                    &iw_event->remote_addr,
 793                                    &cm_id->remote_addr,
 794                                    RDMA_NL_IWCM);
 795         if (ret) {
 796                 cm_id->remote_addr = iw_event->remote_addr;
 797         } else {
 798                 iw_cm_check_wildcard(&listen_id_priv->id.m_local_addr,
 799                                      &iw_event->local_addr,
 800                                      &cm_id->local_addr);
 801                 iw_event->local_addr = cm_id->local_addr;
 802                 iw_event->remote_addr = cm_id->remote_addr;
 803         }
 804
 805         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 806         cm_id_priv->state = IW_CM_STATE_CONN_RECV;
 807
 808         /*
 809          * We could be destroying the listening id. If so, ignore this
 810          * upcall.
 811          */
 812         spin_lock_irqsave(&listen_id_priv->lock, flags);
 813         if (listen_id_priv->state != IW_CM_STATE_LISTEN) {
 814                 spin_unlock_irqrestore(&listen_id_priv->lock, flags);
 815                 iw_cm_reject(cm_id, NULL, 0);
 816                 iw_destroy_cm_id(cm_id);
 817                 goto out;
 818         }
 819         spin_unlock_irqrestore(&listen_id_priv->lock, flags);
 820
 821         ret = alloc_work_entries(cm_id_priv, 3);
 822         if (ret) {
 823                 iw_cm_reject(cm_id, NULL, 0);
 824                 iw_destroy_cm_id(cm_id);
 825                 goto out;
 826         }
 827
 828         /* Call the client CM handler */
 829         ret = cm_id->cm_handler(cm_id, iw_event);
 830         if (ret) {
 831                 iw_cm_reject(cm_id, NULL, 0);
 832                 iw_destroy_cm_id(cm_id);
 833         }
 834
 835 out:
 836         if (iw_event->private_data_len)
 837                 kfree(iw_event->private_data);
 838 }
 839
 840 /*
 841  * Passive Side: CM_ID <-- ESTABLISHED
 842  *
 843  * The provider generated an ESTABLISHED event which means that
 844  * the MPA negotion has completed successfully and we are now in MPA
 845  * FPDU mode.
 846  *
 847  * This event can only be received in the CONN_RECV state. If the
 848  * remote peer closed, the ESTABLISHED event would be received followed
 849  * by the CLOSE event. If the app closes, it will block until we wake
 850  * it up after processing this event.
 851  */
 852 static int cm_conn_est_handler(struct iwcm_id_private *cm_id_priv,
 853                                struct iw_cm_event *iw_event)
 854 {
 855         unsigned long flags;
 856         int ret;
 857
 858         spin_lock_irqsave(&cm_id_priv->lock, flags);
 859
 860         /*
 861          * We clear the CONNECT_WAIT bit here to allow the callback
 862          * function to call iw_cm_disconnect. Calling iw_destroy_cm_id
 863          * from a callback handler is not allowed.
 864          */
 865         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 866         BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
 867         cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
 868         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 869         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 870         wake_up_all(&cm_id_priv->connect_wait);
 871
 872         return ret;
 873 }
 874
 875 /*
 876  * Active Side: CM_ID <-- ESTABLISHED
 877  *
 878  * The app has called connect and is waiting for the established event to
 879  * post it's requests to the server. This event will wake up anyone
 880  * blocked in iw_cm_disconnect or iw_destroy_id.
 881  */
 882 static int cm_conn_rep_handler(struct iwcm_id_private *cm_id_priv,
 883                                struct iw_cm_event *iw_event)
 884 {
 885         struct ib_qp *qp = NULL;
 886         unsigned long flags;
 887         int ret;
 888
 889         spin_lock_irqsave(&cm_id_priv->lock, flags);
 890         /*
 891          * Clear the connect wait bit so a callback function calling
 892          * iw_cm_disconnect will not wait and deadlock this thread
 893          */
 894         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 895         BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
 896         if (iw_event->status == 0) {
 897                 cm_id_priv->id.m_local_addr = iw_event->local_addr;
 898                 cm_id_priv->id.m_remote_addr = iw_event->remote_addr;
 899                 iw_event->local_addr = cm_id_priv->id.local_addr;
 900                 iw_event->remote_addr = cm_id_priv->id.remote_addr;
 901                 cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
 902         } else {
 903                 /* REJECTED or RESET */
 904                 qp = cm_id_priv->qp;
 905                 cm_id_priv->qp = NULL;
 906                 cm_id_priv->state = IW_CM_STATE_IDLE;
 907         }
 908         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 909         if (qp)
 910                 cm_id_priv->id.device->ops.iw_rem_ref(qp);
 911         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 912
 913         if (iw_event->private_data_len)
 914                 kfree(iw_event->private_data);
 915
 916         /* Wake up waiters on connect complete */
 917         wake_up_all(&cm_id_priv->connect_wait);
 918
 919         return ret;
 920 }
 921
 922 /*
 923  * CM_ID <-- CLOSING
 924  *
 925  * If in the ESTABLISHED state, move to CLOSING.
 926  */
 927 static void cm_disconnect_handler(struct iwcm_id_private *cm_id_priv,
 928                                   struct iw_cm_event *iw_event)
 929 {
 930         unsigned long flags;
 931
 932         spin_lock_irqsave(&cm_id_priv->lock, flags);
 933         if (cm_id_priv->state == IW_CM_STATE_ESTABLISHED)
 934                 cm_id_priv->state = IW_CM_STATE_CLOSING;
 935         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 936 }
 937
 938 /*
 939  * CM_ID <-- IDLE
 940  *
 941  * If in the ESTBLISHED or CLOSING states, the QP will have have been
 942  * moved by the provider to the ERR state. Disassociate the CM_ID from
 943  * the QP,  move to IDLE, and remove the 'connected' reference.
 944  *
 945  * If in some other state, the cm_id was destroyed asynchronously.
 946  * This is the last reference that will result in waking up
 947  * the app thread blocked in iw_destroy_cm_id.
 948  */
 949 static int cm_close_handler(struct iwcm_id_private *cm_id_priv,
 950                                   struct iw_cm_event *iw_event)
 951 {
 952         struct ib_qp *qp;
 953         unsigned long flags;
 954         int ret = 0, notify_event = 0;
 955         spin_lock_irqsave(&cm_id_priv->lock, flags);
 956         qp = cm_id_priv->qp;
 957         cm_id_priv->qp = NULL;
 958
 959         switch (cm_id_priv->state) {
 960         case IW_CM_STATE_ESTABLISHED:
 961         case IW_CM_STATE_CLOSING:
 962                 cm_id_priv->state = IW_CM_STATE_IDLE;
 963                 notify_event = 1;
 964                 break;
 965         case IW_CM_STATE_DESTROYING:
 966                 break;
 967         default:
 968                 BUG();
 969         }
 970         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 971
 972         if (qp)
 973                 cm_id_priv->id.device->ops.iw_rem_ref(qp);
 974         if (notify_event)
 975                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 976         return ret;
 977 }
 978
 979 static int process_event(struct iwcm_id_private *cm_id_priv,
 980                          struct iw_cm_event *iw_event)
 981 {
 982         int ret = 0;
 983
 984         switch (iw_event->event) {
 985         case IW_CM_EVENT_CONNECT_REQUEST:
 986                 cm_conn_req_handler(cm_id_priv, iw_event);
 987                 break;
 988         case IW_CM_EVENT_CONNECT_REPLY:
 989                 ret = cm_conn_rep_handler(cm_id_priv, iw_event);
 990                 break;
 991         case IW_CM_EVENT_ESTABLISHED:
 992                 ret = cm_conn_est_handler(cm_id_priv, iw_event);
 993                 break;
 994         case IW_CM_EVENT_DISCONNECT:
 995                 cm_disconnect_handler(cm_id_priv, iw_event);
 996                 break;
 997         case IW_CM_EVENT_CLOSE:
 998                 ret = cm_close_handler(cm_id_priv, iw_event);
 999                 break;
1000         default:
1001                 BUG();
1002         }
1003
1004         return ret;
1005 }
1006
1007 /*
1008  * Process events on the work_list for the cm_id. If the callback
1009  * function requests that the cm_id be deleted, a flag is set in the
1010  * cm_id flags to indicate that when the last reference is
1011  * removed, the cm_id is to be destroyed. This is necessary to
1012  * distinguish between an object that will be destroyed by the app
1013  * thread asleep on the destroy_comp list vs. an object destroyed
1014  * here synchronously when the last reference is removed.
1015  */
1016 static void cm_work_handler(struct work_struct *_work)
1017 {
1018         struct iwcm_work *work = container_of(_work, struct iwcm_work, work);
1019         struct iw_cm_event levent;
1020         struct iwcm_id_private *cm_id_priv = work->cm_id;
1021         unsigned long flags;
1022         int empty;
1023         int ret = 0;
1024
1025         spin_lock_irqsave(&cm_id_priv->lock, flags);
1026         empty = list_empty(&cm_id_priv->work_list);
1027         while (!empty) {
1028                 work = list_entry(cm_id_priv->work_list.next,
1029                                   struct iwcm_work, list);
1030                 list_del_init(&work->list);
1031                 empty = list_empty(&cm_id_priv->work_list);
1032                 levent = work->event;
1033                 put_work(work);
1034                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1035
1036                 if (!test_bit(IWCM_F_DROP_EVENTS, &cm_id_priv->flags)) {
1037                         ret = process_event(cm_id_priv, &levent);
1038                         if (ret)
1039                                 destroy_cm_id(&cm_id_priv->id);
1040                 } else
1041                         pr_debug("dropping event %d\n", levent.event);
1042                 if (iwcm_deref_id(cm_id_priv))
1043                         return;
1044                 if (empty)
1045                         return;
1046                 spin_lock_irqsave(&cm_id_priv->lock, flags);
1047         }
1048         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1049 }
1050
1051 /*
1052  * This function is called on interrupt context. Schedule events on
1053  * the iwcm_wq thread to allow callback functions to downcall into
1054  * the CM and/or block.  Events are queued to a per-CM_ID
1055  * work_list. If this is the first event on the work_list, the work
1056  * element is also queued on the iwcm_wq thread.
1057  *
1058  * Each event holds a reference on the cm_id. Until the last posted
1059  * event has been delivered and processed, the cm_id cannot be
1060  * deleted.
1061  *
1062  * Returns:
1063  *            0 - the event was handled.
1064  *      -ENOMEM - the event was not handled due to lack of resources.
1065  */
1066 static int cm_event_handler(struct iw_cm_id *cm_id,
1067                              struct iw_cm_event *iw_event)
1068 {
1069         struct iwcm_work *work;
1070         struct iwcm_id_private *cm_id_priv;
1071         unsigned long flags;
1072         int ret = 0;
1073
1074         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1075
1076         spin_lock_irqsave(&cm_id_priv->lock, flags);
1077         work = get_work(cm_id_priv);
1078         if (!work) {
1079                 ret = -ENOMEM;
1080                 goto out;
1081         }
1082
1083         INIT_WORK(&work->work, cm_work_handler);
1084         work->cm_id = cm_id_priv;
1085         work->event = *iw_event;
1086
1087         if ((work->event.event == IW_CM_EVENT_CONNECT_REQUEST ||
1088              work->event.event == IW_CM_EVENT_CONNECT_REPLY) &&
1089             work->event.private_data_len) {
1090                 ret = copy_private_data(&work->event);
1091                 if (ret) {
1092                         put_work(work);
1093                         goto out;
1094                 }
1095         }
1096
1097         atomic_inc(&cm_id_priv->refcount);
1098         if (list_empty(&cm_id_priv->work_list)) {
1099                 list_add_tail(&work->list, &cm_id_priv->work_list);
1100                 queue_work(iwcm_wq, &work->work);
1101         } else
1102                 list_add_tail(&work->list, &cm_id_priv->work_list);
1103 out:
1104         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1105         return ret;
1106 }
1107
1108 static int iwcm_init_qp_init_attr(struct iwcm_id_private *cm_id_priv,
1109                                   struct ib_qp_attr *qp_attr,
1110                                   int *qp_attr_mask)
1111 {
1112         unsigned long flags;
1113         int ret;
1114
1115         spin_lock_irqsave(&cm_id_priv->lock, flags);
1116         switch (cm_id_priv->state) {
1117         case IW_CM_STATE_IDLE:
1118         case IW_CM_STATE_CONN_SENT:
1119         case IW_CM_STATE_CONN_RECV:
1120         case IW_CM_STATE_ESTABLISHED:
1121                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1122                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE|
1123                                            IB_ACCESS_REMOTE_READ;
1124                 ret = 0;
1125                 break;
1126         default:
1127                 ret = -EINVAL;
1128                 break;
1129         }
1130         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1131         return ret;
1132 }
1133
1134 static int iwcm_init_qp_rts_attr(struct iwcm_id_private *cm_id_priv,
1135                                   struct ib_qp_attr *qp_attr,
1136                                   int *qp_attr_mask)
1137 {
1138         unsigned long flags;
1139         int ret;
1140
1141         spin_lock_irqsave(&cm_id_priv->lock, flags);
1142         switch (cm_id_priv->state) {
1143         case IW_CM_STATE_IDLE:
1144         case IW_CM_STATE_CONN_SENT:
1145         case IW_CM_STATE_CONN_RECV:
1146         case IW_CM_STATE_ESTABLISHED:
1147                 *qp_attr_mask = 0;
1148                 ret = 0;
1149                 break;
1150         default:
1151                 ret = -EINVAL;
1152                 break;
1153         }
1154         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1155         return ret;
1156 }
1157
1158 int iw_cm_init_qp_attr(struct iw_cm_id *cm_id,
1159                        struct ib_qp_attr *qp_attr,
1160                        int *qp_attr_mask)
1161 {
1162         struct iwcm_id_private *cm_id_priv;
1163         int ret;
1164
1165         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1166         switch (qp_attr->qp_state) {
1167         case IB_QPS_INIT:
1168         case IB_QPS_RTR:
1169                 ret = iwcm_init_qp_init_attr(cm_id_priv,
1170                                              qp_attr, qp_attr_mask);
1171                 break;
1172         case IB_QPS_RTS:
1173                 ret = iwcm_init_qp_rts_attr(cm_id_priv,
1174                                             qp_attr, qp_attr_mask);
1175                 break;
1176         default:
1177                 ret = -EINVAL;
1178                 break;
1179         }
1180         return ret;
1181 }
1182 EXPORT_SYMBOL(iw_cm_init_qp_attr);
1183
1184 static int __init iw_cm_init(void)
1185 {
1186         int ret;
1187
1188         ret = iwpm_init(RDMA_NL_IWCM);
1189         if (ret)
1190                 pr_err("iw_cm: couldn't init iwpm\n");
1191         else
1192                 rdma_nl_register(RDMA_NL_IWCM, iwcm_nl_cb_table);
1193         iwcm_wq = alloc_ordered_workqueue("iw_cm_wq", 0);
1194         if (!iwcm_wq)
1195                 return -ENOMEM;
1196
1197         iwcm_ctl_table_hdr = register_net_sysctl(&init_net, "net/iw_cm",
1198                                                  iwcm_ctl_table);
1199         if (!iwcm_ctl_table_hdr) {
1200                 pr_err("iw_cm: couldn't register sysctl paths\n");
1201                 destroy_workqueue(iwcm_wq);
1202                 return -ENOMEM;
1203         }
1204
1205         return 0;
1206 }
1207
1208 static void __exit iw_cm_cleanup(void)
1209 {
1210         unregister_net_sysctl_table(iwcm_ctl_table_hdr);
1211         destroy_workqueue(iwcm_wq);
1212         rdma_nl_unregister(RDMA_NL_IWCM);
1213         iwpm_exit(RDMA_NL_IWCM);
1214 }
1215
1216 MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_IWCM, 2);
1217
1218 module_init(iw_cm_init);
1219 module_exit(iw_cm_cleanup);