Merge remote-tracking branches 'asoc/topic/cs47l24', 'asoc/topic/cx20442', 'asoc...
[sfrench/cifs-2.6.git] / drivers / infiniband / core / iwcm.c
1 /*
2  * Copyright (c) 2004, 2005 Intel Corporation.  All rights reserved.
3  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
4  * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
5  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
6  * Copyright (c) 2005 Open Grid Computing, Inc. All rights reserved.
7  * Copyright (c) 2005 Network Appliance, Inc. All rights reserved.
8  *
9  * This software is available to you under a choice of one of two
10  * licenses.  You may choose to be licensed under the terms of the GNU
11  * General Public License (GPL) Version 2, available from the file
12  * COPYING in the main directory of this source tree, or the
13  * OpenIB.org BSD license below:
14  *
15  *     Redistribution and use in source and binary forms, with or
16  *     without modification, are permitted provided that the following
17  *     conditions are met:
18  *
19  *      - Redistributions of source code must retain the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer.
22  *
23  *      - Redistributions in binary form must reproduce the above
24  *        copyright notice, this list of conditions and the following
25  *        disclaimer in the documentation and/or other materials
26  *        provided with the distribution.
27  *
28  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
29  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
30  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
31  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
32  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
33  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
34  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
35  * SOFTWARE.
36  *
37  */
38 #include <linux/dma-mapping.h>
39 #include <linux/err.h>
40 #include <linux/idr.h>
41 #include <linux/interrupt.h>
42 #include <linux/rbtree.h>
43 #include <linux/sched.h>
44 #include <linux/spinlock.h>
45 #include <linux/workqueue.h>
46 #include <linux/completion.h>
47 #include <linux/slab.h>
48 #include <linux/module.h>
49 #include <linux/sysctl.h>
50
51 #include <rdma/iw_cm.h>
52 #include <rdma/ib_addr.h>
53 #include <rdma/iw_portmap.h>
54 #include <rdma/rdma_netlink.h>
55
56 #include "iwcm.h"
57
58 MODULE_AUTHOR("Tom Tucker");
59 MODULE_DESCRIPTION("iWARP CM");
60 MODULE_LICENSE("Dual BSD/GPL");
61
62 static const char * const iwcm_rej_reason_strs[] = {
63         [ECONNRESET]                    = "reset by remote host",
64         [ECONNREFUSED]                  = "refused by remote application",
65         [ETIMEDOUT]                     = "setup timeout",
66 };
67
68 const char *__attribute_const__ iwcm_reject_msg(int reason)
69 {
70         size_t index;
71
72         /* iWARP uses negative errnos */
73         index = -reason;
74
75         if (index < ARRAY_SIZE(iwcm_rej_reason_strs) &&
76             iwcm_rej_reason_strs[index])
77                 return iwcm_rej_reason_strs[index];
78         else
79                 return "unrecognized reason";
80 }
81 EXPORT_SYMBOL(iwcm_reject_msg);
82
83 static struct rdma_nl_cbs iwcm_nl_cb_table[RDMA_NL_IWPM_NUM_OPS] = {
84         [RDMA_NL_IWPM_REG_PID] = {.dump = iwpm_register_pid_cb},
85         [RDMA_NL_IWPM_ADD_MAPPING] = {.dump = iwpm_add_mapping_cb},
86         [RDMA_NL_IWPM_QUERY_MAPPING] = {.dump = iwpm_add_and_query_mapping_cb},
87         [RDMA_NL_IWPM_REMOTE_INFO] = {.dump = iwpm_remote_info_cb},
88         [RDMA_NL_IWPM_HANDLE_ERR] = {.dump = iwpm_mapping_error_cb},
89         [RDMA_NL_IWPM_MAPINFO] = {.dump = iwpm_mapping_info_cb},
90         [RDMA_NL_IWPM_MAPINFO_NUM] = {.dump = iwpm_ack_mapping_info_cb}
91 };
92
93 static struct workqueue_struct *iwcm_wq;
94 struct iwcm_work {
95         struct work_struct work;
96         struct iwcm_id_private *cm_id;
97         struct list_head list;
98         struct iw_cm_event event;
99         struct list_head free_list;
100 };
101
102 static unsigned int default_backlog = 256;
103
104 static struct ctl_table_header *iwcm_ctl_table_hdr;
105 static struct ctl_table iwcm_ctl_table[] = {
106         {
107                 .procname       = "default_backlog",
108                 .data           = &default_backlog,
109                 .maxlen         = sizeof(default_backlog),
110                 .mode           = 0644,
111                 .proc_handler   = proc_dointvec,
112         },
113         { }
114 };
115
116 /*
117  * The following services provide a mechanism for pre-allocating iwcm_work
118  * elements.  The design pre-allocates them  based on the cm_id type:
119  *      LISTENING IDS:  Get enough elements preallocated to handle the
120  *                      listen backlog.
121  *      ACTIVE IDS:     4: CONNECT_REPLY, ESTABLISHED, DISCONNECT, CLOSE
122  *      PASSIVE IDS:    3: ESTABLISHED, DISCONNECT, CLOSE
123  *
124  * Allocating them in connect and listen avoids having to deal
125  * with allocation failures on the event upcall from the provider (which
126  * is called in the interrupt context).
127  *
128  * One exception is when creating the cm_id for incoming connection requests.
129  * There are two cases:
130  * 1) in the event upcall, cm_event_handler(), for a listening cm_id.  If
131  *    the backlog is exceeded, then no more connection request events will
132  *    be processed.  cm_event_handler() returns -ENOMEM in this case.  Its up
133  *    to the provider to reject the connection request.
134  * 2) in the connection request workqueue handler, cm_conn_req_handler().
135  *    If work elements cannot be allocated for the new connect request cm_id,
136  *    then IWCM will call the provider reject method.  This is ok since
137  *    cm_conn_req_handler() runs in the workqueue thread context.
138  */
139
140 static struct iwcm_work *get_work(struct iwcm_id_private *cm_id_priv)
141 {
142         struct iwcm_work *work;
143
144         if (list_empty(&cm_id_priv->work_free_list))
145                 return NULL;
146         work = list_entry(cm_id_priv->work_free_list.next, struct iwcm_work,
147                           free_list);
148         list_del_init(&work->free_list);
149         return work;
150 }
151
152 static void put_work(struct iwcm_work *work)
153 {
154         list_add(&work->free_list, &work->cm_id->work_free_list);
155 }
156
157 static void dealloc_work_entries(struct iwcm_id_private *cm_id_priv)
158 {
159         struct list_head *e, *tmp;
160
161         list_for_each_safe(e, tmp, &cm_id_priv->work_free_list)
162                 kfree(list_entry(e, struct iwcm_work, free_list));
163 }
164
165 static int alloc_work_entries(struct iwcm_id_private *cm_id_priv, int count)
166 {
167         struct iwcm_work *work;
168
169         BUG_ON(!list_empty(&cm_id_priv->work_free_list));
170         while (count--) {
171                 work = kmalloc(sizeof(struct iwcm_work), GFP_KERNEL);
172                 if (!work) {
173                         dealloc_work_entries(cm_id_priv);
174                         return -ENOMEM;
175                 }
176                 work->cm_id = cm_id_priv;
177                 INIT_LIST_HEAD(&work->list);
178                 put_work(work);
179         }
180         return 0;
181 }
182
183 /*
184  * Save private data from incoming connection requests to
185  * iw_cm_event, so the low level driver doesn't have to. Adjust
186  * the event ptr to point to the local copy.
187  */
188 static int copy_private_data(struct iw_cm_event *event)
189 {
190         void *p;
191
192         p = kmemdup(event->private_data, event->private_data_len, GFP_ATOMIC);
193         if (!p)
194                 return -ENOMEM;
195         event->private_data = p;
196         return 0;
197 }
198
199 static void free_cm_id(struct iwcm_id_private *cm_id_priv)
200 {
201         dealloc_work_entries(cm_id_priv);
202         kfree(cm_id_priv);
203 }
204
205 /*
206  * Release a reference on cm_id. If the last reference is being
207  * released, free the cm_id and return 1.
208  */
209 static int iwcm_deref_id(struct iwcm_id_private *cm_id_priv)
210 {
211         BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
212         if (atomic_dec_and_test(&cm_id_priv->refcount)) {
213                 BUG_ON(!list_empty(&cm_id_priv->work_list));
214                 free_cm_id(cm_id_priv);
215                 return 1;
216         }
217
218         return 0;
219 }
220
221 static void add_ref(struct iw_cm_id *cm_id)
222 {
223         struct iwcm_id_private *cm_id_priv;
224         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
225         atomic_inc(&cm_id_priv->refcount);
226 }
227
228 static void rem_ref(struct iw_cm_id *cm_id)
229 {
230         struct iwcm_id_private *cm_id_priv;
231
232         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
233
234         (void)iwcm_deref_id(cm_id_priv);
235 }
236
237 static int cm_event_handler(struct iw_cm_id *cm_id, struct iw_cm_event *event);
238
239 struct iw_cm_id *iw_create_cm_id(struct ib_device *device,
240                                  iw_cm_handler cm_handler,
241                                  void *context)
242 {
243         struct iwcm_id_private *cm_id_priv;
244
245         cm_id_priv = kzalloc(sizeof(*cm_id_priv), GFP_KERNEL);
246         if (!cm_id_priv)
247                 return ERR_PTR(-ENOMEM);
248
249         cm_id_priv->state = IW_CM_STATE_IDLE;
250         cm_id_priv->id.device = device;
251         cm_id_priv->id.cm_handler = cm_handler;
252         cm_id_priv->id.context = context;
253         cm_id_priv->id.event_handler = cm_event_handler;
254         cm_id_priv->id.add_ref = add_ref;
255         cm_id_priv->id.rem_ref = rem_ref;
256         spin_lock_init(&cm_id_priv->lock);
257         atomic_set(&cm_id_priv->refcount, 1);
258         init_waitqueue_head(&cm_id_priv->connect_wait);
259         init_completion(&cm_id_priv->destroy_comp);
260         INIT_LIST_HEAD(&cm_id_priv->work_list);
261         INIT_LIST_HEAD(&cm_id_priv->work_free_list);
262
263         return &cm_id_priv->id;
264 }
265 EXPORT_SYMBOL(iw_create_cm_id);
266
267
268 static int iwcm_modify_qp_err(struct ib_qp *qp)
269 {
270         struct ib_qp_attr qp_attr;
271
272         if (!qp)
273                 return -EINVAL;
274
275         qp_attr.qp_state = IB_QPS_ERR;
276         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
277 }
278
279 /*
280  * This is really the RDMAC CLOSING state. It is most similar to the
281  * IB SQD QP state.
282  */
283 static int iwcm_modify_qp_sqd(struct ib_qp *qp)
284 {
285         struct ib_qp_attr qp_attr;
286
287         BUG_ON(qp == NULL);
288         qp_attr.qp_state = IB_QPS_SQD;
289         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
290 }
291
292 /*
293  * CM_ID <-- CLOSING
294  *
295  * Block if a passive or active connection is currently being processed. Then
296  * process the event as follows:
297  * - If we are ESTABLISHED, move to CLOSING and modify the QP state
298  *   based on the abrupt flag
299  * - If the connection is already in the CLOSING or IDLE state, the peer is
300  *   disconnecting concurrently with us and we've already seen the
301  *   DISCONNECT event -- ignore the request and return 0
302  * - Disconnect on a listening endpoint returns -EINVAL
303  */
304 int iw_cm_disconnect(struct iw_cm_id *cm_id, int abrupt)
305 {
306         struct iwcm_id_private *cm_id_priv;
307         unsigned long flags;
308         int ret = 0;
309         struct ib_qp *qp = NULL;
310
311         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
312         /* Wait if we're currently in a connect or accept downcall */
313         wait_event(cm_id_priv->connect_wait,
314                    !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
315
316         spin_lock_irqsave(&cm_id_priv->lock, flags);
317         switch (cm_id_priv->state) {
318         case IW_CM_STATE_ESTABLISHED:
319                 cm_id_priv->state = IW_CM_STATE_CLOSING;
320
321                 /* QP could be <nul> for user-mode client */
322                 if (cm_id_priv->qp)
323                         qp = cm_id_priv->qp;
324                 else
325                         ret = -EINVAL;
326                 break;
327         case IW_CM_STATE_LISTEN:
328                 ret = -EINVAL;
329                 break;
330         case IW_CM_STATE_CLOSING:
331                 /* remote peer closed first */
332         case IW_CM_STATE_IDLE:
333                 /* accept or connect returned !0 */
334                 break;
335         case IW_CM_STATE_CONN_RECV:
336                 /*
337                  * App called disconnect before/without calling accept after
338                  * connect_request event delivered.
339                  */
340                 break;
341         case IW_CM_STATE_CONN_SENT:
342                 /* Can only get here if wait above fails */
343         default:
344                 BUG();
345         }
346         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
347
348         if (qp) {
349                 if (abrupt)
350                         ret = iwcm_modify_qp_err(qp);
351                 else
352                         ret = iwcm_modify_qp_sqd(qp);
353
354                 /*
355                  * If both sides are disconnecting the QP could
356                  * already be in ERR or SQD states
357                  */
358                 ret = 0;
359         }
360
361         return ret;
362 }
363 EXPORT_SYMBOL(iw_cm_disconnect);
364
365 /*
366  * CM_ID <-- DESTROYING
367  *
368  * Clean up all resources associated with the connection and release
369  * the initial reference taken by iw_create_cm_id.
370  */
371 static void destroy_cm_id(struct iw_cm_id *cm_id)
372 {
373         struct iwcm_id_private *cm_id_priv;
374         unsigned long flags;
375
376         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
377         /*
378          * Wait if we're currently in a connect or accept downcall. A
379          * listening endpoint should never block here.
380          */
381         wait_event(cm_id_priv->connect_wait,
382                    !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
383
384         /*
385          * Since we're deleting the cm_id, drop any events that
386          * might arrive before the last dereference.
387          */
388         set_bit(IWCM_F_DROP_EVENTS, &cm_id_priv->flags);
389
390         spin_lock_irqsave(&cm_id_priv->lock, flags);
391         switch (cm_id_priv->state) {
392         case IW_CM_STATE_LISTEN:
393                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
394                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
395                 /* destroy the listening endpoint */
396                 cm_id->device->iwcm->destroy_listen(cm_id);
397                 spin_lock_irqsave(&cm_id_priv->lock, flags);
398                 break;
399         case IW_CM_STATE_ESTABLISHED:
400                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
401                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
402                 /* Abrupt close of the connection */
403                 (void)iwcm_modify_qp_err(cm_id_priv->qp);
404                 spin_lock_irqsave(&cm_id_priv->lock, flags);
405                 break;
406         case IW_CM_STATE_IDLE:
407         case IW_CM_STATE_CLOSING:
408                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
409                 break;
410         case IW_CM_STATE_CONN_RECV:
411                 /*
412                  * App called destroy before/without calling accept after
413                  * receiving connection request event notification or
414                  * returned non zero from the event callback function.
415                  * In either case, must tell the provider to reject.
416                  */
417                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
418                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
419                 cm_id->device->iwcm->reject(cm_id, NULL, 0);
420                 spin_lock_irqsave(&cm_id_priv->lock, flags);
421                 break;
422         case IW_CM_STATE_CONN_SENT:
423         case IW_CM_STATE_DESTROYING:
424         default:
425                 BUG();
426                 break;
427         }
428         if (cm_id_priv->qp) {
429                 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
430                 cm_id_priv->qp = NULL;
431         }
432         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
433
434         if (cm_id->mapped) {
435                 iwpm_remove_mapinfo(&cm_id->local_addr, &cm_id->m_local_addr);
436                 iwpm_remove_mapping(&cm_id->local_addr, RDMA_NL_IWCM);
437         }
438
439         (void)iwcm_deref_id(cm_id_priv);
440 }
441
442 /*
443  * This function is only called by the application thread and cannot
444  * be called by the event thread. The function will wait for all
445  * references to be released on the cm_id and then kfree the cm_id
446  * object.
447  */
448 void iw_destroy_cm_id(struct iw_cm_id *cm_id)
449 {
450         destroy_cm_id(cm_id);
451 }
452 EXPORT_SYMBOL(iw_destroy_cm_id);
453
454 /**
455  * iw_cm_check_wildcard - If IP address is 0 then use original
456  * @pm_addr: sockaddr containing the ip to check for wildcard
457  * @cm_addr: sockaddr containing the actual IP address
458  * @cm_outaddr: sockaddr to set IP addr which leaving port
459  *
460  *  Checks the pm_addr for wildcard and then sets cm_outaddr's
461  *  IP to the actual (cm_addr).
462  */
463 static void iw_cm_check_wildcard(struct sockaddr_storage *pm_addr,
464                                  struct sockaddr_storage *cm_addr,
465                                  struct sockaddr_storage *cm_outaddr)
466 {
467         if (pm_addr->ss_family == AF_INET) {
468                 struct sockaddr_in *pm4_addr = (struct sockaddr_in *)pm_addr;
469
470                 if (pm4_addr->sin_addr.s_addr == htonl(INADDR_ANY)) {
471                         struct sockaddr_in *cm4_addr =
472                                 (struct sockaddr_in *)cm_addr;
473                         struct sockaddr_in *cm4_outaddr =
474                                 (struct sockaddr_in *)cm_outaddr;
475
476                         cm4_outaddr->sin_addr = cm4_addr->sin_addr;
477                 }
478         } else {
479                 struct sockaddr_in6 *pm6_addr = (struct sockaddr_in6 *)pm_addr;
480
481                 if (ipv6_addr_type(&pm6_addr->sin6_addr) == IPV6_ADDR_ANY) {
482                         struct sockaddr_in6 *cm6_addr =
483                                 (struct sockaddr_in6 *)cm_addr;
484                         struct sockaddr_in6 *cm6_outaddr =
485                                 (struct sockaddr_in6 *)cm_outaddr;
486
487                         cm6_outaddr->sin6_addr = cm6_addr->sin6_addr;
488                 }
489         }
490 }
491
492 /**
493  * iw_cm_map - Use portmapper to map the ports
494  * @cm_id: connection manager pointer
495  * @active: Indicates the active side when true
496  * returns nonzero for error only if iwpm_create_mapinfo() fails
497  *
498  * Tries to add a mapping for a port using the Portmapper. If
499  * successful in mapping the IP/Port it will check the remote
500  * mapped IP address for a wildcard IP address and replace the
501  * zero IP address with the remote_addr.
502  */
503 static int iw_cm_map(struct iw_cm_id *cm_id, bool active)
504 {
505         struct iwpm_dev_data pm_reg_msg;
506         struct iwpm_sa_data pm_msg;
507         int status;
508
509         cm_id->m_local_addr = cm_id->local_addr;
510         cm_id->m_remote_addr = cm_id->remote_addr;
511
512         memcpy(pm_reg_msg.dev_name, cm_id->device->name,
513                sizeof(pm_reg_msg.dev_name));
514         memcpy(pm_reg_msg.if_name, cm_id->device->iwcm->ifname,
515                sizeof(pm_reg_msg.if_name));
516
517         if (iwpm_register_pid(&pm_reg_msg, RDMA_NL_IWCM) ||
518             !iwpm_valid_pid())
519                 return 0;
520
521         cm_id->mapped = true;
522         pm_msg.loc_addr = cm_id->local_addr;
523         pm_msg.rem_addr = cm_id->remote_addr;
524         if (active)
525                 status = iwpm_add_and_query_mapping(&pm_msg,
526                                                     RDMA_NL_IWCM);
527         else
528                 status = iwpm_add_mapping(&pm_msg, RDMA_NL_IWCM);
529
530         if (!status) {
531                 cm_id->m_local_addr = pm_msg.mapped_loc_addr;
532                 if (active) {
533                         cm_id->m_remote_addr = pm_msg.mapped_rem_addr;
534                         iw_cm_check_wildcard(&pm_msg.mapped_rem_addr,
535                                              &cm_id->remote_addr,
536                                              &cm_id->m_remote_addr);
537                 }
538         }
539
540         return iwpm_create_mapinfo(&cm_id->local_addr,
541                                    &cm_id->m_local_addr,
542                                    RDMA_NL_IWCM);
543 }
544
545 /*
546  * CM_ID <-- LISTEN
547  *
548  * Start listening for connect requests. Generates one CONNECT_REQUEST
549  * event for each inbound connect request.
550  */
551 int iw_cm_listen(struct iw_cm_id *cm_id, int backlog)
552 {
553         struct iwcm_id_private *cm_id_priv;
554         unsigned long flags;
555         int ret;
556
557         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
558
559         if (!backlog)
560                 backlog = default_backlog;
561
562         ret = alloc_work_entries(cm_id_priv, backlog);
563         if (ret)
564                 return ret;
565
566         spin_lock_irqsave(&cm_id_priv->lock, flags);
567         switch (cm_id_priv->state) {
568         case IW_CM_STATE_IDLE:
569                 cm_id_priv->state = IW_CM_STATE_LISTEN;
570                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
571                 ret = iw_cm_map(cm_id, false);
572                 if (!ret)
573                         ret = cm_id->device->iwcm->create_listen(cm_id, backlog);
574                 if (ret)
575                         cm_id_priv->state = IW_CM_STATE_IDLE;
576                 spin_lock_irqsave(&cm_id_priv->lock, flags);
577                 break;
578         default:
579                 ret = -EINVAL;
580         }
581         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
582
583         return ret;
584 }
585 EXPORT_SYMBOL(iw_cm_listen);
586
587 /*
588  * CM_ID <-- IDLE
589  *
590  * Rejects an inbound connection request. No events are generated.
591  */
592 int iw_cm_reject(struct iw_cm_id *cm_id,
593                  const void *private_data,
594                  u8 private_data_len)
595 {
596         struct iwcm_id_private *cm_id_priv;
597         unsigned long flags;
598         int ret;
599
600         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
601         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
602
603         spin_lock_irqsave(&cm_id_priv->lock, flags);
604         if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
605                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
606                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
607                 wake_up_all(&cm_id_priv->connect_wait);
608                 return -EINVAL;
609         }
610         cm_id_priv->state = IW_CM_STATE_IDLE;
611         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
612
613         ret = cm_id->device->iwcm->reject(cm_id, private_data,
614                                           private_data_len);
615
616         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
617         wake_up_all(&cm_id_priv->connect_wait);
618
619         return ret;
620 }
621 EXPORT_SYMBOL(iw_cm_reject);
622
623 /*
624  * CM_ID <-- ESTABLISHED
625  *
626  * Accepts an inbound connection request and generates an ESTABLISHED
627  * event. Callers of iw_cm_disconnect and iw_destroy_cm_id will block
628  * until the ESTABLISHED event is received from the provider.
629  */
630 int iw_cm_accept(struct iw_cm_id *cm_id,
631                  struct iw_cm_conn_param *iw_param)
632 {
633         struct iwcm_id_private *cm_id_priv;
634         struct ib_qp *qp;
635         unsigned long flags;
636         int ret;
637
638         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
639         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
640
641         spin_lock_irqsave(&cm_id_priv->lock, flags);
642         if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
643                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
644                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
645                 wake_up_all(&cm_id_priv->connect_wait);
646                 return -EINVAL;
647         }
648         /* Get the ib_qp given the QPN */
649         qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
650         if (!qp) {
651                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
652                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
653                 wake_up_all(&cm_id_priv->connect_wait);
654                 return -EINVAL;
655         }
656         cm_id->device->iwcm->add_ref(qp);
657         cm_id_priv->qp = qp;
658         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
659
660         ret = cm_id->device->iwcm->accept(cm_id, iw_param);
661         if (ret) {
662                 /* An error on accept precludes provider events */
663                 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
664                 cm_id_priv->state = IW_CM_STATE_IDLE;
665                 spin_lock_irqsave(&cm_id_priv->lock, flags);
666                 if (cm_id_priv->qp) {
667                         cm_id->device->iwcm->rem_ref(qp);
668                         cm_id_priv->qp = NULL;
669                 }
670                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
671                 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
672                 wake_up_all(&cm_id_priv->connect_wait);
673         }
674
675         return ret;
676 }
677 EXPORT_SYMBOL(iw_cm_accept);
678
679 /*
680  * Active Side: CM_ID <-- CONN_SENT
681  *
682  * If successful, results in the generation of a CONNECT_REPLY
683  * event. iw_cm_disconnect and iw_cm_destroy will block until the
684  * CONNECT_REPLY event is received from the provider.
685  */
686 int iw_cm_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *iw_param)
687 {
688         struct iwcm_id_private *cm_id_priv;
689         int ret;
690         unsigned long flags;
691         struct ib_qp *qp;
692
693         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
694
695         ret = alloc_work_entries(cm_id_priv, 4);
696         if (ret)
697                 return ret;
698
699         set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
700         spin_lock_irqsave(&cm_id_priv->lock, flags);
701
702         if (cm_id_priv->state != IW_CM_STATE_IDLE) {
703                 ret = -EINVAL;
704                 goto err;
705         }
706
707         /* Get the ib_qp given the QPN */
708         qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
709         if (!qp) {
710                 ret = -EINVAL;
711                 goto err;
712         }
713         cm_id->device->iwcm->add_ref(qp);
714         cm_id_priv->qp = qp;
715         cm_id_priv->state = IW_CM_STATE_CONN_SENT;
716         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
717
718         ret = iw_cm_map(cm_id, true);
719         if (!ret)
720                 ret = cm_id->device->iwcm->connect(cm_id, iw_param);
721         if (!ret)
722                 return 0;       /* success */
723
724         spin_lock_irqsave(&cm_id_priv->lock, flags);
725         if (cm_id_priv->qp) {
726                 cm_id->device->iwcm->rem_ref(qp);
727                 cm_id_priv->qp = NULL;
728         }
729         cm_id_priv->state = IW_CM_STATE_IDLE;
730 err:
731         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
732         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
733         wake_up_all(&cm_id_priv->connect_wait);
734         return ret;
735 }
736 EXPORT_SYMBOL(iw_cm_connect);
737
738 /*
739  * Passive Side: new CM_ID <-- CONN_RECV
740  *
741  * Handles an inbound connect request. The function creates a new
742  * iw_cm_id to represent the new connection and inherits the client
743  * callback function and other attributes from the listening parent.
744  *
745  * The work item contains a pointer to the listen_cm_id and the event. The
746  * listen_cm_id contains the client cm_handler, context and
747  * device. These are copied when the device is cloned. The event
748  * contains the new four tuple.
749  *
750  * An error on the child should not affect the parent, so this
751  * function does not return a value.
752  */
753 static void cm_conn_req_handler(struct iwcm_id_private *listen_id_priv,
754                                 struct iw_cm_event *iw_event)
755 {
756         unsigned long flags;
757         struct iw_cm_id *cm_id;
758         struct iwcm_id_private *cm_id_priv;
759         int ret;
760
761         /*
762          * The provider should never generate a connection request
763          * event with a bad status.
764          */
765         BUG_ON(iw_event->status);
766
767         cm_id = iw_create_cm_id(listen_id_priv->id.device,
768                                 listen_id_priv->id.cm_handler,
769                                 listen_id_priv->id.context);
770         /* If the cm_id could not be created, ignore the request */
771         if (IS_ERR(cm_id))
772                 goto out;
773
774         cm_id->provider_data = iw_event->provider_data;
775         cm_id->m_local_addr = iw_event->local_addr;
776         cm_id->m_remote_addr = iw_event->remote_addr;
777         cm_id->local_addr = listen_id_priv->id.local_addr;
778
779         ret = iwpm_get_remote_info(&listen_id_priv->id.m_local_addr,
780                                    &iw_event->remote_addr,
781                                    &cm_id->remote_addr,
782                                    RDMA_NL_IWCM);
783         if (ret) {
784                 cm_id->remote_addr = iw_event->remote_addr;
785         } else {
786                 iw_cm_check_wildcard(&listen_id_priv->id.m_local_addr,
787                                      &iw_event->local_addr,
788                                      &cm_id->local_addr);
789                 iw_event->local_addr = cm_id->local_addr;
790                 iw_event->remote_addr = cm_id->remote_addr;
791         }
792
793         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
794         cm_id_priv->state = IW_CM_STATE_CONN_RECV;
795
796         /*
797          * We could be destroying the listening id. If so, ignore this
798          * upcall.
799          */
800         spin_lock_irqsave(&listen_id_priv->lock, flags);
801         if (listen_id_priv->state != IW_CM_STATE_LISTEN) {
802                 spin_unlock_irqrestore(&listen_id_priv->lock, flags);
803                 iw_cm_reject(cm_id, NULL, 0);
804                 iw_destroy_cm_id(cm_id);
805                 goto out;
806         }
807         spin_unlock_irqrestore(&listen_id_priv->lock, flags);
808
809         ret = alloc_work_entries(cm_id_priv, 3);
810         if (ret) {
811                 iw_cm_reject(cm_id, NULL, 0);
812                 iw_destroy_cm_id(cm_id);
813                 goto out;
814         }
815
816         /* Call the client CM handler */
817         ret = cm_id->cm_handler(cm_id, iw_event);
818         if (ret) {
819                 iw_cm_reject(cm_id, NULL, 0);
820                 iw_destroy_cm_id(cm_id);
821         }
822
823 out:
824         if (iw_event->private_data_len)
825                 kfree(iw_event->private_data);
826 }
827
828 /*
829  * Passive Side: CM_ID <-- ESTABLISHED
830  *
831  * The provider generated an ESTABLISHED event which means that
832  * the MPA negotion has completed successfully and we are now in MPA
833  * FPDU mode.
834  *
835  * This event can only be received in the CONN_RECV state. If the
836  * remote peer closed, the ESTABLISHED event would be received followed
837  * by the CLOSE event. If the app closes, it will block until we wake
838  * it up after processing this event.
839  */
840 static int cm_conn_est_handler(struct iwcm_id_private *cm_id_priv,
841                                struct iw_cm_event *iw_event)
842 {
843         unsigned long flags;
844         int ret;
845
846         spin_lock_irqsave(&cm_id_priv->lock, flags);
847
848         /*
849          * We clear the CONNECT_WAIT bit here to allow the callback
850          * function to call iw_cm_disconnect. Calling iw_destroy_cm_id
851          * from a callback handler is not allowed.
852          */
853         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
854         BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
855         cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
856         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
857         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
858         wake_up_all(&cm_id_priv->connect_wait);
859
860         return ret;
861 }
862
863 /*
864  * Active Side: CM_ID <-- ESTABLISHED
865  *
866  * The app has called connect and is waiting for the established event to
867  * post it's requests to the server. This event will wake up anyone
868  * blocked in iw_cm_disconnect or iw_destroy_id.
869  */
870 static int cm_conn_rep_handler(struct iwcm_id_private *cm_id_priv,
871                                struct iw_cm_event *iw_event)
872 {
873         unsigned long flags;
874         int ret;
875
876         spin_lock_irqsave(&cm_id_priv->lock, flags);
877         /*
878          * Clear the connect wait bit so a callback function calling
879          * iw_cm_disconnect will not wait and deadlock this thread
880          */
881         clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
882         BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
883         if (iw_event->status == 0) {
884                 cm_id_priv->id.m_local_addr = iw_event->local_addr;
885                 cm_id_priv->id.m_remote_addr = iw_event->remote_addr;
886                 iw_event->local_addr = cm_id_priv->id.local_addr;
887                 iw_event->remote_addr = cm_id_priv->id.remote_addr;
888                 cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
889         } else {
890                 /* REJECTED or RESET */
891                 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
892                 cm_id_priv->qp = NULL;
893                 cm_id_priv->state = IW_CM_STATE_IDLE;
894         }
895         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
896         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
897
898         if (iw_event->private_data_len)
899                 kfree(iw_event->private_data);
900
901         /* Wake up waiters on connect complete */
902         wake_up_all(&cm_id_priv->connect_wait);
903
904         return ret;
905 }
906
907 /*
908  * CM_ID <-- CLOSING
909  *
910  * If in the ESTABLISHED state, move to CLOSING.
911  */
912 static void cm_disconnect_handler(struct iwcm_id_private *cm_id_priv,
913                                   struct iw_cm_event *iw_event)
914 {
915         unsigned long flags;
916
917         spin_lock_irqsave(&cm_id_priv->lock, flags);
918         if (cm_id_priv->state == IW_CM_STATE_ESTABLISHED)
919                 cm_id_priv->state = IW_CM_STATE_CLOSING;
920         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
921 }
922
923 /*
924  * CM_ID <-- IDLE
925  *
926  * If in the ESTBLISHED or CLOSING states, the QP will have have been
927  * moved by the provider to the ERR state. Disassociate the CM_ID from
928  * the QP,  move to IDLE, and remove the 'connected' reference.
929  *
930  * If in some other state, the cm_id was destroyed asynchronously.
931  * This is the last reference that will result in waking up
932  * the app thread blocked in iw_destroy_cm_id.
933  */
934 static int cm_close_handler(struct iwcm_id_private *cm_id_priv,
935                                   struct iw_cm_event *iw_event)
936 {
937         unsigned long flags;
938         int ret = 0;
939         spin_lock_irqsave(&cm_id_priv->lock, flags);
940
941         if (cm_id_priv->qp) {
942                 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
943                 cm_id_priv->qp = NULL;
944         }
945         switch (cm_id_priv->state) {
946         case IW_CM_STATE_ESTABLISHED:
947         case IW_CM_STATE_CLOSING:
948                 cm_id_priv->state = IW_CM_STATE_IDLE;
949                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
950                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
951                 spin_lock_irqsave(&cm_id_priv->lock, flags);
952                 break;
953         case IW_CM_STATE_DESTROYING:
954                 break;
955         default:
956                 BUG();
957         }
958         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
959
960         return ret;
961 }
962
963 static int process_event(struct iwcm_id_private *cm_id_priv,
964                          struct iw_cm_event *iw_event)
965 {
966         int ret = 0;
967
968         switch (iw_event->event) {
969         case IW_CM_EVENT_CONNECT_REQUEST:
970                 cm_conn_req_handler(cm_id_priv, iw_event);
971                 break;
972         case IW_CM_EVENT_CONNECT_REPLY:
973                 ret = cm_conn_rep_handler(cm_id_priv, iw_event);
974                 break;
975         case IW_CM_EVENT_ESTABLISHED:
976                 ret = cm_conn_est_handler(cm_id_priv, iw_event);
977                 break;
978         case IW_CM_EVENT_DISCONNECT:
979                 cm_disconnect_handler(cm_id_priv, iw_event);
980                 break;
981         case IW_CM_EVENT_CLOSE:
982                 ret = cm_close_handler(cm_id_priv, iw_event);
983                 break;
984         default:
985                 BUG();
986         }
987
988         return ret;
989 }
990
991 /*
992  * Process events on the work_list for the cm_id. If the callback
993  * function requests that the cm_id be deleted, a flag is set in the
994  * cm_id flags to indicate that when the last reference is
995  * removed, the cm_id is to be destroyed. This is necessary to
996  * distinguish between an object that will be destroyed by the app
997  * thread asleep on the destroy_comp list vs. an object destroyed
998  * here synchronously when the last reference is removed.
999  */
1000 static void cm_work_handler(struct work_struct *_work)
1001 {
1002         struct iwcm_work *work = container_of(_work, struct iwcm_work, work);
1003         struct iw_cm_event levent;
1004         struct iwcm_id_private *cm_id_priv = work->cm_id;
1005         unsigned long flags;
1006         int empty;
1007         int ret = 0;
1008
1009         spin_lock_irqsave(&cm_id_priv->lock, flags);
1010         empty = list_empty(&cm_id_priv->work_list);
1011         while (!empty) {
1012                 work = list_entry(cm_id_priv->work_list.next,
1013                                   struct iwcm_work, list);
1014                 list_del_init(&work->list);
1015                 empty = list_empty(&cm_id_priv->work_list);
1016                 levent = work->event;
1017                 put_work(work);
1018                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1019
1020                 if (!test_bit(IWCM_F_DROP_EVENTS, &cm_id_priv->flags)) {
1021                         ret = process_event(cm_id_priv, &levent);
1022                         if (ret)
1023                                 destroy_cm_id(&cm_id_priv->id);
1024                 } else
1025                         pr_debug("dropping event %d\n", levent.event);
1026                 if (iwcm_deref_id(cm_id_priv))
1027                         return;
1028                 if (empty)
1029                         return;
1030                 spin_lock_irqsave(&cm_id_priv->lock, flags);
1031         }
1032         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1033 }
1034
1035 /*
1036  * This function is called on interrupt context. Schedule events on
1037  * the iwcm_wq thread to allow callback functions to downcall into
1038  * the CM and/or block.  Events are queued to a per-CM_ID
1039  * work_list. If this is the first event on the work_list, the work
1040  * element is also queued on the iwcm_wq thread.
1041  *
1042  * Each event holds a reference on the cm_id. Until the last posted
1043  * event has been delivered and processed, the cm_id cannot be
1044  * deleted.
1045  *
1046  * Returns:
1047  *            0 - the event was handled.
1048  *      -ENOMEM - the event was not handled due to lack of resources.
1049  */
1050 static int cm_event_handler(struct iw_cm_id *cm_id,
1051                              struct iw_cm_event *iw_event)
1052 {
1053         struct iwcm_work *work;
1054         struct iwcm_id_private *cm_id_priv;
1055         unsigned long flags;
1056         int ret = 0;
1057
1058         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1059
1060         spin_lock_irqsave(&cm_id_priv->lock, flags);
1061         work = get_work(cm_id_priv);
1062         if (!work) {
1063                 ret = -ENOMEM;
1064                 goto out;
1065         }
1066
1067         INIT_WORK(&work->work, cm_work_handler);
1068         work->cm_id = cm_id_priv;
1069         work->event = *iw_event;
1070
1071         if ((work->event.event == IW_CM_EVENT_CONNECT_REQUEST ||
1072              work->event.event == IW_CM_EVENT_CONNECT_REPLY) &&
1073             work->event.private_data_len) {
1074                 ret = copy_private_data(&work->event);
1075                 if (ret) {
1076                         put_work(work);
1077                         goto out;
1078                 }
1079         }
1080
1081         atomic_inc(&cm_id_priv->refcount);
1082         if (list_empty(&cm_id_priv->work_list)) {
1083                 list_add_tail(&work->list, &cm_id_priv->work_list);
1084                 queue_work(iwcm_wq, &work->work);
1085         } else
1086                 list_add_tail(&work->list, &cm_id_priv->work_list);
1087 out:
1088         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1089         return ret;
1090 }
1091
1092 static int iwcm_init_qp_init_attr(struct iwcm_id_private *cm_id_priv,
1093                                   struct ib_qp_attr *qp_attr,
1094                                   int *qp_attr_mask)
1095 {
1096         unsigned long flags;
1097         int ret;
1098
1099         spin_lock_irqsave(&cm_id_priv->lock, flags);
1100         switch (cm_id_priv->state) {
1101         case IW_CM_STATE_IDLE:
1102         case IW_CM_STATE_CONN_SENT:
1103         case IW_CM_STATE_CONN_RECV:
1104         case IW_CM_STATE_ESTABLISHED:
1105                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1106                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE|
1107                                            IB_ACCESS_REMOTE_READ;
1108                 ret = 0;
1109                 break;
1110         default:
1111                 ret = -EINVAL;
1112                 break;
1113         }
1114         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1115         return ret;
1116 }
1117
1118 static int iwcm_init_qp_rts_attr(struct iwcm_id_private *cm_id_priv,
1119                                   struct ib_qp_attr *qp_attr,
1120                                   int *qp_attr_mask)
1121 {
1122         unsigned long flags;
1123         int ret;
1124
1125         spin_lock_irqsave(&cm_id_priv->lock, flags);
1126         switch (cm_id_priv->state) {
1127         case IW_CM_STATE_IDLE:
1128         case IW_CM_STATE_CONN_SENT:
1129         case IW_CM_STATE_CONN_RECV:
1130         case IW_CM_STATE_ESTABLISHED:
1131                 *qp_attr_mask = 0;
1132                 ret = 0;
1133                 break;
1134         default:
1135                 ret = -EINVAL;
1136                 break;
1137         }
1138         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1139         return ret;
1140 }
1141
1142 int iw_cm_init_qp_attr(struct iw_cm_id *cm_id,
1143                        struct ib_qp_attr *qp_attr,
1144                        int *qp_attr_mask)
1145 {
1146         struct iwcm_id_private *cm_id_priv;
1147         int ret;
1148
1149         cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1150         switch (qp_attr->qp_state) {
1151         case IB_QPS_INIT:
1152         case IB_QPS_RTR:
1153                 ret = iwcm_init_qp_init_attr(cm_id_priv,
1154                                              qp_attr, qp_attr_mask);
1155                 break;
1156         case IB_QPS_RTS:
1157                 ret = iwcm_init_qp_rts_attr(cm_id_priv,
1158                                             qp_attr, qp_attr_mask);
1159                 break;
1160         default:
1161                 ret = -EINVAL;
1162                 break;
1163         }
1164         return ret;
1165 }
1166 EXPORT_SYMBOL(iw_cm_init_qp_attr);
1167
1168 static int __init iw_cm_init(void)
1169 {
1170         int ret;
1171
1172         ret = iwpm_init(RDMA_NL_IWCM);
1173         if (ret)
1174                 pr_err("iw_cm: couldn't init iwpm\n");
1175         else
1176                 rdma_nl_register(RDMA_NL_IWCM, iwcm_nl_cb_table);
1177         iwcm_wq = alloc_ordered_workqueue("iw_cm_wq", 0);
1178         if (!iwcm_wq)
1179                 return -ENOMEM;
1180
1181         iwcm_ctl_table_hdr = register_net_sysctl(&init_net, "net/iw_cm",
1182                                                  iwcm_ctl_table);
1183         if (!iwcm_ctl_table_hdr) {
1184                 pr_err("iw_cm: couldn't register sysctl paths\n");
1185                 destroy_workqueue(iwcm_wq);
1186                 return -ENOMEM;
1187         }
1188
1189         return 0;
1190 }
1191
1192 static void __exit iw_cm_cleanup(void)
1193 {
1194         unregister_net_sysctl_table(iwcm_ctl_table_hdr);
1195         destroy_workqueue(iwcm_wq);
1196         rdma_nl_unregister(RDMA_NL_IWCM);
1197         iwpm_exit(RDMA_NL_IWCM);
1198 }
1199
1200 MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_IWCM, 2);
1201
1202 module_init(iw_cm_init);
1203 module_exit(iw_cm_cleanup);