Linux Audio

Check our new training course

Loading...
v3.1
   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
  49#include <rdma/iw_cm.h>
  50#include <rdma/ib_addr.h>
 
 
  51
  52#include "iwcm.h"
  53
  54MODULE_AUTHOR("Tom Tucker");
  55MODULE_DESCRIPTION("iWARP CM");
  56MODULE_LICENSE("Dual BSD/GPL");
  57
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  58static struct workqueue_struct *iwcm_wq;
  59struct iwcm_work {
  60	struct work_struct work;
  61	struct iwcm_id_private *cm_id;
  62	struct list_head list;
  63	struct iw_cm_event event;
  64	struct list_head free_list;
  65};
  66
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  67/*
  68 * The following services provide a mechanism for pre-allocating iwcm_work
  69 * elements.  The design pre-allocates them  based on the cm_id type:
  70 *	LISTENING IDS: 	Get enough elements preallocated to handle the
  71 *			listen backlog.
  72 *	ACTIVE IDS:	4: CONNECT_REPLY, ESTABLISHED, DISCONNECT, CLOSE
  73 *	PASSIVE IDS:	3: ESTABLISHED, DISCONNECT, CLOSE
  74 *
  75 * Allocating them in connect and listen avoids having to deal
  76 * with allocation failures on the event upcall from the provider (which
  77 * is called in the interrupt context).
  78 *
  79 * One exception is when creating the cm_id for incoming connection requests.
  80 * There are two cases:
  81 * 1) in the event upcall, cm_event_handler(), for a listening cm_id.  If
  82 *    the backlog is exceeded, then no more connection request events will
  83 *    be processed.  cm_event_handler() returns -ENOMEM in this case.  Its up
  84 *    to the provider to reject the connection request.
  85 * 2) in the connection request workqueue handler, cm_conn_req_handler().
  86 *    If work elements cannot be allocated for the new connect request cm_id,
  87 *    then IWCM will call the provider reject method.  This is ok since
  88 *    cm_conn_req_handler() runs in the workqueue thread context.
  89 */
  90
  91static struct iwcm_work *get_work(struct iwcm_id_private *cm_id_priv)
  92{
  93	struct iwcm_work *work;
  94
  95	if (list_empty(&cm_id_priv->work_free_list))
  96		return NULL;
  97	work = list_entry(cm_id_priv->work_free_list.next, struct iwcm_work,
  98			  free_list);
  99	list_del_init(&work->free_list);
 100	return work;
 101}
 102
 103static void put_work(struct iwcm_work *work)
 104{
 105	list_add(&work->free_list, &work->cm_id->work_free_list);
 106}
 107
 108static void dealloc_work_entries(struct iwcm_id_private *cm_id_priv)
 109{
 110	struct list_head *e, *tmp;
 111
 112	list_for_each_safe(e, tmp, &cm_id_priv->work_free_list)
 
 113		kfree(list_entry(e, struct iwcm_work, free_list));
 
 114}
 115
 116static int alloc_work_entries(struct iwcm_id_private *cm_id_priv, int count)
 117{
 118	struct iwcm_work *work;
 119
 120	BUG_ON(!list_empty(&cm_id_priv->work_free_list));
 121	while (count--) {
 122		work = kmalloc(sizeof(struct iwcm_work), GFP_KERNEL);
 123		if (!work) {
 124			dealloc_work_entries(cm_id_priv);
 125			return -ENOMEM;
 126		}
 127		work->cm_id = cm_id_priv;
 128		INIT_LIST_HEAD(&work->list);
 129		put_work(work);
 130	}
 131	return 0;
 132}
 133
 134/*
 135 * Save private data from incoming connection requests to
 136 * iw_cm_event, so the low level driver doesn't have to. Adjust
 137 * the event ptr to point to the local copy.
 138 */
 139static int copy_private_data(struct iw_cm_event *event)
 140{
 141	void *p;
 142
 143	p = kmemdup(event->private_data, event->private_data_len, GFP_ATOMIC);
 144	if (!p)
 145		return -ENOMEM;
 146	event->private_data = p;
 147	return 0;
 148}
 149
 150static void free_cm_id(struct iwcm_id_private *cm_id_priv)
 151{
 152	dealloc_work_entries(cm_id_priv);
 153	kfree(cm_id_priv);
 154}
 155
 156/*
 157 * Release a reference on cm_id. If the last reference is being
 158 * released, enable the waiting thread (in iw_destroy_cm_id) to
 159 * get woken up, and return 1 if a thread is already waiting.
 160 */
 161static int iwcm_deref_id(struct iwcm_id_private *cm_id_priv)
 162{
 163	BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
 164	if (atomic_dec_and_test(&cm_id_priv->refcount)) {
 165		BUG_ON(!list_empty(&cm_id_priv->work_list));
 166		complete(&cm_id_priv->destroy_comp);
 167		return 1;
 168	}
 169
 170	return 0;
 171}
 172
 173static void add_ref(struct iw_cm_id *cm_id)
 174{
 175	struct iwcm_id_private *cm_id_priv;
 176	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 177	atomic_inc(&cm_id_priv->refcount);
 178}
 179
 180static void rem_ref(struct iw_cm_id *cm_id)
 181{
 182	struct iwcm_id_private *cm_id_priv;
 
 183	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 184	if (iwcm_deref_id(cm_id_priv) &&
 185	    test_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags)) {
 186		BUG_ON(!list_empty(&cm_id_priv->work_list));
 187		free_cm_id(cm_id_priv);
 188	}
 189}
 190
 191static int cm_event_handler(struct iw_cm_id *cm_id, struct iw_cm_event *event);
 192
 193struct iw_cm_id *iw_create_cm_id(struct ib_device *device,
 194				 iw_cm_handler cm_handler,
 195				 void *context)
 196{
 197	struct iwcm_id_private *cm_id_priv;
 198
 199	cm_id_priv = kzalloc(sizeof(*cm_id_priv), GFP_KERNEL);
 200	if (!cm_id_priv)
 201		return ERR_PTR(-ENOMEM);
 202
 203	cm_id_priv->state = IW_CM_STATE_IDLE;
 204	cm_id_priv->id.device = device;
 205	cm_id_priv->id.cm_handler = cm_handler;
 206	cm_id_priv->id.context = context;
 207	cm_id_priv->id.event_handler = cm_event_handler;
 208	cm_id_priv->id.add_ref = add_ref;
 209	cm_id_priv->id.rem_ref = rem_ref;
 210	spin_lock_init(&cm_id_priv->lock);
 211	atomic_set(&cm_id_priv->refcount, 1);
 212	init_waitqueue_head(&cm_id_priv->connect_wait);
 213	init_completion(&cm_id_priv->destroy_comp);
 214	INIT_LIST_HEAD(&cm_id_priv->work_list);
 215	INIT_LIST_HEAD(&cm_id_priv->work_free_list);
 216
 217	return &cm_id_priv->id;
 218}
 219EXPORT_SYMBOL(iw_create_cm_id);
 220
 221
 222static int iwcm_modify_qp_err(struct ib_qp *qp)
 223{
 224	struct ib_qp_attr qp_attr;
 225
 226	if (!qp)
 227		return -EINVAL;
 228
 229	qp_attr.qp_state = IB_QPS_ERR;
 230	return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 231}
 232
 233/*
 234 * This is really the RDMAC CLOSING state. It is most similar to the
 235 * IB SQD QP state.
 236 */
 237static int iwcm_modify_qp_sqd(struct ib_qp *qp)
 238{
 239	struct ib_qp_attr qp_attr;
 240
 241	BUG_ON(qp == NULL);
 242	qp_attr.qp_state = IB_QPS_SQD;
 243	return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 244}
 245
 246/*
 247 * CM_ID <-- CLOSING
 248 *
 249 * Block if a passive or active connection is currently being processed. Then
 250 * process the event as follows:
 251 * - If we are ESTABLISHED, move to CLOSING and modify the QP state
 252 *   based on the abrupt flag
 253 * - If the connection is already in the CLOSING or IDLE state, the peer is
 254 *   disconnecting concurrently with us and we've already seen the
 255 *   DISCONNECT event -- ignore the request and return 0
 256 * - Disconnect on a listening endpoint returns -EINVAL
 257 */
 258int iw_cm_disconnect(struct iw_cm_id *cm_id, int abrupt)
 259{
 260	struct iwcm_id_private *cm_id_priv;
 261	unsigned long flags;
 262	int ret = 0;
 263	struct ib_qp *qp = NULL;
 264
 265	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 266	/* Wait if we're currently in a connect or accept downcall */
 267	wait_event(cm_id_priv->connect_wait,
 268		   !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
 269
 270	spin_lock_irqsave(&cm_id_priv->lock, flags);
 271	switch (cm_id_priv->state) {
 272	case IW_CM_STATE_ESTABLISHED:
 273		cm_id_priv->state = IW_CM_STATE_CLOSING;
 274
 275		/* QP could be <nul> for user-mode client */
 276		if (cm_id_priv->qp)
 277			qp = cm_id_priv->qp;
 278		else
 279			ret = -EINVAL;
 280		break;
 281	case IW_CM_STATE_LISTEN:
 282		ret = -EINVAL;
 283		break;
 284	case IW_CM_STATE_CLOSING:
 285		/* remote peer closed first */
 286	case IW_CM_STATE_IDLE:
 287		/* accept or connect returned !0 */
 288		break;
 289	case IW_CM_STATE_CONN_RECV:
 290		/*
 291		 * App called disconnect before/without calling accept after
 292		 * connect_request event delivered.
 293		 */
 294		break;
 295	case IW_CM_STATE_CONN_SENT:
 296		/* Can only get here if wait above fails */
 297	default:
 298		BUG();
 299	}
 300	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 301
 302	if (qp) {
 303		if (abrupt)
 304			ret = iwcm_modify_qp_err(qp);
 305		else
 306			ret = iwcm_modify_qp_sqd(qp);
 307
 308		/*
 309		 * If both sides are disconnecting the QP could
 310		 * already be in ERR or SQD states
 311		 */
 312		ret = 0;
 313	}
 314
 315	return ret;
 316}
 317EXPORT_SYMBOL(iw_cm_disconnect);
 318
 319/*
 320 * CM_ID <-- DESTROYING
 321 *
 322 * Clean up all resources associated with the connection and release
 323 * the initial reference taken by iw_create_cm_id.
 324 */
 325static void destroy_cm_id(struct iw_cm_id *cm_id)
 326{
 327	struct iwcm_id_private *cm_id_priv;
 
 328	unsigned long flags;
 329	int ret;
 330
 331	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 332	/*
 333	 * Wait if we're currently in a connect or accept downcall. A
 334	 * listening endpoint should never block here.
 335	 */
 336	wait_event(cm_id_priv->connect_wait,
 337		   !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
 338
 
 
 
 
 
 
 339	spin_lock_irqsave(&cm_id_priv->lock, flags);
 
 
 
 340	switch (cm_id_priv->state) {
 341	case IW_CM_STATE_LISTEN:
 342		cm_id_priv->state = IW_CM_STATE_DESTROYING;
 343		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 344		/* destroy the listening endpoint */
 345		ret = cm_id->device->iwcm->destroy_listen(cm_id);
 346		spin_lock_irqsave(&cm_id_priv->lock, flags);
 347		break;
 348	case IW_CM_STATE_ESTABLISHED:
 349		cm_id_priv->state = IW_CM_STATE_DESTROYING;
 350		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 351		/* Abrupt close of the connection */
 352		(void)iwcm_modify_qp_err(cm_id_priv->qp);
 353		spin_lock_irqsave(&cm_id_priv->lock, flags);
 354		break;
 355	case IW_CM_STATE_IDLE:
 356	case IW_CM_STATE_CLOSING:
 357		cm_id_priv->state = IW_CM_STATE_DESTROYING;
 358		break;
 359	case IW_CM_STATE_CONN_RECV:
 360		/*
 361		 * App called destroy before/without calling accept after
 362		 * receiving connection request event notification or
 363		 * returned non zero from the event callback function.
 364		 * In either case, must tell the provider to reject.
 365		 */
 366		cm_id_priv->state = IW_CM_STATE_DESTROYING;
 367		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 368		cm_id->device->iwcm->reject(cm_id, NULL, 0);
 369		spin_lock_irqsave(&cm_id_priv->lock, flags);
 370		break;
 371	case IW_CM_STATE_CONN_SENT:
 372	case IW_CM_STATE_DESTROYING:
 373	default:
 374		BUG();
 375		break;
 376	}
 377	if (cm_id_priv->qp) {
 378		cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
 379		cm_id_priv->qp = NULL;
 380	}
 381	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 
 
 
 
 
 
 
 382
 383	(void)iwcm_deref_id(cm_id_priv);
 384}
 385
 386/*
 387 * This function is only called by the application thread and cannot
 388 * be called by the event thread. The function will wait for all
 389 * references to be released on the cm_id and then kfree the cm_id
 390 * object.
 391 */
 392void iw_destroy_cm_id(struct iw_cm_id *cm_id)
 393{
 394	struct iwcm_id_private *cm_id_priv;
 
 
 395
 396	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 397	BUG_ON(test_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 398
 399	destroy_cm_id(cm_id);
 
 
 
 400
 401	wait_for_completion(&cm_id_priv->destroy_comp);
 
 
 
 
 402
 403	free_cm_id(cm_id_priv);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 404}
 405EXPORT_SYMBOL(iw_destroy_cm_id);
 406
 407/*
 408 * CM_ID <-- LISTEN
 409 *
 410 * Start listening for connect requests. Generates one CONNECT_REQUEST
 411 * event for each inbound connect request.
 412 */
 413int iw_cm_listen(struct iw_cm_id *cm_id, int backlog)
 414{
 415	struct iwcm_id_private *cm_id_priv;
 416	unsigned long flags;
 417	int ret;
 418
 419	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 420
 
 
 
 421	ret = alloc_work_entries(cm_id_priv, backlog);
 422	if (ret)
 423		return ret;
 424
 425	spin_lock_irqsave(&cm_id_priv->lock, flags);
 426	switch (cm_id_priv->state) {
 427	case IW_CM_STATE_IDLE:
 428		cm_id_priv->state = IW_CM_STATE_LISTEN;
 429		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 430		ret = cm_id->device->iwcm->create_listen(cm_id, backlog);
 
 
 
 431		if (ret)
 432			cm_id_priv->state = IW_CM_STATE_IDLE;
 433		spin_lock_irqsave(&cm_id_priv->lock, flags);
 434		break;
 435	default:
 436		ret = -EINVAL;
 437	}
 438	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 439
 440	return ret;
 441}
 442EXPORT_SYMBOL(iw_cm_listen);
 443
 444/*
 445 * CM_ID <-- IDLE
 446 *
 447 * Rejects an inbound connection request. No events are generated.
 448 */
 449int iw_cm_reject(struct iw_cm_id *cm_id,
 450		 const void *private_data,
 451		 u8 private_data_len)
 452{
 453	struct iwcm_id_private *cm_id_priv;
 454	unsigned long flags;
 455	int ret;
 456
 457	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 458	set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 459
 460	spin_lock_irqsave(&cm_id_priv->lock, flags);
 461	if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
 462		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 463		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 464		wake_up_all(&cm_id_priv->connect_wait);
 465		return -EINVAL;
 466	}
 467	cm_id_priv->state = IW_CM_STATE_IDLE;
 468	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 469
 470	ret = cm_id->device->iwcm->reject(cm_id, private_data,
 471					  private_data_len);
 472
 473	clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 474	wake_up_all(&cm_id_priv->connect_wait);
 475
 476	return ret;
 477}
 478EXPORT_SYMBOL(iw_cm_reject);
 479
 480/*
 481 * CM_ID <-- ESTABLISHED
 482 *
 483 * Accepts an inbound connection request and generates an ESTABLISHED
 484 * event. Callers of iw_cm_disconnect and iw_destroy_cm_id will block
 485 * until the ESTABLISHED event is received from the provider.
 486 */
 487int iw_cm_accept(struct iw_cm_id *cm_id,
 488		 struct iw_cm_conn_param *iw_param)
 489{
 490	struct iwcm_id_private *cm_id_priv;
 491	struct ib_qp *qp;
 492	unsigned long flags;
 493	int ret;
 494
 495	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 496	set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 497
 498	spin_lock_irqsave(&cm_id_priv->lock, flags);
 499	if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
 500		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 501		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 502		wake_up_all(&cm_id_priv->connect_wait);
 503		return -EINVAL;
 504	}
 505	/* Get the ib_qp given the QPN */
 506	qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
 507	if (!qp) {
 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	cm_id->device->iwcm->add_ref(qp);
 514	cm_id_priv->qp = qp;
 515	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 516
 517	ret = cm_id->device->iwcm->accept(cm_id, iw_param);
 518	if (ret) {
 519		/* An error on accept precludes provider events */
 520		BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
 521		cm_id_priv->state = IW_CM_STATE_IDLE;
 522		spin_lock_irqsave(&cm_id_priv->lock, flags);
 523		if (cm_id_priv->qp) {
 524			cm_id->device->iwcm->rem_ref(qp);
 525			cm_id_priv->qp = NULL;
 526		}
 527		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 
 
 528		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 529		wake_up_all(&cm_id_priv->connect_wait);
 530	}
 531
 532	return ret;
 533}
 534EXPORT_SYMBOL(iw_cm_accept);
 535
 536/*
 537 * Active Side: CM_ID <-- CONN_SENT
 538 *
 539 * If successful, results in the generation of a CONNECT_REPLY
 540 * event. iw_cm_disconnect and iw_cm_destroy will block until the
 541 * CONNECT_REPLY event is received from the provider.
 542 */
 543int iw_cm_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *iw_param)
 544{
 545	struct iwcm_id_private *cm_id_priv;
 546	int ret;
 547	unsigned long flags;
 548	struct ib_qp *qp;
 549
 550	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 551
 552	ret = alloc_work_entries(cm_id_priv, 4);
 553	if (ret)
 554		return ret;
 555
 556	set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 557	spin_lock_irqsave(&cm_id_priv->lock, flags);
 558
 559	if (cm_id_priv->state != IW_CM_STATE_IDLE) {
 560		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 561		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 562		wake_up_all(&cm_id_priv->connect_wait);
 563		return -EINVAL;
 564	}
 565
 566	/* Get the ib_qp given the QPN */
 567	qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
 568	if (!qp) {
 569		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 570		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 571		wake_up_all(&cm_id_priv->connect_wait);
 572		return -EINVAL;
 573	}
 574	cm_id->device->iwcm->add_ref(qp);
 575	cm_id_priv->qp = qp;
 576	cm_id_priv->state = IW_CM_STATE_CONN_SENT;
 577	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 578
 579	ret = cm_id->device->iwcm->connect(cm_id, iw_param);
 580	if (ret) {
 581		spin_lock_irqsave(&cm_id_priv->lock, flags);
 582		if (cm_id_priv->qp) {
 583			cm_id->device->iwcm->rem_ref(qp);
 584			cm_id_priv->qp = NULL;
 585		}
 586		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 587		BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
 588		cm_id_priv->state = IW_CM_STATE_IDLE;
 589		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 590		wake_up_all(&cm_id_priv->connect_wait);
 591	}
 592
 
 
 
 
 
 
 
 
 
 
 593	return ret;
 594}
 595EXPORT_SYMBOL(iw_cm_connect);
 596
 597/*
 598 * Passive Side: new CM_ID <-- CONN_RECV
 599 *
 600 * Handles an inbound connect request. The function creates a new
 601 * iw_cm_id to represent the new connection and inherits the client
 602 * callback function and other attributes from the listening parent.
 603 *
 604 * The work item contains a pointer to the listen_cm_id and the event. The
 605 * listen_cm_id contains the client cm_handler, context and
 606 * device. These are copied when the device is cloned. The event
 607 * contains the new four tuple.
 608 *
 609 * An error on the child should not affect the parent, so this
 610 * function does not return a value.
 611 */
 612static void cm_conn_req_handler(struct iwcm_id_private *listen_id_priv,
 613				struct iw_cm_event *iw_event)
 614{
 615	unsigned long flags;
 616	struct iw_cm_id *cm_id;
 617	struct iwcm_id_private *cm_id_priv;
 618	int ret;
 619
 620	/*
 621	 * The provider should never generate a connection request
 622	 * event with a bad status.
 623	 */
 624	BUG_ON(iw_event->status);
 625
 626	/*
 627	 * We could be destroying the listening id. If so, ignore this
 628	 * upcall.
 629	 */
 630	spin_lock_irqsave(&listen_id_priv->lock, flags);
 631	if (listen_id_priv->state != IW_CM_STATE_LISTEN) {
 632		spin_unlock_irqrestore(&listen_id_priv->lock, flags);
 633		goto out;
 634	}
 635	spin_unlock_irqrestore(&listen_id_priv->lock, flags);
 636
 637	cm_id = iw_create_cm_id(listen_id_priv->id.device,
 638				listen_id_priv->id.cm_handler,
 639				listen_id_priv->id.context);
 640	/* If the cm_id could not be created, ignore the request */
 641	if (IS_ERR(cm_id))
 642		goto out;
 643
 644	cm_id->provider_data = iw_event->provider_data;
 645	cm_id->local_addr = iw_event->local_addr;
 646	cm_id->remote_addr = iw_event->remote_addr;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 647
 648	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 649	cm_id_priv->state = IW_CM_STATE_CONN_RECV;
 650
 
 
 
 
 
 
 
 
 
 
 
 
 
 651	ret = alloc_work_entries(cm_id_priv, 3);
 652	if (ret) {
 653		iw_cm_reject(cm_id, NULL, 0);
 654		iw_destroy_cm_id(cm_id);
 655		goto out;
 656	}
 657
 658	/* Call the client CM handler */
 659	ret = cm_id->cm_handler(cm_id, iw_event);
 660	if (ret) {
 661		iw_cm_reject(cm_id, NULL, 0);
 662		set_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
 663		destroy_cm_id(cm_id);
 664		if (atomic_read(&cm_id_priv->refcount)==0)
 665			free_cm_id(cm_id_priv);
 666	}
 667
 668out:
 669	if (iw_event->private_data_len)
 670		kfree(iw_event->private_data);
 671}
 672
 673/*
 674 * Passive Side: CM_ID <-- ESTABLISHED
 675 *
 676 * The provider generated an ESTABLISHED event which means that
 677 * the MPA negotion has completed successfully and we are now in MPA
 678 * FPDU mode.
 679 *
 680 * This event can only be received in the CONN_RECV state. If the
 681 * remote peer closed, the ESTABLISHED event would be received followed
 682 * by the CLOSE event. If the app closes, it will block until we wake
 683 * it up after processing this event.
 684 */
 685static int cm_conn_est_handler(struct iwcm_id_private *cm_id_priv,
 686			       struct iw_cm_event *iw_event)
 687{
 688	unsigned long flags;
 689	int ret;
 690
 691	spin_lock_irqsave(&cm_id_priv->lock, flags);
 692
 693	/*
 694	 * We clear the CONNECT_WAIT bit here to allow the callback
 695	 * function to call iw_cm_disconnect. Calling iw_destroy_cm_id
 696	 * from a callback handler is not allowed.
 697	 */
 698	clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 699	BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
 700	cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
 701	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 702	ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 703	wake_up_all(&cm_id_priv->connect_wait);
 704
 705	return ret;
 706}
 707
 708/*
 709 * Active Side: CM_ID <-- ESTABLISHED
 710 *
 711 * The app has called connect and is waiting for the established event to
 712 * post it's requests to the server. This event will wake up anyone
 713 * blocked in iw_cm_disconnect or iw_destroy_id.
 714 */
 715static int cm_conn_rep_handler(struct iwcm_id_private *cm_id_priv,
 716			       struct iw_cm_event *iw_event)
 717{
 
 718	unsigned long flags;
 719	int ret;
 720
 721	spin_lock_irqsave(&cm_id_priv->lock, flags);
 722	/*
 723	 * Clear the connect wait bit so a callback function calling
 724	 * iw_cm_disconnect will not wait and deadlock this thread
 725	 */
 726	clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 727	BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
 728	if (iw_event->status == 0) {
 729		cm_id_priv->id.local_addr = iw_event->local_addr;
 730		cm_id_priv->id.remote_addr = iw_event->remote_addr;
 
 
 731		cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
 732	} else {
 733		/* REJECTED or RESET */
 734		cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
 735		cm_id_priv->qp = NULL;
 736		cm_id_priv->state = IW_CM_STATE_IDLE;
 737	}
 738	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 
 
 739	ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 740
 741	if (iw_event->private_data_len)
 742		kfree(iw_event->private_data);
 743
 744	/* Wake up waiters on connect complete */
 745	wake_up_all(&cm_id_priv->connect_wait);
 746
 747	return ret;
 748}
 749
 750/*
 751 * CM_ID <-- CLOSING
 752 *
 753 * If in the ESTABLISHED state, move to CLOSING.
 754 */
 755static void cm_disconnect_handler(struct iwcm_id_private *cm_id_priv,
 756				  struct iw_cm_event *iw_event)
 757{
 758	unsigned long flags;
 759
 760	spin_lock_irqsave(&cm_id_priv->lock, flags);
 761	if (cm_id_priv->state == IW_CM_STATE_ESTABLISHED)
 762		cm_id_priv->state = IW_CM_STATE_CLOSING;
 763	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 764}
 765
 766/*
 767 * CM_ID <-- IDLE
 768 *
 769 * If in the ESTBLISHED or CLOSING states, the QP will have have been
 770 * moved by the provider to the ERR state. Disassociate the CM_ID from
 771 * the QP,  move to IDLE, and remove the 'connected' reference.
 772 *
 773 * If in some other state, the cm_id was destroyed asynchronously.
 774 * This is the last reference that will result in waking up
 775 * the app thread blocked in iw_destroy_cm_id.
 776 */
 777static int cm_close_handler(struct iwcm_id_private *cm_id_priv,
 778				  struct iw_cm_event *iw_event)
 779{
 
 780	unsigned long flags;
 781	int ret = 0;
 782	spin_lock_irqsave(&cm_id_priv->lock, flags);
 
 
 783
 784	if (cm_id_priv->qp) {
 785		cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
 786		cm_id_priv->qp = NULL;
 787	}
 788	switch (cm_id_priv->state) {
 789	case IW_CM_STATE_ESTABLISHED:
 790	case IW_CM_STATE_CLOSING:
 791		cm_id_priv->state = IW_CM_STATE_IDLE;
 792		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 793		ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 794		spin_lock_irqsave(&cm_id_priv->lock, flags);
 795		break;
 796	case IW_CM_STATE_DESTROYING:
 797		break;
 798	default:
 799		BUG();
 800	}
 801	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 802
 
 
 
 
 803	return ret;
 804}
 805
 806static int process_event(struct iwcm_id_private *cm_id_priv,
 807			 struct iw_cm_event *iw_event)
 808{
 809	int ret = 0;
 810
 811	switch (iw_event->event) {
 812	case IW_CM_EVENT_CONNECT_REQUEST:
 813		cm_conn_req_handler(cm_id_priv, iw_event);
 814		break;
 815	case IW_CM_EVENT_CONNECT_REPLY:
 816		ret = cm_conn_rep_handler(cm_id_priv, iw_event);
 817		break;
 818	case IW_CM_EVENT_ESTABLISHED:
 819		ret = cm_conn_est_handler(cm_id_priv, iw_event);
 820		break;
 821	case IW_CM_EVENT_DISCONNECT:
 822		cm_disconnect_handler(cm_id_priv, iw_event);
 823		break;
 824	case IW_CM_EVENT_CLOSE:
 825		ret = cm_close_handler(cm_id_priv, iw_event);
 826		break;
 827	default:
 828		BUG();
 829	}
 830
 831	return ret;
 832}
 833
 834/*
 835 * Process events on the work_list for the cm_id. If the callback
 836 * function requests that the cm_id be deleted, a flag is set in the
 837 * cm_id flags to indicate that when the last reference is
 838 * removed, the cm_id is to be destroyed. This is necessary to
 839 * distinguish between an object that will be destroyed by the app
 840 * thread asleep on the destroy_comp list vs. an object destroyed
 841 * here synchronously when the last reference is removed.
 842 */
 843static void cm_work_handler(struct work_struct *_work)
 844{
 845	struct iwcm_work *work = container_of(_work, struct iwcm_work, work);
 846	struct iw_cm_event levent;
 847	struct iwcm_id_private *cm_id_priv = work->cm_id;
 848	unsigned long flags;
 849	int empty;
 850	int ret = 0;
 851	int destroy_id;
 852
 853	spin_lock_irqsave(&cm_id_priv->lock, flags);
 854	empty = list_empty(&cm_id_priv->work_list);
 855	while (!empty) {
 856		work = list_entry(cm_id_priv->work_list.next,
 857				  struct iwcm_work, list);
 858		list_del_init(&work->list);
 859		empty = list_empty(&cm_id_priv->work_list);
 860		levent = work->event;
 861		put_work(work);
 862		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 863
 864		ret = process_event(cm_id_priv, &levent);
 865		if (ret) {
 866			set_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
 867			destroy_cm_id(&cm_id_priv->id);
 868		}
 869		BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
 870		destroy_id = test_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
 871		if (iwcm_deref_id(cm_id_priv)) {
 872			if (destroy_id) {
 873				BUG_ON(!list_empty(&cm_id_priv->work_list));
 874				free_cm_id(cm_id_priv);
 875			}
 876			return;
 877		}
 878		spin_lock_irqsave(&cm_id_priv->lock, flags);
 879	}
 880	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 881}
 882
 883/*
 884 * This function is called on interrupt context. Schedule events on
 885 * the iwcm_wq thread to allow callback functions to downcall into
 886 * the CM and/or block.  Events are queued to a per-CM_ID
 887 * work_list. If this is the first event on the work_list, the work
 888 * element is also queued on the iwcm_wq thread.
 889 *
 890 * Each event holds a reference on the cm_id. Until the last posted
 891 * event has been delivered and processed, the cm_id cannot be
 892 * deleted.
 893 *
 894 * Returns:
 895 * 	      0	- the event was handled.
 896 *	-ENOMEM	- the event was not handled due to lack of resources.
 897 */
 898static int cm_event_handler(struct iw_cm_id *cm_id,
 899			     struct iw_cm_event *iw_event)
 900{
 901	struct iwcm_work *work;
 902	struct iwcm_id_private *cm_id_priv;
 903	unsigned long flags;
 904	int ret = 0;
 905
 906	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 907
 908	spin_lock_irqsave(&cm_id_priv->lock, flags);
 909	work = get_work(cm_id_priv);
 910	if (!work) {
 911		ret = -ENOMEM;
 912		goto out;
 913	}
 914
 915	INIT_WORK(&work->work, cm_work_handler);
 916	work->cm_id = cm_id_priv;
 917	work->event = *iw_event;
 918
 919	if ((work->event.event == IW_CM_EVENT_CONNECT_REQUEST ||
 920	     work->event.event == IW_CM_EVENT_CONNECT_REPLY) &&
 921	    work->event.private_data_len) {
 922		ret = copy_private_data(&work->event);
 923		if (ret) {
 924			put_work(work);
 925			goto out;
 926		}
 927	}
 928
 929	atomic_inc(&cm_id_priv->refcount);
 930	if (list_empty(&cm_id_priv->work_list)) {
 931		list_add_tail(&work->list, &cm_id_priv->work_list);
 932		queue_work(iwcm_wq, &work->work);
 933	} else
 934		list_add_tail(&work->list, &cm_id_priv->work_list);
 935out:
 936	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 937	return ret;
 938}
 939
 940static int iwcm_init_qp_init_attr(struct iwcm_id_private *cm_id_priv,
 941				  struct ib_qp_attr *qp_attr,
 942				  int *qp_attr_mask)
 943{
 944	unsigned long flags;
 945	int ret;
 946
 947	spin_lock_irqsave(&cm_id_priv->lock, flags);
 948	switch (cm_id_priv->state) {
 949	case IW_CM_STATE_IDLE:
 950	case IW_CM_STATE_CONN_SENT:
 951	case IW_CM_STATE_CONN_RECV:
 952	case IW_CM_STATE_ESTABLISHED:
 953		*qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 954		qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE|
 955					   IB_ACCESS_REMOTE_READ;
 956		ret = 0;
 957		break;
 958	default:
 959		ret = -EINVAL;
 960		break;
 961	}
 962	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 963	return ret;
 964}
 965
 966static int iwcm_init_qp_rts_attr(struct iwcm_id_private *cm_id_priv,
 967				  struct ib_qp_attr *qp_attr,
 968				  int *qp_attr_mask)
 969{
 970	unsigned long flags;
 971	int ret;
 972
 973	spin_lock_irqsave(&cm_id_priv->lock, flags);
 974	switch (cm_id_priv->state) {
 975	case IW_CM_STATE_IDLE:
 976	case IW_CM_STATE_CONN_SENT:
 977	case IW_CM_STATE_CONN_RECV:
 978	case IW_CM_STATE_ESTABLISHED:
 979		*qp_attr_mask = 0;
 980		ret = 0;
 981		break;
 982	default:
 983		ret = -EINVAL;
 984		break;
 985	}
 986	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 987	return ret;
 988}
 989
 990int iw_cm_init_qp_attr(struct iw_cm_id *cm_id,
 991		       struct ib_qp_attr *qp_attr,
 992		       int *qp_attr_mask)
 993{
 994	struct iwcm_id_private *cm_id_priv;
 995	int ret;
 996
 997	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 998	switch (qp_attr->qp_state) {
 999	case IB_QPS_INIT:
1000	case IB_QPS_RTR:
1001		ret = iwcm_init_qp_init_attr(cm_id_priv,
1002					     qp_attr, qp_attr_mask);
1003		break;
1004	case IB_QPS_RTS:
1005		ret = iwcm_init_qp_rts_attr(cm_id_priv,
1006					    qp_attr, qp_attr_mask);
1007		break;
1008	default:
1009		ret = -EINVAL;
1010		break;
1011	}
1012	return ret;
1013}
1014EXPORT_SYMBOL(iw_cm_init_qp_attr);
1015
1016static int __init iw_cm_init(void)
1017{
1018	iwcm_wq = create_singlethread_workqueue("iw_cm_wq");
 
 
 
 
 
 
1019	if (!iwcm_wq)
1020		return -ENOMEM;
 
 
 
 
 
 
 
1021
 
1022	return 0;
 
 
 
 
 
 
1023}
1024
1025static void __exit iw_cm_cleanup(void)
1026{
 
 
1027	destroy_workqueue(iwcm_wq);
 
1028}
 
 
1029
1030module_init(iw_cm_init);
1031module_exit(iw_cm_cleanup);
v5.14.15
   1/*
   2 * Copyright (c) 2004, 2005 Intel Corporation.  All rights reserved.
   3 * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
   4 * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
   5 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   6 * Copyright (c) 2005 Open Grid Computing, Inc. All rights reserved.
   7 * Copyright (c) 2005 Network Appliance, Inc. All rights reserved.
   8 *
   9 * This software is available to you under a choice of one of two
  10 * licenses.  You may choose to be licensed under the terms of the GNU
  11 * General Public License (GPL) Version 2, available from the file
  12 * COPYING in the main directory of this source tree, or the
  13 * OpenIB.org BSD license below:
  14 *
  15 *     Redistribution and use in source and binary forms, with or
  16 *     without modification, are permitted provided that the following
  17 *     conditions are met:
  18 *
  19 *      - Redistributions of source code must retain the above
  20 *        copyright notice, this list of conditions and the following
  21 *        disclaimer.
  22 *
  23 *      - Redistributions in binary form must reproduce the above
  24 *        copyright notice, this list of conditions and the following
  25 *        disclaimer in the documentation and/or other materials
  26 *        provided with the distribution.
  27 *
  28 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  29 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  30 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  31 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  32 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  33 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  34 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  35 * SOFTWARE.
  36 *
  37 */
  38#include <linux/dma-mapping.h>
  39#include <linux/err.h>
  40#include <linux/idr.h>
  41#include <linux/interrupt.h>
  42#include <linux/rbtree.h>
  43#include <linux/sched.h>
  44#include <linux/spinlock.h>
  45#include <linux/workqueue.h>
  46#include <linux/completion.h>
  47#include <linux/slab.h>
  48#include <linux/module.h>
  49#include <linux/sysctl.h>
  50
  51#include <rdma/iw_cm.h>
  52#include <rdma/ib_addr.h>
  53#include <rdma/iw_portmap.h>
  54#include <rdma/rdma_netlink.h>
  55
  56#include "iwcm.h"
  57
  58MODULE_AUTHOR("Tom Tucker");
  59MODULE_DESCRIPTION("iWARP CM");
  60MODULE_LICENSE("Dual BSD/GPL");
  61
  62static const char * const iwcm_rej_reason_strs[] = {
  63	[ECONNRESET]			= "reset by remote host",
  64	[ECONNREFUSED]			= "refused by remote application",
  65	[ETIMEDOUT]			= "setup timeout",
  66};
  67
  68const char *__attribute_const__ iwcm_reject_msg(int reason)
  69{
  70	size_t index;
  71
  72	/* iWARP uses negative errnos */
  73	index = -reason;
  74
  75	if (index < ARRAY_SIZE(iwcm_rej_reason_strs) &&
  76	    iwcm_rej_reason_strs[index])
  77		return iwcm_rej_reason_strs[index];
  78	else
  79		return "unrecognized reason";
  80}
  81EXPORT_SYMBOL(iwcm_reject_msg);
  82
  83static struct rdma_nl_cbs iwcm_nl_cb_table[RDMA_NL_IWPM_NUM_OPS] = {
  84	[RDMA_NL_IWPM_REG_PID] = {.dump = iwpm_register_pid_cb},
  85	[RDMA_NL_IWPM_ADD_MAPPING] = {.dump = iwpm_add_mapping_cb},
  86	[RDMA_NL_IWPM_QUERY_MAPPING] = {.dump = iwpm_add_and_query_mapping_cb},
  87	[RDMA_NL_IWPM_REMOTE_INFO] = {.dump = iwpm_remote_info_cb},
  88	[RDMA_NL_IWPM_HANDLE_ERR] = {.dump = iwpm_mapping_error_cb},
  89	[RDMA_NL_IWPM_MAPINFO] = {.dump = iwpm_mapping_info_cb},
  90	[RDMA_NL_IWPM_MAPINFO_NUM] = {.dump = iwpm_ack_mapping_info_cb},
  91	[RDMA_NL_IWPM_HELLO] = {.dump = iwpm_hello_cb}
  92};
  93
  94static struct workqueue_struct *iwcm_wq;
  95struct iwcm_work {
  96	struct work_struct work;
  97	struct iwcm_id_private *cm_id;
  98	struct list_head list;
  99	struct iw_cm_event event;
 100	struct list_head free_list;
 101};
 102
 103static unsigned int default_backlog = 256;
 104
 105static struct ctl_table_header *iwcm_ctl_table_hdr;
 106static struct ctl_table iwcm_ctl_table[] = {
 107	{
 108		.procname	= "default_backlog",
 109		.data		= &default_backlog,
 110		.maxlen		= sizeof(default_backlog),
 111		.mode		= 0644,
 112		.proc_handler	= proc_dointvec,
 113	},
 114	{ }
 115};
 116
 117/*
 118 * The following services provide a mechanism for pre-allocating iwcm_work
 119 * elements.  The design pre-allocates them  based on the cm_id type:
 120 *	LISTENING IDS: 	Get enough elements preallocated to handle the
 121 *			listen backlog.
 122 *	ACTIVE IDS:	4: CONNECT_REPLY, ESTABLISHED, DISCONNECT, CLOSE
 123 *	PASSIVE IDS:	3: ESTABLISHED, DISCONNECT, CLOSE
 124 *
 125 * Allocating them in connect and listen avoids having to deal
 126 * with allocation failures on the event upcall from the provider (which
 127 * is called in the interrupt context).
 128 *
 129 * One exception is when creating the cm_id for incoming connection requests.
 130 * There are two cases:
 131 * 1) in the event upcall, cm_event_handler(), for a listening cm_id.  If
 132 *    the backlog is exceeded, then no more connection request events will
 133 *    be processed.  cm_event_handler() returns -ENOMEM in this case.  Its up
 134 *    to the provider to reject the connection request.
 135 * 2) in the connection request workqueue handler, cm_conn_req_handler().
 136 *    If work elements cannot be allocated for the new connect request cm_id,
 137 *    then IWCM will call the provider reject method.  This is ok since
 138 *    cm_conn_req_handler() runs in the workqueue thread context.
 139 */
 140
 141static struct iwcm_work *get_work(struct iwcm_id_private *cm_id_priv)
 142{
 143	struct iwcm_work *work;
 144
 145	if (list_empty(&cm_id_priv->work_free_list))
 146		return NULL;
 147	work = list_entry(cm_id_priv->work_free_list.next, struct iwcm_work,
 148			  free_list);
 149	list_del_init(&work->free_list);
 150	return work;
 151}
 152
 153static void put_work(struct iwcm_work *work)
 154{
 155	list_add(&work->free_list, &work->cm_id->work_free_list);
 156}
 157
 158static void dealloc_work_entries(struct iwcm_id_private *cm_id_priv)
 159{
 160	struct list_head *e, *tmp;
 161
 162	list_for_each_safe(e, tmp, &cm_id_priv->work_free_list) {
 163		list_del(e);
 164		kfree(list_entry(e, struct iwcm_work, free_list));
 165	}
 166}
 167
 168static int alloc_work_entries(struct iwcm_id_private *cm_id_priv, int count)
 169{
 170	struct iwcm_work *work;
 171
 172	BUG_ON(!list_empty(&cm_id_priv->work_free_list));
 173	while (count--) {
 174		work = kmalloc(sizeof(struct iwcm_work), GFP_KERNEL);
 175		if (!work) {
 176			dealloc_work_entries(cm_id_priv);
 177			return -ENOMEM;
 178		}
 179		work->cm_id = cm_id_priv;
 180		INIT_LIST_HEAD(&work->list);
 181		put_work(work);
 182	}
 183	return 0;
 184}
 185
 186/*
 187 * Save private data from incoming connection requests to
 188 * iw_cm_event, so the low level driver doesn't have to. Adjust
 189 * the event ptr to point to the local copy.
 190 */
 191static int copy_private_data(struct iw_cm_event *event)
 192{
 193	void *p;
 194
 195	p = kmemdup(event->private_data, event->private_data_len, GFP_ATOMIC);
 196	if (!p)
 197		return -ENOMEM;
 198	event->private_data = p;
 199	return 0;
 200}
 201
 202static void free_cm_id(struct iwcm_id_private *cm_id_priv)
 203{
 204	dealloc_work_entries(cm_id_priv);
 205	kfree(cm_id_priv);
 206}
 207
 208/*
 209 * Release a reference on cm_id. If the last reference is being
 210 * released, free the cm_id and return 1.
 
 211 */
 212static int iwcm_deref_id(struct iwcm_id_private *cm_id_priv)
 213{
 214	if (refcount_dec_and_test(&cm_id_priv->refcount)) {
 
 215		BUG_ON(!list_empty(&cm_id_priv->work_list));
 216		free_cm_id(cm_id_priv);
 217		return 1;
 218	}
 219
 220	return 0;
 221}
 222
 223static void add_ref(struct iw_cm_id *cm_id)
 224{
 225	struct iwcm_id_private *cm_id_priv;
 226	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 227	refcount_inc(&cm_id_priv->refcount);
 228}
 229
 230static void rem_ref(struct iw_cm_id *cm_id)
 231{
 232	struct iwcm_id_private *cm_id_priv;
 233
 234	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 235
 236	(void)iwcm_deref_id(cm_id_priv);
 
 
 
 237}
 238
 239static int cm_event_handler(struct iw_cm_id *cm_id, struct iw_cm_event *event);
 240
 241struct iw_cm_id *iw_create_cm_id(struct ib_device *device,
 242				 iw_cm_handler cm_handler,
 243				 void *context)
 244{
 245	struct iwcm_id_private *cm_id_priv;
 246
 247	cm_id_priv = kzalloc(sizeof(*cm_id_priv), GFP_KERNEL);
 248	if (!cm_id_priv)
 249		return ERR_PTR(-ENOMEM);
 250
 251	cm_id_priv->state = IW_CM_STATE_IDLE;
 252	cm_id_priv->id.device = device;
 253	cm_id_priv->id.cm_handler = cm_handler;
 254	cm_id_priv->id.context = context;
 255	cm_id_priv->id.event_handler = cm_event_handler;
 256	cm_id_priv->id.add_ref = add_ref;
 257	cm_id_priv->id.rem_ref = rem_ref;
 258	spin_lock_init(&cm_id_priv->lock);
 259	refcount_set(&cm_id_priv->refcount, 1);
 260	init_waitqueue_head(&cm_id_priv->connect_wait);
 261	init_completion(&cm_id_priv->destroy_comp);
 262	INIT_LIST_HEAD(&cm_id_priv->work_list);
 263	INIT_LIST_HEAD(&cm_id_priv->work_free_list);
 264
 265	return &cm_id_priv->id;
 266}
 267EXPORT_SYMBOL(iw_create_cm_id);
 268
 269
 270static int iwcm_modify_qp_err(struct ib_qp *qp)
 271{
 272	struct ib_qp_attr qp_attr;
 273
 274	if (!qp)
 275		return -EINVAL;
 276
 277	qp_attr.qp_state = IB_QPS_ERR;
 278	return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 279}
 280
 281/*
 282 * This is really the RDMAC CLOSING state. It is most similar to the
 283 * IB SQD QP state.
 284 */
 285static int iwcm_modify_qp_sqd(struct ib_qp *qp)
 286{
 287	struct ib_qp_attr qp_attr;
 288
 289	BUG_ON(qp == NULL);
 290	qp_attr.qp_state = IB_QPS_SQD;
 291	return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 292}
 293
 294/*
 295 * CM_ID <-- CLOSING
 296 *
 297 * Block if a passive or active connection is currently being processed. Then
 298 * process the event as follows:
 299 * - If we are ESTABLISHED, move to CLOSING and modify the QP state
 300 *   based on the abrupt flag
 301 * - If the connection is already in the CLOSING or IDLE state, the peer is
 302 *   disconnecting concurrently with us and we've already seen the
 303 *   DISCONNECT event -- ignore the request and return 0
 304 * - Disconnect on a listening endpoint returns -EINVAL
 305 */
 306int iw_cm_disconnect(struct iw_cm_id *cm_id, int abrupt)
 307{
 308	struct iwcm_id_private *cm_id_priv;
 309	unsigned long flags;
 310	int ret = 0;
 311	struct ib_qp *qp = NULL;
 312
 313	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 314	/* Wait if we're currently in a connect or accept downcall */
 315	wait_event(cm_id_priv->connect_wait,
 316		   !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
 317
 318	spin_lock_irqsave(&cm_id_priv->lock, flags);
 319	switch (cm_id_priv->state) {
 320	case IW_CM_STATE_ESTABLISHED:
 321		cm_id_priv->state = IW_CM_STATE_CLOSING;
 322
 323		/* QP could be <nul> for user-mode client */
 324		if (cm_id_priv->qp)
 325			qp = cm_id_priv->qp;
 326		else
 327			ret = -EINVAL;
 328		break;
 329	case IW_CM_STATE_LISTEN:
 330		ret = -EINVAL;
 331		break;
 332	case IW_CM_STATE_CLOSING:
 333		/* remote peer closed first */
 334	case IW_CM_STATE_IDLE:
 335		/* accept or connect returned !0 */
 336		break;
 337	case IW_CM_STATE_CONN_RECV:
 338		/*
 339		 * App called disconnect before/without calling accept after
 340		 * connect_request event delivered.
 341		 */
 342		break;
 343	case IW_CM_STATE_CONN_SENT:
 344		/* Can only get here if wait above fails */
 345	default:
 346		BUG();
 347	}
 348	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 349
 350	if (qp) {
 351		if (abrupt)
 352			ret = iwcm_modify_qp_err(qp);
 353		else
 354			ret = iwcm_modify_qp_sqd(qp);
 355
 356		/*
 357		 * If both sides are disconnecting the QP could
 358		 * already be in ERR or SQD states
 359		 */
 360		ret = 0;
 361	}
 362
 363	return ret;
 364}
 365EXPORT_SYMBOL(iw_cm_disconnect);
 366
 367/*
 368 * CM_ID <-- DESTROYING
 369 *
 370 * Clean up all resources associated with the connection and release
 371 * the initial reference taken by iw_create_cm_id.
 372 */
 373static void destroy_cm_id(struct iw_cm_id *cm_id)
 374{
 375	struct iwcm_id_private *cm_id_priv;
 376	struct ib_qp *qp;
 377	unsigned long flags;
 
 378
 379	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 380	/*
 381	 * Wait if we're currently in a connect or accept downcall. A
 382	 * listening endpoint should never block here.
 383	 */
 384	wait_event(cm_id_priv->connect_wait,
 385		   !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
 386
 387	/*
 388	 * Since we're deleting the cm_id, drop any events that
 389	 * might arrive before the last dereference.
 390	 */
 391	set_bit(IWCM_F_DROP_EVENTS, &cm_id_priv->flags);
 392
 393	spin_lock_irqsave(&cm_id_priv->lock, flags);
 394	qp = cm_id_priv->qp;
 395	cm_id_priv->qp = NULL;
 396
 397	switch (cm_id_priv->state) {
 398	case IW_CM_STATE_LISTEN:
 399		cm_id_priv->state = IW_CM_STATE_DESTROYING;
 400		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 401		/* destroy the listening endpoint */
 402		cm_id->device->ops.iw_destroy_listen(cm_id);
 403		spin_lock_irqsave(&cm_id_priv->lock, flags);
 404		break;
 405	case IW_CM_STATE_ESTABLISHED:
 406		cm_id_priv->state = IW_CM_STATE_DESTROYING;
 407		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 408		/* Abrupt close of the connection */
 409		(void)iwcm_modify_qp_err(qp);
 410		spin_lock_irqsave(&cm_id_priv->lock, flags);
 411		break;
 412	case IW_CM_STATE_IDLE:
 413	case IW_CM_STATE_CLOSING:
 414		cm_id_priv->state = IW_CM_STATE_DESTROYING;
 415		break;
 416	case IW_CM_STATE_CONN_RECV:
 417		/*
 418		 * App called destroy before/without calling accept after
 419		 * receiving connection request event notification or
 420		 * returned non zero from the event callback function.
 421		 * In either case, must tell the provider to reject.
 422		 */
 423		cm_id_priv->state = IW_CM_STATE_DESTROYING;
 424		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 425		cm_id->device->ops.iw_reject(cm_id, NULL, 0);
 426		spin_lock_irqsave(&cm_id_priv->lock, flags);
 427		break;
 428	case IW_CM_STATE_CONN_SENT:
 429	case IW_CM_STATE_DESTROYING:
 430	default:
 431		BUG();
 432		break;
 433	}
 
 
 
 
 434	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 435	if (qp)
 436		cm_id_priv->id.device->ops.iw_rem_ref(qp);
 437
 438	if (cm_id->mapped) {
 439		iwpm_remove_mapinfo(&cm_id->local_addr, &cm_id->m_local_addr);
 440		iwpm_remove_mapping(&cm_id->local_addr, RDMA_NL_IWCM);
 441	}
 442
 443	(void)iwcm_deref_id(cm_id_priv);
 444}
 445
 446/*
 447 * This function is only called by the application thread and cannot
 448 * be called by the event thread. The function will wait for all
 449 * references to be released on the cm_id and then kfree the cm_id
 450 * object.
 451 */
 452void iw_destroy_cm_id(struct iw_cm_id *cm_id)
 453{
 454	destroy_cm_id(cm_id);
 455}
 456EXPORT_SYMBOL(iw_destroy_cm_id);
 457
 458/**
 459 * iw_cm_check_wildcard - If IP address is 0 then use original
 460 * @pm_addr: sockaddr containing the ip to check for wildcard
 461 * @cm_addr: sockaddr containing the actual IP address
 462 * @cm_outaddr: sockaddr to set IP addr which leaving port
 463 *
 464 *  Checks the pm_addr for wildcard and then sets cm_outaddr's
 465 *  IP to the actual (cm_addr).
 466 */
 467static void iw_cm_check_wildcard(struct sockaddr_storage *pm_addr,
 468				 struct sockaddr_storage *cm_addr,
 469				 struct sockaddr_storage *cm_outaddr)
 470{
 471	if (pm_addr->ss_family == AF_INET) {
 472		struct sockaddr_in *pm4_addr = (struct sockaddr_in *)pm_addr;
 473
 474		if (pm4_addr->sin_addr.s_addr == htonl(INADDR_ANY)) {
 475			struct sockaddr_in *cm4_addr =
 476				(struct sockaddr_in *)cm_addr;
 477			struct sockaddr_in *cm4_outaddr =
 478				(struct sockaddr_in *)cm_outaddr;
 479
 480			cm4_outaddr->sin_addr = cm4_addr->sin_addr;
 481		}
 482	} else {
 483		struct sockaddr_in6 *pm6_addr = (struct sockaddr_in6 *)pm_addr;
 484
 485		if (ipv6_addr_type(&pm6_addr->sin6_addr) == IPV6_ADDR_ANY) {
 486			struct sockaddr_in6 *cm6_addr =
 487				(struct sockaddr_in6 *)cm_addr;
 488			struct sockaddr_in6 *cm6_outaddr =
 489				(struct sockaddr_in6 *)cm_outaddr;
 490
 491			cm6_outaddr->sin6_addr = cm6_addr->sin6_addr;
 492		}
 493	}
 494}
 495
 496/**
 497 * iw_cm_map - Use portmapper to map the ports
 498 * @cm_id: connection manager pointer
 499 * @active: Indicates the active side when true
 500 * returns nonzero for error only if iwpm_create_mapinfo() fails
 501 *
 502 * Tries to add a mapping for a port using the Portmapper. If
 503 * successful in mapping the IP/Port it will check the remote
 504 * mapped IP address for a wildcard IP address and replace the
 505 * zero IP address with the remote_addr.
 506 */
 507static int iw_cm_map(struct iw_cm_id *cm_id, bool active)
 508{
 509	const char *devname = dev_name(&cm_id->device->dev);
 510	const char *ifname = cm_id->device->iw_ifname;
 511	struct iwpm_dev_data pm_reg_msg = {};
 512	struct iwpm_sa_data pm_msg;
 513	int status;
 514
 515	if (strlen(devname) >= sizeof(pm_reg_msg.dev_name) ||
 516	    strlen(ifname) >= sizeof(pm_reg_msg.if_name))
 517		return -EINVAL;
 518
 519	cm_id->m_local_addr = cm_id->local_addr;
 520	cm_id->m_remote_addr = cm_id->remote_addr;
 521
 522	strcpy(pm_reg_msg.dev_name, devname);
 523	strcpy(pm_reg_msg.if_name, ifname);
 524
 525	if (iwpm_register_pid(&pm_reg_msg, RDMA_NL_IWCM) ||
 526	    !iwpm_valid_pid())
 527		return 0;
 528
 529	cm_id->mapped = true;
 530	pm_msg.loc_addr = cm_id->local_addr;
 531	pm_msg.rem_addr = cm_id->remote_addr;
 532	pm_msg.flags = (cm_id->device->iw_driver_flags & IW_F_NO_PORT_MAP) ?
 533		       IWPM_FLAGS_NO_PORT_MAP : 0;
 534	if (active)
 535		status = iwpm_add_and_query_mapping(&pm_msg,
 536						    RDMA_NL_IWCM);
 537	else
 538		status = iwpm_add_mapping(&pm_msg, RDMA_NL_IWCM);
 539
 540	if (!status) {
 541		cm_id->m_local_addr = pm_msg.mapped_loc_addr;
 542		if (active) {
 543			cm_id->m_remote_addr = pm_msg.mapped_rem_addr;
 544			iw_cm_check_wildcard(&pm_msg.mapped_rem_addr,
 545					     &cm_id->remote_addr,
 546					     &cm_id->m_remote_addr);
 547		}
 548	}
 549
 550	return iwpm_create_mapinfo(&cm_id->local_addr,
 551				   &cm_id->m_local_addr,
 552				   RDMA_NL_IWCM, pm_msg.flags);
 553}
 
 554
 555/*
 556 * CM_ID <-- LISTEN
 557 *
 558 * Start listening for connect requests. Generates one CONNECT_REQUEST
 559 * event for each inbound connect request.
 560 */
 561int iw_cm_listen(struct iw_cm_id *cm_id, int backlog)
 562{
 563	struct iwcm_id_private *cm_id_priv;
 564	unsigned long flags;
 565	int ret;
 566
 567	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 568
 569	if (!backlog)
 570		backlog = default_backlog;
 571
 572	ret = alloc_work_entries(cm_id_priv, backlog);
 573	if (ret)
 574		return ret;
 575
 576	spin_lock_irqsave(&cm_id_priv->lock, flags);
 577	switch (cm_id_priv->state) {
 578	case IW_CM_STATE_IDLE:
 579		cm_id_priv->state = IW_CM_STATE_LISTEN;
 580		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 581		ret = iw_cm_map(cm_id, false);
 582		if (!ret)
 583			ret = cm_id->device->ops.iw_create_listen(cm_id,
 584								  backlog);
 585		if (ret)
 586			cm_id_priv->state = IW_CM_STATE_IDLE;
 587		spin_lock_irqsave(&cm_id_priv->lock, flags);
 588		break;
 589	default:
 590		ret = -EINVAL;
 591	}
 592	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 593
 594	return ret;
 595}
 596EXPORT_SYMBOL(iw_cm_listen);
 597
 598/*
 599 * CM_ID <-- IDLE
 600 *
 601 * Rejects an inbound connection request. No events are generated.
 602 */
 603int iw_cm_reject(struct iw_cm_id *cm_id,
 604		 const void *private_data,
 605		 u8 private_data_len)
 606{
 607	struct iwcm_id_private *cm_id_priv;
 608	unsigned long flags;
 609	int ret;
 610
 611	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 612	set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 613
 614	spin_lock_irqsave(&cm_id_priv->lock, flags);
 615	if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
 616		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 617		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 618		wake_up_all(&cm_id_priv->connect_wait);
 619		return -EINVAL;
 620	}
 621	cm_id_priv->state = IW_CM_STATE_IDLE;
 622	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 623
 624	ret = cm_id->device->ops.iw_reject(cm_id, private_data,
 625					  private_data_len);
 626
 627	clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 628	wake_up_all(&cm_id_priv->connect_wait);
 629
 630	return ret;
 631}
 632EXPORT_SYMBOL(iw_cm_reject);
 633
 634/*
 635 * CM_ID <-- ESTABLISHED
 636 *
 637 * Accepts an inbound connection request and generates an ESTABLISHED
 638 * event. Callers of iw_cm_disconnect and iw_destroy_cm_id will block
 639 * until the ESTABLISHED event is received from the provider.
 640 */
 641int iw_cm_accept(struct iw_cm_id *cm_id,
 642		 struct iw_cm_conn_param *iw_param)
 643{
 644	struct iwcm_id_private *cm_id_priv;
 645	struct ib_qp *qp;
 646	unsigned long flags;
 647	int ret;
 648
 649	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 650	set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 651
 652	spin_lock_irqsave(&cm_id_priv->lock, flags);
 653	if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
 654		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 655		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 656		wake_up_all(&cm_id_priv->connect_wait);
 657		return -EINVAL;
 658	}
 659	/* Get the ib_qp given the QPN */
 660	qp = cm_id->device->ops.iw_get_qp(cm_id->device, iw_param->qpn);
 661	if (!qp) {
 662		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 663		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 664		wake_up_all(&cm_id_priv->connect_wait);
 665		return -EINVAL;
 666	}
 667	cm_id->device->ops.iw_add_ref(qp);
 668	cm_id_priv->qp = qp;
 669	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 670
 671	ret = cm_id->device->ops.iw_accept(cm_id, iw_param);
 672	if (ret) {
 673		/* An error on accept precludes provider events */
 674		BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
 675		cm_id_priv->state = IW_CM_STATE_IDLE;
 676		spin_lock_irqsave(&cm_id_priv->lock, flags);
 677		qp = cm_id_priv->qp;
 678		cm_id_priv->qp = NULL;
 
 
 679		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 680		if (qp)
 681			cm_id->device->ops.iw_rem_ref(qp);
 682		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 683		wake_up_all(&cm_id_priv->connect_wait);
 684	}
 685
 686	return ret;
 687}
 688EXPORT_SYMBOL(iw_cm_accept);
 689
 690/*
 691 * Active Side: CM_ID <-- CONN_SENT
 692 *
 693 * If successful, results in the generation of a CONNECT_REPLY
 694 * event. iw_cm_disconnect and iw_cm_destroy will block until the
 695 * CONNECT_REPLY event is received from the provider.
 696 */
 697int iw_cm_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *iw_param)
 698{
 699	struct iwcm_id_private *cm_id_priv;
 700	int ret;
 701	unsigned long flags;
 702	struct ib_qp *qp = NULL;
 703
 704	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 705
 706	ret = alloc_work_entries(cm_id_priv, 4);
 707	if (ret)
 708		return ret;
 709
 710	set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 711	spin_lock_irqsave(&cm_id_priv->lock, flags);
 712
 713	if (cm_id_priv->state != IW_CM_STATE_IDLE) {
 714		ret = -EINVAL;
 715		goto err;
 
 
 716	}
 717
 718	/* Get the ib_qp given the QPN */
 719	qp = cm_id->device->ops.iw_get_qp(cm_id->device, iw_param->qpn);
 720	if (!qp) {
 721		ret = -EINVAL;
 722		goto err;
 
 
 723	}
 724	cm_id->device->ops.iw_add_ref(qp);
 725	cm_id_priv->qp = qp;
 726	cm_id_priv->state = IW_CM_STATE_CONN_SENT;
 727	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 728
 729	ret = iw_cm_map(cm_id, true);
 730	if (!ret)
 731		ret = cm_id->device->ops.iw_connect(cm_id, iw_param);
 732	if (!ret)
 733		return 0;	/* success */
 
 
 
 
 
 
 
 
 734
 735	spin_lock_irqsave(&cm_id_priv->lock, flags);
 736	qp = cm_id_priv->qp;
 737	cm_id_priv->qp = NULL;
 738	cm_id_priv->state = IW_CM_STATE_IDLE;
 739err:
 740	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 741	if (qp)
 742		cm_id->device->ops.iw_rem_ref(qp);
 743	clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 744	wake_up_all(&cm_id_priv->connect_wait);
 745	return ret;
 746}
 747EXPORT_SYMBOL(iw_cm_connect);
 748
 749/*
 750 * Passive Side: new CM_ID <-- CONN_RECV
 751 *
 752 * Handles an inbound connect request. The function creates a new
 753 * iw_cm_id to represent the new connection and inherits the client
 754 * callback function and other attributes from the listening parent.
 755 *
 756 * The work item contains a pointer to the listen_cm_id and the event. The
 757 * listen_cm_id contains the client cm_handler, context and
 758 * device. These are copied when the device is cloned. The event
 759 * contains the new four tuple.
 760 *
 761 * An error on the child should not affect the parent, so this
 762 * function does not return a value.
 763 */
 764static void cm_conn_req_handler(struct iwcm_id_private *listen_id_priv,
 765				struct iw_cm_event *iw_event)
 766{
 767	unsigned long flags;
 768	struct iw_cm_id *cm_id;
 769	struct iwcm_id_private *cm_id_priv;
 770	int ret;
 771
 772	/*
 773	 * The provider should never generate a connection request
 774	 * event with a bad status.
 775	 */
 776	BUG_ON(iw_event->status);
 777
 
 
 
 
 
 
 
 
 
 
 
 778	cm_id = iw_create_cm_id(listen_id_priv->id.device,
 779				listen_id_priv->id.cm_handler,
 780				listen_id_priv->id.context);
 781	/* If the cm_id could not be created, ignore the request */
 782	if (IS_ERR(cm_id))
 783		goto out;
 784
 785	cm_id->provider_data = iw_event->provider_data;
 786	cm_id->m_local_addr = iw_event->local_addr;
 787	cm_id->m_remote_addr = iw_event->remote_addr;
 788	cm_id->local_addr = listen_id_priv->id.local_addr;
 789
 790	ret = iwpm_get_remote_info(&listen_id_priv->id.m_local_addr,
 791				   &iw_event->remote_addr,
 792				   &cm_id->remote_addr,
 793				   RDMA_NL_IWCM);
 794	if (ret) {
 795		cm_id->remote_addr = iw_event->remote_addr;
 796	} else {
 797		iw_cm_check_wildcard(&listen_id_priv->id.m_local_addr,
 798				     &iw_event->local_addr,
 799				     &cm_id->local_addr);
 800		iw_event->local_addr = cm_id->local_addr;
 801		iw_event->remote_addr = cm_id->remote_addr;
 802	}
 803
 804	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
 805	cm_id_priv->state = IW_CM_STATE_CONN_RECV;
 806
 807	/*
 808	 * We could be destroying the listening id. If so, ignore this
 809	 * upcall.
 810	 */
 811	spin_lock_irqsave(&listen_id_priv->lock, flags);
 812	if (listen_id_priv->state != IW_CM_STATE_LISTEN) {
 813		spin_unlock_irqrestore(&listen_id_priv->lock, flags);
 814		iw_cm_reject(cm_id, NULL, 0);
 815		iw_destroy_cm_id(cm_id);
 816		goto out;
 817	}
 818	spin_unlock_irqrestore(&listen_id_priv->lock, flags);
 819
 820	ret = alloc_work_entries(cm_id_priv, 3);
 821	if (ret) {
 822		iw_cm_reject(cm_id, NULL, 0);
 823		iw_destroy_cm_id(cm_id);
 824		goto out;
 825	}
 826
 827	/* Call the client CM handler */
 828	ret = cm_id->cm_handler(cm_id, iw_event);
 829	if (ret) {
 830		iw_cm_reject(cm_id, NULL, 0);
 831		iw_destroy_cm_id(cm_id);
 
 
 
 832	}
 833
 834out:
 835	if (iw_event->private_data_len)
 836		kfree(iw_event->private_data);
 837}
 838
 839/*
 840 * Passive Side: CM_ID <-- ESTABLISHED
 841 *
 842 * The provider generated an ESTABLISHED event which means that
 843 * the MPA negotion has completed successfully and we are now in MPA
 844 * FPDU mode.
 845 *
 846 * This event can only be received in the CONN_RECV state. If the
 847 * remote peer closed, the ESTABLISHED event would be received followed
 848 * by the CLOSE event. If the app closes, it will block until we wake
 849 * it up after processing this event.
 850 */
 851static int cm_conn_est_handler(struct iwcm_id_private *cm_id_priv,
 852			       struct iw_cm_event *iw_event)
 853{
 854	unsigned long flags;
 855	int ret;
 856
 857	spin_lock_irqsave(&cm_id_priv->lock, flags);
 858
 859	/*
 860	 * We clear the CONNECT_WAIT bit here to allow the callback
 861	 * function to call iw_cm_disconnect. Calling iw_destroy_cm_id
 862	 * from a callback handler is not allowed.
 863	 */
 864	clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 865	BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
 866	cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
 867	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 868	ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 869	wake_up_all(&cm_id_priv->connect_wait);
 870
 871	return ret;
 872}
 873
 874/*
 875 * Active Side: CM_ID <-- ESTABLISHED
 876 *
 877 * The app has called connect and is waiting for the established event to
 878 * post it's requests to the server. This event will wake up anyone
 879 * blocked in iw_cm_disconnect or iw_destroy_id.
 880 */
 881static int cm_conn_rep_handler(struct iwcm_id_private *cm_id_priv,
 882			       struct iw_cm_event *iw_event)
 883{
 884	struct ib_qp *qp = NULL;
 885	unsigned long flags;
 886	int ret;
 887
 888	spin_lock_irqsave(&cm_id_priv->lock, flags);
 889	/*
 890	 * Clear the connect wait bit so a callback function calling
 891	 * iw_cm_disconnect will not wait and deadlock this thread
 892	 */
 893	clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
 894	BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
 895	if (iw_event->status == 0) {
 896		cm_id_priv->id.m_local_addr = iw_event->local_addr;
 897		cm_id_priv->id.m_remote_addr = iw_event->remote_addr;
 898		iw_event->local_addr = cm_id_priv->id.local_addr;
 899		iw_event->remote_addr = cm_id_priv->id.remote_addr;
 900		cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
 901	} else {
 902		/* REJECTED or RESET */
 903		qp = cm_id_priv->qp;
 904		cm_id_priv->qp = NULL;
 905		cm_id_priv->state = IW_CM_STATE_IDLE;
 906	}
 907	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 908	if (qp)
 909		cm_id_priv->id.device->ops.iw_rem_ref(qp);
 910	ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 911
 912	if (iw_event->private_data_len)
 913		kfree(iw_event->private_data);
 914
 915	/* Wake up waiters on connect complete */
 916	wake_up_all(&cm_id_priv->connect_wait);
 917
 918	return ret;
 919}
 920
 921/*
 922 * CM_ID <-- CLOSING
 923 *
 924 * If in the ESTABLISHED state, move to CLOSING.
 925 */
 926static void cm_disconnect_handler(struct iwcm_id_private *cm_id_priv,
 927				  struct iw_cm_event *iw_event)
 928{
 929	unsigned long flags;
 930
 931	spin_lock_irqsave(&cm_id_priv->lock, flags);
 932	if (cm_id_priv->state == IW_CM_STATE_ESTABLISHED)
 933		cm_id_priv->state = IW_CM_STATE_CLOSING;
 934	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 935}
 936
 937/*
 938 * CM_ID <-- IDLE
 939 *
 940 * If in the ESTBLISHED or CLOSING states, the QP will have have been
 941 * moved by the provider to the ERR state. Disassociate the CM_ID from
 942 * the QP,  move to IDLE, and remove the 'connected' reference.
 943 *
 944 * If in some other state, the cm_id was destroyed asynchronously.
 945 * This is the last reference that will result in waking up
 946 * the app thread blocked in iw_destroy_cm_id.
 947 */
 948static int cm_close_handler(struct iwcm_id_private *cm_id_priv,
 949				  struct iw_cm_event *iw_event)
 950{
 951	struct ib_qp *qp;
 952	unsigned long flags;
 953	int ret = 0, notify_event = 0;
 954	spin_lock_irqsave(&cm_id_priv->lock, flags);
 955	qp = cm_id_priv->qp;
 956	cm_id_priv->qp = NULL;
 957
 
 
 
 
 958	switch (cm_id_priv->state) {
 959	case IW_CM_STATE_ESTABLISHED:
 960	case IW_CM_STATE_CLOSING:
 961		cm_id_priv->state = IW_CM_STATE_IDLE;
 962		notify_event = 1;
 
 
 963		break;
 964	case IW_CM_STATE_DESTROYING:
 965		break;
 966	default:
 967		BUG();
 968	}
 969	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
 970
 971	if (qp)
 972		cm_id_priv->id.device->ops.iw_rem_ref(qp);
 973	if (notify_event)
 974		ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
 975	return ret;
 976}
 977
 978static int process_event(struct iwcm_id_private *cm_id_priv,
 979			 struct iw_cm_event *iw_event)
 980{
 981	int ret = 0;
 982
 983	switch (iw_event->event) {
 984	case IW_CM_EVENT_CONNECT_REQUEST:
 985		cm_conn_req_handler(cm_id_priv, iw_event);
 986		break;
 987	case IW_CM_EVENT_CONNECT_REPLY:
 988		ret = cm_conn_rep_handler(cm_id_priv, iw_event);
 989		break;
 990	case IW_CM_EVENT_ESTABLISHED:
 991		ret = cm_conn_est_handler(cm_id_priv, iw_event);
 992		break;
 993	case IW_CM_EVENT_DISCONNECT:
 994		cm_disconnect_handler(cm_id_priv, iw_event);
 995		break;
 996	case IW_CM_EVENT_CLOSE:
 997		ret = cm_close_handler(cm_id_priv, iw_event);
 998		break;
 999	default:
1000		BUG();
1001	}
1002
1003	return ret;
1004}
1005
1006/*
1007 * Process events on the work_list for the cm_id. If the callback
1008 * function requests that the cm_id be deleted, a flag is set in the
1009 * cm_id flags to indicate that when the last reference is
1010 * removed, the cm_id is to be destroyed. This is necessary to
1011 * distinguish between an object that will be destroyed by the app
1012 * thread asleep on the destroy_comp list vs. an object destroyed
1013 * here synchronously when the last reference is removed.
1014 */
1015static void cm_work_handler(struct work_struct *_work)
1016{
1017	struct iwcm_work *work = container_of(_work, struct iwcm_work, work);
1018	struct iw_cm_event levent;
1019	struct iwcm_id_private *cm_id_priv = work->cm_id;
1020	unsigned long flags;
1021	int empty;
1022	int ret = 0;
 
1023
1024	spin_lock_irqsave(&cm_id_priv->lock, flags);
1025	empty = list_empty(&cm_id_priv->work_list);
1026	while (!empty) {
1027		work = list_entry(cm_id_priv->work_list.next,
1028				  struct iwcm_work, list);
1029		list_del_init(&work->list);
1030		empty = list_empty(&cm_id_priv->work_list);
1031		levent = work->event;
1032		put_work(work);
1033		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1034
1035		if (!test_bit(IWCM_F_DROP_EVENTS, &cm_id_priv->flags)) {
1036			ret = process_event(cm_id_priv, &levent);
1037			if (ret)
1038				destroy_cm_id(&cm_id_priv->id);
1039		} else
1040			pr_debug("dropping event %d\n", levent.event);
1041		if (iwcm_deref_id(cm_id_priv))
1042			return;
1043		if (empty)
 
 
 
1044			return;
 
1045		spin_lock_irqsave(&cm_id_priv->lock, flags);
1046	}
1047	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1048}
1049
1050/*
1051 * This function is called on interrupt context. Schedule events on
1052 * the iwcm_wq thread to allow callback functions to downcall into
1053 * the CM and/or block.  Events are queued to a per-CM_ID
1054 * work_list. If this is the first event on the work_list, the work
1055 * element is also queued on the iwcm_wq thread.
1056 *
1057 * Each event holds a reference on the cm_id. Until the last posted
1058 * event has been delivered and processed, the cm_id cannot be
1059 * deleted.
1060 *
1061 * Returns:
1062 * 	      0	- the event was handled.
1063 *	-ENOMEM	- the event was not handled due to lack of resources.
1064 */
1065static int cm_event_handler(struct iw_cm_id *cm_id,
1066			     struct iw_cm_event *iw_event)
1067{
1068	struct iwcm_work *work;
1069	struct iwcm_id_private *cm_id_priv;
1070	unsigned long flags;
1071	int ret = 0;
1072
1073	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1074
1075	spin_lock_irqsave(&cm_id_priv->lock, flags);
1076	work = get_work(cm_id_priv);
1077	if (!work) {
1078		ret = -ENOMEM;
1079		goto out;
1080	}
1081
1082	INIT_WORK(&work->work, cm_work_handler);
1083	work->cm_id = cm_id_priv;
1084	work->event = *iw_event;
1085
1086	if ((work->event.event == IW_CM_EVENT_CONNECT_REQUEST ||
1087	     work->event.event == IW_CM_EVENT_CONNECT_REPLY) &&
1088	    work->event.private_data_len) {
1089		ret = copy_private_data(&work->event);
1090		if (ret) {
1091			put_work(work);
1092			goto out;
1093		}
1094	}
1095
1096	refcount_inc(&cm_id_priv->refcount);
1097	if (list_empty(&cm_id_priv->work_list)) {
1098		list_add_tail(&work->list, &cm_id_priv->work_list);
1099		queue_work(iwcm_wq, &work->work);
1100	} else
1101		list_add_tail(&work->list, &cm_id_priv->work_list);
1102out:
1103	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1104	return ret;
1105}
1106
1107static int iwcm_init_qp_init_attr(struct iwcm_id_private *cm_id_priv,
1108				  struct ib_qp_attr *qp_attr,
1109				  int *qp_attr_mask)
1110{
1111	unsigned long flags;
1112	int ret;
1113
1114	spin_lock_irqsave(&cm_id_priv->lock, flags);
1115	switch (cm_id_priv->state) {
1116	case IW_CM_STATE_IDLE:
1117	case IW_CM_STATE_CONN_SENT:
1118	case IW_CM_STATE_CONN_RECV:
1119	case IW_CM_STATE_ESTABLISHED:
1120		*qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1121		qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE|
1122					   IB_ACCESS_REMOTE_READ;
1123		ret = 0;
1124		break;
1125	default:
1126		ret = -EINVAL;
1127		break;
1128	}
1129	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1130	return ret;
1131}
1132
1133static int iwcm_init_qp_rts_attr(struct iwcm_id_private *cm_id_priv,
1134				  struct ib_qp_attr *qp_attr,
1135				  int *qp_attr_mask)
1136{
1137	unsigned long flags;
1138	int ret;
1139
1140	spin_lock_irqsave(&cm_id_priv->lock, flags);
1141	switch (cm_id_priv->state) {
1142	case IW_CM_STATE_IDLE:
1143	case IW_CM_STATE_CONN_SENT:
1144	case IW_CM_STATE_CONN_RECV:
1145	case IW_CM_STATE_ESTABLISHED:
1146		*qp_attr_mask = 0;
1147		ret = 0;
1148		break;
1149	default:
1150		ret = -EINVAL;
1151		break;
1152	}
1153	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1154	return ret;
1155}
1156
1157int iw_cm_init_qp_attr(struct iw_cm_id *cm_id,
1158		       struct ib_qp_attr *qp_attr,
1159		       int *qp_attr_mask)
1160{
1161	struct iwcm_id_private *cm_id_priv;
1162	int ret;
1163
1164	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1165	switch (qp_attr->qp_state) {
1166	case IB_QPS_INIT:
1167	case IB_QPS_RTR:
1168		ret = iwcm_init_qp_init_attr(cm_id_priv,
1169					     qp_attr, qp_attr_mask);
1170		break;
1171	case IB_QPS_RTS:
1172		ret = iwcm_init_qp_rts_attr(cm_id_priv,
1173					    qp_attr, qp_attr_mask);
1174		break;
1175	default:
1176		ret = -EINVAL;
1177		break;
1178	}
1179	return ret;
1180}
1181EXPORT_SYMBOL(iw_cm_init_qp_attr);
1182
1183static int __init iw_cm_init(void)
1184{
1185	int ret;
1186
1187	ret = iwpm_init(RDMA_NL_IWCM);
1188	if (ret)
1189		return ret;
1190
1191	iwcm_wq = alloc_ordered_workqueue("iw_cm_wq", 0);
1192	if (!iwcm_wq)
1193		goto err_alloc;
1194
1195	iwcm_ctl_table_hdr = register_net_sysctl(&init_net, "net/iw_cm",
1196						 iwcm_ctl_table);
1197	if (!iwcm_ctl_table_hdr) {
1198		pr_err("iw_cm: couldn't register sysctl paths\n");
1199		goto err_sysctl;
1200	}
1201
1202	rdma_nl_register(RDMA_NL_IWCM, iwcm_nl_cb_table);
1203	return 0;
1204
1205err_sysctl:
1206	destroy_workqueue(iwcm_wq);
1207err_alloc:
1208	iwpm_exit(RDMA_NL_IWCM);
1209	return -ENOMEM;
1210}
1211
1212static void __exit iw_cm_cleanup(void)
1213{
1214	rdma_nl_unregister(RDMA_NL_IWCM);
1215	unregister_net_sysctl_table(iwcm_ctl_table_hdr);
1216	destroy_workqueue(iwcm_wq);
1217	iwpm_exit(RDMA_NL_IWCM);
1218}
1219
1220MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_IWCM, 2);
1221
1222module_init(iw_cm_init);
1223module_exit(iw_cm_cleanup);