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
54#include "iwcm.h"
55
56MODULE_AUTHOR("Tom Tucker");
57MODULE_DESCRIPTION("iWARP CM");
58MODULE_LICENSE("Dual BSD/GPL");
59
60static struct workqueue_struct *iwcm_wq;
61struct iwcm_work {
62	struct work_struct work;
63	struct iwcm_id_private *cm_id;
64	struct list_head list;
65	struct iw_cm_event event;
66	struct list_head free_list;
67};
68
69static unsigned int default_backlog = 256;
70
71static struct ctl_table_header *iwcm_ctl_table_hdr;
72static struct ctl_table iwcm_ctl_table[] = {
73	{
74		.procname	= "default_backlog",
75		.data		= &default_backlog,
76		.maxlen		= sizeof(default_backlog),
77		.mode		= 0644,
78		.proc_handler	= proc_dointvec,
79	},
80	{ }
81};
82
83/*
84 * The following services provide a mechanism for pre-allocating iwcm_work
85 * elements.  The design pre-allocates them  based on the cm_id type:
86 *	LISTENING IDS: 	Get enough elements preallocated to handle the
87 *			listen backlog.
88 *	ACTIVE IDS:	4: CONNECT_REPLY, ESTABLISHED, DISCONNECT, CLOSE
89 *	PASSIVE IDS:	3: ESTABLISHED, DISCONNECT, CLOSE
90 *
91 * Allocating them in connect and listen avoids having to deal
92 * with allocation failures on the event upcall from the provider (which
93 * is called in the interrupt context).
94 *
95 * One exception is when creating the cm_id for incoming connection requests.
96 * There are two cases:
97 * 1) in the event upcall, cm_event_handler(), for a listening cm_id.  If
98 *    the backlog is exceeded, then no more connection request events will
99 *    be processed.  cm_event_handler() returns -ENOMEM in this case.  Its up
100 *    to the provider to reject the connection request.
101 * 2) in the connection request workqueue handler, cm_conn_req_handler().
102 *    If work elements cannot be allocated for the new connect request cm_id,
103 *    then IWCM will call the provider reject method.  This is ok since
104 *    cm_conn_req_handler() runs in the workqueue thread context.
105 */
106
107static struct iwcm_work *get_work(struct iwcm_id_private *cm_id_priv)
108{
109	struct iwcm_work *work;
110
111	if (list_empty(&cm_id_priv->work_free_list))
112		return NULL;
113	work = list_entry(cm_id_priv->work_free_list.next, struct iwcm_work,
114			  free_list);
115	list_del_init(&work->free_list);
116	return work;
117}
118
119static void put_work(struct iwcm_work *work)
120{
121	list_add(&work->free_list, &work->cm_id->work_free_list);
122}
123
124static void dealloc_work_entries(struct iwcm_id_private *cm_id_priv)
125{
126	struct list_head *e, *tmp;
127
128	list_for_each_safe(e, tmp, &cm_id_priv->work_free_list)
129		kfree(list_entry(e, struct iwcm_work, free_list));
130}
131
132static int alloc_work_entries(struct iwcm_id_private *cm_id_priv, int count)
133{
134	struct iwcm_work *work;
135
136	BUG_ON(!list_empty(&cm_id_priv->work_free_list));
137	while (count--) {
138		work = kmalloc(sizeof(struct iwcm_work), GFP_KERNEL);
139		if (!work) {
140			dealloc_work_entries(cm_id_priv);
141			return -ENOMEM;
142		}
143		work->cm_id = cm_id_priv;
144		INIT_LIST_HEAD(&work->list);
145		put_work(work);
146	}
147	return 0;
148}
149
150/*
151 * Save private data from incoming connection requests to
152 * iw_cm_event, so the low level driver doesn't have to. Adjust
153 * the event ptr to point to the local copy.
154 */
155static int copy_private_data(struct iw_cm_event *event)
156{
157	void *p;
158
159	p = kmemdup(event->private_data, event->private_data_len, GFP_ATOMIC);
160	if (!p)
161		return -ENOMEM;
162	event->private_data = p;
163	return 0;
164}
165
166static void free_cm_id(struct iwcm_id_private *cm_id_priv)
167{
168	dealloc_work_entries(cm_id_priv);
169	kfree(cm_id_priv);
170}
171
172/*
173 * Release a reference on cm_id. If the last reference is being
174 * released, enable the waiting thread (in iw_destroy_cm_id) to
175 * get woken up, and return 1 if a thread is already waiting.
176 */
177static int iwcm_deref_id(struct iwcm_id_private *cm_id_priv)
178{
179	BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
180	if (atomic_dec_and_test(&cm_id_priv->refcount)) {
181		BUG_ON(!list_empty(&cm_id_priv->work_list));
182		complete(&cm_id_priv->destroy_comp);
183		return 1;
184	}
185
186	return 0;
187}
188
189static void add_ref(struct iw_cm_id *cm_id)
190{
191	struct iwcm_id_private *cm_id_priv;
192	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
193	atomic_inc(&cm_id_priv->refcount);
194}
195
196static void rem_ref(struct iw_cm_id *cm_id)
197{
198	struct iwcm_id_private *cm_id_priv;
199	int cb_destroy;
200
201	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
202
203	/*
204	 * Test bit before deref in case the cm_id gets freed on another
205	 * thread.
206	 */
207	cb_destroy = test_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
208	if (iwcm_deref_id(cm_id_priv) && cb_destroy) {
209		BUG_ON(!list_empty(&cm_id_priv->work_list));
210		free_cm_id(cm_id_priv);
211	}
212}
213
214static int cm_event_handler(struct iw_cm_id *cm_id, struct iw_cm_event *event);
215
216struct iw_cm_id *iw_create_cm_id(struct ib_device *device,
217				 iw_cm_handler cm_handler,
218				 void *context)
219{
220	struct iwcm_id_private *cm_id_priv;
221
222	cm_id_priv = kzalloc(sizeof(*cm_id_priv), GFP_KERNEL);
223	if (!cm_id_priv)
224		return ERR_PTR(-ENOMEM);
225
226	cm_id_priv->state = IW_CM_STATE_IDLE;
227	cm_id_priv->id.device = device;
228	cm_id_priv->id.cm_handler = cm_handler;
229	cm_id_priv->id.context = context;
230	cm_id_priv->id.event_handler = cm_event_handler;
231	cm_id_priv->id.add_ref = add_ref;
232	cm_id_priv->id.rem_ref = rem_ref;
233	spin_lock_init(&cm_id_priv->lock);
234	atomic_set(&cm_id_priv->refcount, 1);
235	init_waitqueue_head(&cm_id_priv->connect_wait);
236	init_completion(&cm_id_priv->destroy_comp);
237	INIT_LIST_HEAD(&cm_id_priv->work_list);
238	INIT_LIST_HEAD(&cm_id_priv->work_free_list);
239
240	return &cm_id_priv->id;
241}
242EXPORT_SYMBOL(iw_create_cm_id);
243
244
245static int iwcm_modify_qp_err(struct ib_qp *qp)
246{
247	struct ib_qp_attr qp_attr;
248
249	if (!qp)
250		return -EINVAL;
251
252	qp_attr.qp_state = IB_QPS_ERR;
253	return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
254}
255
256/*
257 * This is really the RDMAC CLOSING state. It is most similar to the
258 * IB SQD QP state.
259 */
260static int iwcm_modify_qp_sqd(struct ib_qp *qp)
261{
262	struct ib_qp_attr qp_attr;
263
264	BUG_ON(qp == NULL);
265	qp_attr.qp_state = IB_QPS_SQD;
266	return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
267}
268
269/*
270 * CM_ID <-- CLOSING
271 *
272 * Block if a passive or active connection is currently being processed. Then
273 * process the event as follows:
274 * - If we are ESTABLISHED, move to CLOSING and modify the QP state
275 *   based on the abrupt flag
276 * - If the connection is already in the CLOSING or IDLE state, the peer is
277 *   disconnecting concurrently with us and we've already seen the
278 *   DISCONNECT event -- ignore the request and return 0
279 * - Disconnect on a listening endpoint returns -EINVAL
280 */
281int iw_cm_disconnect(struct iw_cm_id *cm_id, int abrupt)
282{
283	struct iwcm_id_private *cm_id_priv;
284	unsigned long flags;
285	int ret = 0;
286	struct ib_qp *qp = NULL;
287
288	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
289	/* Wait if we're currently in a connect or accept downcall */
290	wait_event(cm_id_priv->connect_wait,
291		   !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
292
293	spin_lock_irqsave(&cm_id_priv->lock, flags);
294	switch (cm_id_priv->state) {
295	case IW_CM_STATE_ESTABLISHED:
296		cm_id_priv->state = IW_CM_STATE_CLOSING;
297
298		/* QP could be <nul> for user-mode client */
299		if (cm_id_priv->qp)
300			qp = cm_id_priv->qp;
301		else
302			ret = -EINVAL;
303		break;
304	case IW_CM_STATE_LISTEN:
305		ret = -EINVAL;
306		break;
307	case IW_CM_STATE_CLOSING:
308		/* remote peer closed first */
309	case IW_CM_STATE_IDLE:
310		/* accept or connect returned !0 */
311		break;
312	case IW_CM_STATE_CONN_RECV:
313		/*
314		 * App called disconnect before/without calling accept after
315		 * connect_request event delivered.
316		 */
317		break;
318	case IW_CM_STATE_CONN_SENT:
319		/* Can only get here if wait above fails */
320	default:
321		BUG();
322	}
323	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
324
325	if (qp) {
326		if (abrupt)
327			ret = iwcm_modify_qp_err(qp);
328		else
329			ret = iwcm_modify_qp_sqd(qp);
330
331		/*
332		 * If both sides are disconnecting the QP could
333		 * already be in ERR or SQD states
334		 */
335		ret = 0;
336	}
337
338	return ret;
339}
340EXPORT_SYMBOL(iw_cm_disconnect);
341
342/*
343 * CM_ID <-- DESTROYING
344 *
345 * Clean up all resources associated with the connection and release
346 * the initial reference taken by iw_create_cm_id.
347 */
348static void destroy_cm_id(struct iw_cm_id *cm_id)
349{
350	struct iwcm_id_private *cm_id_priv;
351	unsigned long flags;
352
353	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
354	/*
355	 * Wait if we're currently in a connect or accept downcall. A
356	 * listening endpoint should never block here.
357	 */
358	wait_event(cm_id_priv->connect_wait,
359		   !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
360
361	spin_lock_irqsave(&cm_id_priv->lock, flags);
362	switch (cm_id_priv->state) {
363	case IW_CM_STATE_LISTEN:
364		cm_id_priv->state = IW_CM_STATE_DESTROYING;
365		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
366		/* destroy the listening endpoint */
367		cm_id->device->iwcm->destroy_listen(cm_id);
368		spin_lock_irqsave(&cm_id_priv->lock, flags);
369		break;
370	case IW_CM_STATE_ESTABLISHED:
371		cm_id_priv->state = IW_CM_STATE_DESTROYING;
372		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
373		/* Abrupt close of the connection */
374		(void)iwcm_modify_qp_err(cm_id_priv->qp);
375		spin_lock_irqsave(&cm_id_priv->lock, flags);
376		break;
377	case IW_CM_STATE_IDLE:
378	case IW_CM_STATE_CLOSING:
379		cm_id_priv->state = IW_CM_STATE_DESTROYING;
380		break;
381	case IW_CM_STATE_CONN_RECV:
382		/*
383		 * App called destroy before/without calling accept after
384		 * receiving connection request event notification or
385		 * returned non zero from the event callback function.
386		 * In either case, must tell the provider to reject.
387		 */
388		cm_id_priv->state = IW_CM_STATE_DESTROYING;
389		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
390		cm_id->device->iwcm->reject(cm_id, NULL, 0);
391		spin_lock_irqsave(&cm_id_priv->lock, flags);
392		break;
393	case IW_CM_STATE_CONN_SENT:
394	case IW_CM_STATE_DESTROYING:
395	default:
396		BUG();
397		break;
398	}
399	if (cm_id_priv->qp) {
400		cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
401		cm_id_priv->qp = NULL;
402	}
403	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
404
405	(void)iwcm_deref_id(cm_id_priv);
406}
407
408/*
409 * This function is only called by the application thread and cannot
410 * be called by the event thread. The function will wait for all
411 * references to be released on the cm_id and then kfree the cm_id
412 * object.
413 */
414void iw_destroy_cm_id(struct iw_cm_id *cm_id)
415{
416	struct iwcm_id_private *cm_id_priv;
417
418	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
419	BUG_ON(test_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags));
420
421	destroy_cm_id(cm_id);
422
423	wait_for_completion(&cm_id_priv->destroy_comp);
424
425	free_cm_id(cm_id_priv);
426}
427EXPORT_SYMBOL(iw_destroy_cm_id);
428
429/*
430 * CM_ID <-- LISTEN
431 *
432 * Start listening for connect requests. Generates one CONNECT_REQUEST
433 * event for each inbound connect request.
434 */
435int iw_cm_listen(struct iw_cm_id *cm_id, int backlog)
436{
437	struct iwcm_id_private *cm_id_priv;
438	unsigned long flags;
439	int ret;
440
441	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
442
443	if (!backlog)
444		backlog = default_backlog;
445
446	ret = alloc_work_entries(cm_id_priv, backlog);
447	if (ret)
448		return ret;
449
450	spin_lock_irqsave(&cm_id_priv->lock, flags);
451	switch (cm_id_priv->state) {
452	case IW_CM_STATE_IDLE:
453		cm_id_priv->state = IW_CM_STATE_LISTEN;
454		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
455		ret = cm_id->device->iwcm->create_listen(cm_id, backlog);
456		if (ret)
457			cm_id_priv->state = IW_CM_STATE_IDLE;
458		spin_lock_irqsave(&cm_id_priv->lock, flags);
459		break;
460	default:
461		ret = -EINVAL;
462	}
463	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
464
465	return ret;
466}
467EXPORT_SYMBOL(iw_cm_listen);
468
469/*
470 * CM_ID <-- IDLE
471 *
472 * Rejects an inbound connection request. No events are generated.
473 */
474int iw_cm_reject(struct iw_cm_id *cm_id,
475		 const void *private_data,
476		 u8 private_data_len)
477{
478	struct iwcm_id_private *cm_id_priv;
479	unsigned long flags;
480	int ret;
481
482	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
483	set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
484
485	spin_lock_irqsave(&cm_id_priv->lock, flags);
486	if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
487		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
488		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
489		wake_up_all(&cm_id_priv->connect_wait);
490		return -EINVAL;
491	}
492	cm_id_priv->state = IW_CM_STATE_IDLE;
493	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
494
495	ret = cm_id->device->iwcm->reject(cm_id, private_data,
496					  private_data_len);
497
498	clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
499	wake_up_all(&cm_id_priv->connect_wait);
500
501	return ret;
502}
503EXPORT_SYMBOL(iw_cm_reject);
504
505/*
506 * CM_ID <-- ESTABLISHED
507 *
508 * Accepts an inbound connection request and generates an ESTABLISHED
509 * event. Callers of iw_cm_disconnect and iw_destroy_cm_id will block
510 * until the ESTABLISHED event is received from the provider.
511 */
512int iw_cm_accept(struct iw_cm_id *cm_id,
513		 struct iw_cm_conn_param *iw_param)
514{
515	struct iwcm_id_private *cm_id_priv;
516	struct ib_qp *qp;
517	unsigned long flags;
518	int ret;
519
520	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
521	set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
522
523	spin_lock_irqsave(&cm_id_priv->lock, flags);
524	if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
525		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
526		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
527		wake_up_all(&cm_id_priv->connect_wait);
528		return -EINVAL;
529	}
530	/* Get the ib_qp given the QPN */
531	qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
532	if (!qp) {
533		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
534		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
535		wake_up_all(&cm_id_priv->connect_wait);
536		return -EINVAL;
537	}
538	cm_id->device->iwcm->add_ref(qp);
539	cm_id_priv->qp = qp;
540	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
541
542	ret = cm_id->device->iwcm->accept(cm_id, iw_param);
543	if (ret) {
544		/* An error on accept precludes provider events */
545		BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
546		cm_id_priv->state = IW_CM_STATE_IDLE;
547		spin_lock_irqsave(&cm_id_priv->lock, flags);
548		if (cm_id_priv->qp) {
549			cm_id->device->iwcm->rem_ref(qp);
550			cm_id_priv->qp = NULL;
551		}
552		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
553		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
554		wake_up_all(&cm_id_priv->connect_wait);
555	}
556
557	return ret;
558}
559EXPORT_SYMBOL(iw_cm_accept);
560
561/*
562 * Active Side: CM_ID <-- CONN_SENT
563 *
564 * If successful, results in the generation of a CONNECT_REPLY
565 * event. iw_cm_disconnect and iw_cm_destroy will block until the
566 * CONNECT_REPLY event is received from the provider.
567 */
568int iw_cm_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *iw_param)
569{
570	struct iwcm_id_private *cm_id_priv;
571	int ret;
572	unsigned long flags;
573	struct ib_qp *qp;
574
575	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
576
577	ret = alloc_work_entries(cm_id_priv, 4);
578	if (ret)
579		return ret;
580
581	set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
582	spin_lock_irqsave(&cm_id_priv->lock, flags);
583
584	if (cm_id_priv->state != IW_CM_STATE_IDLE) {
585		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
586		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
587		wake_up_all(&cm_id_priv->connect_wait);
588		return -EINVAL;
589	}
590
591	/* Get the ib_qp given the QPN */
592	qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
593	if (!qp) {
594		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
595		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
596		wake_up_all(&cm_id_priv->connect_wait);
597		return -EINVAL;
598	}
599	cm_id->device->iwcm->add_ref(qp);
600	cm_id_priv->qp = qp;
601	cm_id_priv->state = IW_CM_STATE_CONN_SENT;
602	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
603
604	ret = cm_id->device->iwcm->connect(cm_id, iw_param);
605	if (ret) {
606		spin_lock_irqsave(&cm_id_priv->lock, flags);
607		if (cm_id_priv->qp) {
608			cm_id->device->iwcm->rem_ref(qp);
609			cm_id_priv->qp = NULL;
610		}
611		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
612		BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
613		cm_id_priv->state = IW_CM_STATE_IDLE;
614		clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
615		wake_up_all(&cm_id_priv->connect_wait);
616	}
617
618	return ret;
619}
620EXPORT_SYMBOL(iw_cm_connect);
621
622/*
623 * Passive Side: new CM_ID <-- CONN_RECV
624 *
625 * Handles an inbound connect request. The function creates a new
626 * iw_cm_id to represent the new connection and inherits the client
627 * callback function and other attributes from the listening parent.
628 *
629 * The work item contains a pointer to the listen_cm_id and the event. The
630 * listen_cm_id contains the client cm_handler, context and
631 * device. These are copied when the device is cloned. The event
632 * contains the new four tuple.
633 *
634 * An error on the child should not affect the parent, so this
635 * function does not return a value.
636 */
637static void cm_conn_req_handler(struct iwcm_id_private *listen_id_priv,
638				struct iw_cm_event *iw_event)
639{
640	unsigned long flags;
641	struct iw_cm_id *cm_id;
642	struct iwcm_id_private *cm_id_priv;
643	int ret;
644
645	/*
646	 * The provider should never generate a connection request
647	 * event with a bad status.
648	 */
649	BUG_ON(iw_event->status);
650
651	cm_id = iw_create_cm_id(listen_id_priv->id.device,
652				listen_id_priv->id.cm_handler,
653				listen_id_priv->id.context);
654	/* If the cm_id could not be created, ignore the request */
655	if (IS_ERR(cm_id))
656		goto out;
657
658	cm_id->provider_data = iw_event->provider_data;
659	cm_id->local_addr = iw_event->local_addr;
660	cm_id->remote_addr = iw_event->remote_addr;
661
662	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
663	cm_id_priv->state = IW_CM_STATE_CONN_RECV;
664
665	/*
666	 * We could be destroying the listening id. If so, ignore this
667	 * upcall.
668	 */
669	spin_lock_irqsave(&listen_id_priv->lock, flags);
670	if (listen_id_priv->state != IW_CM_STATE_LISTEN) {
671		spin_unlock_irqrestore(&listen_id_priv->lock, flags);
672		iw_cm_reject(cm_id, NULL, 0);
673		iw_destroy_cm_id(cm_id);
674		goto out;
675	}
676	spin_unlock_irqrestore(&listen_id_priv->lock, flags);
677
678	ret = alloc_work_entries(cm_id_priv, 3);
679	if (ret) {
680		iw_cm_reject(cm_id, NULL, 0);
681		iw_destroy_cm_id(cm_id);
682		goto out;
683	}
684
685	/* Call the client CM handler */
686	ret = cm_id->cm_handler(cm_id, iw_event);
687	if (ret) {
688		iw_cm_reject(cm_id, NULL, 0);
689		set_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
690		destroy_cm_id(cm_id);
691		if (atomic_read(&cm_id_priv->refcount)==0)
692			free_cm_id(cm_id_priv);
693	}
694
695out:
696	if (iw_event->private_data_len)
697		kfree(iw_event->private_data);
698}
699
700/*
701 * Passive Side: CM_ID <-- ESTABLISHED
702 *
703 * The provider generated an ESTABLISHED event which means that
704 * the MPA negotion has completed successfully and we are now in MPA
705 * FPDU mode.
706 *
707 * This event can only be received in the CONN_RECV state. If the
708 * remote peer closed, the ESTABLISHED event would be received followed
709 * by the CLOSE event. If the app closes, it will block until we wake
710 * it up after processing this event.
711 */
712static int cm_conn_est_handler(struct iwcm_id_private *cm_id_priv,
713			       struct iw_cm_event *iw_event)
714{
715	unsigned long flags;
716	int ret;
717
718	spin_lock_irqsave(&cm_id_priv->lock, flags);
719
720	/*
721	 * We clear the CONNECT_WAIT bit here to allow the callback
722	 * function to call iw_cm_disconnect. Calling iw_destroy_cm_id
723	 * from a callback handler is not allowed.
724	 */
725	clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
726	BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
727	cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
728	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
729	ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
730	wake_up_all(&cm_id_priv->connect_wait);
731
732	return ret;
733}
734
735/*
736 * Active Side: CM_ID <-- ESTABLISHED
737 *
738 * The app has called connect and is waiting for the established event to
739 * post it's requests to the server. This event will wake up anyone
740 * blocked in iw_cm_disconnect or iw_destroy_id.
741 */
742static int cm_conn_rep_handler(struct iwcm_id_private *cm_id_priv,
743			       struct iw_cm_event *iw_event)
744{
745	unsigned long flags;
746	int ret;
747
748	spin_lock_irqsave(&cm_id_priv->lock, flags);
749	/*
750	 * Clear the connect wait bit so a callback function calling
751	 * iw_cm_disconnect will not wait and deadlock this thread
752	 */
753	clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
754	BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
755	if (iw_event->status == 0) {
756		cm_id_priv->id.local_addr = iw_event->local_addr;
757		cm_id_priv->id.remote_addr = iw_event->remote_addr;
758		cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
759	} else {
760		/* REJECTED or RESET */
761		cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
762		cm_id_priv->qp = NULL;
763		cm_id_priv->state = IW_CM_STATE_IDLE;
764	}
765	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
766	ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
767
768	if (iw_event->private_data_len)
769		kfree(iw_event->private_data);
770
771	/* Wake up waiters on connect complete */
772	wake_up_all(&cm_id_priv->connect_wait);
773
774	return ret;
775}
776
777/*
778 * CM_ID <-- CLOSING
779 *
780 * If in the ESTABLISHED state, move to CLOSING.
781 */
782static void cm_disconnect_handler(struct iwcm_id_private *cm_id_priv,
783				  struct iw_cm_event *iw_event)
784{
785	unsigned long flags;
786
787	spin_lock_irqsave(&cm_id_priv->lock, flags);
788	if (cm_id_priv->state == IW_CM_STATE_ESTABLISHED)
789		cm_id_priv->state = IW_CM_STATE_CLOSING;
790	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
791}
792
793/*
794 * CM_ID <-- IDLE
795 *
796 * If in the ESTBLISHED or CLOSING states, the QP will have have been
797 * moved by the provider to the ERR state. Disassociate the CM_ID from
798 * the QP,  move to IDLE, and remove the 'connected' reference.
799 *
800 * If in some other state, the cm_id was destroyed asynchronously.
801 * This is the last reference that will result in waking up
802 * the app thread blocked in iw_destroy_cm_id.
803 */
804static int cm_close_handler(struct iwcm_id_private *cm_id_priv,
805				  struct iw_cm_event *iw_event)
806{
807	unsigned long flags;
808	int ret = 0;
809	spin_lock_irqsave(&cm_id_priv->lock, flags);
810
811	if (cm_id_priv->qp) {
812		cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
813		cm_id_priv->qp = NULL;
814	}
815	switch (cm_id_priv->state) {
816	case IW_CM_STATE_ESTABLISHED:
817	case IW_CM_STATE_CLOSING:
818		cm_id_priv->state = IW_CM_STATE_IDLE;
819		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
820		ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
821		spin_lock_irqsave(&cm_id_priv->lock, flags);
822		break;
823	case IW_CM_STATE_DESTROYING:
824		break;
825	default:
826		BUG();
827	}
828	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
829
830	return ret;
831}
832
833static int process_event(struct iwcm_id_private *cm_id_priv,
834			 struct iw_cm_event *iw_event)
835{
836	int ret = 0;
837
838	switch (iw_event->event) {
839	case IW_CM_EVENT_CONNECT_REQUEST:
840		cm_conn_req_handler(cm_id_priv, iw_event);
841		break;
842	case IW_CM_EVENT_CONNECT_REPLY:
843		ret = cm_conn_rep_handler(cm_id_priv, iw_event);
844		break;
845	case IW_CM_EVENT_ESTABLISHED:
846		ret = cm_conn_est_handler(cm_id_priv, iw_event);
847		break;
848	case IW_CM_EVENT_DISCONNECT:
849		cm_disconnect_handler(cm_id_priv, iw_event);
850		break;
851	case IW_CM_EVENT_CLOSE:
852		ret = cm_close_handler(cm_id_priv, iw_event);
853		break;
854	default:
855		BUG();
856	}
857
858	return ret;
859}
860
861/*
862 * Process events on the work_list for the cm_id. If the callback
863 * function requests that the cm_id be deleted, a flag is set in the
864 * cm_id flags to indicate that when the last reference is
865 * removed, the cm_id is to be destroyed. This is necessary to
866 * distinguish between an object that will be destroyed by the app
867 * thread asleep on the destroy_comp list vs. an object destroyed
868 * here synchronously when the last reference is removed.
869 */
870static void cm_work_handler(struct work_struct *_work)
871{
872	struct iwcm_work *work = container_of(_work, struct iwcm_work, work);
873	struct iw_cm_event levent;
874	struct iwcm_id_private *cm_id_priv = work->cm_id;
875	unsigned long flags;
876	int empty;
877	int ret = 0;
878	int destroy_id;
879
880	spin_lock_irqsave(&cm_id_priv->lock, flags);
881	empty = list_empty(&cm_id_priv->work_list);
882	while (!empty) {
883		work = list_entry(cm_id_priv->work_list.next,
884				  struct iwcm_work, list);
885		list_del_init(&work->list);
886		empty = list_empty(&cm_id_priv->work_list);
887		levent = work->event;
888		put_work(work);
889		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
890
891		ret = process_event(cm_id_priv, &levent);
892		if (ret) {
893			set_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
894			destroy_cm_id(&cm_id_priv->id);
895		}
896		BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
897		destroy_id = test_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
898		if (iwcm_deref_id(cm_id_priv)) {
899			if (destroy_id) {
900				BUG_ON(!list_empty(&cm_id_priv->work_list));
901				free_cm_id(cm_id_priv);
902			}
903			return;
904		}
905		if (empty)
906			return;
907		spin_lock_irqsave(&cm_id_priv->lock, flags);
908	}
909	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
910}
911
912/*
913 * This function is called on interrupt context. Schedule events on
914 * the iwcm_wq thread to allow callback functions to downcall into
915 * the CM and/or block.  Events are queued to a per-CM_ID
916 * work_list. If this is the first event on the work_list, the work
917 * element is also queued on the iwcm_wq thread.
918 *
919 * Each event holds a reference on the cm_id. Until the last posted
920 * event has been delivered and processed, the cm_id cannot be
921 * deleted.
922 *
923 * Returns:
924 * 	      0	- the event was handled.
925 *	-ENOMEM	- the event was not handled due to lack of resources.
926 */
927static int cm_event_handler(struct iw_cm_id *cm_id,
928			     struct iw_cm_event *iw_event)
929{
930	struct iwcm_work *work;
931	struct iwcm_id_private *cm_id_priv;
932	unsigned long flags;
933	int ret = 0;
934
935	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
936
937	spin_lock_irqsave(&cm_id_priv->lock, flags);
938	work = get_work(cm_id_priv);
939	if (!work) {
940		ret = -ENOMEM;
941		goto out;
942	}
943
944	INIT_WORK(&work->work, cm_work_handler);
945	work->cm_id = cm_id_priv;
946	work->event = *iw_event;
947
948	if ((work->event.event == IW_CM_EVENT_CONNECT_REQUEST ||
949	     work->event.event == IW_CM_EVENT_CONNECT_REPLY) &&
950	    work->event.private_data_len) {
951		ret = copy_private_data(&work->event);
952		if (ret) {
953			put_work(work);
954			goto out;
955		}
956	}
957
958	atomic_inc(&cm_id_priv->refcount);
959	if (list_empty(&cm_id_priv->work_list)) {
960		list_add_tail(&work->list, &cm_id_priv->work_list);
961		queue_work(iwcm_wq, &work->work);
962	} else
963		list_add_tail(&work->list, &cm_id_priv->work_list);
964out:
965	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
966	return ret;
967}
968
969static int iwcm_init_qp_init_attr(struct iwcm_id_private *cm_id_priv,
970				  struct ib_qp_attr *qp_attr,
971				  int *qp_attr_mask)
972{
973	unsigned long flags;
974	int ret;
975
976	spin_lock_irqsave(&cm_id_priv->lock, flags);
977	switch (cm_id_priv->state) {
978	case IW_CM_STATE_IDLE:
979	case IW_CM_STATE_CONN_SENT:
980	case IW_CM_STATE_CONN_RECV:
981	case IW_CM_STATE_ESTABLISHED:
982		*qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
983		qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE|
984					   IB_ACCESS_REMOTE_READ;
985		ret = 0;
986		break;
987	default:
988		ret = -EINVAL;
989		break;
990	}
991	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
992	return ret;
993}
994
995static int iwcm_init_qp_rts_attr(struct iwcm_id_private *cm_id_priv,
996				  struct ib_qp_attr *qp_attr,
997				  int *qp_attr_mask)
998{
999	unsigned long flags;
1000	int ret;
1001
1002	spin_lock_irqsave(&cm_id_priv->lock, flags);
1003	switch (cm_id_priv->state) {
1004	case IW_CM_STATE_IDLE:
1005	case IW_CM_STATE_CONN_SENT:
1006	case IW_CM_STATE_CONN_RECV:
1007	case IW_CM_STATE_ESTABLISHED:
1008		*qp_attr_mask = 0;
1009		ret = 0;
1010		break;
1011	default:
1012		ret = -EINVAL;
1013		break;
1014	}
1015	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1016	return ret;
1017}
1018
1019int iw_cm_init_qp_attr(struct iw_cm_id *cm_id,
1020		       struct ib_qp_attr *qp_attr,
1021		       int *qp_attr_mask)
1022{
1023	struct iwcm_id_private *cm_id_priv;
1024	int ret;
1025
1026	cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1027	switch (qp_attr->qp_state) {
1028	case IB_QPS_INIT:
1029	case IB_QPS_RTR:
1030		ret = iwcm_init_qp_init_attr(cm_id_priv,
1031					     qp_attr, qp_attr_mask);
1032		break;
1033	case IB_QPS_RTS:
1034		ret = iwcm_init_qp_rts_attr(cm_id_priv,
1035					    qp_attr, qp_attr_mask);
1036		break;
1037	default:
1038		ret = -EINVAL;
1039		break;
1040	}
1041	return ret;
1042}
1043EXPORT_SYMBOL(iw_cm_init_qp_attr);
1044
1045static int __init iw_cm_init(void)
1046{
1047	iwcm_wq = create_singlethread_workqueue("iw_cm_wq");
1048	if (!iwcm_wq)
1049		return -ENOMEM;
1050
1051	iwcm_ctl_table_hdr = register_net_sysctl(&init_net, "net/iw_cm",
1052						 iwcm_ctl_table);
1053	if (!iwcm_ctl_table_hdr) {
1054		pr_err("iw_cm: couldn't register sysctl paths\n");
1055		destroy_workqueue(iwcm_wq);
1056		return -ENOMEM;
1057	}
1058
1059	return 0;
1060}
1061
1062static void __exit iw_cm_cleanup(void)
1063{
1064	unregister_net_sysctl_table(iwcm_ctl_table_hdr);
1065	destroy_workqueue(iwcm_wq);
1066}
1067
1068module_init(iw_cm_init);
1069module_exit(iw_cm_cleanup);
1070