1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0
3  *
4  * Copyright (c) 2004, 2005 Intel Corporation.  All rights reserved.
5  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
6  * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
7  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
8  * Copyright (c) 2005 Open Grid Computing, Inc. All rights reserved.
9  * Copyright (c) 2005 Network Appliance, Inc. All rights reserved.
10  *
11  * This software is available to you under a choice of one of two
12  * licenses.  You may choose to be licensed under the terms of the GNU
13  * General Public License (GPL) Version 2, available from the file
14  * COPYING in the main directory of this source tree, or the
15  * OpenIB.org BSD license below:
16  *
17  *     Redistribution and use in source and binary forms, with or
18  *     without modification, are permitted provided that the following
19  *     conditions are met:
20  *
21  *      - Redistributions of source code must retain the above
22  *        copyright notice, this list of conditions and the following
23  *        disclaimer.
24  *
25  *      - Redistributions in binary form must reproduce the above
26  *        copyright notice, this list of conditions and the following
27  *        disclaimer in the documentation and/or other materials
28  *        provided with the distribution.
29  *
30  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
31  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
32  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
33  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
34  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
35  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
36  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
37  * SOFTWARE.
38  */
39 
40 #include <sys/cdefs.h>
41 __FBSDID("$FreeBSD$");
42 
43 #include <linux/dma-mapping.h>
44 #include <linux/err.h>
45 #include <linux/idr.h>
46 #include <linux/interrupt.h>
47 #include <linux/rbtree.h>
48 #include <linux/sched.h>
49 #include <linux/spinlock.h>
50 #include <linux/workqueue.h>
51 #include <linux/completion.h>
52 #include <linux/slab.h>
53 #include <linux/module.h>
54 #include <linux/wait.h>
55 
56 #include <rdma/iw_cm.h>
57 #include <rdma/ib_addr.h>
58 #include <rdma/iw_portmap.h>
59 
60 #include "iwcm.h"
61 
62 MODULE_AUTHOR("Tom Tucker");
63 MODULE_DESCRIPTION("iWARP CM");
64 MODULE_LICENSE("Dual BSD/GPL");
65 
66 static const char * const iwcm_rej_reason_strs[] = {
67 	[ECONNRESET]			= "reset by remote host",
68 	[ECONNREFUSED]			= "refused by remote application",
69 	[ETIMEDOUT]			= "setup timeout",
70 };
71 
iwcm_reject_msg(int reason)72 const char *__attribute_const__ iwcm_reject_msg(int reason)
73 {
74 	size_t index;
75 
76 	/* iWARP uses negative errnos */
77 	index = -reason;
78 
79 	if (index < ARRAY_SIZE(iwcm_rej_reason_strs) &&
80 	    iwcm_rej_reason_strs[index])
81 		return iwcm_rej_reason_strs[index];
82 	else
83 		return "unrecognized reason";
84 }
85 EXPORT_SYMBOL(iwcm_reject_msg);
86 
87 static struct workqueue_struct *iwcm_wq;
88 struct iwcm_work {
89 	struct work_struct work;
90 	struct iwcm_id_private *cm_id;
91 	struct list_head list;
92 	struct iw_cm_event event;
93 	struct list_head free_list;
94 };
95 
96 static unsigned int default_backlog = 256;
97 
98 /*
99  * The following services provide a mechanism for pre-allocating iwcm_work
100  * elements.  The design pre-allocates them  based on the cm_id type:
101  *	LISTENING IDS: 	Get enough elements preallocated to handle the
102  *			listen backlog.
103  *	ACTIVE IDS:	4: CONNECT_REPLY, ESTABLISHED, DISCONNECT, CLOSE
104  *	PASSIVE IDS:	3: ESTABLISHED, DISCONNECT, CLOSE
105  *
106  * Allocating them in connect and listen avoids having to deal
107  * with allocation failures on the event upcall from the provider (which
108  * is called in the interrupt context).
109  *
110  * One exception is when creating the cm_id for incoming connection requests.
111  * There are two cases:
112  * 1) in the event upcall, cm_event_handler(), for a listening cm_id.  If
113  *    the backlog is exceeded, then no more connection request events will
114  *    be processed.  cm_event_handler() returns -ENOMEM in this case.  Its up
115  *    to the provider to reject the connection request.
116  * 2) in the connection request workqueue handler, cm_conn_req_handler().
117  *    If work elements cannot be allocated for the new connect request cm_id,
118  *    then IWCM will call the provider reject method.  This is ok since
119  *    cm_conn_req_handler() runs in the workqueue thread context.
120  */
121 
get_work(struct iwcm_id_private * cm_id_priv)122 static struct iwcm_work *get_work(struct iwcm_id_private *cm_id_priv)
123 {
124 	struct iwcm_work *work;
125 
126 	if (list_empty(&cm_id_priv->work_free_list))
127 		return NULL;
128 	work = list_entry(cm_id_priv->work_free_list.next, struct iwcm_work,
129 			  free_list);
130 	list_del_init(&work->free_list);
131 	return work;
132 }
133 
put_work(struct iwcm_work * work)134 static void put_work(struct iwcm_work *work)
135 {
136 	list_add(&work->free_list, &work->cm_id->work_free_list);
137 }
138 
dealloc_work_entries(struct iwcm_id_private * cm_id_priv)139 static void dealloc_work_entries(struct iwcm_id_private *cm_id_priv)
140 {
141 	struct list_head *e, *tmp;
142 
143 	list_for_each_safe(e, tmp, &cm_id_priv->work_free_list)
144 		kfree(list_entry(e, struct iwcm_work, free_list));
145 }
146 
alloc_work_entries(struct iwcm_id_private * cm_id_priv,int count)147 static int alloc_work_entries(struct iwcm_id_private *cm_id_priv, int count)
148 {
149 	struct iwcm_work *work;
150 
151 	BUG_ON(!list_empty(&cm_id_priv->work_free_list));
152 	while (count--) {
153 		work = kmalloc(sizeof(struct iwcm_work), GFP_KERNEL);
154 		if (!work) {
155 			dealloc_work_entries(cm_id_priv);
156 			return -ENOMEM;
157 		}
158 		work->cm_id = cm_id_priv;
159 		INIT_LIST_HEAD(&work->list);
160 		put_work(work);
161 	}
162 	return 0;
163 }
164 
165 /*
166  * Save private data from incoming connection requests to
167  * iw_cm_event, so the low level driver doesn't have to. Adjust
168  * the event ptr to point to the local copy.
169  */
copy_private_data(struct iw_cm_event * event)170 static int copy_private_data(struct iw_cm_event *event)
171 {
172 	void *p;
173 
174 	p = kmemdup(event->private_data, event->private_data_len, GFP_ATOMIC);
175 	if (!p)
176 		return -ENOMEM;
177 	event->private_data = p;
178 	return 0;
179 }
180 
free_cm_id(struct iwcm_id_private * cm_id_priv)181 static void free_cm_id(struct iwcm_id_private *cm_id_priv)
182 {
183 	dealloc_work_entries(cm_id_priv);
184 	kfree(cm_id_priv);
185 }
186 
187 /*
188  * Release a reference on cm_id. If the last reference is being
189  * released, free the cm_id and return 1.
190  */
iwcm_deref_id(struct iwcm_id_private * cm_id_priv)191 static int iwcm_deref_id(struct iwcm_id_private *cm_id_priv)
192 {
193 	BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
194 	if (atomic_dec_and_test(&cm_id_priv->refcount)) {
195 		BUG_ON(!list_empty(&cm_id_priv->work_list));
196 		free_cm_id(cm_id_priv);
197 		return 1;
198 	}
199 
200 	return 0;
201 }
202 
add_ref(struct iw_cm_id * cm_id)203 static void add_ref(struct iw_cm_id *cm_id)
204 {
205 	struct iwcm_id_private *cm_id_priv;
206 	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
207 	atomic_inc(&cm_id_priv->refcount);
208 }
209 
rem_ref(struct iw_cm_id * cm_id)210 static void rem_ref(struct iw_cm_id *cm_id)
211 {
212 	struct iwcm_id_private *cm_id_priv;
213 
214 	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
215 
216 	(void)iwcm_deref_id(cm_id_priv);
217 }
218 
219 static int cm_event_handler(struct iw_cm_id *cm_id, struct iw_cm_event *event);
220 
iw_create_cm_id(struct ib_device * device,iw_cm_handler cm_handler,void * context)221 struct iw_cm_id *iw_create_cm_id(struct ib_device *device,
222 				 iw_cm_handler cm_handler,
223 				 void *context)
224 {
225 	struct iwcm_id_private *cm_id_priv;
226 
227 	cm_id_priv = kzalloc(sizeof(*cm_id_priv), GFP_KERNEL);
228 	if (!cm_id_priv)
229 		return ERR_PTR(-ENOMEM);
230 
231 	cm_id_priv->state = IW_CM_STATE_IDLE;
232 	cm_id_priv->id.device = device;
233 	cm_id_priv->id.cm_handler = cm_handler;
234 	cm_id_priv->id.context = context;
235 	cm_id_priv->id.event_handler = cm_event_handler;
236 	cm_id_priv->id.add_ref = add_ref;
237 	cm_id_priv->id.rem_ref = rem_ref;
238 	spin_lock_init(&cm_id_priv->lock);
239 	atomic_set(&cm_id_priv->refcount, 1);
240 	init_waitqueue_head(&cm_id_priv->connect_wait);
241 	init_completion(&cm_id_priv->destroy_comp);
242 	INIT_LIST_HEAD(&cm_id_priv->work_list);
243 	INIT_LIST_HEAD(&cm_id_priv->work_free_list);
244 
245 	return &cm_id_priv->id;
246 }
247 EXPORT_SYMBOL(iw_create_cm_id);
248 
249 
iwcm_modify_qp_err(struct ib_qp * qp)250 static int iwcm_modify_qp_err(struct ib_qp *qp)
251 {
252 	struct ib_qp_attr qp_attr;
253 
254 	if (!qp)
255 		return -EINVAL;
256 
257 	qp_attr.qp_state = IB_QPS_ERR;
258 	return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
259 }
260 
261 /*
262  * This is really the RDMAC CLOSING state. It is most similar to the
263  * IB SQD QP state.
264  */
iwcm_modify_qp_sqd(struct ib_qp * qp)265 static int iwcm_modify_qp_sqd(struct ib_qp *qp)
266 {
267 	struct ib_qp_attr qp_attr;
268 
269 	BUG_ON(qp == NULL);
270 	qp_attr.qp_state = IB_QPS_SQD;
271 	return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
272 }
273 
274 /*
275  * CM_ID <-- CLOSING
276  *
277  * Block if a passive or active connection is currently being processed. Then
278  * process the event as follows:
279  * - If we are ESTABLISHED, move to CLOSING and modify the QP state
280  *   based on the abrupt flag
281  * - If the connection is already in the CLOSING or IDLE state, the peer is
282  *   disconnecting concurrently with us and we've already seen the
283  *   DISCONNECT event -- ignore the request and return 0
284  * - Disconnect on a listening endpoint returns -EINVAL
285  */
iw_cm_disconnect(struct iw_cm_id * cm_id,int abrupt)286 int iw_cm_disconnect(struct iw_cm_id *cm_id, int abrupt)
287 {
288 	struct iwcm_id_private *cm_id_priv;
289 	unsigned long flags;
290 	int ret = 0;
291 	struct ib_qp *qp = NULL;
292 
293 	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
294 	/* Wait if we're currently in a connect or accept downcall */
295 	wait_event(cm_id_priv->connect_wait,
296 		   !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
297 
298 	spin_lock_irqsave(&cm_id_priv->lock, flags);
299 	switch (cm_id_priv->state) {
300 	case IW_CM_STATE_ESTABLISHED:
301 		cm_id_priv->state = IW_CM_STATE_CLOSING;
302 
303 		/* QP could be <nul> for user-mode client */
304 		if (cm_id_priv->qp)
305 			qp = cm_id_priv->qp;
306 		else
307 			ret = -EINVAL;
308 		break;
309 	case IW_CM_STATE_LISTEN:
310 		ret = -EINVAL;
311 		break;
312 	case IW_CM_STATE_CLOSING:
313 		/* remote peer closed first */
314 	case IW_CM_STATE_IDLE:
315 		/* accept or connect returned !0 */
316 		break;
317 	case IW_CM_STATE_CONN_RECV:
318 		/*
319 		 * App called disconnect before/without calling accept after
320 		 * connect_request event delivered.
321 		 */
322 		break;
323 	case IW_CM_STATE_CONN_SENT:
324 		/* Can only get here if wait above fails */
325 	default:
326 		BUG();
327 	}
328 	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
329 
330 	if (qp) {
331 		if (abrupt)
332 			(void) iwcm_modify_qp_err(qp);
333 		else
334 			(void) iwcm_modify_qp_sqd(qp);
335 
336 		/*
337 		 * If both sides are disconnecting the QP could
338 		 * already be in ERR or SQD states
339 		 */
340 		ret = 0;
341 	}
342 
343 	return ret;
344 }
345 EXPORT_SYMBOL(iw_cm_disconnect);
346 
347 /*
348  * CM_ID <-- DESTROYING
349  *
350  * Clean up all resources associated with the connection and release
351  * the initial reference taken by iw_create_cm_id.
352  */
destroy_cm_id(struct iw_cm_id * cm_id)353 static void destroy_cm_id(struct iw_cm_id *cm_id)
354 {
355 	struct iwcm_id_private *cm_id_priv;
356 	unsigned long flags;
357 
358 	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
359 	/*
360 	 * Wait if we're currently in a connect or accept downcall. A
361 	 * listening endpoint should never block here.
362 	 */
363 	wait_event(cm_id_priv->connect_wait,
364 		   !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
365 
366 	/*
367 	 * Since we're deleting the cm_id, drop any events that
368 	 * might arrive before the last dereference.
369 	 */
370 	set_bit(IWCM_F_DROP_EVENTS, &cm_id_priv->flags);
371 
372 	spin_lock_irqsave(&cm_id_priv->lock, flags);
373 	switch (cm_id_priv->state) {
374 	case IW_CM_STATE_LISTEN:
375 		cm_id_priv->state = IW_CM_STATE_DESTROYING;
376 		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
377 		/* destroy the listening endpoint */
378 		cm_id->device->iwcm->destroy_listen(cm_id);
379 		spin_lock_irqsave(&cm_id_priv->lock, flags);
380 		break;
381 	case IW_CM_STATE_ESTABLISHED:
382 		cm_id_priv->state = IW_CM_STATE_DESTROYING;
383 		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
384 		/* Abrupt close of the connection */
385 		(void)iwcm_modify_qp_err(cm_id_priv->qp);
386 		spin_lock_irqsave(&cm_id_priv->lock, flags);
387 		break;
388 	case IW_CM_STATE_IDLE:
389 	case IW_CM_STATE_CLOSING:
390 		cm_id_priv->state = IW_CM_STATE_DESTROYING;
391 		break;
392 	case IW_CM_STATE_CONN_RECV:
393 		/*
394 		 * App called destroy before/without calling accept after
395 		 * receiving connection request event notification or
396 		 * returned non zero from the event callback function.
397 		 * In either case, must tell the provider to reject.
398 		 */
399 		cm_id_priv->state = IW_CM_STATE_DESTROYING;
400 		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
401 		cm_id->device->iwcm->reject(cm_id, NULL, 0);
402 		spin_lock_irqsave(&cm_id_priv->lock, flags);
403 		break;
404 	case IW_CM_STATE_CONN_SENT:
405 	case IW_CM_STATE_DESTROYING:
406 	default:
407 		BUG();
408 		break;
409 	}
410 	if (cm_id_priv->qp) {
411 		cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
412 		cm_id_priv->qp = NULL;
413 	}
414 	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
415 
416 	(void)iwcm_deref_id(cm_id_priv);
417 }
418 
419 /*
420  * This function is only called by the application thread and cannot
421  * be called by the event thread. The function will wait for all
422  * references to be released on the cm_id and then kfree the cm_id
423  * object.
424  */
iw_destroy_cm_id(struct iw_cm_id * cm_id)425 void iw_destroy_cm_id(struct iw_cm_id *cm_id)
426 {
427 	struct iwcm_id_private *cm_id_priv;
428 
429 	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
430 	destroy_cm_id(cm_id);
431 }
432 EXPORT_SYMBOL(iw_destroy_cm_id);
433 
434 /**
435  * iw_cm_map - Use portmapper to map the ports
436  * @cm_id: connection manager pointer
437  * @active: Indicates the active side when true
438  * returns nonzero for error only if iwpm_create_mapinfo() fails
439  *
440  * Tries to add a mapping for a port using the Portmapper. If
441  * successful in mapping the IP/Port it will check the remote
442  * mapped IP address for a wildcard IP address and replace the
443  * zero IP address with the remote_addr.
444  */
iw_cm_map(struct iw_cm_id * cm_id,bool active)445 static int iw_cm_map(struct iw_cm_id *cm_id, bool active)
446 {
447 	cm_id->m_local_addr = cm_id->local_addr;
448 	cm_id->m_remote_addr = cm_id->remote_addr;
449 
450 	return 0;
451 }
452 
453 /*
454  * CM_ID <-- LISTEN
455  *
456  * Start listening for connect requests. Generates one CONNECT_REQUEST
457  * event for each inbound connect request.
458  */
iw_cm_listen(struct iw_cm_id * cm_id,int backlog)459 int iw_cm_listen(struct iw_cm_id *cm_id, int backlog)
460 {
461 	struct iwcm_id_private *cm_id_priv;
462 	unsigned long flags;
463 	int ret;
464 
465 	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
466 
467 	if (!backlog)
468 		backlog = default_backlog;
469 
470 	ret = alloc_work_entries(cm_id_priv, backlog);
471 	if (ret)
472 		return ret;
473 
474 	spin_lock_irqsave(&cm_id_priv->lock, flags);
475 	switch (cm_id_priv->state) {
476 	case IW_CM_STATE_IDLE:
477 		cm_id_priv->state = IW_CM_STATE_LISTEN;
478 		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
479 		ret = iw_cm_map(cm_id, false);
480 		if (!ret)
481 			ret = cm_id->device->iwcm->create_listen(cm_id, backlog);
482 		if (ret)
483 			cm_id_priv->state = IW_CM_STATE_IDLE;
484 		spin_lock_irqsave(&cm_id_priv->lock, flags);
485 		break;
486 	default:
487 		ret = -EINVAL;
488 	}
489 	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
490 
491 	return ret;
492 }
493 EXPORT_SYMBOL(iw_cm_listen);
494 
495 /*
496  * CM_ID <-- IDLE
497  *
498  * Rejects an inbound connection request. No events are generated.
499  */
iw_cm_reject(struct iw_cm_id * cm_id,const void * private_data,u8 private_data_len)500 int iw_cm_reject(struct iw_cm_id *cm_id,
501 		 const void *private_data,
502 		 u8 private_data_len)
503 {
504 	struct iwcm_id_private *cm_id_priv;
505 	unsigned long flags;
506 	int ret;
507 
508 	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
509 	set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
510 
511 	spin_lock_irqsave(&cm_id_priv->lock, flags);
512 	if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
513 		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
514 		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
515 		wake_up_all(&cm_id_priv->connect_wait);
516 		return -EINVAL;
517 	}
518 	cm_id_priv->state = IW_CM_STATE_IDLE;
519 	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
520 
521 	ret = cm_id->device->iwcm->reject(cm_id, private_data,
522 					  private_data_len);
523 
524 	clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
525 	wake_up_all(&cm_id_priv->connect_wait);
526 
527 	return ret;
528 }
529 EXPORT_SYMBOL(iw_cm_reject);
530 
531 /*
532  * CM_ID <-- ESTABLISHED
533  *
534  * Accepts an inbound connection request and generates an ESTABLISHED
535  * event. Callers of iw_cm_disconnect and iw_destroy_cm_id will block
536  * until the ESTABLISHED event is received from the provider.
537  */
iw_cm_accept(struct iw_cm_id * cm_id,struct iw_cm_conn_param * iw_param)538 int iw_cm_accept(struct iw_cm_id *cm_id,
539 		 struct iw_cm_conn_param *iw_param)
540 {
541 	struct iwcm_id_private *cm_id_priv;
542 	struct ib_qp *qp;
543 	unsigned long flags;
544 	int ret;
545 
546 	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
547 	set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
548 
549 	spin_lock_irqsave(&cm_id_priv->lock, flags);
550 	if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
551 		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
552 		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
553 		wake_up_all(&cm_id_priv->connect_wait);
554 		return -EINVAL;
555 	}
556 	/* Get the ib_qp given the QPN */
557 	qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
558 	if (!qp) {
559 		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
560 		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
561 		wake_up_all(&cm_id_priv->connect_wait);
562 		return -EINVAL;
563 	}
564 	cm_id->device->iwcm->add_ref(qp);
565 	cm_id_priv->qp = qp;
566 	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
567 
568 	ret = cm_id->device->iwcm->accept(cm_id, iw_param);
569 	if (ret) {
570 		/* An error on accept precludes provider events */
571 		BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
572 		cm_id_priv->state = IW_CM_STATE_IDLE;
573 		spin_lock_irqsave(&cm_id_priv->lock, flags);
574 		if (cm_id_priv->qp) {
575 			cm_id->device->iwcm->rem_ref(qp);
576 			cm_id_priv->qp = NULL;
577 		}
578 		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
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_accept);
586 
587 /*
588  * Active Side: CM_ID <-- CONN_SENT
589  *
590  * If successful, results in the generation of a CONNECT_REPLY
591  * event. iw_cm_disconnect and iw_cm_destroy will block until the
592  * CONNECT_REPLY event is received from the provider.
593  */
iw_cm_connect(struct iw_cm_id * cm_id,struct iw_cm_conn_param * iw_param)594 int iw_cm_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *iw_param)
595 {
596 	struct iwcm_id_private *cm_id_priv;
597 	int ret;
598 	unsigned long flags;
599 	struct ib_qp *qp;
600 
601 	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
602 
603 	ret = alloc_work_entries(cm_id_priv, 4);
604 	if (ret)
605 		return ret;
606 
607 	set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
608 	spin_lock_irqsave(&cm_id_priv->lock, flags);
609 
610 	if (cm_id_priv->state != IW_CM_STATE_IDLE) {
611 		ret = -EINVAL;
612 		goto err;
613 	}
614 
615 	/* Get the ib_qp given the QPN */
616 	qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
617 	if (!qp) {
618 		ret = -EINVAL;
619 		goto err;
620 	}
621 	cm_id->device->iwcm->add_ref(qp);
622 	cm_id_priv->qp = qp;
623 	cm_id_priv->state = IW_CM_STATE_CONN_SENT;
624 	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
625 
626 	ret = iw_cm_map(cm_id, true);
627 	if (!ret)
628 		ret = cm_id->device->iwcm->connect(cm_id, iw_param);
629 	if (!ret)
630 		return 0;	/* success */
631 
632 	spin_lock_irqsave(&cm_id_priv->lock, flags);
633 	if (cm_id_priv->qp) {
634 		cm_id->device->iwcm->rem_ref(qp);
635 		cm_id_priv->qp = NULL;
636 	}
637 	cm_id_priv->state = IW_CM_STATE_IDLE;
638 err:
639 	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
640 	clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
641 	wake_up_all(&cm_id_priv->connect_wait);
642 	return ret;
643 }
644 EXPORT_SYMBOL(iw_cm_connect);
645 
646 /*
647  * Passive Side: new CM_ID <-- CONN_RECV
648  *
649  * Handles an inbound connect request. The function creates a new
650  * iw_cm_id to represent the new connection and inherits the client
651  * callback function and other attributes from the listening parent.
652  *
653  * The work item contains a pointer to the listen_cm_id and the event. The
654  * listen_cm_id contains the client cm_handler, context and
655  * device. These are copied when the device is cloned. The event
656  * contains the new four tuple.
657  *
658  * An error on the child should not affect the parent, so this
659  * function does not return a value.
660  */
cm_conn_req_handler(struct iwcm_id_private * listen_id_priv,struct iw_cm_event * iw_event)661 static void cm_conn_req_handler(struct iwcm_id_private *listen_id_priv,
662 				struct iw_cm_event *iw_event)
663 {
664 	unsigned long flags;
665 	struct iw_cm_id *cm_id;
666 	struct iwcm_id_private *cm_id_priv;
667 	int ret;
668 
669 	/*
670 	 * The provider should never generate a connection request
671 	 * event with a bad status.
672 	 */
673 	BUG_ON(iw_event->status);
674 
675 	cm_id = iw_create_cm_id(listen_id_priv->id.device,
676 				listen_id_priv->id.cm_handler,
677 				listen_id_priv->id.context);
678 	/* If the cm_id could not be created, ignore the request */
679 	if (IS_ERR(cm_id))
680 		goto out;
681 
682 	cm_id->provider_data = iw_event->provider_data;
683 	cm_id->m_local_addr = iw_event->local_addr;
684 	cm_id->m_remote_addr = iw_event->remote_addr;
685 	cm_id->local_addr = listen_id_priv->id.local_addr;
686 	cm_id->remote_addr = iw_event->remote_addr;
687 	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
688 	cm_id_priv->state = IW_CM_STATE_CONN_RECV;
689 
690 	/*
691 	 * We could be destroying the listening id. If so, ignore this
692 	 * upcall.
693 	 */
694 	spin_lock_irqsave(&listen_id_priv->lock, flags);
695 	if (listen_id_priv->state != IW_CM_STATE_LISTEN) {
696 		spin_unlock_irqrestore(&listen_id_priv->lock, flags);
697 		iw_cm_reject(cm_id, NULL, 0);
698 		iw_destroy_cm_id(cm_id);
699 		goto out;
700 	}
701 	spin_unlock_irqrestore(&listen_id_priv->lock, flags);
702 
703 	ret = alloc_work_entries(cm_id_priv, 3);
704 	if (ret) {
705 		iw_cm_reject(cm_id, NULL, 0);
706 		iw_destroy_cm_id(cm_id);
707 		goto out;
708 	}
709 
710 	/* Call the client CM handler */
711 	ret = cm_id->cm_handler(cm_id, iw_event);
712 	if (ret) {
713 		iw_cm_reject(cm_id, NULL, 0);
714 		iw_destroy_cm_id(cm_id);
715 	}
716 
717 out:
718 	if (iw_event->private_data_len)
719 		kfree(iw_event->private_data);
720 }
721 
722 /*
723  * Passive Side: CM_ID <-- ESTABLISHED
724  *
725  * The provider generated an ESTABLISHED event which means that
726  * the MPA negotion has completed successfully and we are now in MPA
727  * FPDU mode.
728  *
729  * This event can only be received in the CONN_RECV state. If the
730  * remote peer closed, the ESTABLISHED event would be received followed
731  * by the CLOSE event. If the app closes, it will block until we wake
732  * it up after processing this event.
733  */
cm_conn_est_handler(struct iwcm_id_private * cm_id_priv,struct iw_cm_event * iw_event)734 static int cm_conn_est_handler(struct iwcm_id_private *cm_id_priv,
735 			       struct iw_cm_event *iw_event)
736 {
737 	unsigned long flags;
738 	int ret;
739 
740 	spin_lock_irqsave(&cm_id_priv->lock, flags);
741 
742 	/*
743 	 * We clear the CONNECT_WAIT bit here to allow the callback
744 	 * function to call iw_cm_disconnect. Calling iw_destroy_cm_id
745 	 * from a callback handler is not allowed.
746 	 */
747 	clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
748 	BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
749 	cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
750 	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
751 	ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
752 	wake_up_all(&cm_id_priv->connect_wait);
753 
754 	return ret;
755 }
756 
757 /*
758  * Active Side: CM_ID <-- ESTABLISHED
759  *
760  * The app has called connect and is waiting for the established event to
761  * post it's requests to the server. This event will wake up anyone
762  * blocked in iw_cm_disconnect or iw_destroy_id.
763  */
cm_conn_rep_handler(struct iwcm_id_private * cm_id_priv,struct iw_cm_event * iw_event)764 static int cm_conn_rep_handler(struct iwcm_id_private *cm_id_priv,
765 			       struct iw_cm_event *iw_event)
766 {
767 	unsigned long flags;
768 	int ret;
769 
770 	spin_lock_irqsave(&cm_id_priv->lock, flags);
771 	/*
772 	 * Clear the connect wait bit so a callback function calling
773 	 * iw_cm_disconnect will not wait and deadlock this thread
774 	 */
775 	clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
776 	BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
777 	if (iw_event->status == 0) {
778 		cm_id_priv->id.m_local_addr = iw_event->local_addr;
779 		cm_id_priv->id.m_remote_addr = iw_event->remote_addr;
780 		iw_event->local_addr = cm_id_priv->id.local_addr;
781 		iw_event->remote_addr = cm_id_priv->id.remote_addr;
782 		cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
783 	} else {
784 		/* REJECTED or RESET */
785 		cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
786 		cm_id_priv->qp = NULL;
787 		cm_id_priv->state = IW_CM_STATE_IDLE;
788 	}
789 	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
790 	ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
791 
792 	if (iw_event->private_data_len)
793 		kfree(iw_event->private_data);
794 
795 	/* Wake up waiters on connect complete */
796 	wake_up_all(&cm_id_priv->connect_wait);
797 
798 	return ret;
799 }
800 
801 /*
802  * CM_ID <-- CLOSING
803  *
804  * If in the ESTABLISHED state, move to CLOSING.
805  */
cm_disconnect_handler(struct iwcm_id_private * cm_id_priv,struct iw_cm_event * iw_event)806 static void cm_disconnect_handler(struct iwcm_id_private *cm_id_priv,
807 				  struct iw_cm_event *iw_event)
808 {
809 	unsigned long flags;
810 
811 	spin_lock_irqsave(&cm_id_priv->lock, flags);
812 	if (cm_id_priv->state == IW_CM_STATE_ESTABLISHED)
813 		cm_id_priv->state = IW_CM_STATE_CLOSING;
814 	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
815 }
816 
817 /*
818  * CM_ID <-- IDLE
819  *
820  * If in the ESTBLISHED or CLOSING states, the QP will have have been
821  * moved by the provider to the ERR state. Disassociate the CM_ID from
822  * the QP,  move to IDLE, and remove the 'connected' reference.
823  *
824  * If in some other state, the cm_id was destroyed asynchronously.
825  * This is the last reference that will result in waking up
826  * the app thread blocked in iw_destroy_cm_id.
827  */
cm_close_handler(struct iwcm_id_private * cm_id_priv,struct iw_cm_event * iw_event)828 static int cm_close_handler(struct iwcm_id_private *cm_id_priv,
829 				  struct iw_cm_event *iw_event)
830 {
831 	unsigned long flags;
832 	int ret = 0;
833 	spin_lock_irqsave(&cm_id_priv->lock, flags);
834 
835 	if (cm_id_priv->qp) {
836 		cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
837 		cm_id_priv->qp = NULL;
838 	}
839 	switch (cm_id_priv->state) {
840 	case IW_CM_STATE_ESTABLISHED:
841 	case IW_CM_STATE_CLOSING:
842 		cm_id_priv->state = IW_CM_STATE_IDLE;
843 		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
844 		ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
845 		spin_lock_irqsave(&cm_id_priv->lock, flags);
846 		break;
847 	case IW_CM_STATE_DESTROYING:
848 		break;
849 	default:
850 		BUG();
851 	}
852 	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
853 
854 	return ret;
855 }
856 
process_event(struct iwcm_id_private * cm_id_priv,struct iw_cm_event * iw_event)857 static int process_event(struct iwcm_id_private *cm_id_priv,
858 			 struct iw_cm_event *iw_event)
859 {
860 	int ret = 0;
861 
862 	switch (iw_event->event) {
863 	case IW_CM_EVENT_CONNECT_REQUEST:
864 		cm_conn_req_handler(cm_id_priv, iw_event);
865 		break;
866 	case IW_CM_EVENT_CONNECT_REPLY:
867 		ret = cm_conn_rep_handler(cm_id_priv, iw_event);
868 		break;
869 	case IW_CM_EVENT_ESTABLISHED:
870 		ret = cm_conn_est_handler(cm_id_priv, iw_event);
871 		break;
872 	case IW_CM_EVENT_DISCONNECT:
873 		cm_disconnect_handler(cm_id_priv, iw_event);
874 		break;
875 	case IW_CM_EVENT_CLOSE:
876 		ret = cm_close_handler(cm_id_priv, iw_event);
877 		break;
878 	default:
879 		BUG();
880 	}
881 
882 	return ret;
883 }
884 
885 /*
886  * Process events on the work_list for the cm_id. If the callback
887  * function requests that the cm_id be deleted, a flag is set in the
888  * cm_id flags to indicate that when the last reference is
889  * removed, the cm_id is to be destroyed. This is necessary to
890  * distinguish between an object that will be destroyed by the app
891  * thread asleep on the destroy_comp list vs. an object destroyed
892  * here synchronously when the last reference is removed.
893  */
cm_work_handler(struct work_struct * _work)894 static void cm_work_handler(struct work_struct *_work)
895 {
896 	struct iwcm_work *work = container_of(_work, struct iwcm_work, work);
897 	struct iw_cm_event levent;
898 	struct iwcm_id_private *cm_id_priv = work->cm_id;
899 	unsigned long flags;
900 	int empty;
901 	int ret = 0;
902 
903 	spin_lock_irqsave(&cm_id_priv->lock, flags);
904 	empty = list_empty(&cm_id_priv->work_list);
905 	while (!empty) {
906 		work = list_entry(cm_id_priv->work_list.next,
907 				  struct iwcm_work, list);
908 		list_del_init(&work->list);
909 		empty = list_empty(&cm_id_priv->work_list);
910 		levent = work->event;
911 		put_work(work);
912 		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
913 
914 		if (!test_bit(IWCM_F_DROP_EVENTS, &cm_id_priv->flags)) {
915 			ret = process_event(cm_id_priv, &levent);
916 			if (ret)
917 				destroy_cm_id(&cm_id_priv->id);
918 		} else
919 			pr_debug("dropping event %d\n", levent.event);
920 		if (iwcm_deref_id(cm_id_priv))
921 			return;
922 		if (empty)
923 			return;
924 		spin_lock_irqsave(&cm_id_priv->lock, flags);
925 	}
926 	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
927 }
928 
929 /*
930  * This function is called on interrupt context. Schedule events on
931  * the iwcm_wq thread to allow callback functions to downcall into
932  * the CM and/or block.  Events are queued to a per-CM_ID
933  * work_list. If this is the first event on the work_list, the work
934  * element is also queued on the iwcm_wq thread.
935  *
936  * Each event holds a reference on the cm_id. Until the last posted
937  * event has been delivered and processed, the cm_id cannot be
938  * deleted.
939  *
940  * Returns:
941  * 	      0	- the event was handled.
942  *	-ENOMEM	- the event was not handled due to lack of resources.
943  */
cm_event_handler(struct iw_cm_id * cm_id,struct iw_cm_event * iw_event)944 static int cm_event_handler(struct iw_cm_id *cm_id,
945 			     struct iw_cm_event *iw_event)
946 {
947 	struct iwcm_work *work;
948 	struct iwcm_id_private *cm_id_priv;
949 	unsigned long flags;
950 	int ret = 0;
951 
952 	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
953 
954 	spin_lock_irqsave(&cm_id_priv->lock, flags);
955 	work = get_work(cm_id_priv);
956 	if (!work) {
957 		ret = -ENOMEM;
958 		goto out;
959 	}
960 
961 	INIT_WORK(&work->work, cm_work_handler);
962 	work->cm_id = cm_id_priv;
963 	work->event = *iw_event;
964 
965 	if ((work->event.event == IW_CM_EVENT_CONNECT_REQUEST ||
966 	     work->event.event == IW_CM_EVENT_CONNECT_REPLY) &&
967 	    work->event.private_data_len) {
968 		ret = copy_private_data(&work->event);
969 		if (ret) {
970 			put_work(work);
971 			goto out;
972 		}
973 	}
974 
975 	atomic_inc(&cm_id_priv->refcount);
976 	if (list_empty(&cm_id_priv->work_list)) {
977 		list_add_tail(&work->list, &cm_id_priv->work_list);
978 		queue_work(iwcm_wq, &work->work);
979 	} else
980 		list_add_tail(&work->list, &cm_id_priv->work_list);
981 out:
982 	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
983 	return ret;
984 }
985 
iwcm_init_qp_init_attr(struct iwcm_id_private * cm_id_priv,struct ib_qp_attr * qp_attr,int * qp_attr_mask)986 static int iwcm_init_qp_init_attr(struct iwcm_id_private *cm_id_priv,
987 				  struct ib_qp_attr *qp_attr,
988 				  int *qp_attr_mask)
989 {
990 	unsigned long flags;
991 	int ret;
992 
993 	spin_lock_irqsave(&cm_id_priv->lock, flags);
994 	switch (cm_id_priv->state) {
995 	case IW_CM_STATE_IDLE:
996 	case IW_CM_STATE_CONN_SENT:
997 	case IW_CM_STATE_CONN_RECV:
998 	case IW_CM_STATE_ESTABLISHED:
999 		*qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1000 		qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE|
1001 					   IB_ACCESS_REMOTE_READ;
1002 		ret = 0;
1003 		break;
1004 	default:
1005 		ret = -EINVAL;
1006 		break;
1007 	}
1008 	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1009 	return ret;
1010 }
1011 
iwcm_init_qp_rts_attr(struct iwcm_id_private * cm_id_priv,struct ib_qp_attr * qp_attr,int * qp_attr_mask)1012 static int iwcm_init_qp_rts_attr(struct iwcm_id_private *cm_id_priv,
1013 				  struct ib_qp_attr *qp_attr,
1014 				  int *qp_attr_mask)
1015 {
1016 	unsigned long flags;
1017 	int ret;
1018 
1019 	spin_lock_irqsave(&cm_id_priv->lock, flags);
1020 	switch (cm_id_priv->state) {
1021 	case IW_CM_STATE_IDLE:
1022 	case IW_CM_STATE_CONN_SENT:
1023 	case IW_CM_STATE_CONN_RECV:
1024 	case IW_CM_STATE_ESTABLISHED:
1025 		*qp_attr_mask = 0;
1026 		ret = 0;
1027 		break;
1028 	default:
1029 		ret = -EINVAL;
1030 		break;
1031 	}
1032 	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1033 	return ret;
1034 }
1035 
iw_cm_init_qp_attr(struct iw_cm_id * cm_id,struct ib_qp_attr * qp_attr,int * qp_attr_mask)1036 int iw_cm_init_qp_attr(struct iw_cm_id *cm_id,
1037 		       struct ib_qp_attr *qp_attr,
1038 		       int *qp_attr_mask)
1039 {
1040 	struct iwcm_id_private *cm_id_priv;
1041 	int ret;
1042 
1043 	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1044 	switch (qp_attr->qp_state) {
1045 	case IB_QPS_INIT:
1046 	case IB_QPS_RTR:
1047 		ret = iwcm_init_qp_init_attr(cm_id_priv,
1048 					     qp_attr, qp_attr_mask);
1049 		break;
1050 	case IB_QPS_RTS:
1051 		ret = iwcm_init_qp_rts_attr(cm_id_priv,
1052 					    qp_attr, qp_attr_mask);
1053 		break;
1054 	default:
1055 		ret = -EINVAL;
1056 		break;
1057 	}
1058 	return ret;
1059 }
1060 EXPORT_SYMBOL(iw_cm_init_qp_attr);
1061 
iw_cm_init(void)1062 static int __init iw_cm_init(void)
1063 {
1064 	iwcm_wq = alloc_ordered_workqueue("iw_cm_wq", WQ_MEM_RECLAIM);
1065 	if (!iwcm_wq)
1066 		return -ENOMEM;
1067 
1068 	return 0;
1069 }
1070 
iw_cm_cleanup(void)1071 static void __exit iw_cm_cleanup(void)
1072 {
1073 	destroy_workqueue(iwcm_wq);
1074 }
1075 
1076 module_init_order(iw_cm_init, SI_ORDER_FIRST);
1077 module_exit_order(iw_cm_cleanup, SI_ORDER_FIRST);
1078