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