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/spinlock.h>
  44 #include <linux/workqueue.h>
  45 #include <linux/completion.h>
  46
  47 #include <rdma/iw_cm.h>
  48 #include <rdma/ib_addr.h>
  49
  50 #include "iwcm.h"
  51
  52 MODULE_AUTHOR("Tom Tucker");
  53 MODULE_DESCRIPTION("iWARP CM");
  54 MODULE_LICENSE("Dual BSD/GPL");
  55
  56 static struct workqueue_struct *iwcm_wq;
  57 struct iwcm_work {
  58         struct work_struct work;
  59         struct iwcm_id_private *cm_id;
  60         struct list_head list;
  61         struct iw_cm_event event;
  62         struct list_head free_list;
  63 };
  64
  65 /*
  66  * The following services provide a mechanism for pre-allocating iwcm_work
  67  * elements.  The design pre-allocates them  based on the cm_id type:
  68  *      LISTENING IDS:  Get enough elements preallocated to handle the
  69  *                      listen backlog.
  70  *      ACTIVE IDS:     4: CONNECT_REPLY, ESTABLISHED, DISCONNECT, CLOSE
  71  *      PASSIVE IDS:    3: ESTABLISHED, DISCONNECT, CLOSE
  72  *
  73  * Allocating them in connect and listen avoids having to deal
  74  * with allocation failures on the event upcall from the provider (which
  75  * is called in the interrupt context).
  76  *
  77  * One exception is when creating the cm_id for incoming connection requests.
  78  * There are two cases:
  79  * 1) in the event upcall, cm_event_handler(), for a listening cm_id.  If
  80  *    the backlog is exceeded, then no more connection request events will
  81  *    be processed.  cm_event_handler() returns -ENOMEM in this case.  Its up
  82  *    to the provider to reject the connection request.
  83  * 2) in the connection request workqueue handler, cm_conn_req_handler().
  84  *    If work elements cannot be allocated for the new connect request cm_id,
  85  *    then IWCM will call the provider reject method.  This is ok since
  86  *    cm_conn_req_handler() runs in the workqueue thread context.
  87  */
  88
  89 static struct iwcm_work *get_work(struct iwcm_id_private *cm_id_priv)
  90 {
  91         struct iwcm_work *work;
  92
  93         if (list_empty(&cm_id_priv->work_free_list))
  94                 return NULL;
  95         work = list_entry(cm_id_priv->work_free_list.next, struct iwcm_work,
  96                           free_list);
  97         list_del_init(&work->free_list);
  98         return work;
  99 }
 100
 101 static void put_work(struct iwcm_work *work)
 102 {
 103         list_add(&work->free_list, &work->cm_id->work_free_list);
 104 }
 105
 106 static void dealloc_work_entries(struct iwcm_id_private *cm_id_priv)
 107 {
 108         struct list_head *e, *tmp;
 109
 110         list_for_each_safe(e, tmp, &cm_id_priv->work_free_list)
 111                 kfree(list_entry(e, struct iwcm_work, free_list));
 112 }
 113
 114 static int alloc_work_entries(struct iwcm_id_private *cm_id_priv, int count)
 115 {
 116         struct iwcm_work *work;
 117
 118         BUG_ON(!list_empty(&cm_id_priv->work_free_list));
 119         while (count--) {
 120                 work = kmalloc(sizeof(struct iwcm_work), GFP_KERNEL);
 121                 if (!work) {
 122                         dealloc_work_entries(cm_id_priv);
 123                         return -ENOMEM;
 124                 }
 125                 work->cm_id = cm_id_priv;
 126                 INIT_LIST_HEAD(&work->list);
 127                 put_work(work);
 128         }
 129         return 0;
 130 }
 131
 132 /*
 133  * Save private data from incoming connection requests to
 134  * iw_cm_event, so the low level driver doesn't have to. Adjust
 135  * the event ptr to point to the local copy.
 136  */
 137 static int copy_private_data(struct iw_cm_event *event)
 138 {
 139         void *p;
 140
 141         p = kmemdup(event->private_data, event->private_data_len, GFP_ATOMIC);
 142         if (!p)
 143                 return -ENOMEM;
 144         event->private_data = p;
 145         return 0;
 146 }
 147
 148 static void free_cm_id(struct iwcm_id_private *cm_id_priv)
 149 {
 150         dealloc_work_entries(cm_id_priv);
 151         kfree(cm_id_priv);
 152 }
 153
 154 /*
 155  * Release a reference on cm_id. If the last reference is being
 156  * released, enable the waiting thread (in iw_destroy_cm_id) to
 157  * get woken up, and return 1 if a thread is already waiting.
 158  */
 159 static int iwcm_deref_id(struct iwcm_id_private *cm_id_priv)
 160 {
 161         BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
 162         if (atomic_dec_and_test(&cm_id_priv->refcount)) {
 163                 BUG_ON(!list_empty(&cm_id_priv->work_list));
 164                 complete(&cm_id_priv->destroy_comp);
 165                 return 1;
 166         }
 167
 168         return 0;
 169 }
 170
 171 static void add_ref(struct iw_cm_id *cm_id)
 172 {
 173         struct iwcm_id_private *cm_id_priv;
 174         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 175         atomic_inc(&cm_id_priv->refcount);
 176 }
 177
 178 static void rem_ref(struct iw_cm_id *cm_id)
 179 {
 180         struct iwcm_id_private *cm_id_priv;
 181         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 182         if (iwcm_deref_id(cm_id_priv) &&
 183             test_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags)) {
 184                 BUG_ON(!list_empty(&cm_id_priv->work_list));
 185                 free_cm_id(cm_id_priv);
 186         }
 187 }
 188
 189 static int cm_event_handler(struct iw_cm_id *cm_id, struct iw_cm_event *event);
 190
 191 struct iw_cm_id *iw_create_cm_id(struct ib_device *device,
 192                                  iw_cm_handler cm_handler,
 193                                  void *context)
 194 {
 195         struct iwcm_id_private *cm_id_priv;
 196
 197         cm_id_priv = kzalloc(sizeof(*cm_id_priv), GFP_KERNEL);
 198         if (!cm_id_priv)
 199                 return ERR_PTR(-ENOMEM);
 200
 201         cm_id_priv->state = IW_CM_STATE_IDLE;
 202         cm_id_priv->id.device = device;
 203         cm_id_priv->id.cm_handler = cm_handler;
 204         cm_id_priv->id.context = context;
 205         cm_id_priv->id.event_handler = cm_event_handler;
 206         cm_id_priv->id.add_ref = add_ref;
 207         cm_id_priv->id.rem_ref = rem_ref;
 208         spin_lock_init(&cm_id_priv->lock);
 209         atomic_set(&cm_id_priv->refcount, 1);
 210         init_waitqueue_head(&cm_id_priv->connect_wait);
 211         init_completion(&cm_id_priv->destroy_comp);
 212         INIT_LIST_HEAD(&cm_id_priv->work_list);
 213         INIT_LIST_HEAD(&cm_id_priv->work_free_list);
 214
 215         return &cm_id_priv->id;
 216 }
 217 EXPORT_SYMBOL(iw_create_cm_id);
 218
 219
 220 static int iwcm_modify_qp_err(struct ib_qp *qp)
 221 {
 222         struct ib_qp_attr qp_attr;
 223
 224         if (!qp)
 225                 return -EINVAL;
 226
 227         qp_attr.qp_state = IB_QPS_ERR;
 228         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 229 }
 230
 231 /*
 232  * This is really the RDMAC CLOSING state. It is most similar to the
 233  * IB SQD QP state.
 234  */
 235 static int iwcm_modify_qp_sqd(struct ib_qp *qp)
 236 {
 237         struct ib_qp_attr qp_attr;
 238
 239         BUG_ON(qp == NULL);
 240         qp_attr.qp_state = IB_QPS_SQD;
 241         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 242 }
 243
 244 /*
 245  * CM_ID <-- CLOSING
 246  *
 247  * Block if a passive or active connection is currently being processed. Then
 248  * process the event as follows:
 249  * - If we are ESTABLISHED, move to CLOSING and modify the QP state
 250  *   based on the abrupt flag
 251  * - If the connection is already in the CLOSING or IDLE state, the peer is
 252  *   disconnecting concurrently with us and we've already seen the
 253  *   DISCONNECT event -- ignore the request and return 0
 254  * - Disconnect on a listening endpoint returns -EINVAL
 255  */
 256 int iw_cm_disconnect(struct iw_cm_id *cm_id, int abrupt)
 257 {
 258         struct iwcm_id_private *cm_id_priv;
 259         unsigned long flags;
 260         int ret = 0;
 261         struct ib_qp *qp = NULL;
 262
 263         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 264         /* Wait if we're currently in a connect or accept downcall */
 265         wait_event(cm_id_priv->connect_wait,
 266                    !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
 267
 268         spin_lock_irqsave(&cm_id_priv->lock, flags);
 269         switch (cm_id_priv->state) {
 270         case IW_CM_STATE_ESTABLISHED:
 271                 cm_id_priv->state = IW_CM_STATE_CLOSING;
 272
 273                 /* QP could be <nul> for user-mode client */
 274                 if (cm_id_priv->qp)
 275                         qp = cm_id_priv->qp;
 276                 else
 277                         ret = -EINVAL;
 278                 break;
 279         case IW_CM_STATE_LISTEN:
 280                 ret = -EINVAL;
 281                 break;
 282         case IW_CM_STATE_CLOSING:
 283                 /* remote peer closed first */
 284         case IW_CM_STATE_IDLE:
 285                 /* accept or connect returned !0 */
 286                 break;
 287         case IW_CM_STATE_CONN_RECV:
 288                 /*
 289                  * App called disconnect before/without calling accept after
 290                  * connect_request event delivered.
 291                  */
 292                 break;
 293         case IW_CM_STATE_CONN_SENT:
 294                 /* Can only get here if wait above fails */
 295         default:
 296                 BUG();
 297         }
 298         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 299
 300         if (qp) {
 301                 if (abrupt)
 302                         ret = iwcm_modify_qp_err(qp);
 303                 else
 304                         ret = iwcm_modify_qp_sqd(qp);
 305
 306                 /*
 307                  * If both sides are disconnecting the QP could
 308                  * already be in ERR or SQD states
 309                  */
 310                 ret = 0;
 311         }
 312
 313         return ret;
 314 }
 315 EXPORT_SYMBOL(iw_cm_disconnect);
 316
 317 /*
 318  * CM_ID <-- DESTROYING
 319  *
 320  * Clean up all resources associated with the connection and release
 321  * the initial reference taken by iw_create_cm_id.
 322  */
 323 static void destroy_cm_id(struct iw_cm_id *cm_id)
 324 {
 325         struct iwcm_id_private *cm_id_priv;
 326         unsigned long flags;
 327         int ret;
 328
 329         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 330         /*
 331          * Wait if we're currently in a connect or accept downcall. A
 332          * listening endpoint should never block here.
 333          */
 334         wait_event(cm_id_priv->connect_wait,
 335                    !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
 336
 337         spin_lock_irqsave(&cm_id_priv->lock, flags);
 338         switch (cm_id_priv->state) {
 339         case IW_CM_STATE_LISTEN:
 340                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 341                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 342                 /* destroy the listening endpoint */
 343                 ret = cm_id->device->iwcm->destroy_listen(cm_id);
 344                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 345                 break;
 346         case IW_CM_STATE_ESTABLISHED:
 347                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 348                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 349                 /* Abrupt close of the connection */
 350                 (void)iwcm_modify_qp_err(cm_id_priv->qp);
 351                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 352                 break;
 353         case IW_CM_STATE_IDLE:
 354         case IW_CM_STATE_CLOSING:
 355                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 356                 break;
 357         case IW_CM_STATE_CONN_RECV:
 358                 /*
 359                  * App called destroy before/without calling accept after
 360                  * receiving connection request event notification or
 361                  * returned non zero from the event callback function.
 362                  * In either case, must tell the provider to reject.
 363                  */
 364                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
 365                 break;
 366         case IW_CM_STATE_CONN_SENT:
 367         case IW_CM_STATE_DESTROYING:
 368         default:
 369                 BUG();
 370                 break;
 371         }
 372         if (cm_id_priv->qp) {
 373                 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
 374                 cm_id_priv->qp = NULL;
 375         }
 376         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 377
 378         (void)iwcm_deref_id(cm_id_priv);
 379 }
 380
 381 /*
 382  * This function is only called by the application thread and cannot
 383  * be called by the event thread. The function will wait for all
 384  * references to be released on the cm_id and then kfree the cm_id
 385  * object.
 386  */
 387 void iw_destroy_cm_id(struct iw_cm_id *cm_id)
 388 {
 389         struct iwcm_id_private *cm_id_priv;
 390
 391         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 392         BUG_ON(test_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags));
 393
 394         destroy_cm_id(cm_id);
 395
 396         wait_for_completion(&cm_id_priv->destroy_comp);
 397
 398         free_cm_id(cm_id_priv);
 399 }
 400 EXPORT_SYMBOL(iw_destroy_cm_id);
 401
 402 /*
 403  * CM_ID <-- LISTEN
 404  *
 405  * Start listening for connect requests. Generates one CONNECT_REQUEST
 406  * event for each inbound connect request.
 407  */
 408 int iw_cm_listen(struct iw_cm_id *cm_id, int backlog)
 409 {
 410         struct iwcm_id_private *cm_id_priv;
 411         unsigned long flags;
 412         int ret;
 413
 414         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 415
 416         ret = alloc_work_entries(cm_id_priv, backlog);
 417         if (ret)
 418                 return ret;
 419
 420         spin_lock_irqsave(&cm_id_priv->lock, flags);
 421         switch (cm_id_priv->state) {
 422         case IW_CM_STATE_IDLE:
 423                 cm_id_priv->state = IW_CM_STATE_LISTEN;
 424                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 425                 ret = cm_id->device->iwcm->create_listen(cm_id, backlog);
 426                 if (ret)
 427                         cm_id_priv->state = IW_CM_STATE_IDLE;
 428                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 429                 break;
 430         default:
 431                 ret = -EINVAL;
 432         }
 433         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 434
 435         return ret;
 436 }
 437 EXPORT_SYMBOL(iw_cm_listen);
 438
 439 /*
 440  * CM_ID <-- IDLE
 441  *
 442  * Rejects an inbound connection request. No events are generated.
 443  */
 444 int iw_cm_reject(struct iw_cm_id *cm_id,
 445                  const void *private_data,
 446                  u8 private_data_len)
 447 {
 448         struct iwcm_id_private *cm_id_priv;
 449         unsigned long flags;
 450         int ret;
 451
 452         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 453         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 454
 455         spin_lock_irqsave(&cm_id_priv->lock, flags);
 456         if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
 457                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 458                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 459                 wake_up_all(&cm_id_priv->connect_wait);
 460                 return -EINVAL;
 461         }
 462         cm_id_priv->state = IW_CM_STATE_IDLE;
 463         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 464
 465         ret = cm_id->device->iwcm->reject(cm_id, private_data,
 466                                           private_data_len);
 467
 468         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 469         wake_up_all(&cm_id_priv->connect_wait);
 470
 471         return ret;
 472 }
 473 EXPORT_SYMBOL(iw_cm_reject);
 474
 475 /*
 476  * CM_ID <-- ESTABLISHED
 477  *
 478  * Accepts an inbound connection request and generates an ESTABLISHED
 479  * event. Callers of iw_cm_disconnect and iw_destroy_cm_id will block
 480  * until the ESTABLISHED event is received from the provider.
 481  */
 482 int iw_cm_accept(struct iw_cm_id *cm_id,
 483                  struct iw_cm_conn_param *iw_param)
 484 {
 485         struct iwcm_id_private *cm_id_priv;
 486         struct ib_qp *qp;
 487         unsigned long flags;
 488         int ret;
 489
 490         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 491         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 492
 493         spin_lock_irqsave(&cm_id_priv->lock, flags);
 494         if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
 495                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 496                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 497                 wake_up_all(&cm_id_priv->connect_wait);
 498                 return -EINVAL;
 499         }
 500         /* Get the ib_qp given the QPN */
 501         qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
 502         if (!qp) {
 503                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 504                 return -EINVAL;
 505         }
 506         cm_id->device->iwcm->add_ref(qp);
 507         cm_id_priv->qp = qp;
 508         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 509
 510         ret = cm_id->device->iwcm->accept(cm_id, iw_param);
 511         if (ret) {
 512                 /* An error on accept precludes provider events */
 513                 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
 514                 cm_id_priv->state = IW_CM_STATE_IDLE;
 515                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 516                 if (cm_id_priv->qp) {
 517                         cm_id->device->iwcm->rem_ref(qp);
 518                         cm_id_priv->qp = NULL;
 519                 }
 520                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 521                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 522                 wake_up_all(&cm_id_priv->connect_wait);
 523         }
 524
 525         return ret;
 526 }
 527 EXPORT_SYMBOL(iw_cm_accept);
 528
 529 /*
 530  * Active Side: CM_ID <-- CONN_SENT
 531  *
 532  * If successful, results in the generation of a CONNECT_REPLY
 533  * event. iw_cm_disconnect and iw_cm_destroy will block until the
 534  * CONNECT_REPLY event is received from the provider.
 535  */
 536 int iw_cm_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *iw_param)
 537 {
 538         struct iwcm_id_private *cm_id_priv;
 539         int ret;
 540         unsigned long flags;
 541         struct ib_qp *qp;
 542
 543         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 544
 545         ret = alloc_work_entries(cm_id_priv, 4);
 546         if (ret)
 547                 return ret;
 548
 549         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 550         spin_lock_irqsave(&cm_id_priv->lock, flags);
 551
 552         if (cm_id_priv->state != IW_CM_STATE_IDLE) {
 553                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 554                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 555                 wake_up_all(&cm_id_priv->connect_wait);
 556                 return -EINVAL;
 557         }
 558
 559         /* Get the ib_qp given the QPN */
 560         qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
 561         if (!qp) {
 562                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 563                 return -EINVAL;
 564         }
 565         cm_id->device->iwcm->add_ref(qp);
 566         cm_id_priv->qp = qp;
 567         cm_id_priv->state = IW_CM_STATE_CONN_SENT;
 568         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 569
 570         ret = cm_id->device->iwcm->connect(cm_id, iw_param);
 571         if (ret) {
 572                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 573                 if (cm_id_priv->qp) {
 574                         cm_id->device->iwcm->rem_ref(qp);
 575                         cm_id_priv->qp = NULL;
 576                 }
 577                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 578                 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
 579                 cm_id_priv->state = IW_CM_STATE_IDLE;
 580                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 581                 wake_up_all(&cm_id_priv->connect_wait);
 582         }
 583
 584         return ret;
 585 }
 586 EXPORT_SYMBOL(iw_cm_connect);
 587
 588 /*
 589  * Passive Side: new CM_ID <-- CONN_RECV
 590  *
 591  * Handles an inbound connect request. The function creates a new
 592  * iw_cm_id to represent the new connection and inherits the client
 593  * callback function and other attributes from the listening parent.
 594  *
 595  * The work item contains a pointer to the listen_cm_id and the event. The
 596  * listen_cm_id contains the client cm_handler, context and
 597  * device. These are copied when the device is cloned. The event
 598  * contains the new four tuple.
 599  *
 600  * An error on the child should not affect the parent, so this
 601  * function does not return a value.
 602  */
 603 static void cm_conn_req_handler(struct iwcm_id_private *listen_id_priv,
 604                                 struct iw_cm_event *iw_event)
 605 {
 606         unsigned long flags;
 607         struct iw_cm_id *cm_id;
 608         struct iwcm_id_private *cm_id_priv;
 609         int ret;
 610
 611         /*
 612          * The provider should never generate a connection request
 613          * event with a bad status.
 614          */
 615         BUG_ON(iw_event->status);
 616
 617         /*
 618          * We could be destroying the listening id. If so, ignore this
 619          * upcall.
 620          */
 621         spin_lock_irqsave(&listen_id_priv->lock, flags);
 622         if (listen_id_priv->state != IW_CM_STATE_LISTEN) {
 623                 spin_unlock_irqrestore(&listen_id_priv->lock, flags);
 624                 goto out;
 625         }
 626         spin_unlock_irqrestore(&listen_id_priv->lock, flags);
 627
 628         cm_id = iw_create_cm_id(listen_id_priv->id.device,
 629                                 listen_id_priv->id.cm_handler,
 630                                 listen_id_priv->id.context);
 631         /* If the cm_id could not be created, ignore the request */
 632         if (IS_ERR(cm_id))
 633                 goto out;
 634
 635         cm_id->provider_data = iw_event->provider_data;
 636         cm_id->local_addr = iw_event->local_addr;
 637         cm_id->remote_addr = iw_event->remote_addr;
 638
 639         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 640         cm_id_priv->state = IW_CM_STATE_CONN_RECV;
 641
 642         ret = alloc_work_entries(cm_id_priv, 3);
 643         if (ret) {
 644                 iw_cm_reject(cm_id, NULL, 0);
 645                 iw_destroy_cm_id(cm_id);
 646                 goto out;
 647         }
 648
 649         /* Call the client CM handler */
 650         ret = cm_id->cm_handler(cm_id, iw_event);
 651         if (ret) {
 652                 iw_cm_reject(cm_id, NULL, 0);
 653                 set_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
 654                 destroy_cm_id(cm_id);
 655                 if (atomic_read(&cm_id_priv->refcount)==0)
 656                         free_cm_id(cm_id_priv);
 657         }
 658
 659 out:
 660         if (iw_event->private_data_len)
 661                 kfree(iw_event->private_data);
 662 }
 663
 664 /*
 665  * Passive Side: CM_ID <-- ESTABLISHED
 666  *
 667  * The provider generated an ESTABLISHED event which means that
 668  * the MPA negotion has completed successfully and we are now in MPA
 669  * FPDU mode.
 670  *
 671  * This event can only be received in the CONN_RECV state. If the
 672  * remote peer closed, the ESTABLISHED event would be received followed
 673  * by the CLOSE event. If the app closes, it will block until we wake
 674  * it up after processing this event.
 675  */
 676 static int cm_conn_est_handler(struct iwcm_id_private *cm_id_priv,
 677                                struct iw_cm_event *iw_event)
 678 {
 679         unsigned long flags;
 680         int ret;
 681
 682         spin_lock_irqsave(&cm_id_priv->lock, flags);
 683
 684         /*
 685          * We clear the CONNECT_WAIT bit here to allow the callback
 686          * function to call iw_cm_disconnect. Calling iw_destroy_cm_id
 687          * from a callback handler is not allowed.
 688          */
 689         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 690         BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
 691         cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
 692         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 693         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 694         wake_up_all(&cm_id_priv->connect_wait);
 695
 696         return ret;
 697 }
 698
 699 /*
 700  * Active Side: CM_ID <-- ESTABLISHED
 701  *
 702  * The app has called connect and is waiting for the established event to
 703  * post it's requests to the server. This event will wake up anyone
 704  * blocked in iw_cm_disconnect or iw_destroy_id.
 705  */
 706 static int cm_conn_rep_handler(struct iwcm_id_private *cm_id_priv,
 707                                struct iw_cm_event *iw_event)
 708 {
 709         unsigned long flags;
 710         int ret;
 711
 712         spin_lock_irqsave(&cm_id_priv->lock, flags);
 713         /*
 714          * Clear the connect wait bit so a callback function calling
 715          * iw_cm_disconnect will not wait and deadlock this thread
 716          */
 717         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 718         BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
 719         if (iw_event->status == IW_CM_EVENT_STATUS_ACCEPTED) {
 720                 cm_id_priv->id.local_addr = iw_event->local_addr;
 721                 cm_id_priv->id.remote_addr = iw_event->remote_addr;
 722                 cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
 723         } else {
 724                 /* REJECTED or RESET */
 725                 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
 726                 cm_id_priv->qp = NULL;
 727                 cm_id_priv->state = IW_CM_STATE_IDLE;
 728         }
 729         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 730         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 731
 732         if (iw_event->private_data_len)
 733                 kfree(iw_event->private_data);
 734
 735         /* Wake up waiters on connect complete */
 736         wake_up_all(&cm_id_priv->connect_wait);
 737
 738         return ret;
 739 }
 740
 741 /*
 742  * CM_ID <-- CLOSING
 743  *
 744  * If in the ESTABLISHED state, move to CLOSING.
 745  */
 746 static void cm_disconnect_handler(struct iwcm_id_private *cm_id_priv,
 747                                   struct iw_cm_event *iw_event)
 748 {
 749         unsigned long flags;
 750
 751         spin_lock_irqsave(&cm_id_priv->lock, flags);
 752         if (cm_id_priv->state == IW_CM_STATE_ESTABLISHED)
 753                 cm_id_priv->state = IW_CM_STATE_CLOSING;
 754         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 755 }
 756
 757 /*
 758  * CM_ID <-- IDLE
 759  *
 760  * If in the ESTBLISHED or CLOSING states, the QP will have have been
 761  * moved by the provider to the ERR state. Disassociate the CM_ID from
 762  * the QP,  move to IDLE, and remove the 'connected' reference.
 763  *
 764  * If in some other state, the cm_id was destroyed asynchronously.
 765  * This is the last reference that will result in waking up
 766  * the app thread blocked in iw_destroy_cm_id.
 767  */
 768 static int cm_close_handler(struct iwcm_id_private *cm_id_priv,
 769                                   struct iw_cm_event *iw_event)
 770 {
 771         unsigned long flags;
 772         int ret = 0;
 773         spin_lock_irqsave(&cm_id_priv->lock, flags);
 774
 775         if (cm_id_priv->qp) {
 776                 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
 777                 cm_id_priv->qp = NULL;
 778         }
 779         switch (cm_id_priv->state) {
 780         case IW_CM_STATE_ESTABLISHED:
 781         case IW_CM_STATE_CLOSING:
 782                 cm_id_priv->state = IW_CM_STATE_IDLE;
 783                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 784                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 785                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 786                 break;
 787         case IW_CM_STATE_DESTROYING:
 788                 break;
 789         default:
 790                 BUG();
 791         }
 792         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 793
 794         return ret;
 795 }
 796
 797 static int process_event(struct iwcm_id_private *cm_id_priv,
 798                          struct iw_cm_event *iw_event)
 799 {
 800         int ret = 0;
 801
 802         switch (iw_event->event) {
 803         case IW_CM_EVENT_CONNECT_REQUEST:
 804                 cm_conn_req_handler(cm_id_priv, iw_event);
 805                 break;
 806         case IW_CM_EVENT_CONNECT_REPLY:
 807                 ret = cm_conn_rep_handler(cm_id_priv, iw_event);
 808                 break;
 809         case IW_CM_EVENT_ESTABLISHED:
 810                 ret = cm_conn_est_handler(cm_id_priv, iw_event);
 811                 break;
 812         case IW_CM_EVENT_DISCONNECT:
 813                 cm_disconnect_handler(cm_id_priv, iw_event);
 814                 break;
 815         case IW_CM_EVENT_CLOSE:
 816                 ret = cm_close_handler(cm_id_priv, iw_event);
 817                 break;
 818         default:
 819                 BUG();
 820         }
 821
 822         return ret;
 823 }
 824
 825 /*
 826  * Process events on the work_list for the cm_id. If the callback
 827  * function requests that the cm_id be deleted, a flag is set in the
 828  * cm_id flags to indicate that when the last reference is
 829  * removed, the cm_id is to be destroyed. This is necessary to
 830  * distinguish between an object that will be destroyed by the app
 831  * thread asleep on the destroy_comp list vs. an object destroyed
 832  * here synchronously when the last reference is removed.
 833  */
 834 static void cm_work_handler(struct work_struct *_work)
 835 {
 836         struct iwcm_work *work = container_of(_work, struct iwcm_work, work);
 837         struct iw_cm_event levent;
 838         struct iwcm_id_private *cm_id_priv = work->cm_id;
 839         unsigned long flags;
 840         int empty;
 841         int ret = 0;
 842
 843         spin_lock_irqsave(&cm_id_priv->lock, flags);
 844         empty = list_empty(&cm_id_priv->work_list);
 845         while (!empty) {
 846                 work = list_entry(cm_id_priv->work_list.next,
 847                                   struct iwcm_work, list);
 848                 list_del_init(&work->list);
 849                 empty = list_empty(&cm_id_priv->work_list);
 850                 levent = work->event;
 851                 put_work(work);
 852                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 853
 854                 ret = process_event(cm_id_priv, &levent);
 855                 if (ret) {
 856                         set_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
 857                         destroy_cm_id(&cm_id_priv->id);
 858                 }
 859                 BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
 860                 if (iwcm_deref_id(cm_id_priv)) {
 861                         if (test_bit(IWCM_F_CALLBACK_DESTROY,
 862                                      &cm_id_priv->flags)) {
 863                                 BUG_ON(!list_empty(&cm_id_priv->work_list));
 864                                 free_cm_id(cm_id_priv);
 865                         }
 866                         return;
 867                 }
 868                 spin_lock_irqsave(&cm_id_priv->lock, flags);
 869         }
 870         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 871 }
 872
 873 /*
 874  * This function is called on interrupt context. Schedule events on
 875  * the iwcm_wq thread to allow callback functions to downcall into
 876  * the CM and/or block.  Events are queued to a per-CM_ID
 877  * work_list. If this is the first event on the work_list, the work
 878  * element is also queued on the iwcm_wq thread.
 879  *
 880  * Each event holds a reference on the cm_id. Until the last posted
 881  * event has been delivered and processed, the cm_id cannot be
 882  * deleted.
 883  *
 884  * Returns:
 885  *            0 - the event was handled.
 886  *      -ENOMEM - the event was not handled due to lack of resources.
 887  */
 888 static int cm_event_handler(struct iw_cm_id *cm_id,
 889                              struct iw_cm_event *iw_event)
 890 {
 891         struct iwcm_work *work;
 892         struct iwcm_id_private *cm_id_priv;
 893         unsigned long flags;
 894         int ret = 0;
 895
 896         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 897
 898         spin_lock_irqsave(&cm_id_priv->lock, flags);
 899         work = get_work(cm_id_priv);
 900         if (!work) {
 901                 ret = -ENOMEM;
 902                 goto out;
 903         }
 904
 905         INIT_WORK(&work->work, cm_work_handler);
 906         work->cm_id = cm_id_priv;
 907         work->event = *iw_event;
 908
 909         if ((work->event.event == IW_CM_EVENT_CONNECT_REQUEST ||
 910              work->event.event == IW_CM_EVENT_CONNECT_REPLY) &&
 911             work->event.private_data_len) {
 912                 ret = copy_private_data(&work->event);
 913                 if (ret) {
 914                         put_work(work);
 915                         goto out;
 916                 }
 917         }
 918
 919         atomic_inc(&cm_id_priv->refcount);
 920         if (list_empty(&cm_id_priv->work_list)) {
 921                 list_add_tail(&work->list, &cm_id_priv->work_list);
 922                 queue_work(iwcm_wq, &work->work);
 923         } else
 924                 list_add_tail(&work->list, &cm_id_priv->work_list);
 925 out:
 926         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 927         return ret;
 928 }
 929
 930 static int iwcm_init_qp_init_attr(struct iwcm_id_private *cm_id_priv,
 931                                   struct ib_qp_attr *qp_attr,
 932                                   int *qp_attr_mask)
 933 {
 934         unsigned long flags;
 935         int ret;
 936
 937         spin_lock_irqsave(&cm_id_priv->lock, flags);
 938         switch (cm_id_priv->state) {
 939         case IW_CM_STATE_IDLE:
 940         case IW_CM_STATE_CONN_SENT:
 941         case IW_CM_STATE_CONN_RECV:
 942         case IW_CM_STATE_ESTABLISHED:
 943                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 944                 qp_attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE |
 945                                            IB_ACCESS_REMOTE_WRITE|
 946                                            IB_ACCESS_REMOTE_READ;
 947                 ret = 0;
 948                 break;
 949         default:
 950                 ret = -EINVAL;
 951                 break;
 952         }
 953         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 954         return ret;
 955 }
 956
 957 static int iwcm_init_qp_rts_attr(struct iwcm_id_private *cm_id_priv,
 958                                   struct ib_qp_attr *qp_attr,
 959                                   int *qp_attr_mask)
 960 {
 961         unsigned long flags;
 962         int ret;
 963
 964         spin_lock_irqsave(&cm_id_priv->lock, flags);
 965         switch (cm_id_priv->state) {
 966         case IW_CM_STATE_IDLE:
 967         case IW_CM_STATE_CONN_SENT:
 968         case IW_CM_STATE_CONN_RECV:
 969         case IW_CM_STATE_ESTABLISHED:
 970                 *qp_attr_mask = 0;
 971                 ret = 0;
 972                 break;
 973         default:
 974                 ret = -EINVAL;
 975                 break;
 976         }
 977         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 978         return ret;
 979 }
 980
 981 int iw_cm_init_qp_attr(struct iw_cm_id *cm_id,
 982                        struct ib_qp_attr *qp_attr,
 983                        int *qp_attr_mask)
 984 {
 985         struct iwcm_id_private *cm_id_priv;
 986         int ret;
 987
 988         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 989         switch (qp_attr->qp_state) {
 990         case IB_QPS_INIT:
 991         case IB_QPS_RTR:
 992                 ret = iwcm_init_qp_init_attr(cm_id_priv,
 993                                              qp_attr, qp_attr_mask);
 994                 break;
 995         case IB_QPS_RTS:
 996                 ret = iwcm_init_qp_rts_attr(cm_id_priv,
 997                                             qp_attr, qp_attr_mask);
 998                 break;
 999         default:
1000                 ret = -EINVAL;
1001                 break;
1002         }
1003         return ret;
1004 }
1005 EXPORT_SYMBOL(iw_cm_init_qp_attr);
1006
1007 static int __init iw_cm_init(void)
1008 {
1009         iwcm_wq = create_singlethread_workqueue("iw_cm_wq");
1010         if (!iwcm_wq)
1011                 return -ENOMEM;
1012
1013         return 0;
1014 }
1015
1016 static void __exit iw_cm_cleanup(void)
1017 {
1018         destroy_workqueue(iwcm_wq);
1019 }
1020
1021 module_init(iw_cm_init);
1022 module_exit(iw_cm_cleanup);