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/string.h>
50
51 #include <rdma/iw_cm.h>
52 #include <rdma/ib_addr.h>
53
54 #include "iwcm.h"
55
56 MODULE_AUTHOR("Tom Tucker");
57 MODULE_DESCRIPTION("iWARP CM");
58 MODULE_LICENSE("Dual BSD/GPL");
59
60 static struct workqueue_struct *iwcm_wq;
61 struct 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
69 /*
70 * The following services provide a mechanism for pre-allocating iwcm_work
71 * elements. The design pre-allocates them based on the cm_id type:
72 * LISTENING IDS: Get enough elements preallocated to handle the
73 * listen backlog.
74 * ACTIVE IDS: 4: CONNECT_REPLY, ESTABLISHED, DISCONNECT, CLOSE
75 * PASSIVE IDS: 3: ESTABLISHED, DISCONNECT, CLOSE
76 *
77 * Allocating them in connect and listen avoids having to deal
78 * with allocation failures on the event upcall from the provider (which
79 * is called in the interrupt context).
80 *
81 * One exception is when creating the cm_id for incoming connection requests.
82 * There are two cases:
83 * 1) in the event upcall, cm_event_handler(), for a listening cm_id. If
84 * the backlog is exceeded, then no more connection request events will
85 * be processed. cm_event_handler() returns -ENOMEM in this case. Its up
86 * to the provider to reject the connection request.
87 * 2) in the connection request workqueue handler, cm_conn_req_handler().
88 * If work elements cannot be allocated for the new connect request cm_id,
89 * then IWCM will call the provider reject method. This is ok since
90 * cm_conn_req_handler() runs in the workqueue thread context.
91 */
92
get_work(struct iwcm_id_private * cm_id_priv)93 static struct iwcm_work *get_work(struct iwcm_id_private *cm_id_priv)
94 {
95 struct iwcm_work *work;
96
97 if (list_empty(&cm_id_priv->work_free_list))
98 return NULL;
99 work = list_entry(cm_id_priv->work_free_list.next, struct iwcm_work,
100 free_list);
101 list_del_init(&work->free_list);
102 return work;
103 }
104
put_work(struct iwcm_work * work)105 static void put_work(struct iwcm_work *work)
106 {
107 list_add(&work->free_list, &work->cm_id->work_free_list);
108 }
109
dealloc_work_entries(struct iwcm_id_private * cm_id_priv)110 static void dealloc_work_entries(struct iwcm_id_private *cm_id_priv)
111 {
112 struct list_head *e, *tmp;
113
114 list_for_each_safe(e, tmp, &cm_id_priv->work_free_list)
115 kfree(list_entry(e, struct iwcm_work, free_list));
116 }
117
alloc_work_entries(struct iwcm_id_private * cm_id_priv,int count)118 static int alloc_work_entries(struct iwcm_id_private *cm_id_priv, int count)
119 {
120 struct iwcm_work *work;
121
122 BUG_ON(!list_empty(&cm_id_priv->work_free_list));
123 while (count--) {
124 work = kmalloc(sizeof(struct iwcm_work), GFP_KERNEL);
125 if (!work) {
126 dealloc_work_entries(cm_id_priv);
127 return -ENOMEM;
128 }
129 work->cm_id = cm_id_priv;
130 INIT_LIST_HEAD(&work->list);
131 put_work(work);
132 }
133 return 0;
134 }
135
136 /*
137 * Save private data from incoming connection requests to
138 * iw_cm_event, so the low level driver doesn't have to. Adjust
139 * the event ptr to point to the local copy.
140 */
copy_private_data(struct iw_cm_event * event)141 static int copy_private_data(struct iw_cm_event *event)
142 {
143 void *p;
144
145 p = kmemdup(event->private_data, event->private_data_len, GFP_ATOMIC);
146 if (!p)
147 return -ENOMEM;
148 event->private_data = p;
149 return 0;
150 }
151
free_cm_id(struct iwcm_id_private * cm_id_priv)152 static void free_cm_id(struct iwcm_id_private *cm_id_priv)
153 {
154 dealloc_work_entries(cm_id_priv);
155 kfree(cm_id_priv);
156 }
157
158 /*
159 * Release a reference on cm_id. If the last reference is being
160 * released, enable the waiting thread (in iw_destroy_cm_id) to
161 * get woken up, and return 1 if a thread is already waiting.
162 */
iwcm_deref_id(struct iwcm_id_private * cm_id_priv)163 static int iwcm_deref_id(struct iwcm_id_private *cm_id_priv)
164 {
165 BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
166 if (atomic_dec_and_test(&cm_id_priv->refcount)) {
167 BUG_ON(!list_empty(&cm_id_priv->work_list));
168 complete(&cm_id_priv->destroy_comp);
169 return 1;
170 }
171
172 return 0;
173 }
174
add_ref(struct iw_cm_id * cm_id)175 static void add_ref(struct iw_cm_id *cm_id)
176 {
177 struct iwcm_id_private *cm_id_priv;
178 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
179 atomic_inc(&cm_id_priv->refcount);
180 }
181
rem_ref(struct iw_cm_id * cm_id)182 static void rem_ref(struct iw_cm_id *cm_id)
183 {
184 struct iwcm_id_private *cm_id_priv;
185 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
186 if (iwcm_deref_id(cm_id_priv) &&
187 test_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags)) {
188 BUG_ON(!list_empty(&cm_id_priv->work_list));
189 free_cm_id(cm_id_priv);
190 }
191 }
192
193 static int cm_event_handler(struct iw_cm_id *cm_id, struct iw_cm_event *event);
194
iw_create_cm_id(struct ib_device * device,struct socket * so,iw_cm_handler cm_handler,void * context)195 struct iw_cm_id *iw_create_cm_id(struct ib_device *device,
196 struct socket *so,
197 iw_cm_handler cm_handler,
198 void *context)
199 {
200 struct iwcm_id_private *cm_id_priv;
201
202 cm_id_priv = kzalloc(sizeof(*cm_id_priv), GFP_KERNEL);
203 if (!cm_id_priv)
204 return ERR_PTR(-ENOMEM);
205
206 cm_id_priv->state = IW_CM_STATE_IDLE;
207 cm_id_priv->id.device = device;
208 cm_id_priv->id.cm_handler = cm_handler;
209 cm_id_priv->id.context = context;
210 cm_id_priv->id.event_handler = cm_event_handler;
211 cm_id_priv->id.add_ref = add_ref;
212 cm_id_priv->id.rem_ref = rem_ref;
213 cm_id_priv->id.so = so;
214 spin_lock_init(&cm_id_priv->lock);
215 atomic_set(&cm_id_priv->refcount, 1);
216 init_waitqueue_head(&cm_id_priv->connect_wait);
217 init_completion(&cm_id_priv->destroy_comp);
218 INIT_LIST_HEAD(&cm_id_priv->work_list);
219 INIT_LIST_HEAD(&cm_id_priv->work_free_list);
220
221 return &cm_id_priv->id;
222 }
223 EXPORT_SYMBOL(iw_create_cm_id);
224
225
iwcm_modify_qp_err(struct ib_qp * qp)226 static int iwcm_modify_qp_err(struct ib_qp *qp)
227 {
228 struct ib_qp_attr qp_attr;
229
230 if (!qp)
231 return -EINVAL;
232
233 qp_attr.qp_state = IB_QPS_ERR;
234 return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
235 }
236
237 /*
238 * This is really the RDMAC CLOSING state. It is most similar to the
239 * IB SQD QP state.
240 */
iwcm_modify_qp_sqd(struct ib_qp * qp)241 static int iwcm_modify_qp_sqd(struct ib_qp *qp)
242 {
243 struct ib_qp_attr qp_attr;
244
245 BUG_ON(qp == NULL);
246 qp_attr.qp_state = IB_QPS_SQD;
247 return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
248 }
249
250 /*
251 * CM_ID <-- CLOSING
252 *
253 * Block if a passive or active connection is currently being processed. Then
254 * process the event as follows:
255 * - If we are ESTABLISHED, move to CLOSING and modify the QP state
256 * based on the abrupt flag
257 * - If the connection is already in the CLOSING or IDLE state, the peer is
258 * disconnecting concurrently with us and we've already seen the
259 * DISCONNECT event -- ignore the request and return 0
260 * - Disconnect on a listening endpoint returns -EINVAL
261 */
iw_cm_disconnect(struct iw_cm_id * cm_id,int abrupt)262 int iw_cm_disconnect(struct iw_cm_id *cm_id, int abrupt)
263 {
264 struct iwcm_id_private *cm_id_priv;
265 unsigned long flags;
266 int ret = 0;
267 struct ib_qp *qp = NULL;
268
269 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
270 /* Wait if we're currently in a connect or accept downcall */
271 wait_event(cm_id_priv->connect_wait,
272 !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
273
274 spin_lock_irqsave(&cm_id_priv->lock, flags);
275 switch (cm_id_priv->state) {
276 case IW_CM_STATE_ESTABLISHED:
277 cm_id_priv->state = IW_CM_STATE_CLOSING;
278
279 /* QP could be <nul> for user-mode client */
280 if (cm_id_priv->qp)
281 qp = cm_id_priv->qp;
282 else
283 ret = -EINVAL;
284 break;
285 case IW_CM_STATE_LISTEN:
286 ret = -EINVAL;
287 break;
288 case IW_CM_STATE_CLOSING:
289 /* remote peer closed first */
290 case IW_CM_STATE_IDLE:
291 /* accept or connect returned !0 */
292 break;
293 case IW_CM_STATE_CONN_RECV:
294 /*
295 * App called disconnect before/without calling accept after
296 * connect_request event delivered.
297 */
298 break;
299 case IW_CM_STATE_CONN_SENT:
300 /* Can only get here if wait above fails */
301 default:
302 BUG();
303 }
304 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
305
306 if (qp) {
307 if (abrupt)
308 ret = iwcm_modify_qp_err(qp);
309 else
310 ret = iwcm_modify_qp_sqd(qp);
311
312 /*
313 * If both sides are disconnecting the QP could
314 * already be in ERR or SQD states
315 */
316 ret = 0;
317 }
318
319 return ret;
320 }
321 EXPORT_SYMBOL(iw_cm_disconnect);
322
323 /*
324 * CM_ID <-- DESTROYING
325 *
326 * Clean up all resources associated with the connection and release
327 * the initial reference taken by iw_create_cm_id.
328 */
destroy_cm_id(struct iw_cm_id * cm_id)329 static void destroy_cm_id(struct iw_cm_id *cm_id)
330 {
331 struct iwcm_id_private *cm_id_priv;
332 unsigned long flags;
333 int ret;
334
335 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
336 /*
337 * Wait if we're currently in a connect or accept downcall. A
338 * listening endpoint should never block here.
339 */
340 wait_event(cm_id_priv->connect_wait,
341 !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
342
343 spin_lock_irqsave(&cm_id_priv->lock, flags);
344 switch (cm_id_priv->state) {
345 case IW_CM_STATE_LISTEN:
346 cm_id_priv->state = IW_CM_STATE_DESTROYING;
347 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
348 /* destroy the listening endpoint */
349 ret = cm_id->device->iwcm->destroy_listen(cm_id);
350 spin_lock_irqsave(&cm_id_priv->lock, flags);
351 break;
352 case IW_CM_STATE_ESTABLISHED:
353 cm_id_priv->state = IW_CM_STATE_DESTROYING;
354 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
355 /* Abrupt close of the connection */
356 (void)iwcm_modify_qp_err(cm_id_priv->qp);
357 spin_lock_irqsave(&cm_id_priv->lock, flags);
358 break;
359 case IW_CM_STATE_IDLE:
360 case IW_CM_STATE_CLOSING:
361 cm_id_priv->state = IW_CM_STATE_DESTROYING;
362 break;
363 case IW_CM_STATE_CONN_RECV:
364 /*
365 * App called destroy before/without calling accept after
366 * receiving connection request event notification or
367 * returned non zero from the event callback function.
368 * In either case, must tell the provider to reject.
369 */
370 cm_id_priv->state = IW_CM_STATE_DESTROYING;
371 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
372 cm_id->device->iwcm->reject(cm_id, NULL, 0);
373 spin_lock_irqsave(&cm_id_priv->lock, flags);
374 break;
375 case IW_CM_STATE_CONN_SENT:
376 case IW_CM_STATE_DESTROYING:
377 default:
378 BUG();
379 break;
380 }
381 if (cm_id_priv->qp) {
382 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
383 cm_id_priv->qp = NULL;
384 }
385 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
386
387 (void)iwcm_deref_id(cm_id_priv);
388 }
389
390 /*
391 * This function is only called by the application thread and cannot
392 * be called by the event thread. The function will wait for all
393 * references to be released on the cm_id and then kfree the cm_id
394 * object.
395 */
iw_destroy_cm_id(struct iw_cm_id * cm_id)396 void iw_destroy_cm_id(struct iw_cm_id *cm_id)
397 {
398 struct iwcm_id_private *cm_id_priv;
399
400 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
401 BUG_ON(test_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags));
402
403 destroy_cm_id(cm_id);
404
405 wait_for_completion(&cm_id_priv->destroy_comp);
406
407 free_cm_id(cm_id_priv);
408 }
409 EXPORT_SYMBOL(iw_destroy_cm_id);
410
411 /*
412 * CM_ID <-- LISTEN
413 *
414 * Start listening for connect requests. Generates one CONNECT_REQUEST
415 * event for each inbound connect request.
416 */
iw_cm_listen(struct iw_cm_id * cm_id,int backlog)417 int iw_cm_listen(struct iw_cm_id *cm_id, int backlog)
418 {
419 struct iwcm_id_private *cm_id_priv;
420 unsigned long flags;
421 int ret;
422
423 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
424
425 ret = alloc_work_entries(cm_id_priv, backlog);
426 if (ret)
427 return ret;
428
429 spin_lock_irqsave(&cm_id_priv->lock, flags);
430 switch (cm_id_priv->state) {
431 case IW_CM_STATE_IDLE:
432 cm_id_priv->state = IW_CM_STATE_LISTEN;
433 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
434 ret = cm_id->device->iwcm->create_listen(cm_id, backlog);
435 if (ret)
436 cm_id_priv->state = IW_CM_STATE_IDLE;
437 spin_lock_irqsave(&cm_id_priv->lock, flags);
438 break;
439 default:
440 ret = -EINVAL;
441 }
442 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
443
444 return ret;
445 }
446 EXPORT_SYMBOL(iw_cm_listen);
447
448 /*
449 * CM_ID <-- IDLE
450 *
451 * Rejects an inbound connection request. No events are generated.
452 */
iw_cm_reject(struct iw_cm_id * cm_id,const void * private_data,u8 private_data_len)453 int iw_cm_reject(struct iw_cm_id *cm_id,
454 const void *private_data,
455 u8 private_data_len)
456 {
457 struct iwcm_id_private *cm_id_priv;
458 unsigned long flags;
459 int ret;
460
461 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
462 set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
463
464 spin_lock_irqsave(&cm_id_priv->lock, flags);
465 if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
466 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
467 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
468 wake_up_all(&cm_id_priv->connect_wait);
469 return -EINVAL;
470 }
471 cm_id_priv->state = IW_CM_STATE_IDLE;
472 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
473
474 ret = cm_id->device->iwcm->reject(cm_id, private_data,
475 private_data_len);
476
477 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
478 wake_up_all(&cm_id_priv->connect_wait);
479
480 return ret;
481 }
482 EXPORT_SYMBOL(iw_cm_reject);
483
484 /*
485 * CM_ID <-- ESTABLISHED
486 *
487 * Accepts an inbound connection request and generates an ESTABLISHED
488 * event. Callers of iw_cm_disconnect and iw_destroy_cm_id will block
489 * until the ESTABLISHED event is received from the provider.
490 */
iw_cm_accept(struct iw_cm_id * cm_id,struct iw_cm_conn_param * iw_param)491 int iw_cm_accept(struct iw_cm_id *cm_id,
492 struct iw_cm_conn_param *iw_param)
493 {
494 struct iwcm_id_private *cm_id_priv;
495 struct ib_qp *qp;
496 unsigned long flags;
497 int ret;
498
499 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
500 set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
501
502 spin_lock_irqsave(&cm_id_priv->lock, flags);
503 if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
504 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
505 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
506 wake_up_all(&cm_id_priv->connect_wait);
507 return -EINVAL;
508 }
509 /* Get the ib_qp given the QPN */
510 qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
511 if (!qp) {
512 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
513 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
514 wake_up_all(&cm_id_priv->connect_wait);
515 return -EINVAL;
516 }
517 cm_id->device->iwcm->add_ref(qp);
518 cm_id_priv->qp = qp;
519 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
520
521 ret = cm_id->device->iwcm->accept(cm_id, iw_param);
522 if (ret) {
523 /* An error on accept precludes provider events */
524 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
525 cm_id_priv->state = IW_CM_STATE_IDLE;
526 spin_lock_irqsave(&cm_id_priv->lock, flags);
527 if (cm_id_priv->qp) {
528 cm_id->device->iwcm->rem_ref(qp);
529 cm_id_priv->qp = NULL;
530 }
531 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
532 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
533 wake_up_all(&cm_id_priv->connect_wait);
534 }
535
536 return ret;
537 }
538 EXPORT_SYMBOL(iw_cm_accept);
539
540 /*
541 * Active Side: CM_ID <-- CONN_SENT
542 *
543 * If successful, results in the generation of a CONNECT_REPLY
544 * event. iw_cm_disconnect and iw_cm_destroy will block until the
545 * CONNECT_REPLY event is received from the provider.
546 */
iw_cm_connect(struct iw_cm_id * cm_id,struct iw_cm_conn_param * iw_param)547 int iw_cm_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *iw_param)
548 {
549 struct iwcm_id_private *cm_id_priv;
550 int ret;
551 unsigned long flags;
552 struct ib_qp *qp;
553
554 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
555
556 ret = alloc_work_entries(cm_id_priv, 4);
557 if (ret)
558 return ret;
559
560 set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
561 spin_lock_irqsave(&cm_id_priv->lock, flags);
562
563 if (cm_id_priv->state != IW_CM_STATE_IDLE) {
564 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
565 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
566 wake_up_all(&cm_id_priv->connect_wait);
567 return -EINVAL;
568 }
569
570 /* Get the ib_qp given the QPN */
571 qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
572 if (!qp) {
573 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
574 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
575 wake_up_all(&cm_id_priv->connect_wait);
576 return -EINVAL;
577 }
578 cm_id->device->iwcm->add_ref(qp);
579 cm_id_priv->qp = qp;
580 cm_id_priv->state = IW_CM_STATE_CONN_SENT;
581 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
582
583 ret = cm_id->device->iwcm->connect(cm_id, iw_param);
584 if (ret) {
585 spin_lock_irqsave(&cm_id_priv->lock, flags);
586 if (cm_id_priv->qp) {
587 cm_id->device->iwcm->rem_ref(qp);
588 cm_id_priv->qp = NULL;
589 }
590 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
591 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
592 cm_id_priv->state = IW_CM_STATE_IDLE;
593 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
594 wake_up_all(&cm_id_priv->connect_wait);
595 }
596
597 return ret;
598 }
599 EXPORT_SYMBOL(iw_cm_connect);
600
601 /*
602 * Passive Side: new CM_ID <-- CONN_RECV
603 *
604 * Handles an inbound connect request. The function creates a new
605 * iw_cm_id to represent the new connection and inherits the client
606 * callback function and other attributes from the listening parent.
607 *
608 * The work item contains a pointer to the listen_cm_id and the event. The
609 * listen_cm_id contains the client cm_handler, context and
610 * device. These are copied when the device is cloned. The event
611 * contains the new four tuple.
612 *
613 * An error on the child should not affect the parent, so this
614 * function does not return a value.
615 */
cm_conn_req_handler(struct iwcm_id_private * listen_id_priv,struct iw_cm_event * iw_event)616 static void cm_conn_req_handler(struct iwcm_id_private *listen_id_priv,
617 struct iw_cm_event *iw_event)
618 {
619 unsigned long flags;
620 struct iw_cm_id *cm_id;
621 struct iwcm_id_private *cm_id_priv;
622 int ret;
623
624 /*
625 * The provider should never generate a connection request
626 * event with a bad status.
627 */
628 BUG_ON(iw_event->status);
629
630 cm_id = iw_create_cm_id(listen_id_priv->id.device,
631 iw_event->so,
632 listen_id_priv->id.cm_handler,
633 listen_id_priv->id.context);
634 /* If the cm_id could not be created, ignore the request */
635 if (IS_ERR(cm_id))
636 goto out;
637
638 cm_id->provider_data = iw_event->provider_data;
639 cm_id->local_addr = iw_event->local_addr;
640 cm_id->remote_addr = iw_event->remote_addr;
641
642 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
643 cm_id_priv->state = IW_CM_STATE_CONN_RECV;
644
645 /*
646 * We could be destroying the listening id. If so, ignore this
647 * upcall.
648 */
649 spin_lock_irqsave(&listen_id_priv->lock, flags);
650 if (listen_id_priv->state != IW_CM_STATE_LISTEN) {
651 spin_unlock_irqrestore(&listen_id_priv->lock, flags);
652 iw_cm_reject(cm_id, NULL, 0);
653 iw_destroy_cm_id(cm_id);
654 goto out;
655 }
656 spin_unlock_irqrestore(&listen_id_priv->lock, flags);
657
658 ret = alloc_work_entries(cm_id_priv, 3);
659 if (ret) {
660 iw_cm_reject(cm_id, NULL, 0);
661 iw_destroy_cm_id(cm_id);
662 goto out;
663 }
664
665 /* Call the client CM handler */
666 ret = cm_id->cm_handler(cm_id, iw_event);
667 if (ret) {
668 iw_cm_reject(cm_id, NULL, 0);
669 set_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
670 destroy_cm_id(cm_id);
671 if (atomic_read(&cm_id_priv->refcount)==0)
672 free_cm_id(cm_id_priv);
673 }
674
675 out:
676 if (iw_event->private_data_len)
677 kfree(iw_event->private_data);
678 }
679
680 /*
681 * Passive Side: CM_ID <-- ESTABLISHED
682 *
683 * The provider generated an ESTABLISHED event which means that
684 * the MPA negotion has completed successfully and we are now in MPA
685 * FPDU mode.
686 *
687 * This event can only be received in the CONN_RECV state. If the
688 * remote peer closed, the ESTABLISHED event would be received followed
689 * by the CLOSE event. If the app closes, it will block until we wake
690 * it up after processing this event.
691 */
cm_conn_est_handler(struct iwcm_id_private * cm_id_priv,struct iw_cm_event * iw_event)692 static int cm_conn_est_handler(struct iwcm_id_private *cm_id_priv,
693 struct iw_cm_event *iw_event)
694 {
695 unsigned long flags;
696 int ret;
697
698 spin_lock_irqsave(&cm_id_priv->lock, flags);
699
700 /*
701 * We clear the CONNECT_WAIT bit here to allow the callback
702 * function to call iw_cm_disconnect. Calling iw_destroy_cm_id
703 * from a callback handler is not allowed.
704 */
705 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
706 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
707 cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
708 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
709 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
710 wake_up_all(&cm_id_priv->connect_wait);
711
712 return ret;
713 }
714
715 /*
716 * Active Side: CM_ID <-- ESTABLISHED
717 *
718 * The app has called connect and is waiting for the established event to
719 * post it's requests to the server. This event will wake up anyone
720 * blocked in iw_cm_disconnect or iw_destroy_id.
721 */
cm_conn_rep_handler(struct iwcm_id_private * cm_id_priv,struct iw_cm_event * iw_event)722 static int cm_conn_rep_handler(struct iwcm_id_private *cm_id_priv,
723 struct iw_cm_event *iw_event)
724 {
725 unsigned long flags;
726 int ret;
727
728 spin_lock_irqsave(&cm_id_priv->lock, flags);
729 /*
730 * Clear the connect wait bit so a callback function calling
731 * iw_cm_disconnect will not wait and deadlock this thread
732 */
733 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
734 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
735 if (iw_event->status == 0) {
736 cm_id_priv->id.local_addr = iw_event->local_addr;
737 cm_id_priv->id.remote_addr = iw_event->remote_addr;
738 cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
739 } else {
740 /* REJECTED or RESET */
741 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
742 cm_id_priv->qp = NULL;
743 cm_id_priv->state = IW_CM_STATE_IDLE;
744 }
745 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
746 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
747
748 if (iw_event->private_data_len)
749 kfree(iw_event->private_data);
750
751 /* Wake up waiters on connect complete */
752 wake_up_all(&cm_id_priv->connect_wait);
753
754 return ret;
755 }
756
757 /*
758 * CM_ID <-- CLOSING
759 *
760 * If in the ESTABLISHED state, move to CLOSING.
761 */
cm_disconnect_handler(struct iwcm_id_private * cm_id_priv,struct iw_cm_event * iw_event)762 static void cm_disconnect_handler(struct iwcm_id_private *cm_id_priv,
763 struct iw_cm_event *iw_event)
764 {
765 unsigned long flags;
766
767 spin_lock_irqsave(&cm_id_priv->lock, flags);
768 if (cm_id_priv->state == IW_CM_STATE_ESTABLISHED)
769 cm_id_priv->state = IW_CM_STATE_CLOSING;
770 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
771 }
772
773 /*
774 * CM_ID <-- IDLE
775 *
776 * If in the ESTBLISHED or CLOSING states, the QP will have have been
777 * moved by the provider to the ERR state. Disassociate the CM_ID from
778 * the QP, move to IDLE, and remove the 'connected' reference.
779 *
780 * If in some other state, the cm_id was destroyed asynchronously.
781 * This is the last reference that will result in waking up
782 * the app thread blocked in iw_destroy_cm_id.
783 */
cm_close_handler(struct iwcm_id_private * cm_id_priv,struct iw_cm_event * iw_event)784 static int cm_close_handler(struct iwcm_id_private *cm_id_priv,
785 struct iw_cm_event *iw_event)
786 {
787 unsigned long flags;
788 int ret = 0;
789 spin_lock_irqsave(&cm_id_priv->lock, flags);
790
791 if (cm_id_priv->qp) {
792 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
793 cm_id_priv->qp = NULL;
794 }
795 switch (cm_id_priv->state) {
796 case IW_CM_STATE_ESTABLISHED:
797 case IW_CM_STATE_CLOSING:
798 cm_id_priv->state = IW_CM_STATE_IDLE;
799 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
800 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
801 spin_lock_irqsave(&cm_id_priv->lock, flags);
802 break;
803 case IW_CM_STATE_DESTROYING:
804 break;
805 default:
806 BUG();
807 }
808 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
809
810 return ret;
811 }
812
process_event(struct iwcm_id_private * cm_id_priv,struct iw_cm_event * iw_event)813 static int process_event(struct iwcm_id_private *cm_id_priv,
814 struct iw_cm_event *iw_event)
815 {
816 int ret = 0;
817
818 switch (iw_event->event) {
819 case IW_CM_EVENT_CONNECT_REQUEST:
820 cm_conn_req_handler(cm_id_priv, iw_event);
821 break;
822 case IW_CM_EVENT_CONNECT_REPLY:
823 ret = cm_conn_rep_handler(cm_id_priv, iw_event);
824 break;
825 case IW_CM_EVENT_ESTABLISHED:
826 ret = cm_conn_est_handler(cm_id_priv, iw_event);
827 break;
828 case IW_CM_EVENT_DISCONNECT:
829 cm_disconnect_handler(cm_id_priv, iw_event);
830 break;
831 case IW_CM_EVENT_CLOSE:
832 ret = cm_close_handler(cm_id_priv, iw_event);
833 break;
834 default:
835 BUG();
836 }
837
838 return ret;
839 }
840
841 /*
842 * Process events on the work_list for the cm_id. If the callback
843 * function requests that the cm_id be deleted, a flag is set in the
844 * cm_id flags to indicate that when the last reference is
845 * removed, the cm_id is to be destroyed. This is necessary to
846 * distinguish between an object that will be destroyed by the app
847 * thread asleep on the destroy_comp list vs. an object destroyed
848 * here synchronously when the last reference is removed.
849 */
cm_work_handler(struct work_struct * _work)850 static void cm_work_handler(struct work_struct *_work)
851 {
852 struct iwcm_work *work = container_of(_work, struct iwcm_work, work);
853 struct iw_cm_event levent;
854 struct iwcm_id_private *cm_id_priv = work->cm_id;
855 unsigned long flags;
856 int empty;
857 int ret = 0;
858 int destroy_id;
859
860 spin_lock_irqsave(&cm_id_priv->lock, flags);
861 empty = list_empty(&cm_id_priv->work_list);
862 while (!empty) {
863 work = list_entry(cm_id_priv->work_list.next,
864 struct iwcm_work, list);
865 list_del_init(&work->list);
866 empty = list_empty(&cm_id_priv->work_list);
867 levent = work->event;
868 put_work(work);
869 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
870
871 ret = process_event(cm_id_priv, &levent);
872 if (ret) {
873 set_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
874 destroy_cm_id(&cm_id_priv->id);
875 }
876 BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
877 destroy_id = test_bit(IWCM_F_CALLBACK_DESTROY, &cm_id_priv->flags);
878 if (iwcm_deref_id(cm_id_priv)) {
879 if (destroy_id) {
880 BUG_ON(!list_empty(&cm_id_priv->work_list));
881 free_cm_id(cm_id_priv);
882 }
883 return;
884 }
885 spin_lock_irqsave(&cm_id_priv->lock, flags);
886 }
887 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
888 }
889
890 /*
891 * This function is called on interrupt context. Schedule events on
892 * the iwcm_wq thread to allow callback functions to downcall into
893 * the CM and/or block. Events are queued to a per-CM_ID
894 * work_list. If this is the first event on the work_list, the work
895 * element is also queued on the iwcm_wq thread.
896 *
897 * Each event holds a reference on the cm_id. Until the last posted
898 * event has been delivered and processed, the cm_id cannot be
899 * deleted.
900 *
901 * Returns:
902 * 0 - the event was handled.
903 * -ENOMEM - the event was not handled due to lack of resources.
904 */
cm_event_handler(struct iw_cm_id * cm_id,struct iw_cm_event * iw_event)905 static int cm_event_handler(struct iw_cm_id *cm_id,
906 struct iw_cm_event *iw_event)
907 {
908 struct iwcm_work *work;
909 struct iwcm_id_private *cm_id_priv;
910 unsigned long flags;
911 int ret = 0;
912
913 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
914
915 spin_lock_irqsave(&cm_id_priv->lock, flags);
916 work = get_work(cm_id_priv);
917 if (!work) {
918 ret = -ENOMEM;
919 goto out;
920 }
921
922 INIT_WORK(&work->work, cm_work_handler);
923 work->cm_id = cm_id_priv;
924 work->event = *iw_event;
925
926 if ((work->event.event == IW_CM_EVENT_CONNECT_REQUEST ||
927 work->event.event == IW_CM_EVENT_CONNECT_REPLY) &&
928 work->event.private_data_len) {
929 ret = copy_private_data(&work->event);
930 if (ret) {
931 put_work(work);
932 goto out;
933 }
934 }
935
936 atomic_inc(&cm_id_priv->refcount);
937 if (list_empty(&cm_id_priv->work_list)) {
938 list_add_tail(&work->list, &cm_id_priv->work_list);
939 queue_work(iwcm_wq, &work->work);
940 } else
941 list_add_tail(&work->list, &cm_id_priv->work_list);
942 out:
943 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
944 return ret;
945 }
946
iwcm_init_qp_init_attr(struct iwcm_id_private * cm_id_priv,struct ib_qp_attr * qp_attr,int * qp_attr_mask)947 static int iwcm_init_qp_init_attr(struct iwcm_id_private *cm_id_priv,
948 struct ib_qp_attr *qp_attr,
949 int *qp_attr_mask)
950 {
951 unsigned long flags;
952 int ret;
953
954 spin_lock_irqsave(&cm_id_priv->lock, flags);
955 switch (cm_id_priv->state) {
956 case IW_CM_STATE_IDLE:
957 case IW_CM_STATE_CONN_SENT:
958 case IW_CM_STATE_CONN_RECV:
959 case IW_CM_STATE_ESTABLISHED:
960 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
961 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE|
962 IB_ACCESS_REMOTE_READ;
963 ret = 0;
964 break;
965 default:
966 ret = -EINVAL;
967 break;
968 }
969 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
970 return ret;
971 }
972
iwcm_init_qp_rts_attr(struct iwcm_id_private * cm_id_priv,struct ib_qp_attr * qp_attr,int * qp_attr_mask)973 static int iwcm_init_qp_rts_attr(struct iwcm_id_private *cm_id_priv,
974 struct ib_qp_attr *qp_attr,
975 int *qp_attr_mask)
976 {
977 unsigned long flags;
978 int ret;
979
980 spin_lock_irqsave(&cm_id_priv->lock, flags);
981 switch (cm_id_priv->state) {
982 case IW_CM_STATE_IDLE:
983 case IW_CM_STATE_CONN_SENT:
984 case IW_CM_STATE_CONN_RECV:
985 case IW_CM_STATE_ESTABLISHED:
986 *qp_attr_mask = 0;
987 ret = 0;
988 break;
989 default:
990 ret = -EINVAL;
991 break;
992 }
993 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
994 return ret;
995 }
996
iw_cm_init_qp_attr(struct iw_cm_id * cm_id,struct ib_qp_attr * qp_attr,int * qp_attr_mask)997 int iw_cm_init_qp_attr(struct iw_cm_id *cm_id,
998 struct ib_qp_attr *qp_attr,
999 int *qp_attr_mask)
1000 {
1001 struct iwcm_id_private *cm_id_priv;
1002 int ret;
1003
1004 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1005 switch (qp_attr->qp_state) {
1006 case IB_QPS_INIT:
1007 case IB_QPS_RTR:
1008 ret = iwcm_init_qp_init_attr(cm_id_priv,
1009 qp_attr, qp_attr_mask);
1010 break;
1011 case IB_QPS_RTS:
1012 ret = iwcm_init_qp_rts_attr(cm_id_priv,
1013 qp_attr, qp_attr_mask);
1014 break;
1015 default:
1016 ret = -EINVAL;
1017 break;
1018 }
1019 return ret;
1020 }
1021 EXPORT_SYMBOL(iw_cm_init_qp_attr);
1022
iw_cm_init(void)1023 static int __init iw_cm_init(void)
1024 {
1025 iwcm_wq = create_singlethread_workqueue("iw_cm_wq");
1026 if (!iwcm_wq)
1027 return -ENOMEM;
1028
1029 return 0;
1030 }
1031
iw_cm_cleanup(void)1032 static void __exit iw_cm_cleanup(void)
1033 {
1034 destroy_workqueue(iwcm_wq);
1035 }
1036
1037 module_init(iw_cm_init);
1038 module_exit(iw_cm_cleanup);
1039