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
  50 #include <rdma/iw_cm.h>
  51 #include <rdma/ib_addr.h>
  52
  53 #include "iwcm.h"
  54
  55 MODULE_AUTHOR("Tom Tucker");
  56 MODULE_DESCRIPTION("iWARP CM");
  57 MODULE_LICENSE("Dual BSD/GPL");
  58
  59 static struct workqueue_struct *iwcm_wq;
  60 struct iwcm_work {
  61         struct work_struct work;
  62         struct iwcm_id_private *cm_id;
  63         struct list_head list;
  64         struct iw_cm_event event;
  65         struct list_head free_list;
  66 };
  67
  68 /*
  69  * The following services provide a mechanism for pre-allocating iwcm_work
  70  * elements.  The design pre-allocates them  based on the cm_id type:
  71  *      LISTENING IDS:  Get enough elements preallocated to handle the
  72  *                      listen backlog.
  73  *      ACTIVE IDS:     4: CONNECT_REPLY, ESTABLISHED, DISCONNECT, CLOSE
  74  *      PASSIVE IDS:    3: ESTABLISHED, DISCONNECT, CLOSE
  75  *
  76  * Allocating them in connect and listen avoids having to deal
  77  * with allocation failures on the event upcall from the provider (which
  78  * is called in the interrupt context).
  79  *
  80  * One exception is when creating the cm_id for incoming connection requests.
  81  * There are two cases:
  82  * 1) in the event upcall, cm_event_handler(), for a listening cm_id.  If
  83  *    the backlog is exceeded, then no more connection request events will
  84  *    be processed.  cm_event_handler() returns -ENOMEM in this case.  Its up
  85  *    to the provider to reject the connection request.
  86  * 2) in the connection request workqueue handler, cm_conn_req_handler().
  87  *    If work elements cannot be allocated for the new connect request cm_id,
  88  *    then IWCM will call the provider reject method.  This is ok since
  89  *    cm_conn_req_handler() runs in the workqueue thread context.
  90  */
  91
  92 static struct iwcm_work *get_work(struct iwcm_id_private *cm_id_priv)
  93 {
  94         struct iwcm_work *work;
  95
  96         if (list_empty(&cm_id_priv->work_free_list))
  97                 return NULL;
  98         work = list_entry(cm_id_priv->work_free_list.next, struct iwcm_work,
  99                           free_list);
 100         list_del_init(&work->free_list);
 101         return work;
 102 }
 103
 104 static void put_work(struct iwcm_work *work)
 105 {
 106         list_add(&work->free_list, &work->cm_id->work_free_list);
 107 }
 108
 109 static void dealloc_work_entries(struct iwcm_id_private *cm_id_priv)
 110 {
 111         struct list_head *e, *tmp;
 112
 113         list_for_each_safe(e, tmp, &cm_id_priv->work_free_list)
 114                 kfree(list_entry(e, struct iwcm_work, free_list));
 115 }
 116
 117 static int alloc_work_entries(struct iwcm_id_private *cm_id_priv, int count)
 118 {
 119         struct iwcm_work *work;
 120
 121         BUG_ON(!list_empty(&cm_id_priv->work_free_list));
 122         while (count--) {
 123                 work = kmalloc(sizeof(struct iwcm_work), GFP_KERNEL);
 124                 if (!work) {
 125                         dealloc_work_entries(cm_id_priv);
 126                         return -ENOMEM;
 127                 }
 128                 work->cm_id = cm_id_priv;
 129                 INIT_LIST_HEAD(&work->list);
 130                 put_work(work);
 131         }
 132         return 0;
 133 }
 134
 135 /*
 136  * Save private data from incoming connection requests to
 137  * iw_cm_event, so the low level driver doesn't have to. Adjust
 138  * the event ptr to point to the local copy.
 139  */
 140 static int copy_private_data(struct iw_cm_event *event)
 141 {
 142         void *p;
 143
 144         p = kmemdup(event->private_data, event->private_data_len, GFP_ATOMIC);
 145         if (!p)
 146                 return -ENOMEM;
 147         event->private_data = p;
 148         return 0;
 149 }
 150
 151 static void free_cm_id(struct iwcm_id_private *cm_id_priv)
 152 {
 153         dealloc_work_entries(cm_id_priv);
 154         kfree(cm_id_priv);
 155 }
 156
 157 /*
 158  * Release a reference on cm_id. If the last reference is being
 159  * released, enable the waiting thread (in iw_destroy_cm_id) to
 160  * get woken up, and return 1 if a thread is already waiting.
 161  */
 162 static int iwcm_deref_id(struct iwcm_id_private *cm_id_priv)
 163 {
 164         BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
 165         if (atomic_dec_and_test(&cm_id_priv->refcount)) {
 166                 BUG_ON(!list_empty(&cm_id_priv->work_list));
 167                 complete(&cm_id_priv->destroy_comp);
 168                 return 1;
 169         }
 170
 171         return 0;
 172 }
 173
 174 static void add_ref(struct iw_cm_id *cm_id)
 175 {
 176         struct iwcm_id_private *cm_id_priv;
 177         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 178         atomic_inc(&cm_id_priv->refcount);
 179 }
 180
 181 static void rem_ref(struct iw_cm_id *cm_id)
 182 {
 183         struct iwcm_id_private *cm_id_priv;
 184         int cb_destroy;
 185
 186         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 187
 188         /*
 189          * Test bit before deref in case the cm_id gets freed on another
 190          * thread.
 191          */
 192         cb_destroy = test_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
 193         if (iwcm_deref_id(cm_id_priv) && cb_destroy) {
 194                 BUG_ON(!list_empty(&cm_id_priv->work_list));
 195                 free_cm_id(cm_id_priv);
 196         }
 197 }
 198
 199 static int cm_event_handler(struct iw_cm_id *cm_id, struct iw_cm_event *event);
 200
 201 struct iw_cm_id *iw_create_cm_id(struct ib_device *device,
 202                                  iw_cm_handler cm_handler,
 203                                  void *context)
 204 {
 205         struct iwcm_id_private *cm_id_priv;
 206
 207         cm_id_priv = kzalloc(sizeof(*cm_id_priv), GFP_KERNEL);
 208         if (!cm_id_priv)
 209                 return ERR_PTR(-ENOMEM);
 210
 211         cm_id_priv->state = IW_CM_STATE_IDLE;
 212         cm_id_priv->id.device = device;
 213         cm_id_priv->id.cm_handler = cm_handler;
 214         cm_id_priv->id.context = context;
 215         cm_id_priv->id.event_handler = cm_event_handler;
 216         cm_id_priv->id.add_ref = add_ref;
 217         cm_id_priv->id.rem_ref = rem_ref;
 218         spin_lock_init(&cm_id_priv->lock);
 219         atomic_set(&cm_id_priv->refcount, 1);
 220         init_waitqueue_head(&cm_id_priv->connect_wait);
 221         init_completion(&cm_id_priv->destroy_comp);
 222         INIT_LIST_HEAD(&cm_id_priv->work_list);
 223         INIT_LIST_HEAD(&cm_id_priv->work_free_list);
 224
 225         return &cm_id_priv->id;
 226 }
 227 EXPORT_SYMBOL(iw_create_cm_id);
 228
 229
 230 static int iwcm_modify_qp_err(struct ib_qp *qp)
 231 {
 232         struct ib_qp_attr qp_attr;
 233
 234         if (!qp)
 235                 return -EINVAL;
 236
 237         qp_attr.qp_state = IB_QPS_ERR;
 238         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 239 }
 240
 241 /*
 242  * This is really the RDMAC CLOSING state. It is most similar to the
 243  * IB SQD QP state.
 244  */
 245 static int iwcm_modify_qp_sqd(struct ib_qp *qp)
 246 {
 247         struct ib_qp_attr qp_attr;
 248
 249         BUG_ON(qp == NULL);
 250         qp_attr.qp_state = IB_QPS_SQD;
 251         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 252 }
 253
 254 /*
 255  * CM_ID <-- CLOSING
 256  *
 257  * Block if a passive or active connection is currently being processed. Then
 258  * process the event as follows:
 259  * - If we are ESTABLISHED, move to CLOSING and modify the QP state
 260  *   based on the abrupt flag
 261  * - If the connection is already in the CLOSING or IDLE state, the peer is
 262  *   disconnecting concurrently with us and we've already seen the
 263  *   DISCONNECT event -- ignore the request and return 0
 264  * - Disconnect on a listening endpoint returns -EINVAL
 265  */
 266 int iw_cm_disconnect(struct iw_cm_id *cm_id, int abrupt)
 267 {
 268         struct iwcm_id_private *cm_id_priv;
 269         unsigned long flags;
 270         int ret = 0;
 271         struct ib_qp *qp = NULL;
 272
 273         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 274         /* Wait if we're currently in a connect or accept downcall */
 275         wait_event(cm_id_priv->connect_wait,
 276                    !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
 277
 278         spin_lock_irqsave(&cm_id_priv->lock, flags);
 279         switch (cm_id_priv->state) {
 280         case IW_CM_STATE_ESTABLISHED:
 281                 cm_id_priv->state = IW_CM_STATE_CLOSING;
 282
 283                 /* QP could be <nul> for user-mode client */
 284                 if (cm_id_priv->qp)
 285                         qp = cm_id_priv->qp;
 286                 else
 287                         ret = -EINVAL;
 288                 break;
 289         case IW_CM_STATE_LISTEN:
 290                 ret = -EINVAL;
 291                 break;
 292         case IW_CM_STATE_CLOSING:
 293                 /* remote peer closed first */
 294         case IW_CM_STATE_IDLE:
 295                 /* accept or connect returned !0 */
 296                 break;
 297         case IW_CM_STATE_CONN_RECV:
 298                 /*
 299                  * App called disconnect before/without calling accept after
 300                  * connect_request event delivered.
 301                  */
 302                 break;
 303         case IW_CM_STATE_CONN_SENT:
 304                 /* Can only get here if wait above fails */
 305         default:
 306                 BUG();
 307         }
 308         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 309
 310         if (qp) {
 311                 if (abrupt)
 312                         ret = iwcm_modify_qp_err(qp);
 313                 else
 314                         ret = iwcm_modify_qp_sqd(qp);
 315
 316                 /*
 317                  * If both sides are disconnecting the QP could
 318                  * already be in ERR or SQD states
 319                  */
 320                 ret = 0;
 321         }
 322
 323         return ret;
 324 }
 325 EXPORT_SYMBOL(iw_cm_disconnect);
 326
 327 /*
 328  * CM_ID <-- DESTROYING
 329  *
 330  * Clean up all resources associated with the connection and release
 331  * the initial reference taken by iw_create_cm_id.
 332  */
 333 static void destroy_cm_id(struct iw_cm_id *cm_id)
 334 {
 335         struct iwcm_id_private *cm_id_priv;
 336         unsigned long flags;
 337         int ret;
 338
 339         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 340         /*
 341          * Wait if we're currently in a connect or accept downcall. A
 342          * listening endpoint should never block here.
 343          */
 344         wait_event(cm_id_priv->connect_wait,
 345                    !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
 346
 347         spin_lock_irqsave(&cm_id_priv->lock, flags);
 348         switch (cm_id_priv->state) {
 349         case IW_CM_STATE_LISTEN:
 350                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 351                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 352                 /* destroy the listening endpoint */
 353                 ret = cm_id->device->iwcm->destroy_listen(cm_id);
 354                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 355                 break;
 356         case IW_CM_STATE_ESTABLISHED:
 357                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 358                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 359                 /* Abrupt close of the connection */
 360                 (void)iwcm_modify_qp_err(cm_id_priv->qp);
 361                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 362                 break;
 363         case IW_CM_STATE_IDLE:
 364         case IW_CM_STATE_CLOSING:
 365                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 366                 break;
 367         case IW_CM_STATE_CONN_RECV:
 368                 /*
 369                  * App called destroy before/without calling accept after
 370                  * receiving connection request event notification or
 371                  * returned non zero from the event callback function.
 372                  * In either case, must tell the provider to reject.
 373                  */
 374                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 375                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 376                 cm_id->device->iwcm->reject(cm_id, NULL, 0);
 377                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 378                 break;
 379         case IW_CM_STATE_CONN_SENT:
 380         case IW_CM_STATE_DESTROYING:
 381         default:
 382                 BUG();
 383                 break;
 384         }
 385         if (cm_id_priv->qp) {
 386                 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
 387                 cm_id_priv->qp = NULL;
 388         }
 389         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 390
 391         (void)iwcm_deref_id(cm_id_priv);
 392 }
 393
 394 /*
 395  * This function is only called by the application thread and cannot
 396  * be called by the event thread. The function will wait for all
 397  * references to be released on the cm_id and then kfree the cm_id
 398  * object.
 399  */
 400 void iw_destroy_cm_id(struct iw_cm_id *cm_id)
 401 {
 402         struct iwcm_id_private *cm_id_priv;
 403
 404         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 405         BUG_ON(test_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags));
 406
 407         destroy_cm_id(cm_id);
 408
 409         wait_for_completion(&cm_id_priv->destroy_comp);
 410
 411         free_cm_id(cm_id_priv);
 412 }
 413 EXPORT_SYMBOL(iw_destroy_cm_id);
 414
 415 /*
 416  * CM_ID <-- LISTEN
 417  *
 418  * Start listening for connect requests. Generates one CONNECT_REQUEST
 419  * event for each inbound connect request.
 420  */
 421 int iw_cm_listen(struct iw_cm_id *cm_id, int backlog)
 422 {
 423         struct iwcm_id_private *cm_id_priv;
 424         unsigned long flags;
 425         int ret;
 426
 427         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 428
 429         ret = alloc_work_entries(cm_id_priv, backlog);
 430         if (ret)
 431                 return ret;
 432
 433         spin_lock_irqsave(&cm_id_priv->lock, flags);
 434         switch (cm_id_priv->state) {
 435         case IW_CM_STATE_IDLE:
 436                 cm_id_priv->state = IW_CM_STATE_LISTEN;
 437                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 438                 ret = cm_id->device->iwcm->create_listen(cm_id, backlog);
 439                 if (ret)
 440                         cm_id_priv->state = IW_CM_STATE_IDLE;
 441                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 442                 break;
 443         default:
 444                 ret = -EINVAL;
 445         }
 446         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 447
 448         return ret;
 449 }
 450 EXPORT_SYMBOL(iw_cm_listen);
 451
 452 /*
 453  * CM_ID <-- IDLE
 454  *
 455  * Rejects an inbound connection request. No events are generated.
 456  */
 457 int iw_cm_reject(struct iw_cm_id *cm_id,
 458                  const void *private_data,
 459                  u8 private_data_len)
 460 {
 461         struct iwcm_id_private *cm_id_priv;
 462         unsigned long flags;
 463         int ret;
 464
 465         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 466         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 467
 468         spin_lock_irqsave(&cm_id_priv->lock, flags);
 469         if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
 470                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 471                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 472                 wake_up_all(&cm_id_priv->connect_wait);
 473                 return -EINVAL;
 474         }
 475         cm_id_priv->state = IW_CM_STATE_IDLE;
 476         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 477
 478         ret = cm_id->device->iwcm->reject(cm_id, private_data,
 479                                           private_data_len);
 480
 481         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 482         wake_up_all(&cm_id_priv->connect_wait);
 483
 484         return ret;
 485 }
 486 EXPORT_SYMBOL(iw_cm_reject);
 487
 488 /*
 489  * CM_ID <-- ESTABLISHED
 490  *
 491  * Accepts an inbound connection request and generates an ESTABLISHED
 492  * event. Callers of iw_cm_disconnect and iw_destroy_cm_id will block
 493  * until the ESTABLISHED event is received from the provider.
 494  */
 495 int iw_cm_accept(struct iw_cm_id *cm_id,
 496                  struct iw_cm_conn_param *iw_param)
 497 {
 498         struct iwcm_id_private *cm_id_priv;
 499         struct ib_qp *qp;
 500         unsigned long flags;
 501         int ret;
 502
 503         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 504         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 505
 506         spin_lock_irqsave(&cm_id_priv->lock, flags);
 507         if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
 508                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 509                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 510                 wake_up_all(&cm_id_priv->connect_wait);
 511                 return -EINVAL;
 512         }
 513         /* Get the ib_qp given the QPN */
 514         qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
 515         if (!qp) {
 516                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 517                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 518                 wake_up_all(&cm_id_priv->connect_wait);
 519                 return -EINVAL;
 520         }
 521         cm_id->device->iwcm->add_ref(qp);
 522         cm_id_priv->qp = qp;
 523         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 524
 525         ret = cm_id->device->iwcm->accept(cm_id, iw_param);
 526         if (ret) {
 527                 /* An error on accept precludes provider events */
 528                 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
 529                 cm_id_priv->state = IW_CM_STATE_IDLE;
 530                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 531                 if (cm_id_priv->qp) {
 532                         cm_id->device->iwcm->rem_ref(qp);
 533                         cm_id_priv->qp = NULL;
 534                 }
 535                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 536                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 537                 wake_up_all(&cm_id_priv->connect_wait);
 538         }
 539
 540         return ret;
 541 }
 542 EXPORT_SYMBOL(iw_cm_accept);
 543
 544 /*
 545  * Active Side: CM_ID <-- CONN_SENT
 546  *
 547  * If successful, results in the generation of a CONNECT_REPLY
 548  * event. iw_cm_disconnect and iw_cm_destroy will block until the
 549  * CONNECT_REPLY event is received from the provider.
 550  */
 551 int iw_cm_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *iw_param)
 552 {
 553         struct iwcm_id_private *cm_id_priv;
 554         int ret;
 555         unsigned long flags;
 556         struct ib_qp *qp;
 557
 558         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 559
 560         ret = alloc_work_entries(cm_id_priv, 4);
 561         if (ret)
 562                 return ret;
 563
 564         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 565         spin_lock_irqsave(&cm_id_priv->lock, flags);
 566
 567         if (cm_id_priv->state != IW_CM_STATE_IDLE) {
 568                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 569                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 570                 wake_up_all(&cm_id_priv->connect_wait);
 571                 return -EINVAL;
 572         }
 573
 574         /* Get the ib_qp given the QPN */
 575         qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
 576         if (!qp) {
 577                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 578                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 579                 wake_up_all(&cm_id_priv->connect_wait);
 580                 return -EINVAL;
 581         }
 582         cm_id->device->iwcm->add_ref(qp);
 583         cm_id_priv->qp = qp;
 584         cm_id_priv->state = IW_CM_STATE_CONN_SENT;
 585         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 586
 587         ret = cm_id->device->iwcm->connect(cm_id, iw_param);
 588         if (ret) {
 589                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 590                 if (cm_id_priv->qp) {
 591                         cm_id->device->iwcm->rem_ref(qp);
 592                         cm_id_priv->qp = NULL;
 593                 }
 594                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 595                 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
 596                 cm_id_priv->state = IW_CM_STATE_IDLE;
 597                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 598                 wake_up_all(&cm_id_priv->connect_wait);
 599         }
 600
 601         return ret;
 602 }
 603 EXPORT_SYMBOL(iw_cm_connect);
 604
 605 /*
 606  * Passive Side: new CM_ID <-- CONN_RECV
 607  *
 608  * Handles an inbound connect request. The function creates a new
 609  * iw_cm_id to represent the new connection and inherits the client
 610  * callback function and other attributes from the listening parent.
 611  *
 612  * The work item contains a pointer to the listen_cm_id and the event. The
 613  * listen_cm_id contains the client cm_handler, context and
 614  * device. These are copied when the device is cloned. The event
 615  * contains the new four tuple.
 616  *
 617  * An error on the child should not affect the parent, so this
 618  * function does not return a value.
 619  */
 620 static void cm_conn_req_handler(struct iwcm_id_private *listen_id_priv,
 621                                 struct iw_cm_event *iw_event)
 622 {
 623         unsigned long flags;
 624         struct iw_cm_id *cm_id;
 625         struct iwcm_id_private *cm_id_priv;
 626         int ret;
 627
 628         /*
 629          * The provider should never generate a connection request
 630          * event with a bad status.
 631          */
 632         BUG_ON(iw_event->status);
 633
 634         cm_id = iw_create_cm_id(listen_id_priv->id.device,
 635                                 listen_id_priv->id.cm_handler,
 636                                 listen_id_priv->id.context);
 637         /* If the cm_id could not be created, ignore the request */
 638         if (IS_ERR(cm_id))
 639                 goto out;
 640
 641         cm_id->provider_data = iw_event->provider_data;
 642         cm_id->local_addr = iw_event->local_addr;
 643         cm_id->remote_addr = iw_event->remote_addr;
 644
 645         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 646         cm_id_priv->state = IW_CM_STATE_CONN_RECV;
 647
 648         /*
 649          * We could be destroying the listening id. If so, ignore this
 650          * upcall.
 651          */
 652         spin_lock_irqsave(&listen_id_priv->lock, flags);
 653         if (listen_id_priv->state != IW_CM_STATE_LISTEN) {
 654                 spin_unlock_irqrestore(&listen_id_priv->lock, flags);
 655                 iw_cm_reject(cm_id, NULL, 0);
 656                 iw_destroy_cm_id(cm_id);
 657                 goto out;
 658         }
 659         spin_unlock_irqrestore(&listen_id_priv->lock, flags);
 660
 661         ret = alloc_work_entries(cm_id_priv, 3);
 662         if (ret) {
 663                 iw_cm_reject(cm_id, NULL, 0);
 664                 iw_destroy_cm_id(cm_id);
 665                 goto out;
 666         }
 667
 668         /* Call the client CM handler */
 669         ret = cm_id->cm_handler(cm_id, iw_event);
 670         if (ret) {
 671                 iw_cm_reject(cm_id, NULL, 0);
 672                 set_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
 673                 destroy_cm_id(cm_id);
 674                 if (atomic_read(&cm_id_priv->refcount)==0)
 675                         free_cm_id(cm_id_priv);
 676         }
 677
 678 out:
 679         if (iw_event->private_data_len)
 680                 kfree(iw_event->private_data);
 681 }
 682
 683 /*
 684  * Passive Side: CM_ID <-- ESTABLISHED
 685  *
 686  * The provider generated an ESTABLISHED event which means that
 687  * the MPA negotion has completed successfully and we are now in MPA
 688  * FPDU mode.
 689  *
 690  * This event can only be received in the CONN_RECV state. If the
 691  * remote peer closed, the ESTABLISHED event would be received followed
 692  * by the CLOSE event. If the app closes, it will block until we wake
 693  * it up after processing this event.
 694  */
 695 static int cm_conn_est_handler(struct iwcm_id_private *cm_id_priv,
 696                                struct iw_cm_event *iw_event)
 697 {
 698         unsigned long flags;
 699         int ret;
 700
 701         spin_lock_irqsave(&cm_id_priv->lock, flags);
 702
 703         /*
 704          * We clear the CONNECT_WAIT bit here to allow the callback
 705          * function to call iw_cm_disconnect. Calling iw_destroy_cm_id
 706          * from a callback handler is not allowed.
 707          */
 708         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 709         BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
 710         cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
 711         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 712         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 713         wake_up_all(&cm_id_priv->connect_wait);
 714
 715         return ret;
 716 }
 717
 718 /*
 719  * Active Side: CM_ID <-- ESTABLISHED
 720  *
 721  * The app has called connect and is waiting for the established event to
 722  * post it's requests to the server. This event will wake up anyone
 723  * blocked in iw_cm_disconnect or iw_destroy_id.
 724  */
 725 static int cm_conn_rep_handler(struct iwcm_id_private *cm_id_priv,
 726                                struct iw_cm_event *iw_event)
 727 {
 728         unsigned long flags;
 729         int ret;
 730
 731         spin_lock_irqsave(&cm_id_priv->lock, flags);
 732         /*
 733          * Clear the connect wait bit so a callback function calling
 734          * iw_cm_disconnect will not wait and deadlock this thread
 735          */
 736         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 737         BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
 738         if (iw_event->status == 0) {
 739                 cm_id_priv->id.local_addr = iw_event->local_addr;
 740                 cm_id_priv->id.remote_addr = iw_event->remote_addr;
 741                 cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
 742         } else {
 743                 /* REJECTED or RESET */
 744                 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
 745                 cm_id_priv->qp = NULL;
 746                 cm_id_priv->state = IW_CM_STATE_IDLE;
 747         }
 748         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 749         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 750
 751         if (iw_event->private_data_len)
 752                 kfree(iw_event->private_data);
 753
 754         /* Wake up waiters on connect complete */
 755         wake_up_all(&cm_id_priv->connect_wait);
 756
 757         return ret;
 758 }
 759
 760 /*
 761  * CM_ID <-- CLOSING
 762  *
 763  * If in the ESTABLISHED state, move to CLOSING.
 764  */
 765 static void cm_disconnect_handler(struct iwcm_id_private *cm_id_priv,
 766                                   struct iw_cm_event *iw_event)
 767 {
 768         unsigned long flags;
 769
 770         spin_lock_irqsave(&cm_id_priv->lock, flags);
 771         if (cm_id_priv->state == IW_CM_STATE_ESTABLISHED)
 772                 cm_id_priv->state = IW_CM_STATE_CLOSING;
 773         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 774 }
 775
 776 /*
 777  * CM_ID <-- IDLE
 778  *
 779  * If in the ESTBLISHED or CLOSING states, the QP will have have been
 780  * moved by the provider to the ERR state. Disassociate the CM_ID from
 781  * the QP,  move to IDLE, and remove the 'connected' reference.
 782  *
 783  * If in some other state, the cm_id was destroyed asynchronously.
 784  * This is the last reference that will result in waking up
 785  * the app thread blocked in iw_destroy_cm_id.
 786  */
 787 static int cm_close_handler(struct iwcm_id_private *cm_id_priv,
 788                                   struct iw_cm_event *iw_event)
 789 {
 790         unsigned long flags;
 791         int ret = 0;
 792         spin_lock_irqsave(&cm_id_priv->lock, flags);
 793
 794         if (cm_id_priv->qp) {
 795                 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
 796                 cm_id_priv->qp = NULL;
 797         }
 798         switch (cm_id_priv->state) {
 799         case IW_CM_STATE_ESTABLISHED:
 800         case IW_CM_STATE_CLOSING:
 801                 cm_id_priv->state = IW_CM_STATE_IDLE;
 802                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 803                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 804                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 805                 break;
 806         case IW_CM_STATE_DESTROYING:
 807                 break;
 808         default:
 809                 BUG();
 810         }
 811         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 812
 813         return ret;
 814 }
 815
 816 static int process_event(struct iwcm_id_private *cm_id_priv,
 817                          struct iw_cm_event *iw_event)
 818 {
 819         int ret = 0;
 820
 821         switch (iw_event->event) {
 822         case IW_CM_EVENT_CONNECT_REQUEST:
 823                 cm_conn_req_handler(cm_id_priv, iw_event);
 824                 break;
 825         case IW_CM_EVENT_CONNECT_REPLY:
 826                 ret = cm_conn_rep_handler(cm_id_priv, iw_event);
 827                 break;
 828         case IW_CM_EVENT_ESTABLISHED:
 829                 ret = cm_conn_est_handler(cm_id_priv, iw_event);
 830                 break;
 831         case IW_CM_EVENT_DISCONNECT:
 832                 cm_disconnect_handler(cm_id_priv, iw_event);
 833                 break;
 834         case IW_CM_EVENT_CLOSE:
 835                 ret = cm_close_handler(cm_id_priv, iw_event);
 836                 break;
 837         default:
 838                 BUG();
 839         }
 840
 841         return ret;
 842 }
 843
 844 /*
 845  * Process events on the work_list for the cm_id. If the callback
 846  * function requests that the cm_id be deleted, a flag is set in the
 847  * cm_id flags to indicate that when the last reference is
 848  * removed, the cm_id is to be destroyed. This is necessary to
 849  * distinguish between an object that will be destroyed by the app
 850  * thread asleep on the destroy_comp list vs. an object destroyed
 851  * here synchronously when the last reference is removed.
 852  */
 853 static void cm_work_handler(struct work_struct *_work)
 854 {
 855         struct iwcm_work *work = container_of(_work, struct iwcm_work, work);
 856         struct iw_cm_event levent;
 857         struct iwcm_id_private *cm_id_priv = work->cm_id;
 858         unsigned long flags;
 859         int empty;
 860         int ret = 0;
 861         int destroy_id;
 862
 863         spin_lock_irqsave(&cm_id_priv->lock, flags);
 864         empty = list_empty(&cm_id_priv->work_list);
 865         while (!empty) {
 866                 work = list_entry(cm_id_priv->work_list.next,
 867                                   struct iwcm_work, list);
 868                 list_del_init(&work->list);
 869                 empty = list_empty(&cm_id_priv->work_list);
 870                 levent = work->event;
 871                 put_work(work);
 872                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 873
 874                 ret = process_event(cm_id_priv, &levent);
 875                 if (ret) {
 876                         set_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
 877                         destroy_cm_id(&cm_id_priv->id);
 878                 }
 879                 BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
 880                 destroy_id = test_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
 881                 if (iwcm_deref_id(cm_id_priv)) {
 882                         if (destroy_id) {
 883                                 BUG_ON(!list_empty(&cm_id_priv->work_list));
 884                                 free_cm_id(cm_id_priv);
 885                         }
 886                         return;
 887                 }
 888                 if (empty)
 889                         return;
 890                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 891         }
 892         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 893 }
 894
 895 /*
 896  * This function is called on interrupt context. Schedule events on
 897  * the iwcm_wq thread to allow callback functions to downcall into
 898  * the CM and/or block.  Events are queued to a per-CM_ID
 899  * work_list. If this is the first event on the work_list, the work
 900  * element is also queued on the iwcm_wq thread.
 901  *
 902  * Each event holds a reference on the cm_id. Until the last posted
 903  * event has been delivered and processed, the cm_id cannot be
 904  * deleted.
 905  *
 906  * Returns:
 907  *            0 - the event was handled.
 908  *      -ENOMEM - the event was not handled due to lack of resources.
 909  */
 910 static int cm_event_handler(struct iw_cm_id *cm_id,
 911                              struct iw_cm_event *iw_event)
 912 {
 913         struct iwcm_work *work;
 914         struct iwcm_id_private *cm_id_priv;
 915         unsigned long flags;
 916         int ret = 0;
 917
 918         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 919
 920         spin_lock_irqsave(&cm_id_priv->lock, flags);
 921         work = get_work(cm_id_priv);
 922         if (!work) {
 923                 ret = -ENOMEM;
 924                 goto out;
 925         }
 926
 927         INIT_WORK(&work->work, cm_work_handler);
 928         work->cm_id = cm_id_priv;
 929         work->event = *iw_event;
 930
 931         if ((work->event.event == IW_CM_EVENT_CONNECT_REQUEST ||
 932              work->event.event == IW_CM_EVENT_CONNECT_REPLY) &&
 933             work->event.private_data_len) {
 934                 ret = copy_private_data(&work->event);
 935                 if (ret) {
 936                         put_work(work);
 937                         goto out;
 938                 }
 939         }
 940
 941         atomic_inc(&cm_id_priv->refcount);
 942         if (list_empty(&cm_id_priv->work_list)) {
 943                 list_add_tail(&work->list, &cm_id_priv->work_list);
 944                 queue_work(iwcm_wq, &work->work);
 945         } else
 946                 list_add_tail(&work->list, &cm_id_priv->work_list);
 947 out:
 948         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 949         return ret;
 950 }
 951
 952 static int iwcm_init_qp_init_attr(struct iwcm_id_private *cm_id_priv,
 953                                   struct ib_qp_attr *qp_attr,
 954                                   int *qp_attr_mask)
 955 {
 956         unsigned long flags;
 957         int ret;
 958
 959         spin_lock_irqsave(&cm_id_priv->lock, flags);
 960         switch (cm_id_priv->state) {
 961         case IW_CM_STATE_IDLE:
 962         case IW_CM_STATE_CONN_SENT:
 963         case IW_CM_STATE_CONN_RECV:
 964         case IW_CM_STATE_ESTABLISHED:
 965                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 966                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE|
 967                                            IB_ACCESS_REMOTE_READ;
 968                 ret = 0;
 969                 break;
 970         default:
 971                 ret = -EINVAL;
 972                 break;
 973         }
 974         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 975         return ret;
 976 }
 977
 978 static int iwcm_init_qp_rts_attr(struct iwcm_id_private *cm_id_priv,
 979                                   struct ib_qp_attr *qp_attr,
 980                                   int *qp_attr_mask)
 981 {
 982         unsigned long flags;
 983         int ret;
 984
 985         spin_lock_irqsave(&cm_id_priv->lock, flags);
 986         switch (cm_id_priv->state) {
 987         case IW_CM_STATE_IDLE:
 988         case IW_CM_STATE_CONN_SENT:
 989         case IW_CM_STATE_CONN_RECV:
 990         case IW_CM_STATE_ESTABLISHED:
 991                 *qp_attr_mask = 0;
 992                 ret = 0;
 993                 break;
 994         default:
 995                 ret = -EINVAL;
 996                 break;
 997         }
 998         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 999         return ret;
1000 }
1001
1002 int iw_cm_init_qp_attr(struct iw_cm_id *cm_id,
1003                        struct ib_qp_attr *qp_attr,
1004                        int *qp_attr_mask)
1005 {
1006         struct iwcm_id_private *cm_id_priv;
1007         int ret;
1008
1009         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1010         switch (qp_attr->qp_state) {
1011         case IB_QPS_INIT:
1012         case IB_QPS_RTR:
1013                 ret = iwcm_init_qp_init_attr(cm_id_priv,
1014                                              qp_attr, qp_attr_mask);
1015                 break;
1016         case IB_QPS_RTS:
1017                 ret = iwcm_init_qp_rts_attr(cm_id_priv,
1018                                             qp_attr, qp_attr_mask);
1019                 break;
1020         default:
1021                 ret = -EINVAL;
1022                 break;
1023         }
1024         return ret;
1025 }
1026 EXPORT_SYMBOL(iw_cm_init_qp_attr);
1027
1028 static int __init iw_cm_init(void)
1029 {
1030         iwcm_wq = create_singlethread_workqueue("iw_cm_wq");
1031         if (!iwcm_wq)
1032                 return -ENOMEM;
1033
1034         return 0;
1035 }
1036
1037 static void __exit iw_cm_cleanup(void)
1038 {
1039         destroy_workqueue(iwcm_wq);
1040 }
1041
1042 module_init(iw_cm_init);
1043 module_exit(iw_cm_cleanup);