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