]> err.no Git - linux-2.6/blob - drivers/infiniband/core/cm.c
[PATCH] IB CM: support CM redir
[linux-2.6] / drivers / infiniband / core / cm.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  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * OpenIB.org BSD license below:
12  *
13  *     Redistribution and use in source and binary forms, with or
14  *     without modification, are permitted provided that the following
15  *     conditions are met:
16  *
17  *      - Redistributions of source code must retain the above
18  *        copyright notice, this list of conditions and the following
19  *        disclaimer.
20  *
21  *      - Redistributions in binary form must reproduce the above
22  *        copyright notice, this list of conditions and the following
23  *        disclaimer in the documentation and/or other materials
24  *        provided with the distribution.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33  * SOFTWARE.
34  *
35  * $Id: cm.c 2821 2005-07-08 17:07:28Z sean.hefty $
36  */
37 #include <linux/dma-mapping.h>
38 #include <linux/err.h>
39 #include <linux/idr.h>
40 #include <linux/interrupt.h>
41 #include <linux/pci.h>
42 #include <linux/rbtree.h>
43 #include <linux/spinlock.h>
44 #include <linux/workqueue.h>
45
46 #include <rdma/ib_cache.h>
47 #include <rdma/ib_cm.h>
48 #include "cm_msgs.h"
49
50 MODULE_AUTHOR("Sean Hefty");
51 MODULE_DESCRIPTION("InfiniBand CM");
52 MODULE_LICENSE("Dual BSD/GPL");
53
54 static void cm_add_one(struct ib_device *device);
55 static void cm_remove_one(struct ib_device *device);
56
57 static struct ib_client cm_client = {
58         .name   = "cm",
59         .add    = cm_add_one,
60         .remove = cm_remove_one
61 };
62
63 static struct ib_cm {
64         spinlock_t lock;
65         struct list_head device_list;
66         rwlock_t device_lock;
67         struct rb_root listen_service_table;
68         u64 listen_service_id;
69         /* struct rb_root peer_service_table; todo: fix peer to peer */
70         struct rb_root remote_qp_table;
71         struct rb_root remote_id_table;
72         struct rb_root remote_sidr_table;
73         struct idr local_id_table;
74         struct workqueue_struct *wq;
75 } cm;
76
77 struct cm_port {
78         struct cm_device *cm_dev;
79         struct ib_mad_agent *mad_agent;
80         u8 port_num;
81 };
82
83 struct cm_device {
84         struct list_head list;
85         struct ib_device *device;
86         __be64 ca_guid;
87         struct cm_port port[0];
88 };
89
90 struct cm_av {
91         struct cm_port *port;
92         union ib_gid dgid;
93         struct ib_ah_attr ah_attr;
94         u16 pkey_index;
95         u8 packet_life_time;
96 };
97
98 struct cm_work {
99         struct work_struct work;
100         struct list_head list;
101         struct cm_port *port;
102         struct ib_mad_recv_wc *mad_recv_wc;     /* Received MADs */
103         __be32 local_id;                        /* Established / timewait */
104         __be32 remote_id;
105         struct ib_cm_event cm_event;
106         struct ib_sa_path_rec path[0];
107 };
108
109 struct cm_timewait_info {
110         struct cm_work work;                    /* Must be first. */
111         struct rb_node remote_qp_node;
112         struct rb_node remote_id_node;
113         __be64 remote_ca_guid;
114         __be32 remote_qpn;
115         u8 inserted_remote_qp;
116         u8 inserted_remote_id;
117 };
118
119 struct cm_id_private {
120         struct ib_cm_id id;
121
122         struct rb_node service_node;
123         struct rb_node sidr_id_node;
124         spinlock_t lock;
125         wait_queue_head_t wait;
126         atomic_t refcount;
127
128         struct ib_mad_send_buf *msg;
129         struct cm_timewait_info *timewait_info;
130         /* todo: use alternate port on send failure */
131         struct cm_av av;
132         struct cm_av alt_av;
133
134         void *private_data;
135         __be64 tid;
136         __be32 local_qpn;
137         __be32 remote_qpn;
138         __be32 sq_psn;
139         __be32 rq_psn;
140         int timeout_ms;
141         enum ib_mtu path_mtu;
142         u8 private_data_len;
143         u8 max_cm_retries;
144         u8 peer_to_peer;
145         u8 responder_resources;
146         u8 initiator_depth;
147         u8 local_ack_timeout;
148         u8 retry_count;
149         u8 rnr_retry_count;
150         u8 service_timeout;
151
152         struct list_head work_list;
153         atomic_t work_count;
154 };
155
156 static void cm_work_handler(void *data);
157
158 static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
159 {
160         if (atomic_dec_and_test(&cm_id_priv->refcount))
161                 wake_up(&cm_id_priv->wait);
162 }
163
164 static int cm_alloc_msg(struct cm_id_private *cm_id_priv,
165                         struct ib_mad_send_buf **msg)
166 {
167         struct ib_mad_agent *mad_agent;
168         struct ib_mad_send_buf *m;
169         struct ib_ah *ah;
170
171         mad_agent = cm_id_priv->av.port->mad_agent;
172         ah = ib_create_ah(mad_agent->qp->pd, &cm_id_priv->av.ah_attr);
173         if (IS_ERR(ah))
174                 return PTR_ERR(ah);
175
176         m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn, 
177                                cm_id_priv->av.pkey_index,
178                                ah, 0, sizeof(struct ib_mad_hdr),
179                                sizeof(struct ib_mad)-sizeof(struct ib_mad_hdr),
180                                GFP_ATOMIC);
181         if (IS_ERR(m)) {
182                 ib_destroy_ah(ah);
183                 return PTR_ERR(m);
184         }
185
186         /* Timeout set by caller if response is expected. */
187         m->send_wr.wr.ud.retries = cm_id_priv->max_cm_retries;
188
189         atomic_inc(&cm_id_priv->refcount);
190         m->context[0] = cm_id_priv;
191         *msg = m;
192         return 0;
193 }
194
195 static int cm_alloc_response_msg(struct cm_port *port,
196                                  struct ib_mad_recv_wc *mad_recv_wc,
197                                  struct ib_mad_send_buf **msg)
198 {
199         struct ib_mad_send_buf *m;
200         struct ib_ah *ah;
201
202         ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
203                                   mad_recv_wc->recv_buf.grh, port->port_num);
204         if (IS_ERR(ah))
205                 return PTR_ERR(ah);
206
207         m = ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
208                                ah, 0, sizeof(struct ib_mad_hdr),
209                                sizeof(struct ib_mad)-sizeof(struct ib_mad_hdr),
210                                GFP_ATOMIC);
211         if (IS_ERR(m)) {
212                 ib_destroy_ah(ah);
213                 return PTR_ERR(m);
214         }
215         *msg = m;
216         return 0;
217 }
218
219 static void cm_free_msg(struct ib_mad_send_buf *msg)
220 {
221         ib_destroy_ah(msg->send_wr.wr.ud.ah);
222         if (msg->context[0])
223                 cm_deref_id(msg->context[0]);
224         ib_free_send_mad(msg);
225 }
226
227 static void * cm_copy_private_data(const void *private_data,
228                                    u8 private_data_len)
229 {
230         void *data;
231
232         if (!private_data || !private_data_len)
233                 return NULL;
234
235         data = kmalloc(private_data_len, GFP_KERNEL);
236         if (!data)
237                 return ERR_PTR(-ENOMEM);
238
239         memcpy(data, private_data, private_data_len);
240         return data;
241 }
242
243 static void cm_set_private_data(struct cm_id_private *cm_id_priv,
244                                  void *private_data, u8 private_data_len)
245 {
246         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
247                 kfree(cm_id_priv->private_data);
248
249         cm_id_priv->private_data = private_data;
250         cm_id_priv->private_data_len = private_data_len;
251 }
252
253 static void cm_set_ah_attr(struct ib_ah_attr *ah_attr, u8 port_num,
254                            u16 dlid, u8 sl, u16 src_path_bits)
255 {
256         memset(ah_attr, 0, sizeof ah_attr);
257         ah_attr->dlid = dlid;
258         ah_attr->sl = sl;
259         ah_attr->src_path_bits = src_path_bits;
260         ah_attr->port_num = port_num;
261 }
262
263 static void cm_init_av_for_response(struct cm_port *port,
264                                     struct ib_wc *wc, struct cm_av *av)
265 {
266         av->port = port;
267         av->pkey_index = wc->pkey_index;
268         cm_set_ah_attr(&av->ah_attr, port->port_num, wc->slid,
269                        wc->sl, wc->dlid_path_bits);
270 }
271
272 static int cm_init_av_by_path(struct ib_sa_path_rec *path, struct cm_av *av)
273 {
274         struct cm_device *cm_dev;
275         struct cm_port *port = NULL;
276         unsigned long flags;
277         int ret;
278         u8 p;
279
280         read_lock_irqsave(&cm.device_lock, flags);
281         list_for_each_entry(cm_dev, &cm.device_list, list) {
282                 if (!ib_find_cached_gid(cm_dev->device, &path->sgid,
283                                         &p, NULL)) {
284                         port = &cm_dev->port[p-1];
285                         break;
286                 }
287         }
288         read_unlock_irqrestore(&cm.device_lock, flags);
289
290         if (!port)
291                 return -EINVAL;
292
293         ret = ib_find_cached_pkey(cm_dev->device, port->port_num,
294                                   be16_to_cpu(path->pkey), &av->pkey_index);
295         if (ret)
296                 return ret;
297
298         av->port = port;
299         cm_set_ah_attr(&av->ah_attr, av->port->port_num,
300                        be16_to_cpu(path->dlid), path->sl,
301                        be16_to_cpu(path->slid) & 0x7F);
302         av->packet_life_time = path->packet_life_time;
303         return 0;
304 }
305
306 static int cm_alloc_id(struct cm_id_private *cm_id_priv)
307 {
308         unsigned long flags;
309         int ret;
310
311         do {
312                 spin_lock_irqsave(&cm.lock, flags);
313                 ret = idr_get_new_above(&cm.local_id_table, cm_id_priv, 1,
314                                         (__force int *) &cm_id_priv->id.local_id);
315                 spin_unlock_irqrestore(&cm.lock, flags);
316         } while( (ret == -EAGAIN) && idr_pre_get(&cm.local_id_table, GFP_KERNEL) );
317         return ret;
318 }
319
320 static void cm_free_id(__be32 local_id)
321 {
322         unsigned long flags;
323
324         spin_lock_irqsave(&cm.lock, flags);
325         idr_remove(&cm.local_id_table, (__force int) local_id);
326         spin_unlock_irqrestore(&cm.lock, flags);
327 }
328
329 static struct cm_id_private * cm_get_id(__be32 local_id, __be32 remote_id)
330 {
331         struct cm_id_private *cm_id_priv;
332
333         cm_id_priv = idr_find(&cm.local_id_table, (__force int) local_id);
334         if (cm_id_priv) {
335                 if (cm_id_priv->id.remote_id == remote_id)
336                         atomic_inc(&cm_id_priv->refcount);
337                 else
338                         cm_id_priv = NULL;
339         }
340
341         return cm_id_priv;
342 }
343
344 static struct cm_id_private * cm_acquire_id(__be32 local_id, __be32 remote_id)
345 {
346         struct cm_id_private *cm_id_priv;
347         unsigned long flags;
348
349         spin_lock_irqsave(&cm.lock, flags);
350         cm_id_priv = cm_get_id(local_id, remote_id);
351         spin_unlock_irqrestore(&cm.lock, flags);
352
353         return cm_id_priv;
354 }
355
356 static struct cm_id_private * cm_insert_listen(struct cm_id_private *cm_id_priv)
357 {
358         struct rb_node **link = &cm.listen_service_table.rb_node;
359         struct rb_node *parent = NULL;
360         struct cm_id_private *cur_cm_id_priv;
361         __be64 service_id = cm_id_priv->id.service_id;
362         __be64 service_mask = cm_id_priv->id.service_mask;
363
364         while (*link) {
365                 parent = *link;
366                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
367                                           service_node);
368                 if ((cur_cm_id_priv->id.service_mask & service_id) ==
369                     (service_mask & cur_cm_id_priv->id.service_id))
370                         return cm_id_priv;
371                 if (service_id < cur_cm_id_priv->id.service_id)
372                         link = &(*link)->rb_left;
373                 else
374                         link = &(*link)->rb_right;
375         }
376         rb_link_node(&cm_id_priv->service_node, parent, link);
377         rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
378         return NULL;
379 }
380
381 static struct cm_id_private * cm_find_listen(__be64 service_id)
382 {
383         struct rb_node *node = cm.listen_service_table.rb_node;
384         struct cm_id_private *cm_id_priv;
385
386         while (node) {
387                 cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
388                 if ((cm_id_priv->id.service_mask & service_id) ==
389                     (cm_id_priv->id.service_mask & cm_id_priv->id.service_id))
390                         return cm_id_priv;
391                 if (service_id < cm_id_priv->id.service_id)
392                         node = node->rb_left;
393                 else
394                         node = node->rb_right;
395         }
396         return NULL;
397 }
398
399 static struct cm_timewait_info * cm_insert_remote_id(struct cm_timewait_info
400                                                      *timewait_info)
401 {
402         struct rb_node **link = &cm.remote_id_table.rb_node;
403         struct rb_node *parent = NULL;
404         struct cm_timewait_info *cur_timewait_info;
405         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
406         __be32 remote_id = timewait_info->work.remote_id;
407
408         while (*link) {
409                 parent = *link;
410                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
411                                              remote_id_node);
412                 if (remote_id < cur_timewait_info->work.remote_id)
413                         link = &(*link)->rb_left;
414                 else if (remote_id > cur_timewait_info->work.remote_id)
415                         link = &(*link)->rb_right;
416                 else if (remote_ca_guid < cur_timewait_info->remote_ca_guid)
417                         link = &(*link)->rb_left;
418                 else if (remote_ca_guid > cur_timewait_info->remote_ca_guid)
419                         link = &(*link)->rb_right;
420                 else
421                         return cur_timewait_info;
422         }
423         timewait_info->inserted_remote_id = 1;
424         rb_link_node(&timewait_info->remote_id_node, parent, link);
425         rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
426         return NULL;
427 }
428
429 static struct cm_timewait_info * cm_find_remote_id(__be64 remote_ca_guid,
430                                                    __be32 remote_id)
431 {
432         struct rb_node *node = cm.remote_id_table.rb_node;
433         struct cm_timewait_info *timewait_info;
434
435         while (node) {
436                 timewait_info = rb_entry(node, struct cm_timewait_info,
437                                          remote_id_node);
438                 if (remote_id < timewait_info->work.remote_id)
439                         node = node->rb_left;
440                 else if (remote_id > timewait_info->work.remote_id)
441                         node = node->rb_right;
442                 else if (remote_ca_guid < timewait_info->remote_ca_guid)
443                         node = node->rb_left;
444                 else if (remote_ca_guid > timewait_info->remote_ca_guid)
445                         node = node->rb_right;
446                 else
447                         return timewait_info;
448         }
449         return NULL;
450 }
451
452 static struct cm_timewait_info * cm_insert_remote_qpn(struct cm_timewait_info
453                                                       *timewait_info)
454 {
455         struct rb_node **link = &cm.remote_qp_table.rb_node;
456         struct rb_node *parent = NULL;
457         struct cm_timewait_info *cur_timewait_info;
458         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
459         __be32 remote_qpn = timewait_info->remote_qpn;
460
461         while (*link) {
462                 parent = *link;
463                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
464                                              remote_qp_node);
465                 if (remote_qpn < cur_timewait_info->remote_qpn)
466                         link = &(*link)->rb_left;
467                 else if (remote_qpn > cur_timewait_info->remote_qpn)
468                         link = &(*link)->rb_right;
469                 else if (remote_ca_guid < cur_timewait_info->remote_ca_guid)
470                         link = &(*link)->rb_left;
471                 else if (remote_ca_guid > cur_timewait_info->remote_ca_guid)
472                         link = &(*link)->rb_right;
473                 else
474                         return cur_timewait_info;
475         }
476         timewait_info->inserted_remote_qp = 1;
477         rb_link_node(&timewait_info->remote_qp_node, parent, link);
478         rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
479         return NULL;
480 }
481
482 static struct cm_id_private * cm_insert_remote_sidr(struct cm_id_private
483                                                     *cm_id_priv)
484 {
485         struct rb_node **link = &cm.remote_sidr_table.rb_node;
486         struct rb_node *parent = NULL;
487         struct cm_id_private *cur_cm_id_priv;
488         union ib_gid *port_gid = &cm_id_priv->av.dgid;
489         __be32 remote_id = cm_id_priv->id.remote_id;
490
491         while (*link) {
492                 parent = *link;
493                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
494                                           sidr_id_node);
495                 if (remote_id < cur_cm_id_priv->id.remote_id)
496                         link = &(*link)->rb_left;
497                 else if (remote_id > cur_cm_id_priv->id.remote_id)
498                         link = &(*link)->rb_right;
499                 else {
500                         int cmp;
501                         cmp = memcmp(port_gid, &cur_cm_id_priv->av.dgid,
502                                      sizeof *port_gid);
503                         if (cmp < 0)
504                                 link = &(*link)->rb_left;
505                         else if (cmp > 0)
506                                 link = &(*link)->rb_right;
507                         else
508                                 return cur_cm_id_priv;
509                 }
510         }
511         rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
512         rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
513         return NULL;
514 }
515
516 static void cm_reject_sidr_req(struct cm_id_private *cm_id_priv,
517                                enum ib_cm_sidr_status status)
518 {
519         struct ib_cm_sidr_rep_param param;
520
521         memset(&param, 0, sizeof param);
522         param.status = status;
523         ib_send_cm_sidr_rep(&cm_id_priv->id, &param);
524 }
525
526 struct ib_cm_id *ib_create_cm_id(ib_cm_handler cm_handler,
527                                  void *context)
528 {
529         struct cm_id_private *cm_id_priv;
530         int ret;
531
532         cm_id_priv = kmalloc(sizeof *cm_id_priv, GFP_KERNEL);
533         if (!cm_id_priv)
534                 return ERR_PTR(-ENOMEM);
535
536         memset(cm_id_priv, 0, sizeof *cm_id_priv);
537         cm_id_priv->id.state = IB_CM_IDLE;
538         cm_id_priv->id.cm_handler = cm_handler;
539         cm_id_priv->id.context = context;
540         cm_id_priv->id.remote_cm_qpn = 1;
541         ret = cm_alloc_id(cm_id_priv);
542         if (ret)
543                 goto error;
544
545         spin_lock_init(&cm_id_priv->lock);
546         init_waitqueue_head(&cm_id_priv->wait);
547         INIT_LIST_HEAD(&cm_id_priv->work_list);
548         atomic_set(&cm_id_priv->work_count, -1);
549         atomic_set(&cm_id_priv->refcount, 1);
550         return &cm_id_priv->id;
551
552 error:
553         kfree(cm_id_priv);
554         return ERR_PTR(-ENOMEM);
555 }
556 EXPORT_SYMBOL(ib_create_cm_id);
557
558 static struct cm_work * cm_dequeue_work(struct cm_id_private *cm_id_priv)
559 {
560         struct cm_work *work;
561
562         if (list_empty(&cm_id_priv->work_list))
563                 return NULL;
564
565         work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
566         list_del(&work->list);
567         return work;
568 }
569
570 static void cm_free_work(struct cm_work *work)
571 {
572         if (work->mad_recv_wc)
573                 ib_free_recv_mad(work->mad_recv_wc);
574         kfree(work);
575 }
576
577 static inline int cm_convert_to_ms(int iba_time)
578 {
579         /* approximate conversion to ms from 4.096us x 2^iba_time */
580         return 1 << max(iba_time - 8, 0);
581 }
582
583 static void cm_cleanup_timewait(struct cm_timewait_info *timewait_info)
584 {
585         unsigned long flags;
586
587         if (!timewait_info->inserted_remote_id &&
588             !timewait_info->inserted_remote_qp)
589             return;
590
591         spin_lock_irqsave(&cm.lock, flags);
592         if (timewait_info->inserted_remote_id) {
593                 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
594                 timewait_info->inserted_remote_id = 0;
595         }
596
597         if (timewait_info->inserted_remote_qp) {
598                 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
599                 timewait_info->inserted_remote_qp = 0;
600         }
601         spin_unlock_irqrestore(&cm.lock, flags);
602 }
603
604 static struct cm_timewait_info * cm_create_timewait_info(__be32 local_id)
605 {
606         struct cm_timewait_info *timewait_info;
607
608         timewait_info = kmalloc(sizeof *timewait_info, GFP_KERNEL);
609         if (!timewait_info)
610                 return ERR_PTR(-ENOMEM);
611         memset(timewait_info, 0, sizeof *timewait_info);
612
613         timewait_info->work.local_id = local_id;
614         INIT_WORK(&timewait_info->work.work, cm_work_handler,
615                   &timewait_info->work);
616         timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
617         return timewait_info;
618 }
619
620 static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
621 {
622         int wait_time;
623
624         /*
625          * The cm_id could be destroyed by the user before we exit timewait.
626          * To protect against this, we search for the cm_id after exiting
627          * timewait before notifying the user that we've exited timewait.
628          */
629         cm_id_priv->id.state = IB_CM_TIMEWAIT;
630         wait_time = cm_convert_to_ms(cm_id_priv->local_ack_timeout);
631         queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
632                            msecs_to_jiffies(wait_time));
633         cm_id_priv->timewait_info = NULL;
634 }
635
636 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
637 {
638         cm_id_priv->id.state = IB_CM_IDLE;
639         if (cm_id_priv->timewait_info) {
640                 cm_cleanup_timewait(cm_id_priv->timewait_info);
641                 kfree(cm_id_priv->timewait_info);
642                 cm_id_priv->timewait_info = NULL;
643         }
644 }
645
646 void ib_destroy_cm_id(struct ib_cm_id *cm_id)
647 {
648         struct cm_id_private *cm_id_priv;
649         struct cm_work *work;
650         unsigned long flags;
651
652         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
653 retest:
654         spin_lock_irqsave(&cm_id_priv->lock, flags);
655         switch (cm_id->state) {
656         case IB_CM_LISTEN:
657                 cm_id->state = IB_CM_IDLE;
658                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
659                 spin_lock_irqsave(&cm.lock, flags);
660                 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
661                 spin_unlock_irqrestore(&cm.lock, flags);
662                 break;
663         case IB_CM_SIDR_REQ_SENT:
664                 cm_id->state = IB_CM_IDLE;
665                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
666                               (unsigned long) cm_id_priv->msg);
667                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
668                 break;
669         case IB_CM_SIDR_REQ_RCVD:
670                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
671                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_REJECT);
672                 break;
673         case IB_CM_REQ_SENT:
674         case IB_CM_MRA_REQ_RCVD:
675         case IB_CM_REP_SENT:
676         case IB_CM_MRA_REP_RCVD:
677                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
678                               (unsigned long) cm_id_priv->msg);
679                 /* Fall through */
680         case IB_CM_REQ_RCVD:
681         case IB_CM_MRA_REQ_SENT:
682         case IB_CM_REP_RCVD:
683         case IB_CM_MRA_REP_SENT:
684                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
685                 ib_send_cm_rej(cm_id, IB_CM_REJ_TIMEOUT,
686                                &cm_id_priv->av.port->cm_dev->ca_guid,
687                                sizeof cm_id_priv->av.port->cm_dev->ca_guid,
688                                NULL, 0);
689                 break;
690         case IB_CM_ESTABLISHED:
691                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
692                 ib_send_cm_dreq(cm_id, NULL, 0);
693                 goto retest;
694         case IB_CM_DREQ_SENT:
695                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
696                               (unsigned long) cm_id_priv->msg);
697                 cm_enter_timewait(cm_id_priv);
698                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
699                 break;
700         case IB_CM_DREQ_RCVD:
701                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
702                 ib_send_cm_drep(cm_id, NULL, 0);
703                 break;
704         default:
705                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
706                 break;
707         }
708
709         cm_free_id(cm_id->local_id);
710         atomic_dec(&cm_id_priv->refcount);
711         wait_event(cm_id_priv->wait, !atomic_read(&cm_id_priv->refcount));
712         while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
713                 cm_free_work(work);
714         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
715                 kfree(cm_id_priv->private_data);
716         kfree(cm_id_priv);
717 }
718 EXPORT_SYMBOL(ib_destroy_cm_id);
719
720 int ib_cm_listen(struct ib_cm_id *cm_id,
721                  __be64 service_id,
722                  __be64 service_mask)
723 {
724         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
725         unsigned long flags;
726         int ret = 0;
727
728         service_mask = service_mask ? service_mask :
729                        __constant_cpu_to_be64(~0ULL);
730         service_id &= service_mask;
731         if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID &&
732             (service_id != IB_CM_ASSIGN_SERVICE_ID))
733                 return -EINVAL;
734
735         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
736         BUG_ON(cm_id->state != IB_CM_IDLE);
737
738         cm_id->state = IB_CM_LISTEN;
739
740         spin_lock_irqsave(&cm.lock, flags);
741         if (service_id == IB_CM_ASSIGN_SERVICE_ID) {
742                 cm_id->service_id = cpu_to_be64(cm.listen_service_id++);
743                 cm_id->service_mask = __constant_cpu_to_be64(~0ULL);
744         } else {
745                 cm_id->service_id = service_id;
746                 cm_id->service_mask = service_mask;
747         }
748         cur_cm_id_priv = cm_insert_listen(cm_id_priv);
749         spin_unlock_irqrestore(&cm.lock, flags);
750
751         if (cur_cm_id_priv) {
752                 cm_id->state = IB_CM_IDLE;
753                 ret = -EBUSY;
754         }
755         return ret;
756 }
757 EXPORT_SYMBOL(ib_cm_listen);
758
759 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv,
760                           enum cm_msg_sequence msg_seq)
761 {
762         u64 hi_tid, low_tid;
763
764         hi_tid   = ((u64) cm_id_priv->av.port->mad_agent->hi_tid) << 32;
765         low_tid  = (u64) ((__force u32)cm_id_priv->id.local_id |
766                           (msg_seq << 30));
767         return cpu_to_be64(hi_tid | low_tid);
768 }
769
770 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
771                               __be16 attr_id, __be64 tid)
772 {
773         hdr->base_version  = IB_MGMT_BASE_VERSION;
774         hdr->mgmt_class    = IB_MGMT_CLASS_CM;
775         hdr->class_version = IB_CM_CLASS_VERSION;
776         hdr->method        = IB_MGMT_METHOD_SEND;
777         hdr->attr_id       = attr_id;
778         hdr->tid           = tid;
779 }
780
781 static void cm_format_req(struct cm_req_msg *req_msg,
782                           struct cm_id_private *cm_id_priv,
783                           struct ib_cm_req_param *param)
784 {
785         cm_format_mad_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
786                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_REQ));
787
788         req_msg->local_comm_id = cm_id_priv->id.local_id;
789         req_msg->service_id = param->service_id;
790         req_msg->local_ca_guid = cm_id_priv->av.port->cm_dev->ca_guid;
791         cm_req_set_local_qpn(req_msg, cpu_to_be32(param->qp_num));
792         cm_req_set_resp_res(req_msg, param->responder_resources);
793         cm_req_set_init_depth(req_msg, param->initiator_depth);
794         cm_req_set_remote_resp_timeout(req_msg,
795                                        param->remote_cm_response_timeout);
796         cm_req_set_qp_type(req_msg, param->qp_type);
797         cm_req_set_flow_ctrl(req_msg, param->flow_control);
798         cm_req_set_starting_psn(req_msg, cpu_to_be32(param->starting_psn));
799         cm_req_set_local_resp_timeout(req_msg,
800                                       param->local_cm_response_timeout);
801         cm_req_set_retry_count(req_msg, param->retry_count);
802         req_msg->pkey = param->primary_path->pkey;
803         cm_req_set_path_mtu(req_msg, param->primary_path->mtu);
804         cm_req_set_rnr_retry_count(req_msg, param->rnr_retry_count);
805         cm_req_set_max_cm_retries(req_msg, param->max_cm_retries);
806         cm_req_set_srq(req_msg, param->srq);
807
808         req_msg->primary_local_lid = param->primary_path->slid;
809         req_msg->primary_remote_lid = param->primary_path->dlid;
810         req_msg->primary_local_gid = param->primary_path->sgid;
811         req_msg->primary_remote_gid = param->primary_path->dgid;
812         cm_req_set_primary_flow_label(req_msg, param->primary_path->flow_label);
813         cm_req_set_primary_packet_rate(req_msg, param->primary_path->rate);
814         req_msg->primary_traffic_class = param->primary_path->traffic_class;
815         req_msg->primary_hop_limit = param->primary_path->hop_limit;
816         cm_req_set_primary_sl(req_msg, param->primary_path->sl);
817         cm_req_set_primary_subnet_local(req_msg, 1); /* local only... */
818         cm_req_set_primary_local_ack_timeout(req_msg,
819                 min(31, param->primary_path->packet_life_time + 1));
820
821         if (param->alternate_path) {
822                 req_msg->alt_local_lid = param->alternate_path->slid;
823                 req_msg->alt_remote_lid = param->alternate_path->dlid;
824                 req_msg->alt_local_gid = param->alternate_path->sgid;
825                 req_msg->alt_remote_gid = param->alternate_path->dgid;
826                 cm_req_set_alt_flow_label(req_msg,
827                                           param->alternate_path->flow_label);
828                 cm_req_set_alt_packet_rate(req_msg, param->alternate_path->rate);
829                 req_msg->alt_traffic_class = param->alternate_path->traffic_class;
830                 req_msg->alt_hop_limit = param->alternate_path->hop_limit;
831                 cm_req_set_alt_sl(req_msg, param->alternate_path->sl);
832                 cm_req_set_alt_subnet_local(req_msg, 1); /* local only... */
833                 cm_req_set_alt_local_ack_timeout(req_msg,
834                         min(31, param->alternate_path->packet_life_time + 1));
835         }
836
837         if (param->private_data && param->private_data_len)
838                 memcpy(req_msg->private_data, param->private_data,
839                        param->private_data_len);
840 }
841
842 static inline int cm_validate_req_param(struct ib_cm_req_param *param)
843 {
844         /* peer-to-peer not supported */
845         if (param->peer_to_peer)
846                 return -EINVAL;
847
848         if (!param->primary_path)
849                 return -EINVAL;
850
851         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC)
852                 return -EINVAL;
853
854         if (param->private_data &&
855             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
856                 return -EINVAL;
857
858         if (param->alternate_path &&
859             (param->alternate_path->pkey != param->primary_path->pkey ||
860              param->alternate_path->mtu != param->primary_path->mtu))
861                 return -EINVAL;
862
863         return 0;
864 }
865
866 int ib_send_cm_req(struct ib_cm_id *cm_id,
867                    struct ib_cm_req_param *param)
868 {
869         struct cm_id_private *cm_id_priv;
870         struct ib_send_wr *bad_send_wr;
871         struct cm_req_msg *req_msg;
872         unsigned long flags;
873         int ret;
874
875         ret = cm_validate_req_param(param);
876         if (ret)
877                 return ret;
878
879         /* Verify that we're not in timewait. */
880         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
881         spin_lock_irqsave(&cm_id_priv->lock, flags);
882         if (cm_id->state != IB_CM_IDLE) {
883                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
884                 ret = -EINVAL;
885                 goto out;
886         }
887         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
888
889         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
890                                                             id.local_id);
891         if (IS_ERR(cm_id_priv->timewait_info))
892                 goto out;
893
894         ret = cm_init_av_by_path(param->primary_path, &cm_id_priv->av);
895         if (ret)
896                 goto error1;
897         if (param->alternate_path) {
898                 ret = cm_init_av_by_path(param->alternate_path,
899                                          &cm_id_priv->alt_av);
900                 if (ret)
901                         goto error1;
902         }
903         cm_id->service_id = param->service_id;
904         cm_id->service_mask = __constant_cpu_to_be64(~0ULL);
905         cm_id_priv->timeout_ms = cm_convert_to_ms(
906                                     param->primary_path->packet_life_time) * 2 +
907                                  cm_convert_to_ms(
908                                     param->remote_cm_response_timeout);
909         cm_id_priv->max_cm_retries = param->max_cm_retries;
910         cm_id_priv->initiator_depth = param->initiator_depth;
911         cm_id_priv->responder_resources = param->responder_resources;
912         cm_id_priv->retry_count = param->retry_count;
913         cm_id_priv->path_mtu = param->primary_path->mtu;
914
915         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
916         if (ret)
917                 goto error1;
918
919         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
920         cm_format_req(req_msg, cm_id_priv, param);
921         cm_id_priv->tid = req_msg->hdr.tid;
922         cm_id_priv->msg->send_wr.wr.ud.timeout_ms = cm_id_priv->timeout_ms;
923         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
924
925         cm_id_priv->local_qpn = cm_req_get_local_qpn(req_msg);
926         cm_id_priv->rq_psn = cm_req_get_starting_psn(req_msg);
927         cm_id_priv->local_ack_timeout =
928                                 cm_req_get_primary_local_ack_timeout(req_msg);
929
930         spin_lock_irqsave(&cm_id_priv->lock, flags);
931         ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
932                                 &cm_id_priv->msg->send_wr, &bad_send_wr);
933         if (ret) {
934                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
935                 goto error2;
936         }
937         BUG_ON(cm_id->state != IB_CM_IDLE);
938         cm_id->state = IB_CM_REQ_SENT;
939         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
940         return 0;
941
942 error2: cm_free_msg(cm_id_priv->msg);
943 error1: kfree(cm_id_priv->timewait_info);
944 out:    return ret;
945 }
946 EXPORT_SYMBOL(ib_send_cm_req);
947
948 static int cm_issue_rej(struct cm_port *port,
949                         struct ib_mad_recv_wc *mad_recv_wc,
950                         enum ib_cm_rej_reason reason,
951                         enum cm_msg_response msg_rejected,
952                         void *ari, u8 ari_length)
953 {
954         struct ib_mad_send_buf *msg = NULL;
955         struct ib_send_wr *bad_send_wr;
956         struct cm_rej_msg *rej_msg, *rcv_msg;
957         int ret;
958
959         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
960         if (ret)
961                 return ret;
962
963         /* We just need common CM header information.  Cast to any message. */
964         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
965         rej_msg = (struct cm_rej_msg *) msg->mad;
966
967         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
968         rej_msg->remote_comm_id = rcv_msg->local_comm_id;
969         rej_msg->local_comm_id = rcv_msg->remote_comm_id;
970         cm_rej_set_msg_rejected(rej_msg, msg_rejected);
971         rej_msg->reason = cpu_to_be16(reason);
972
973         if (ari && ari_length) {
974                 cm_rej_set_reject_info_len(rej_msg, ari_length);
975                 memcpy(rej_msg->ari, ari, ari_length);
976         }
977
978         ret = ib_post_send_mad(port->mad_agent, &msg->send_wr, &bad_send_wr);
979         if (ret)
980                 cm_free_msg(msg);
981
982         return ret;
983 }
984
985 static inline int cm_is_active_peer(__be64 local_ca_guid, __be64 remote_ca_guid,
986                                     __be32 local_qpn, __be32 remote_qpn)
987 {
988         return (be64_to_cpu(local_ca_guid) > be64_to_cpu(remote_ca_guid) ||
989                 ((local_ca_guid == remote_ca_guid) &&
990                  (be32_to_cpu(local_qpn) > be32_to_cpu(remote_qpn))));
991 }
992
993 static inline void cm_format_paths_from_req(struct cm_req_msg *req_msg,
994                                             struct ib_sa_path_rec *primary_path,
995                                             struct ib_sa_path_rec *alt_path)
996 {
997         memset(primary_path, 0, sizeof *primary_path);
998         primary_path->dgid = req_msg->primary_local_gid;
999         primary_path->sgid = req_msg->primary_remote_gid;
1000         primary_path->dlid = req_msg->primary_local_lid;
1001         primary_path->slid = req_msg->primary_remote_lid;
1002         primary_path->flow_label = cm_req_get_primary_flow_label(req_msg);
1003         primary_path->hop_limit = req_msg->primary_hop_limit;
1004         primary_path->traffic_class = req_msg->primary_traffic_class;
1005         primary_path->reversible = 1;
1006         primary_path->pkey = req_msg->pkey;
1007         primary_path->sl = cm_req_get_primary_sl(req_msg);
1008         primary_path->mtu_selector = IB_SA_EQ;
1009         primary_path->mtu = cm_req_get_path_mtu(req_msg);
1010         primary_path->rate_selector = IB_SA_EQ;
1011         primary_path->rate = cm_req_get_primary_packet_rate(req_msg);
1012         primary_path->packet_life_time_selector = IB_SA_EQ;
1013         primary_path->packet_life_time =
1014                 cm_req_get_primary_local_ack_timeout(req_msg);
1015         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1016
1017         if (req_msg->alt_local_lid) {
1018                 memset(alt_path, 0, sizeof *alt_path);
1019                 alt_path->dgid = req_msg->alt_local_gid;
1020                 alt_path->sgid = req_msg->alt_remote_gid;
1021                 alt_path->dlid = req_msg->alt_local_lid;
1022                 alt_path->slid = req_msg->alt_remote_lid;
1023                 alt_path->flow_label = cm_req_get_alt_flow_label(req_msg);
1024                 alt_path->hop_limit = req_msg->alt_hop_limit;
1025                 alt_path->traffic_class = req_msg->alt_traffic_class;
1026                 alt_path->reversible = 1;
1027                 alt_path->pkey = req_msg->pkey;
1028                 alt_path->sl = cm_req_get_alt_sl(req_msg);
1029                 alt_path->mtu_selector = IB_SA_EQ;
1030                 alt_path->mtu = cm_req_get_path_mtu(req_msg);
1031                 alt_path->rate_selector = IB_SA_EQ;
1032                 alt_path->rate = cm_req_get_alt_packet_rate(req_msg);
1033                 alt_path->packet_life_time_selector = IB_SA_EQ;
1034                 alt_path->packet_life_time =
1035                         cm_req_get_alt_local_ack_timeout(req_msg);
1036                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1037         }
1038 }
1039
1040 static void cm_format_req_event(struct cm_work *work,
1041                                 struct cm_id_private *cm_id_priv,
1042                                 struct ib_cm_id *listen_id)
1043 {
1044         struct cm_req_msg *req_msg;
1045         struct ib_cm_req_event_param *param;
1046
1047         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1048         param = &work->cm_event.param.req_rcvd;
1049         param->listen_id = listen_id;
1050         param->device = cm_id_priv->av.port->mad_agent->device;
1051         param->port = cm_id_priv->av.port->port_num;
1052         param->primary_path = &work->path[0];
1053         if (req_msg->alt_local_lid)
1054                 param->alternate_path = &work->path[1];
1055         else
1056                 param->alternate_path = NULL;
1057         param->remote_ca_guid = req_msg->local_ca_guid;
1058         param->remote_qkey = be32_to_cpu(req_msg->local_qkey);
1059         param->remote_qpn = be32_to_cpu(cm_req_get_local_qpn(req_msg));
1060         param->qp_type = cm_req_get_qp_type(req_msg);
1061         param->starting_psn = be32_to_cpu(cm_req_get_starting_psn(req_msg));
1062         param->responder_resources = cm_req_get_init_depth(req_msg);
1063         param->initiator_depth = cm_req_get_resp_res(req_msg);
1064         param->local_cm_response_timeout =
1065                                         cm_req_get_remote_resp_timeout(req_msg);
1066         param->flow_control = cm_req_get_flow_ctrl(req_msg);
1067         param->remote_cm_response_timeout =
1068                                         cm_req_get_local_resp_timeout(req_msg);
1069         param->retry_count = cm_req_get_retry_count(req_msg);
1070         param->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1071         param->srq = cm_req_get_srq(req_msg);
1072         work->cm_event.private_data = &req_msg->private_data;
1073 }
1074
1075 static void cm_process_work(struct cm_id_private *cm_id_priv,
1076                             struct cm_work *work)
1077 {
1078         unsigned long flags;
1079         int ret;
1080
1081         /* We will typically only have the current event to report. */
1082         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1083         cm_free_work(work);
1084
1085         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1086                 spin_lock_irqsave(&cm_id_priv->lock, flags);
1087                 work = cm_dequeue_work(cm_id_priv);
1088                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1089                 BUG_ON(!work);
1090                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1091                                                 &work->cm_event);
1092                 cm_free_work(work);
1093         }
1094         cm_deref_id(cm_id_priv);
1095         if (ret)
1096                 ib_destroy_cm_id(&cm_id_priv->id);
1097 }
1098
1099 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1100                           struct cm_id_private *cm_id_priv,
1101                           enum cm_msg_response msg_mraed, u8 service_timeout,
1102                           const void *private_data, u8 private_data_len)
1103 {
1104         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1105         cm_mra_set_msg_mraed(mra_msg, msg_mraed);
1106         mra_msg->local_comm_id = cm_id_priv->id.local_id;
1107         mra_msg->remote_comm_id = cm_id_priv->id.remote_id;
1108         cm_mra_set_service_timeout(mra_msg, service_timeout);
1109
1110         if (private_data && private_data_len)
1111                 memcpy(mra_msg->private_data, private_data, private_data_len);
1112 }
1113
1114 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1115                           struct cm_id_private *cm_id_priv,
1116                           enum ib_cm_rej_reason reason,
1117                           void *ari,
1118                           u8 ari_length,
1119                           const void *private_data,
1120                           u8 private_data_len)
1121 {
1122         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1123         rej_msg->remote_comm_id = cm_id_priv->id.remote_id;
1124
1125         switch(cm_id_priv->id.state) {
1126         case IB_CM_REQ_RCVD:
1127                 rej_msg->local_comm_id = 0;
1128                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1129                 break;
1130         case IB_CM_MRA_REQ_SENT:
1131                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1132                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1133                 break;
1134         case IB_CM_REP_RCVD:
1135         case IB_CM_MRA_REP_SENT:
1136                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1137                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REP);
1138                 break;
1139         default:
1140                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1141                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_OTHER);
1142                 break;
1143         }
1144
1145         rej_msg->reason = cpu_to_be16(reason);
1146         if (ari && ari_length) {
1147                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1148                 memcpy(rej_msg->ari, ari, ari_length);
1149         }
1150
1151         if (private_data && private_data_len)
1152                 memcpy(rej_msg->private_data, private_data, private_data_len);
1153 }
1154
1155 static void cm_dup_req_handler(struct cm_work *work,
1156                                struct cm_id_private *cm_id_priv)
1157 {
1158         struct ib_mad_send_buf *msg = NULL;
1159         struct ib_send_wr *bad_send_wr;
1160         unsigned long flags;
1161         int ret;
1162
1163         /* Quick state check to discard duplicate REQs. */
1164         if (cm_id_priv->id.state == IB_CM_REQ_RCVD)
1165                 return;
1166
1167         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1168         if (ret)
1169                 return;
1170
1171         spin_lock_irqsave(&cm_id_priv->lock, flags);
1172         switch (cm_id_priv->id.state) {
1173         case IB_CM_MRA_REQ_SENT:
1174                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1175                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1176                               cm_id_priv->private_data,
1177                               cm_id_priv->private_data_len);
1178                 break;
1179         case IB_CM_TIMEWAIT:
1180                 cm_format_rej((struct cm_rej_msg *) msg->mad, cm_id_priv,
1181                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0);
1182                 break;
1183         default:
1184                 goto unlock;
1185         }
1186         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1187
1188         ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent, &msg->send_wr,
1189                                &bad_send_wr);
1190         if (ret)
1191                 goto free;
1192         return;
1193
1194 unlock: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1195 free:   cm_free_msg(msg);
1196 }
1197
1198 static struct cm_id_private * cm_match_req(struct cm_work *work,
1199                                            struct cm_id_private *cm_id_priv)
1200 {
1201         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1202         struct cm_timewait_info *timewait_info;
1203         struct cm_req_msg *req_msg;
1204         unsigned long flags;
1205
1206         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1207
1208         /* Check for duplicate REQ and stale connections. */
1209         spin_lock_irqsave(&cm.lock, flags);
1210         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1211         if (!timewait_info)
1212                 timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
1213
1214         if (timewait_info) {
1215                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1216                                            timewait_info->work.remote_id);
1217                 spin_unlock_irqrestore(&cm.lock, flags);
1218                 if (cur_cm_id_priv) {
1219                         cm_dup_req_handler(work, cur_cm_id_priv);
1220                         cm_deref_id(cur_cm_id_priv);
1221                 } else
1222                         cm_issue_rej(work->port, work->mad_recv_wc,
1223                                      IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
1224                                      NULL, 0);
1225                 goto error;
1226         }
1227
1228         /* Find matching listen request. */
1229         listen_cm_id_priv = cm_find_listen(req_msg->service_id);
1230         if (!listen_cm_id_priv) {
1231                 spin_unlock_irqrestore(&cm.lock, flags);
1232                 cm_issue_rej(work->port, work->mad_recv_wc,
1233                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
1234                              NULL, 0);
1235                 goto error;
1236         }
1237         atomic_inc(&listen_cm_id_priv->refcount);
1238         atomic_inc(&cm_id_priv->refcount);
1239         cm_id_priv->id.state = IB_CM_REQ_RCVD;
1240         atomic_inc(&cm_id_priv->work_count);
1241         spin_unlock_irqrestore(&cm.lock, flags);
1242         return listen_cm_id_priv;
1243
1244 error:  cm_cleanup_timewait(cm_id_priv->timewait_info);
1245         return NULL;
1246 }
1247
1248 static int cm_req_handler(struct cm_work *work)
1249 {
1250         struct ib_cm_id *cm_id;
1251         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
1252         struct cm_req_msg *req_msg;
1253         int ret;
1254
1255         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1256
1257         cm_id = ib_create_cm_id(NULL, NULL);
1258         if (IS_ERR(cm_id))
1259                 return PTR_ERR(cm_id);
1260
1261         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1262         cm_id_priv->id.remote_id = req_msg->local_comm_id;
1263         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
1264                                 &cm_id_priv->av);
1265         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1266                                                             id.local_id);
1267         if (IS_ERR(cm_id_priv->timewait_info)) {
1268                 ret = PTR_ERR(cm_id_priv->timewait_info);
1269                 goto error1;
1270         }
1271         cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
1272         cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
1273         cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
1274
1275         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
1276         if (!listen_cm_id_priv) {
1277                 ret = -EINVAL;
1278                 goto error2;
1279         }
1280
1281         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
1282         cm_id_priv->id.context = listen_cm_id_priv->id.context;
1283         cm_id_priv->id.service_id = req_msg->service_id;
1284         cm_id_priv->id.service_mask = __constant_cpu_to_be64(~0ULL);
1285
1286         cm_format_paths_from_req(req_msg, &work->path[0], &work->path[1]);
1287         ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av);
1288         if (ret)
1289                 goto error3;
1290         if (req_msg->alt_local_lid) {
1291                 ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av);
1292                 if (ret)
1293                         goto error3;
1294         }
1295         cm_id_priv->tid = req_msg->hdr.tid;
1296         cm_id_priv->timeout_ms = cm_convert_to_ms(
1297                                         cm_req_get_local_resp_timeout(req_msg));
1298         cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
1299         cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
1300         cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
1301         cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
1302         cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
1303         cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
1304         cm_id_priv->local_ack_timeout =
1305                                 cm_req_get_primary_local_ack_timeout(req_msg);
1306         cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
1307         cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1308
1309         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
1310         cm_process_work(cm_id_priv, work);
1311         cm_deref_id(listen_cm_id_priv);
1312         return 0;
1313
1314 error3: atomic_dec(&cm_id_priv->refcount);
1315         cm_deref_id(listen_cm_id_priv);
1316         cm_cleanup_timewait(cm_id_priv->timewait_info);
1317 error2: kfree(cm_id_priv->timewait_info);
1318 error1: ib_destroy_cm_id(&cm_id_priv->id);
1319         return ret;
1320 }
1321
1322 static void cm_format_rep(struct cm_rep_msg *rep_msg,
1323                           struct cm_id_private *cm_id_priv,
1324                           struct ib_cm_rep_param *param)
1325 {
1326         cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
1327         rep_msg->local_comm_id = cm_id_priv->id.local_id;
1328         rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1329         cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
1330         cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
1331         rep_msg->resp_resources = param->responder_resources;
1332         rep_msg->initiator_depth = param->initiator_depth;
1333         cm_rep_set_target_ack_delay(rep_msg, param->target_ack_delay);
1334         cm_rep_set_failover(rep_msg, param->failover_accepted);
1335         cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
1336         cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
1337         cm_rep_set_srq(rep_msg, param->srq);
1338         rep_msg->local_ca_guid = cm_id_priv->av.port->cm_dev->ca_guid;
1339
1340         if (param->private_data && param->private_data_len)
1341                 memcpy(rep_msg->private_data, param->private_data,
1342                        param->private_data_len);
1343 }
1344
1345 int ib_send_cm_rep(struct ib_cm_id *cm_id,
1346                    struct ib_cm_rep_param *param)
1347 {
1348         struct cm_id_private *cm_id_priv;
1349         struct ib_mad_send_buf *msg;
1350         struct cm_rep_msg *rep_msg;
1351         struct ib_send_wr *bad_send_wr;
1352         unsigned long flags;
1353         int ret;
1354
1355         if (param->private_data &&
1356             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
1357                 return -EINVAL;
1358
1359         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1360         spin_lock_irqsave(&cm_id_priv->lock, flags);
1361         if (cm_id->state != IB_CM_REQ_RCVD &&
1362             cm_id->state != IB_CM_MRA_REQ_SENT) {
1363                 ret = -EINVAL;
1364                 goto out;
1365         }
1366
1367         ret = cm_alloc_msg(cm_id_priv, &msg);
1368         if (ret)
1369                 goto out;
1370
1371         rep_msg = (struct cm_rep_msg *) msg->mad;
1372         cm_format_rep(rep_msg, cm_id_priv, param);
1373         msg->send_wr.wr.ud.timeout_ms = cm_id_priv->timeout_ms;
1374         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
1375
1376         ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
1377                                &msg->send_wr, &bad_send_wr);
1378         if (ret) {
1379                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1380                 cm_free_msg(msg);
1381                 return ret;
1382         }
1383
1384         cm_id->state = IB_CM_REP_SENT;
1385         cm_id_priv->msg = msg;
1386         cm_id_priv->initiator_depth = param->initiator_depth;
1387         cm_id_priv->responder_resources = param->responder_resources;
1388         cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
1389         cm_id_priv->local_qpn = cm_rep_get_local_qpn(rep_msg);
1390
1391 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1392         return ret;
1393 }
1394 EXPORT_SYMBOL(ib_send_cm_rep);
1395
1396 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
1397                           struct cm_id_private *cm_id_priv,
1398                           const void *private_data,
1399                           u8 private_data_len)
1400 {
1401         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
1402         rtu_msg->local_comm_id = cm_id_priv->id.local_id;
1403         rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
1404
1405         if (private_data && private_data_len)
1406                 memcpy(rtu_msg->private_data, private_data, private_data_len);
1407 }
1408
1409 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
1410                    const void *private_data,
1411                    u8 private_data_len)
1412 {
1413         struct cm_id_private *cm_id_priv;
1414         struct ib_mad_send_buf *msg;
1415         struct ib_send_wr *bad_send_wr;
1416         unsigned long flags;
1417         void *data;
1418         int ret;
1419
1420         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
1421                 return -EINVAL;
1422
1423         data = cm_copy_private_data(private_data, private_data_len);
1424         if (IS_ERR(data))
1425                 return PTR_ERR(data);
1426
1427         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1428         spin_lock_irqsave(&cm_id_priv->lock, flags);
1429         if (cm_id->state != IB_CM_REP_RCVD &&
1430             cm_id->state != IB_CM_MRA_REP_SENT) {
1431                 ret = -EINVAL;
1432                 goto error;
1433         }
1434
1435         ret = cm_alloc_msg(cm_id_priv, &msg);
1436         if (ret)
1437                 goto error;
1438
1439         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1440                       private_data, private_data_len);
1441
1442         ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
1443                                &msg->send_wr, &bad_send_wr);
1444         if (ret) {
1445                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1446                 cm_free_msg(msg);
1447                 kfree(data);
1448                 return ret;
1449         }
1450
1451         cm_id->state = IB_CM_ESTABLISHED;
1452         cm_set_private_data(cm_id_priv, data, private_data_len);
1453         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1454         return 0;
1455
1456 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1457         kfree(data);
1458         return ret;
1459 }
1460 EXPORT_SYMBOL(ib_send_cm_rtu);
1461
1462 static void cm_format_rep_event(struct cm_work *work)
1463 {
1464         struct cm_rep_msg *rep_msg;
1465         struct ib_cm_rep_event_param *param;
1466
1467         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1468         param = &work->cm_event.param.rep_rcvd;
1469         param->remote_ca_guid = rep_msg->local_ca_guid;
1470         param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
1471         param->remote_qpn = be32_to_cpu(cm_rep_get_local_qpn(rep_msg));
1472         param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
1473         param->responder_resources = rep_msg->initiator_depth;
1474         param->initiator_depth = rep_msg->resp_resources;
1475         param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1476         param->failover_accepted = cm_rep_get_failover(rep_msg);
1477         param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
1478         param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1479         param->srq = cm_rep_get_srq(rep_msg);
1480         work->cm_event.private_data = &rep_msg->private_data;
1481 }
1482
1483 static void cm_dup_rep_handler(struct cm_work *work)
1484 {
1485         struct cm_id_private *cm_id_priv;
1486         struct cm_rep_msg *rep_msg;
1487         struct ib_mad_send_buf *msg = NULL;
1488         struct ib_send_wr *bad_send_wr;
1489         unsigned long flags;
1490         int ret;
1491
1492         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
1493         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
1494                                    rep_msg->local_comm_id);
1495         if (!cm_id_priv)
1496                 return;
1497
1498         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1499         if (ret)
1500                 goto deref;
1501
1502         spin_lock_irqsave(&cm_id_priv->lock, flags);
1503         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
1504                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1505                               cm_id_priv->private_data,
1506                               cm_id_priv->private_data_len);
1507         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
1508                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1509                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
1510                               cm_id_priv->private_data,
1511                               cm_id_priv->private_data_len);
1512         else
1513                 goto unlock;
1514         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1515
1516         ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent, &msg->send_wr,
1517                                &bad_send_wr);
1518         if (ret)
1519                 goto free;
1520         goto deref;
1521
1522 unlock: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1523 free:   cm_free_msg(msg);
1524 deref:  cm_deref_id(cm_id_priv);
1525 }
1526
1527 static int cm_rep_handler(struct cm_work *work)
1528 {
1529         struct cm_id_private *cm_id_priv;
1530         struct cm_rep_msg *rep_msg;
1531         unsigned long flags;
1532         int ret;
1533
1534         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1535         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
1536         if (!cm_id_priv) {
1537                 cm_dup_rep_handler(work);
1538                 return -EINVAL;
1539         }
1540
1541         cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
1542         cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
1543         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_local_qpn(rep_msg);
1544
1545         spin_lock_irqsave(&cm.lock, flags);
1546         /* Check for duplicate REP. */
1547         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
1548                 spin_unlock_irqrestore(&cm.lock, flags);
1549                 ret = -EINVAL;
1550                 goto error;
1551         }
1552         /* Check for a stale connection. */
1553         if (cm_insert_remote_qpn(cm_id_priv->timewait_info)) {
1554                 spin_unlock_irqrestore(&cm.lock, flags);
1555                 cm_issue_rej(work->port, work->mad_recv_wc,
1556                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
1557                              NULL, 0);
1558                 ret = -EINVAL;
1559                 goto error;
1560         }
1561         spin_unlock_irqrestore(&cm.lock, flags);
1562
1563         cm_format_rep_event(work);
1564
1565         spin_lock_irqsave(&cm_id_priv->lock, flags);
1566         switch (cm_id_priv->id.state) {
1567         case IB_CM_REQ_SENT:
1568         case IB_CM_MRA_REQ_RCVD:
1569                 break;
1570         default:
1571                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1572                 ret = -EINVAL;
1573                 goto error;
1574         }
1575         cm_id_priv->id.state = IB_CM_REP_RCVD;
1576         cm_id_priv->id.remote_id = rep_msg->local_comm_id;
1577         cm_id_priv->remote_qpn = cm_rep_get_local_qpn(rep_msg);
1578         cm_id_priv->initiator_depth = rep_msg->resp_resources;
1579         cm_id_priv->responder_resources = rep_msg->initiator_depth;
1580         cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
1581         cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1582
1583         /* todo: handle peer_to_peer */
1584
1585         ib_cancel_mad(cm_id_priv->av.port->mad_agent,
1586                       (unsigned long) cm_id_priv->msg);
1587         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1588         if (!ret)
1589                 list_add_tail(&work->list, &cm_id_priv->work_list);
1590         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1591
1592         if (ret)
1593                 cm_process_work(cm_id_priv, work);
1594         else
1595                 cm_deref_id(cm_id_priv);
1596         return 0;
1597
1598 error:  cm_cleanup_timewait(cm_id_priv->timewait_info);
1599         cm_deref_id(cm_id_priv);
1600         return ret;
1601 }
1602
1603 static int cm_establish_handler(struct cm_work *work)
1604 {
1605         struct cm_id_private *cm_id_priv;
1606         unsigned long flags;
1607         int ret;
1608
1609         /* See comment in ib_cm_establish about lookup. */
1610         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
1611         if (!cm_id_priv)
1612                 return -EINVAL;
1613
1614         spin_lock_irqsave(&cm_id_priv->lock, flags);
1615         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
1616                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1617                 goto out;
1618         }
1619
1620         ib_cancel_mad(cm_id_priv->av.port->mad_agent,
1621                       (unsigned long) cm_id_priv->msg);
1622         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1623         if (!ret)
1624                 list_add_tail(&work->list, &cm_id_priv->work_list);
1625         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1626
1627         if (ret)
1628                 cm_process_work(cm_id_priv, work);
1629         else
1630                 cm_deref_id(cm_id_priv);
1631         return 0;
1632 out:
1633         cm_deref_id(cm_id_priv);
1634         return -EINVAL;
1635 }
1636
1637 static int cm_rtu_handler(struct cm_work *work)
1638 {
1639         struct cm_id_private *cm_id_priv;
1640         struct cm_rtu_msg *rtu_msg;
1641         unsigned long flags;
1642         int ret;
1643
1644         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
1645         cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
1646                                    rtu_msg->local_comm_id);
1647         if (!cm_id_priv)
1648                 return -EINVAL;
1649
1650         work->cm_event.private_data = &rtu_msg->private_data;
1651
1652         spin_lock_irqsave(&cm_id_priv->lock, flags);
1653         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
1654             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
1655                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1656                 goto out;
1657         }
1658         cm_id_priv->id.state = IB_CM_ESTABLISHED;
1659
1660         ib_cancel_mad(cm_id_priv->av.port->mad_agent,
1661                       (unsigned long) cm_id_priv->msg);
1662         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1663         if (!ret)
1664                 list_add_tail(&work->list, &cm_id_priv->work_list);
1665         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1666
1667         if (ret)
1668                 cm_process_work(cm_id_priv, work);
1669         else
1670                 cm_deref_id(cm_id_priv);
1671         return 0;
1672 out:
1673         cm_deref_id(cm_id_priv);
1674         return -EINVAL;
1675 }
1676
1677 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
1678                           struct cm_id_private *cm_id_priv,
1679                           const void *private_data,
1680                           u8 private_data_len)
1681 {
1682         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
1683                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
1684         dreq_msg->local_comm_id = cm_id_priv->id.local_id;
1685         dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
1686         cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
1687
1688         if (private_data && private_data_len)
1689                 memcpy(dreq_msg->private_data, private_data, private_data_len);
1690 }
1691
1692 int ib_send_cm_dreq(struct ib_cm_id *cm_id,
1693                     const void *private_data,
1694                     u8 private_data_len)
1695 {
1696         struct cm_id_private *cm_id_priv;
1697         struct ib_mad_send_buf *msg;
1698         struct ib_send_wr *bad_send_wr;
1699         unsigned long flags;
1700         int ret;
1701
1702         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
1703                 return -EINVAL;
1704
1705         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1706         spin_lock_irqsave(&cm_id_priv->lock, flags);
1707         if (cm_id->state != IB_CM_ESTABLISHED) {
1708                 ret = -EINVAL;
1709                 goto out;
1710         }
1711
1712         ret = cm_alloc_msg(cm_id_priv, &msg);
1713         if (ret) {
1714                 cm_enter_timewait(cm_id_priv);
1715                 goto out;
1716         }
1717
1718         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
1719                        private_data, private_data_len);
1720         msg->send_wr.wr.ud.timeout_ms = cm_id_priv->timeout_ms;
1721         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
1722
1723         ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
1724                                &msg->send_wr, &bad_send_wr);
1725         if (ret) {
1726                 cm_enter_timewait(cm_id_priv);
1727                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1728                 cm_free_msg(msg);
1729                 return ret;
1730         }
1731
1732         cm_id->state = IB_CM_DREQ_SENT;
1733         cm_id_priv->msg = msg;
1734 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1735         return ret;
1736 }
1737 EXPORT_SYMBOL(ib_send_cm_dreq);
1738
1739 static void cm_format_drep(struct cm_drep_msg *drep_msg,
1740                           struct cm_id_private *cm_id_priv,
1741                           const void *private_data,
1742                           u8 private_data_len)
1743 {
1744         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
1745         drep_msg->local_comm_id = cm_id_priv->id.local_id;
1746         drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1747
1748         if (private_data && private_data_len)
1749                 memcpy(drep_msg->private_data, private_data, private_data_len);
1750 }
1751
1752 int ib_send_cm_drep(struct ib_cm_id *cm_id,
1753                     const void *private_data,
1754                     u8 private_data_len)
1755 {
1756         struct cm_id_private *cm_id_priv;
1757         struct ib_mad_send_buf *msg;
1758         struct ib_send_wr *bad_send_wr;
1759         unsigned long flags;
1760         void *data;
1761         int ret;
1762
1763         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
1764                 return -EINVAL;
1765
1766         data = cm_copy_private_data(private_data, private_data_len);
1767         if (IS_ERR(data))
1768                 return PTR_ERR(data);
1769
1770         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1771         spin_lock_irqsave(&cm_id_priv->lock, flags);
1772         if (cm_id->state != IB_CM_DREQ_RCVD) {
1773                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1774                 kfree(data);
1775                 return -EINVAL;
1776         }
1777
1778         cm_set_private_data(cm_id_priv, data, private_data_len);
1779         cm_enter_timewait(cm_id_priv);
1780
1781         ret = cm_alloc_msg(cm_id_priv, &msg);
1782         if (ret)
1783                 goto out;
1784
1785         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
1786                        private_data, private_data_len);
1787
1788         ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent, &msg->send_wr,
1789                                &bad_send_wr);
1790         if (ret) {
1791                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1792                 cm_free_msg(msg);
1793                 return ret;
1794         }
1795
1796 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1797         return ret;
1798 }
1799 EXPORT_SYMBOL(ib_send_cm_drep);
1800
1801 static int cm_dreq_handler(struct cm_work *work)
1802 {
1803         struct cm_id_private *cm_id_priv;
1804         struct cm_dreq_msg *dreq_msg;
1805         struct ib_mad_send_buf *msg = NULL;
1806         struct ib_send_wr *bad_send_wr;
1807         unsigned long flags;
1808         int ret;
1809
1810         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
1811         cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
1812                                    dreq_msg->local_comm_id);
1813         if (!cm_id_priv)
1814                 return -EINVAL;
1815
1816         work->cm_event.private_data = &dreq_msg->private_data;
1817
1818         spin_lock_irqsave(&cm_id_priv->lock, flags);
1819         if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
1820                 goto unlock;
1821
1822         switch (cm_id_priv->id.state) {
1823         case IB_CM_REP_SENT:
1824         case IB_CM_DREQ_SENT:
1825                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
1826                               (unsigned long) cm_id_priv->msg);
1827                 break;
1828         case IB_CM_ESTABLISHED:
1829         case IB_CM_MRA_REP_RCVD:
1830                 break;
1831         case IB_CM_TIMEWAIT:
1832                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
1833                         goto unlock;
1834
1835                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
1836                                cm_id_priv->private_data,
1837                                cm_id_priv->private_data_len);
1838                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1839
1840                 if (ib_post_send_mad(cm_id_priv->av.port->mad_agent,
1841                                      &msg->send_wr, &bad_send_wr))
1842                         cm_free_msg(msg);
1843                 goto deref;
1844         default:
1845                 goto unlock;
1846         }
1847         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
1848         cm_id_priv->tid = dreq_msg->hdr.tid;
1849         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1850         if (!ret)
1851                 list_add_tail(&work->list, &cm_id_priv->work_list);
1852         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1853
1854         if (ret)
1855                 cm_process_work(cm_id_priv, work);
1856         else
1857                 cm_deref_id(cm_id_priv);
1858         return 0;
1859
1860 unlock: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1861 deref:  cm_deref_id(cm_id_priv);
1862         return -EINVAL;
1863 }
1864
1865 static int cm_drep_handler(struct cm_work *work)
1866 {
1867         struct cm_id_private *cm_id_priv;
1868         struct cm_drep_msg *drep_msg;
1869         unsigned long flags;
1870         int ret;
1871
1872         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
1873         cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
1874                                    drep_msg->local_comm_id);
1875         if (!cm_id_priv)
1876                 return -EINVAL;
1877
1878         work->cm_event.private_data = &drep_msg->private_data;
1879
1880         spin_lock_irqsave(&cm_id_priv->lock, flags);
1881         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
1882             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
1883                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1884                 goto out;
1885         }
1886         cm_enter_timewait(cm_id_priv);
1887
1888         ib_cancel_mad(cm_id_priv->av.port->mad_agent,
1889                       (unsigned long) cm_id_priv->msg);
1890         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1891         if (!ret)
1892                 list_add_tail(&work->list, &cm_id_priv->work_list);
1893         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1894
1895         if (ret)
1896                 cm_process_work(cm_id_priv, work);
1897         else
1898                 cm_deref_id(cm_id_priv);
1899         return 0;
1900 out:
1901         cm_deref_id(cm_id_priv);
1902         return -EINVAL;
1903 }
1904
1905 int ib_send_cm_rej(struct ib_cm_id *cm_id,
1906                    enum ib_cm_rej_reason reason,
1907                    void *ari,
1908                    u8 ari_length,
1909                    const void *private_data,
1910                    u8 private_data_len)
1911 {
1912         struct cm_id_private *cm_id_priv;
1913         struct ib_mad_send_buf *msg;
1914         struct ib_send_wr *bad_send_wr;
1915         unsigned long flags;
1916         int ret;
1917
1918         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
1919             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
1920                 return -EINVAL;
1921
1922         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1923
1924         spin_lock_irqsave(&cm_id_priv->lock, flags);
1925         switch (cm_id->state) {
1926         case IB_CM_REQ_SENT:
1927         case IB_CM_MRA_REQ_RCVD:
1928         case IB_CM_REQ_RCVD:
1929         case IB_CM_MRA_REQ_SENT:
1930         case IB_CM_REP_RCVD:
1931         case IB_CM_MRA_REP_SENT:
1932                 ret = cm_alloc_msg(cm_id_priv, &msg);
1933                 if (!ret)
1934                         cm_format_rej((struct cm_rej_msg *) msg->mad,
1935                                       cm_id_priv, reason, ari, ari_length,
1936                                       private_data, private_data_len);
1937
1938                 cm_reset_to_idle(cm_id_priv);
1939                 break;
1940         case IB_CM_REP_SENT:
1941         case IB_CM_MRA_REP_RCVD:
1942                 ret = cm_alloc_msg(cm_id_priv, &msg);
1943                 if (!ret)
1944                         cm_format_rej((struct cm_rej_msg *) msg->mad,
1945                                       cm_id_priv, reason, ari, ari_length,
1946                                       private_data, private_data_len);
1947
1948                 cm_enter_timewait(cm_id_priv);
1949                 break;
1950         default:
1951                 ret = -EINVAL;
1952                 goto out;
1953         }
1954
1955         if (ret)
1956                 goto out;
1957
1958         ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
1959                                &msg->send_wr, &bad_send_wr);
1960         if (ret)
1961                 cm_free_msg(msg);
1962
1963 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1964         return ret;
1965 }
1966 EXPORT_SYMBOL(ib_send_cm_rej);
1967
1968 static void cm_format_rej_event(struct cm_work *work)
1969 {
1970         struct cm_rej_msg *rej_msg;
1971         struct ib_cm_rej_event_param *param;
1972
1973         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
1974         param = &work->cm_event.param.rej_rcvd;
1975         param->ari = rej_msg->ari;
1976         param->ari_length = cm_rej_get_reject_info_len(rej_msg);
1977         param->reason = __be16_to_cpu(rej_msg->reason);
1978         work->cm_event.private_data = &rej_msg->private_data;
1979 }
1980
1981 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
1982 {
1983         struct cm_timewait_info *timewait_info;
1984         struct cm_id_private *cm_id_priv;
1985         unsigned long flags;
1986         __be32 remote_id;
1987
1988         remote_id = rej_msg->local_comm_id;
1989
1990         if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_TIMEOUT) {
1991                 spin_lock_irqsave(&cm.lock, flags);
1992                 timewait_info = cm_find_remote_id( *((__be64 *) rej_msg->ari),
1993                                                   remote_id);
1994                 if (!timewait_info) {
1995                         spin_unlock_irqrestore(&cm.lock, flags);
1996                         return NULL;
1997                 }
1998                 cm_id_priv = idr_find(&cm.local_id_table,
1999                                       (__force int) timewait_info->work.local_id);
2000                 if (cm_id_priv) {
2001                         if (cm_id_priv->id.remote_id == remote_id)
2002                                 atomic_inc(&cm_id_priv->refcount);
2003                         else
2004                                 cm_id_priv = NULL;
2005                 }
2006                 spin_unlock_irqrestore(&cm.lock, flags);
2007         } else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
2008                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
2009         else
2010                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
2011
2012         return cm_id_priv;
2013 }
2014
2015 static int cm_rej_handler(struct cm_work *work)
2016 {
2017         struct cm_id_private *cm_id_priv;
2018         struct cm_rej_msg *rej_msg;
2019         unsigned long flags;
2020         int ret;
2021
2022         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2023         cm_id_priv = cm_acquire_rejected_id(rej_msg);
2024         if (!cm_id_priv)
2025                 return -EINVAL;
2026
2027         cm_format_rej_event(work);
2028
2029         spin_lock_irqsave(&cm_id_priv->lock, flags);
2030         switch (cm_id_priv->id.state) {
2031         case IB_CM_REQ_SENT:
2032         case IB_CM_MRA_REQ_RCVD:
2033         case IB_CM_REP_SENT:
2034         case IB_CM_MRA_REP_RCVD:
2035                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
2036                               (unsigned long) cm_id_priv->msg);
2037                 /* fall through */
2038         case IB_CM_REQ_RCVD:
2039         case IB_CM_MRA_REQ_SENT:
2040                 if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_STALE_CONN)
2041                         cm_enter_timewait(cm_id_priv);
2042                 else
2043                         cm_reset_to_idle(cm_id_priv);
2044                 break;
2045         case IB_CM_DREQ_SENT:
2046                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
2047                               (unsigned long) cm_id_priv->msg);
2048                 /* fall through */
2049         case IB_CM_REP_RCVD:
2050         case IB_CM_MRA_REP_SENT:
2051         case IB_CM_ESTABLISHED:
2052                 cm_enter_timewait(cm_id_priv);
2053                 break;
2054         default:
2055                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2056                 ret = -EINVAL;
2057                 goto out;
2058         }
2059
2060         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2061         if (!ret)
2062                 list_add_tail(&work->list, &cm_id_priv->work_list);
2063         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2064
2065         if (ret)
2066                 cm_process_work(cm_id_priv, work);
2067         else
2068                 cm_deref_id(cm_id_priv);
2069         return 0;
2070 out:
2071         cm_deref_id(cm_id_priv);
2072         return -EINVAL;
2073 }
2074
2075 int ib_send_cm_mra(struct ib_cm_id *cm_id,
2076                    u8 service_timeout,
2077                    const void *private_data,
2078                    u8 private_data_len)
2079 {
2080         struct cm_id_private *cm_id_priv;
2081         struct ib_mad_send_buf *msg;
2082         struct ib_send_wr *bad_send_wr;
2083         void *data;
2084         unsigned long flags;
2085         int ret;
2086
2087         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
2088                 return -EINVAL;
2089
2090         data = cm_copy_private_data(private_data, private_data_len);
2091         if (IS_ERR(data))
2092                 return PTR_ERR(data);
2093
2094         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2095
2096         spin_lock_irqsave(&cm_id_priv->lock, flags);
2097         switch(cm_id_priv->id.state) {
2098         case IB_CM_REQ_RCVD:
2099                 ret = cm_alloc_msg(cm_id_priv, &msg);
2100                 if (ret)
2101                         goto error1;
2102
2103                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2104                               CM_MSG_RESPONSE_REQ, service_timeout,
2105                               private_data, private_data_len);
2106                 ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
2107                                        &msg->send_wr, &bad_send_wr);
2108                 if (ret)
2109                         goto error2;
2110                 cm_id->state = IB_CM_MRA_REQ_SENT;
2111                 break;
2112         case IB_CM_REP_RCVD:
2113                 ret = cm_alloc_msg(cm_id_priv, &msg);
2114                 if (ret)
2115                         goto error1;
2116
2117                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2118                               CM_MSG_RESPONSE_REP, service_timeout,
2119                               private_data, private_data_len);
2120                 ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
2121                                        &msg->send_wr, &bad_send_wr);
2122                 if (ret)
2123                         goto error2;
2124                 cm_id->state = IB_CM_MRA_REP_SENT;
2125                 break;
2126         case IB_CM_ESTABLISHED:
2127                 ret = cm_alloc_msg(cm_id_priv, &msg);
2128                 if (ret)
2129                         goto error1;
2130
2131                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2132                               CM_MSG_RESPONSE_OTHER, service_timeout,
2133                               private_data, private_data_len);
2134                 ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
2135                                        &msg->send_wr, &bad_send_wr);
2136                 if (ret)
2137                         goto error2;
2138                 cm_id->lap_state = IB_CM_MRA_LAP_SENT;
2139                 break;
2140         default:
2141                 ret = -EINVAL;
2142                 goto error1;
2143         }
2144         cm_id_priv->service_timeout = service_timeout;
2145         cm_set_private_data(cm_id_priv, data, private_data_len);
2146         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2147         return 0;
2148
2149 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2150         kfree(data);
2151         return ret;
2152
2153 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2154         kfree(data);
2155         cm_free_msg(msg);
2156         return ret;
2157 }
2158 EXPORT_SYMBOL(ib_send_cm_mra);
2159
2160 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2161 {
2162         switch (cm_mra_get_msg_mraed(mra_msg)) {
2163         case CM_MSG_RESPONSE_REQ:
2164                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2165         case CM_MSG_RESPONSE_REP:
2166         case CM_MSG_RESPONSE_OTHER:
2167                 return cm_acquire_id(mra_msg->remote_comm_id,
2168                                      mra_msg->local_comm_id);
2169         default:
2170                 return NULL;
2171         }
2172 }
2173
2174 static int cm_mra_handler(struct cm_work *work)
2175 {
2176         struct cm_id_private *cm_id_priv;
2177         struct cm_mra_msg *mra_msg;
2178         unsigned long flags;
2179         int timeout, ret;
2180
2181         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2182         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2183         if (!cm_id_priv)
2184                 return -EINVAL;
2185
2186         work->cm_event.private_data = &mra_msg->private_data;
2187         work->cm_event.param.mra_rcvd.service_timeout =
2188                                         cm_mra_get_service_timeout(mra_msg);
2189         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2190                   cm_convert_to_ms(cm_id_priv->av.packet_life_time);
2191
2192         spin_lock_irqsave(&cm_id_priv->lock, flags);
2193         switch (cm_id_priv->id.state) {
2194         case IB_CM_REQ_SENT:
2195                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2196                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2197                                   (unsigned long) cm_id_priv->msg, timeout))
2198                         goto out;
2199                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2200                 break;
2201         case IB_CM_REP_SENT:
2202                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2203                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2204                                   (unsigned long) cm_id_priv->msg, timeout))
2205                         goto out;
2206                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2207                 break;
2208         case IB_CM_ESTABLISHED:
2209                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2210                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2211                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2212                                   (unsigned long) cm_id_priv->msg, timeout))
2213                         goto out;
2214                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
2215                 break;
2216         default:
2217                 goto out;
2218         }
2219
2220         cm_id_priv->msg->context[1] = (void *) (unsigned long)
2221                                       cm_id_priv->id.state;
2222         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2223         if (!ret)
2224                 list_add_tail(&work->list, &cm_id_priv->work_list);
2225         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2226
2227         if (ret)
2228                 cm_process_work(cm_id_priv, work);
2229         else
2230                 cm_deref_id(cm_id_priv);
2231         return 0;
2232 out:
2233         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2234         cm_deref_id(cm_id_priv);
2235         return -EINVAL;
2236 }
2237
2238 static void cm_format_lap(struct cm_lap_msg *lap_msg,
2239                           struct cm_id_private *cm_id_priv,
2240                           struct ib_sa_path_rec *alternate_path,
2241                           const void *private_data,
2242                           u8 private_data_len)
2243 {
2244         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
2245                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
2246         lap_msg->local_comm_id = cm_id_priv->id.local_id;
2247         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
2248         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
2249         /* todo: need remote CM response timeout */
2250         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
2251         lap_msg->alt_local_lid = alternate_path->slid;
2252         lap_msg->alt_remote_lid = alternate_path->dlid;
2253         lap_msg->alt_local_gid = alternate_path->sgid;
2254         lap_msg->alt_remote_gid = alternate_path->dgid;
2255         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
2256         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
2257         lap_msg->alt_hop_limit = alternate_path->hop_limit;
2258         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
2259         cm_lap_set_sl(lap_msg, alternate_path->sl);
2260         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
2261         cm_lap_set_local_ack_timeout(lap_msg,
2262                 min(31, alternate_path->packet_life_time + 1));
2263
2264         if (private_data && private_data_len)
2265                 memcpy(lap_msg->private_data, private_data, private_data_len);
2266 }
2267
2268 int ib_send_cm_lap(struct ib_cm_id *cm_id,
2269                    struct ib_sa_path_rec *alternate_path,
2270                    const void *private_data,
2271                    u8 private_data_len)
2272 {
2273         struct cm_id_private *cm_id_priv;
2274         struct ib_mad_send_buf *msg;
2275         struct ib_send_wr *bad_send_wr;
2276         unsigned long flags;
2277         int ret;
2278
2279         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
2280                 return -EINVAL;
2281
2282         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2283         spin_lock_irqsave(&cm_id_priv->lock, flags);
2284         if (cm_id->state != IB_CM_ESTABLISHED ||
2285             cm_id->lap_state != IB_CM_LAP_IDLE) {
2286                 ret = -EINVAL;
2287                 goto out;
2288         }
2289
2290         ret = cm_alloc_msg(cm_id_priv, &msg);
2291         if (ret)
2292                 goto out;
2293
2294         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
2295                       alternate_path, private_data, private_data_len);
2296         msg->send_wr.wr.ud.timeout_ms = cm_id_priv->timeout_ms;
2297         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
2298
2299         ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
2300                                &msg->send_wr, &bad_send_wr);
2301         if (ret) {
2302                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2303                 cm_free_msg(msg);
2304                 return ret;
2305         }
2306
2307         cm_id->lap_state = IB_CM_LAP_SENT;
2308         cm_id_priv->msg = msg;
2309
2310 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2311         return ret;
2312 }
2313 EXPORT_SYMBOL(ib_send_cm_lap);
2314
2315 static void cm_format_path_from_lap(struct ib_sa_path_rec *path,
2316                                     struct cm_lap_msg *lap_msg)
2317 {
2318         memset(path, 0, sizeof *path);
2319         path->dgid = lap_msg->alt_local_gid;
2320         path->sgid = lap_msg->alt_remote_gid;
2321         path->dlid = lap_msg->alt_local_lid;
2322         path->slid = lap_msg->alt_remote_lid;
2323         path->flow_label = cm_lap_get_flow_label(lap_msg);
2324         path->hop_limit = lap_msg->alt_hop_limit;
2325         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
2326         path->reversible = 1;
2327         /* pkey is same as in REQ */
2328         path->sl = cm_lap_get_sl(lap_msg);
2329         path->mtu_selector = IB_SA_EQ;
2330         /* mtu is same as in REQ */
2331         path->rate_selector = IB_SA_EQ;
2332         path->rate = cm_lap_get_packet_rate(lap_msg);
2333         path->packet_life_time_selector = IB_SA_EQ;
2334         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
2335         path->packet_life_time -= (path->packet_life_time > 0);
2336 }
2337
2338 static int cm_lap_handler(struct cm_work *work)
2339 {
2340         struct cm_id_private *cm_id_priv;
2341         struct cm_lap_msg *lap_msg;
2342         struct ib_cm_lap_event_param *param;
2343         struct ib_mad_send_buf *msg = NULL;
2344         struct ib_send_wr *bad_send_wr;
2345         unsigned long flags;
2346         int ret;
2347
2348         /* todo: verify LAP request and send reject APR if invalid. */
2349         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
2350         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
2351                                    lap_msg->local_comm_id);
2352         if (!cm_id_priv)
2353                 return -EINVAL;
2354
2355         param = &work->cm_event.param.lap_rcvd;
2356         param->alternate_path = &work->path[0];
2357         cm_format_path_from_lap(param->alternate_path, lap_msg);
2358         work->cm_event.private_data = &lap_msg->private_data;
2359
2360         spin_lock_irqsave(&cm_id_priv->lock, flags);
2361         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
2362                 goto unlock;
2363
2364         switch (cm_id_priv->id.lap_state) {
2365         case IB_CM_LAP_IDLE:
2366                 break;
2367         case IB_CM_MRA_LAP_SENT:
2368                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2369                         goto unlock;
2370
2371                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2372                               CM_MSG_RESPONSE_OTHER,
2373                               cm_id_priv->service_timeout,
2374                               cm_id_priv->private_data,
2375                               cm_id_priv->private_data_len);
2376                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2377
2378                 if (ib_post_send_mad(cm_id_priv->av.port->mad_agent,
2379                                      &msg->send_wr, &bad_send_wr))
2380                         cm_free_msg(msg);
2381                 goto deref;
2382         default:
2383                 goto unlock;
2384         }
2385
2386         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
2387         cm_id_priv->tid = lap_msg->hdr.tid;
2388         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2389         if (!ret)
2390                 list_add_tail(&work->list, &cm_id_priv->work_list);
2391         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2392
2393         if (ret)
2394                 cm_process_work(cm_id_priv, work);
2395         else
2396                 cm_deref_id(cm_id_priv);
2397         return 0;
2398
2399 unlock: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2400 deref:  cm_deref_id(cm_id_priv);
2401         return -EINVAL;
2402 }
2403
2404 static void cm_format_apr(struct cm_apr_msg *apr_msg,
2405                           struct cm_id_private *cm_id_priv,
2406                           enum ib_cm_apr_status status,
2407                           void *info,
2408                           u8 info_length,
2409                           const void *private_data,
2410                           u8 private_data_len)
2411 {
2412         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
2413         apr_msg->local_comm_id = cm_id_priv->id.local_id;
2414         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
2415         apr_msg->ap_status = (u8) status;
2416
2417         if (info && info_length) {
2418                 apr_msg->info_length = info_length;
2419                 memcpy(apr_msg->info, info, info_length);
2420         }
2421
2422         if (private_data && private_data_len)
2423                 memcpy(apr_msg->private_data, private_data, private_data_len);
2424 }
2425
2426 int ib_send_cm_apr(struct ib_cm_id *cm_id,
2427                    enum ib_cm_apr_status status,
2428                    void *info,
2429                    u8 info_length,
2430                    const void *private_data,
2431                    u8 private_data_len)
2432 {
2433         struct cm_id_private *cm_id_priv;
2434         struct ib_mad_send_buf *msg;
2435         struct ib_send_wr *bad_send_wr;
2436         unsigned long flags;
2437         int ret;
2438
2439         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
2440             (info && info_length > IB_CM_APR_INFO_LENGTH))
2441                 return -EINVAL;
2442
2443         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2444         spin_lock_irqsave(&cm_id_priv->lock, flags);
2445         if (cm_id->state != IB_CM_ESTABLISHED ||
2446             (cm_id->lap_state != IB_CM_LAP_RCVD &&
2447              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
2448                 ret = -EINVAL;
2449                 goto out;
2450         }
2451
2452         ret = cm_alloc_msg(cm_id_priv, &msg);
2453         if (ret)
2454                 goto out;
2455
2456         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
2457                       info, info_length, private_data, private_data_len);
2458         ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
2459                                &msg->send_wr, &bad_send_wr);
2460         if (ret) {
2461                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2462                 cm_free_msg(msg);
2463                 return ret;
2464         }
2465
2466         cm_id->lap_state = IB_CM_LAP_IDLE;
2467 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2468         return ret;
2469 }
2470 EXPORT_SYMBOL(ib_send_cm_apr);
2471
2472 static int cm_apr_handler(struct cm_work *work)
2473 {
2474         struct cm_id_private *cm_id_priv;
2475         struct cm_apr_msg *apr_msg;
2476         unsigned long flags;
2477         int ret;
2478
2479         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
2480         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
2481                                    apr_msg->local_comm_id);
2482         if (!cm_id_priv)
2483                 return -EINVAL; /* Unmatched reply. */
2484
2485         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
2486         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
2487         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
2488         work->cm_event.private_data = &apr_msg->private_data;
2489
2490         spin_lock_irqsave(&cm_id_priv->lock, flags);
2491         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
2492             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
2493              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
2494                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2495                 goto out;
2496         }
2497         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
2498         ib_cancel_mad(cm_id_priv->av.port->mad_agent,
2499                       (unsigned long) cm_id_priv->msg);
2500         cm_id_priv->msg = NULL;
2501
2502         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2503         if (!ret)
2504                 list_add_tail(&work->list, &cm_id_priv->work_list);
2505         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2506
2507         if (ret)
2508                 cm_process_work(cm_id_priv, work);
2509         else
2510                 cm_deref_id(cm_id_priv);
2511         return 0;
2512 out:
2513         cm_deref_id(cm_id_priv);
2514         return -EINVAL;
2515 }
2516
2517 static int cm_timewait_handler(struct cm_work *work)
2518 {
2519         struct cm_timewait_info *timewait_info;
2520         struct cm_id_private *cm_id_priv;
2521         unsigned long flags;
2522         int ret;
2523
2524         timewait_info = (struct cm_timewait_info *)work;
2525         cm_cleanup_timewait(timewait_info);
2526
2527         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2528                                    timewait_info->work.remote_id);
2529         if (!cm_id_priv)
2530                 return -EINVAL;
2531
2532         spin_lock_irqsave(&cm_id_priv->lock, flags);
2533         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
2534             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
2535                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2536                 goto out;
2537         }
2538         cm_id_priv->id.state = IB_CM_IDLE;
2539         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2540         if (!ret)
2541                 list_add_tail(&work->list, &cm_id_priv->work_list);
2542         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2543
2544         if (ret)
2545                 cm_process_work(cm_id_priv, work);
2546         else
2547                 cm_deref_id(cm_id_priv);
2548         return 0;
2549 out:
2550         cm_deref_id(cm_id_priv);
2551         return -EINVAL;
2552 }
2553
2554 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
2555                                struct cm_id_private *cm_id_priv,
2556                                struct ib_cm_sidr_req_param *param)
2557 {
2558         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
2559                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
2560         sidr_req_msg->request_id = cm_id_priv->id.local_id;
2561         sidr_req_msg->pkey = cpu_to_be16(param->pkey);
2562         sidr_req_msg->service_id = param->service_id;
2563
2564         if (param->private_data && param->private_data_len)
2565                 memcpy(sidr_req_msg->private_data, param->private_data,
2566                        param->private_data_len);
2567 }
2568
2569 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
2570                         struct ib_cm_sidr_req_param *param)
2571 {
2572         struct cm_id_private *cm_id_priv;
2573         struct ib_mad_send_buf *msg;
2574         struct ib_send_wr *bad_send_wr;
2575         unsigned long flags;
2576         int ret;
2577
2578         if (!param->path || (param->private_data &&
2579              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
2580                 return -EINVAL;
2581
2582         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2583         ret = cm_init_av_by_path(param->path, &cm_id_priv->av);
2584         if (ret)
2585                 goto out;
2586
2587         cm_id->service_id = param->service_id;
2588         cm_id->service_mask = __constant_cpu_to_be64(~0ULL);
2589         cm_id_priv->timeout_ms = param->timeout_ms;
2590         cm_id_priv->max_cm_retries = param->max_cm_retries;
2591         ret = cm_alloc_msg(cm_id_priv, &msg);
2592         if (ret)
2593                 goto out;
2594
2595         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
2596                            param);
2597         msg->send_wr.wr.ud.timeout_ms = cm_id_priv->timeout_ms;
2598         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
2599
2600         spin_lock_irqsave(&cm_id_priv->lock, flags);
2601         if (cm_id->state == IB_CM_IDLE)
2602                 ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
2603                                        &msg->send_wr, &bad_send_wr);
2604         else
2605                 ret = -EINVAL;
2606
2607         if (ret) {
2608                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2609                 cm_free_msg(msg);
2610                 goto out;
2611         }
2612         cm_id->state = IB_CM_SIDR_REQ_SENT;
2613         cm_id_priv->msg = msg;
2614         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2615 out:
2616         return ret;
2617 }
2618 EXPORT_SYMBOL(ib_send_cm_sidr_req);
2619
2620 static void cm_format_sidr_req_event(struct cm_work *work,
2621                                      struct ib_cm_id *listen_id)
2622 {
2623         struct cm_sidr_req_msg *sidr_req_msg;
2624         struct ib_cm_sidr_req_event_param *param;
2625
2626         sidr_req_msg = (struct cm_sidr_req_msg *)
2627                                 work->mad_recv_wc->recv_buf.mad;
2628         param = &work->cm_event.param.sidr_req_rcvd;
2629         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
2630         param->listen_id = listen_id;
2631         param->device = work->port->mad_agent->device;
2632         param->port = work->port->port_num;
2633         work->cm_event.private_data = &sidr_req_msg->private_data;
2634 }
2635
2636 static int cm_sidr_req_handler(struct cm_work *work)
2637 {
2638         struct ib_cm_id *cm_id;
2639         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
2640         struct cm_sidr_req_msg *sidr_req_msg;
2641         struct ib_wc *wc;
2642         unsigned long flags;
2643
2644         cm_id = ib_create_cm_id(NULL, NULL);
2645         if (IS_ERR(cm_id))
2646                 return PTR_ERR(cm_id);
2647         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2648
2649         /* Record SGID/SLID and request ID for lookup. */
2650         sidr_req_msg = (struct cm_sidr_req_msg *)
2651                                 work->mad_recv_wc->recv_buf.mad;
2652         wc = work->mad_recv_wc->wc;
2653         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
2654         cm_id_priv->av.dgid.global.interface_id = 0;
2655         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2656                                 &cm_id_priv->av);
2657         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
2658         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
2659         cm_id_priv->tid = sidr_req_msg->hdr.tid;
2660         atomic_inc(&cm_id_priv->work_count);
2661
2662         spin_lock_irqsave(&cm.lock, flags);
2663         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
2664         if (cur_cm_id_priv) {
2665                 spin_unlock_irqrestore(&cm.lock, flags);
2666                 goto out; /* Duplicate message. */
2667         }
2668         cur_cm_id_priv = cm_find_listen(sidr_req_msg->service_id);
2669         if (!cur_cm_id_priv) {
2670                 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
2671                 spin_unlock_irqrestore(&cm.lock, flags);
2672                 /* todo: reply with no match */
2673                 goto out; /* No match. */
2674         }
2675         atomic_inc(&cur_cm_id_priv->refcount);
2676         spin_unlock_irqrestore(&cm.lock, flags);
2677
2678         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
2679         cm_id_priv->id.context = cur_cm_id_priv->id.context;
2680         cm_id_priv->id.service_id = sidr_req_msg->service_id;
2681         cm_id_priv->id.service_mask = __constant_cpu_to_be64(~0ULL);
2682
2683         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
2684         cm_process_work(cm_id_priv, work);
2685         cm_deref_id(cur_cm_id_priv);
2686         return 0;
2687 out:
2688         ib_destroy_cm_id(&cm_id_priv->id);
2689         return -EINVAL;
2690 }
2691
2692 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
2693                                struct cm_id_private *cm_id_priv,
2694                                struct ib_cm_sidr_rep_param *param)
2695 {
2696         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
2697                           cm_id_priv->tid);
2698         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
2699         sidr_rep_msg->status = param->status;
2700         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
2701         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
2702         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
2703
2704         if (param->info && param->info_length)
2705                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
2706
2707         if (param->private_data && param->private_data_len)
2708                 memcpy(sidr_rep_msg->private_data, param->private_data,
2709                        param->private_data_len);
2710 }
2711
2712 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
2713                         struct ib_cm_sidr_rep_param *param)
2714 {
2715         struct cm_id_private *cm_id_priv;
2716         struct ib_mad_send_buf *msg;
2717         struct ib_send_wr *bad_send_wr;
2718         unsigned long flags;
2719         int ret;
2720
2721         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
2722             (param->private_data &&
2723              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
2724                 return -EINVAL;
2725
2726         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2727         spin_lock_irqsave(&cm_id_priv->lock, flags);
2728         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
2729                 ret = -EINVAL;
2730                 goto error;
2731         }
2732
2733         ret = cm_alloc_msg(cm_id_priv, &msg);
2734         if (ret)
2735                 goto error;
2736
2737         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
2738                            param);
2739         ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
2740                                &msg->send_wr, &bad_send_wr);
2741         if (ret) {
2742                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2743                 cm_free_msg(msg);
2744                 return ret;
2745         }
2746         cm_id->state = IB_CM_IDLE;
2747         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2748
2749         spin_lock_irqsave(&cm.lock, flags);
2750         rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
2751         spin_unlock_irqrestore(&cm.lock, flags);
2752         return 0;
2753
2754 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2755         return ret;
2756 }
2757 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
2758
2759 static void cm_format_sidr_rep_event(struct cm_work *work)
2760 {
2761         struct cm_sidr_rep_msg *sidr_rep_msg;
2762         struct ib_cm_sidr_rep_event_param *param;
2763
2764         sidr_rep_msg = (struct cm_sidr_rep_msg *)
2765                                 work->mad_recv_wc->recv_buf.mad;
2766         param = &work->cm_event.param.sidr_rep_rcvd;
2767         param->status = sidr_rep_msg->status;
2768         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
2769         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
2770         param->info = &sidr_rep_msg->info;
2771         param->info_len = sidr_rep_msg->info_length;
2772         work->cm_event.private_data = &sidr_rep_msg->private_data;
2773 }
2774
2775 static int cm_sidr_rep_handler(struct cm_work *work)
2776 {
2777         struct cm_sidr_rep_msg *sidr_rep_msg;
2778         struct cm_id_private *cm_id_priv;
2779         unsigned long flags;
2780
2781         sidr_rep_msg = (struct cm_sidr_rep_msg *)
2782                                 work->mad_recv_wc->recv_buf.mad;
2783         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
2784         if (!cm_id_priv)
2785                 return -EINVAL; /* Unmatched reply. */
2786
2787         spin_lock_irqsave(&cm_id_priv->lock, flags);
2788         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
2789                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2790                 goto out;
2791         }
2792         cm_id_priv->id.state = IB_CM_IDLE;
2793         ib_cancel_mad(cm_id_priv->av.port->mad_agent,
2794                       (unsigned long) cm_id_priv->msg);
2795         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2796
2797         cm_format_sidr_rep_event(work);
2798         cm_process_work(cm_id_priv, work);
2799         return 0;
2800 out:
2801         cm_deref_id(cm_id_priv);
2802         return -EINVAL;
2803 }
2804
2805 static void cm_process_send_error(struct ib_mad_send_buf *msg,
2806                                   enum ib_wc_status wc_status)
2807 {
2808         struct cm_id_private *cm_id_priv;
2809         struct ib_cm_event cm_event;
2810         enum ib_cm_state state;
2811         unsigned long flags;
2812         int ret;
2813
2814         memset(&cm_event, 0, sizeof cm_event);
2815         cm_id_priv = msg->context[0];
2816
2817         /* Discard old sends or ones without a response. */
2818         spin_lock_irqsave(&cm_id_priv->lock, flags);
2819         state = (enum ib_cm_state) (unsigned long) msg->context[1];
2820         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
2821                 goto discard;
2822
2823         switch (state) {
2824         case IB_CM_REQ_SENT:
2825         case IB_CM_MRA_REQ_RCVD:
2826                 cm_reset_to_idle(cm_id_priv);
2827                 cm_event.event = IB_CM_REQ_ERROR;
2828                 break;
2829         case IB_CM_REP_SENT:
2830         case IB_CM_MRA_REP_RCVD:
2831                 cm_reset_to_idle(cm_id_priv);
2832                 cm_event.event = IB_CM_REP_ERROR;
2833                 break;
2834         case IB_CM_DREQ_SENT:
2835                 cm_enter_timewait(cm_id_priv);
2836                 cm_event.event = IB_CM_DREQ_ERROR;
2837                 break;
2838         case IB_CM_SIDR_REQ_SENT:
2839                 cm_id_priv->id.state = IB_CM_IDLE;
2840                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
2841                 break;
2842         default:
2843                 goto discard;
2844         }
2845         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2846         cm_event.param.send_status = wc_status;
2847
2848         /* No other events can occur on the cm_id at this point. */
2849         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
2850         cm_free_msg(msg);
2851         if (ret)
2852                 ib_destroy_cm_id(&cm_id_priv->id);
2853         return;
2854 discard:
2855         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2856         cm_free_msg(msg);
2857 }
2858
2859 static void cm_send_handler(struct ib_mad_agent *mad_agent,
2860                             struct ib_mad_send_wc *mad_send_wc)
2861 {
2862         struct ib_mad_send_buf *msg;
2863
2864         msg = (struct ib_mad_send_buf *)(unsigned long)mad_send_wc->wr_id;
2865
2866         switch (mad_send_wc->status) {
2867         case IB_WC_SUCCESS:
2868         case IB_WC_WR_FLUSH_ERR:
2869                 cm_free_msg(msg);
2870                 break;
2871         default:
2872                 if (msg->context[0] && msg->context[1])
2873                         cm_process_send_error(msg, mad_send_wc->status);
2874                 else
2875                         cm_free_msg(msg);
2876                 break;
2877         }
2878 }
2879
2880 static void cm_work_handler(void *data)
2881 {
2882         struct cm_work *work = data;
2883         int ret;
2884
2885         switch (work->cm_event.event) {
2886         case IB_CM_REQ_RECEIVED:
2887                 ret = cm_req_handler(work);
2888                 break;
2889         case IB_CM_MRA_RECEIVED:
2890                 ret = cm_mra_handler(work);
2891                 break;
2892         case IB_CM_REJ_RECEIVED:
2893                 ret = cm_rej_handler(work);
2894                 break;
2895         case IB_CM_REP_RECEIVED:
2896                 ret = cm_rep_handler(work);
2897                 break;
2898         case IB_CM_RTU_RECEIVED:
2899                 ret = cm_rtu_handler(work);
2900                 break;
2901         case IB_CM_USER_ESTABLISHED:
2902                 ret = cm_establish_handler(work);
2903                 break;
2904         case IB_CM_DREQ_RECEIVED:
2905                 ret = cm_dreq_handler(work);
2906                 break;
2907         case IB_CM_DREP_RECEIVED:
2908                 ret = cm_drep_handler(work);
2909                 break;
2910         case IB_CM_SIDR_REQ_RECEIVED:
2911                 ret = cm_sidr_req_handler(work);
2912                 break;
2913         case IB_CM_SIDR_REP_RECEIVED:
2914                 ret = cm_sidr_rep_handler(work);
2915                 break;
2916         case IB_CM_LAP_RECEIVED:
2917                 ret = cm_lap_handler(work);
2918                 break;
2919         case IB_CM_APR_RECEIVED:
2920                 ret = cm_apr_handler(work);
2921                 break;
2922         case IB_CM_TIMEWAIT_EXIT:
2923                 ret = cm_timewait_handler(work);
2924                 break;
2925         default:
2926                 ret = -EINVAL;
2927                 break;
2928         }
2929         if (ret)
2930                 cm_free_work(work);
2931 }
2932
2933 int ib_cm_establish(struct ib_cm_id *cm_id)
2934 {
2935         struct cm_id_private *cm_id_priv;
2936         struct cm_work *work;
2937         unsigned long flags;
2938         int ret = 0;
2939
2940         work = kmalloc(sizeof *work, GFP_ATOMIC);
2941         if (!work)
2942                 return -ENOMEM;
2943
2944         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2945         spin_lock_irqsave(&cm_id_priv->lock, flags);
2946         switch (cm_id->state)
2947         {
2948         case IB_CM_REP_SENT:
2949         case IB_CM_MRA_REP_RCVD:
2950                 cm_id->state = IB_CM_ESTABLISHED;
2951                 break;
2952         case IB_CM_ESTABLISHED:
2953                 ret = -EISCONN;
2954                 break;
2955         default:
2956                 ret = -EINVAL;
2957                 break;
2958         }
2959         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2960
2961         if (ret) {
2962                 kfree(work);
2963                 goto out;
2964         }
2965
2966         /*
2967          * The CM worker thread may try to destroy the cm_id before it
2968          * can execute this work item.  To prevent potential deadlock,
2969          * we need to find the cm_id once we're in the context of the
2970          * worker thread, rather than holding a reference on it.
2971          */
2972         INIT_WORK(&work->work, cm_work_handler, work);
2973         work->local_id = cm_id->local_id;
2974         work->remote_id = cm_id->remote_id;
2975         work->mad_recv_wc = NULL;
2976         work->cm_event.event = IB_CM_USER_ESTABLISHED;
2977         queue_work(cm.wq, &work->work);
2978 out:
2979         return ret;
2980 }
2981 EXPORT_SYMBOL(ib_cm_establish);
2982
2983 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
2984                             struct ib_mad_recv_wc *mad_recv_wc)
2985 {
2986         struct cm_work *work;
2987         enum ib_cm_event_type event;
2988         int paths = 0;
2989
2990         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
2991         case CM_REQ_ATTR_ID:
2992                 paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
2993                                                     alt_local_lid != 0);
2994                 event = IB_CM_REQ_RECEIVED;
2995                 break;
2996         case CM_MRA_ATTR_ID:
2997                 event = IB_CM_MRA_RECEIVED;
2998                 break;
2999         case CM_REJ_ATTR_ID:
3000                 event = IB_CM_REJ_RECEIVED;
3001                 break;
3002         case CM_REP_ATTR_ID:
3003                 event = IB_CM_REP_RECEIVED;
3004                 break;
3005         case CM_RTU_ATTR_ID:
3006                 event = IB_CM_RTU_RECEIVED;
3007                 break;
3008         case CM_DREQ_ATTR_ID:
3009                 event = IB_CM_DREQ_RECEIVED;
3010                 break;
3011         case CM_DREP_ATTR_ID:
3012                 event = IB_CM_DREP_RECEIVED;
3013                 break;
3014         case CM_SIDR_REQ_ATTR_ID:
3015                 event = IB_CM_SIDR_REQ_RECEIVED;
3016                 break;
3017         case CM_SIDR_REP_ATTR_ID:
3018                 event = IB_CM_SIDR_REP_RECEIVED;
3019                 break;
3020         case CM_LAP_ATTR_ID:
3021                 paths = 1;
3022                 event = IB_CM_LAP_RECEIVED;
3023                 break;
3024         case CM_APR_ATTR_ID:
3025                 event = IB_CM_APR_RECEIVED;
3026                 break;
3027         default:
3028                 ib_free_recv_mad(mad_recv_wc);
3029                 return;
3030         }
3031
3032         work = kmalloc(sizeof *work + sizeof(struct ib_sa_path_rec) * paths,
3033                        GFP_KERNEL);
3034         if (!work) {
3035                 ib_free_recv_mad(mad_recv_wc);
3036                 return;
3037         }
3038
3039         INIT_WORK(&work->work, cm_work_handler, work);
3040         work->cm_event.event = event;
3041         work->mad_recv_wc = mad_recv_wc;
3042         work->port = (struct cm_port *)mad_agent->context;
3043         queue_work(cm.wq, &work->work);
3044 }
3045
3046 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
3047                                 struct ib_qp_attr *qp_attr,
3048                                 int *qp_attr_mask)
3049 {
3050         unsigned long flags;
3051         int ret;
3052
3053         spin_lock_irqsave(&cm_id_priv->lock, flags);
3054         switch (cm_id_priv->id.state) {
3055         case IB_CM_REQ_SENT:
3056         case IB_CM_MRA_REQ_RCVD:
3057         case IB_CM_REQ_RCVD:
3058         case IB_CM_MRA_REQ_SENT:
3059         case IB_CM_REP_RCVD:
3060         case IB_CM_MRA_REP_SENT:
3061         case IB_CM_REP_SENT:
3062         case IB_CM_MRA_REP_RCVD:
3063         case IB_CM_ESTABLISHED:
3064                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
3065                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
3066                 qp_attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE;
3067                 if (cm_id_priv->responder_resources)
3068                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_WRITE |
3069                                                     IB_ACCESS_REMOTE_READ;
3070                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
3071                 qp_attr->port_num = cm_id_priv->av.port->port_num;
3072                 ret = 0;
3073                 break;
3074         default:
3075                 ret = -EINVAL;
3076                 break;
3077         }
3078         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3079         return ret;
3080 }
3081
3082 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
3083                                struct ib_qp_attr *qp_attr,
3084                                int *qp_attr_mask)
3085 {
3086         unsigned long flags;
3087         int ret;
3088
3089         spin_lock_irqsave(&cm_id_priv->lock, flags);
3090         switch (cm_id_priv->id.state) {
3091         case IB_CM_REQ_RCVD:
3092         case IB_CM_MRA_REQ_SENT:
3093         case IB_CM_REP_RCVD:
3094         case IB_CM_MRA_REP_SENT:
3095         case IB_CM_REP_SENT:
3096         case IB_CM_MRA_REP_RCVD:
3097         case IB_CM_ESTABLISHED:
3098                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
3099                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN |
3100                                 IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
3101                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
3102                 qp_attr->path_mtu = cm_id_priv->path_mtu;
3103                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
3104                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
3105                 qp_attr->max_dest_rd_atomic = cm_id_priv->responder_resources;
3106                 qp_attr->min_rnr_timer = 0;
3107                 if (cm_id_priv->alt_av.ah_attr.dlid) {
3108                         *qp_attr_mask |= IB_QP_ALT_PATH;
3109                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3110                 }
3111                 ret = 0;
3112                 break;
3113         default:
3114                 ret = -EINVAL;
3115                 break;
3116         }
3117         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3118         return ret;
3119 }
3120
3121 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
3122                                struct ib_qp_attr *qp_attr,
3123                                int *qp_attr_mask)
3124 {
3125         unsigned long flags;
3126         int ret;
3127
3128         spin_lock_irqsave(&cm_id_priv->lock, flags);
3129         switch (cm_id_priv->id.state) {
3130         case IB_CM_REP_RCVD:
3131         case IB_CM_MRA_REP_SENT:
3132         case IB_CM_REP_SENT:
3133         case IB_CM_MRA_REP_RCVD:
3134         case IB_CM_ESTABLISHED:
3135                 *qp_attr_mask = IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
3136                                 IB_QP_RNR_RETRY | IB_QP_SQ_PSN |
3137                                 IB_QP_MAX_QP_RD_ATOMIC;
3138                 qp_attr->timeout = cm_id_priv->local_ack_timeout;
3139                 qp_attr->retry_cnt = cm_id_priv->retry_count;
3140                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
3141                 qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
3142                 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
3143                 if (cm_id_priv->alt_av.ah_attr.dlid) {
3144                         *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
3145                         qp_attr->path_mig_state = IB_MIG_REARM;
3146                 }
3147                 ret = 0;
3148                 break;
3149         default:
3150                 ret = -EINVAL;
3151                 break;
3152         }
3153         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3154         return ret;
3155 }
3156
3157 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
3158                        struct ib_qp_attr *qp_attr,
3159                        int *qp_attr_mask)
3160 {
3161         struct cm_id_private *cm_id_priv;
3162         int ret;
3163
3164         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3165         switch (qp_attr->qp_state) {
3166         case IB_QPS_INIT:
3167                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
3168                 break;
3169         case IB_QPS_RTR:
3170                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
3171                 break;
3172         case IB_QPS_RTS:
3173                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
3174                 break;
3175         default:
3176                 ret = -EINVAL;
3177                 break;
3178         }
3179         return ret;
3180 }
3181 EXPORT_SYMBOL(ib_cm_init_qp_attr);
3182
3183 static __be64 cm_get_ca_guid(struct ib_device *device)
3184 {
3185         struct ib_device_attr *device_attr;
3186         __be64 guid;
3187         int ret;
3188
3189         device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
3190         if (!device_attr)
3191                 return 0;
3192
3193         ret = ib_query_device(device, device_attr);
3194         guid = ret ? 0 : device_attr->node_guid;
3195         kfree(device_attr);
3196         return guid;
3197 }
3198
3199 static void cm_add_one(struct ib_device *device)
3200 {
3201         struct cm_device *cm_dev;
3202         struct cm_port *port;
3203         struct ib_mad_reg_req reg_req = {
3204                 .mgmt_class = IB_MGMT_CLASS_CM,
3205                 .mgmt_class_version = IB_CM_CLASS_VERSION
3206         };
3207         struct ib_port_modify port_modify = {
3208                 .set_port_cap_mask = IB_PORT_CM_SUP
3209         };
3210         unsigned long flags;
3211         int ret;
3212         u8 i;
3213
3214         cm_dev = kmalloc(sizeof(*cm_dev) + sizeof(*port) *
3215                          device->phys_port_cnt, GFP_KERNEL);
3216         if (!cm_dev)
3217                 return;
3218
3219         cm_dev->device = device;
3220         cm_dev->ca_guid = cm_get_ca_guid(device);
3221         if (!cm_dev->ca_guid)
3222                 goto error1;
3223
3224         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
3225         for (i = 1; i <= device->phys_port_cnt; i++) {
3226                 port = &cm_dev->port[i-1];
3227                 port->cm_dev = cm_dev;
3228                 port->port_num = i;
3229                 port->mad_agent = ib_register_mad_agent(device, i,
3230                                                         IB_QPT_GSI,
3231                                                         &reg_req,
3232                                                         0,
3233                                                         cm_send_handler,
3234                                                         cm_recv_handler,
3235                                                         port);
3236                 if (IS_ERR(port->mad_agent))
3237                         goto error2;
3238
3239                 ret = ib_modify_port(device, i, 0, &port_modify);
3240                 if (ret)
3241                         goto error3;
3242         }
3243         ib_set_client_data(device, &cm_client, cm_dev);
3244
3245         write_lock_irqsave(&cm.device_lock, flags);
3246         list_add_tail(&cm_dev->list, &cm.device_list);
3247         write_unlock_irqrestore(&cm.device_lock, flags);
3248         return;
3249
3250 error3:
3251         ib_unregister_mad_agent(port->mad_agent);
3252 error2:
3253         port_modify.set_port_cap_mask = 0;
3254         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
3255         while (--i) {
3256                 port = &cm_dev->port[i-1];
3257                 ib_modify_port(device, port->port_num, 0, &port_modify);
3258                 ib_unregister_mad_agent(port->mad_agent);
3259         }
3260 error1:
3261         kfree(cm_dev);
3262 }
3263
3264 static void cm_remove_one(struct ib_device *device)
3265 {
3266         struct cm_device *cm_dev;
3267         struct cm_port *port;
3268         struct ib_port_modify port_modify = {
3269                 .clr_port_cap_mask = IB_PORT_CM_SUP
3270         };
3271         unsigned long flags;
3272         int i;
3273
3274         cm_dev = ib_get_client_data(device, &cm_client);
3275         if (!cm_dev)
3276                 return;
3277
3278         write_lock_irqsave(&cm.device_lock, flags);
3279         list_del(&cm_dev->list);
3280         write_unlock_irqrestore(&cm.device_lock, flags);
3281
3282         for (i = 1; i <= device->phys_port_cnt; i++) {
3283                 port = &cm_dev->port[i-1];
3284                 ib_modify_port(device, port->port_num, 0, &port_modify);
3285                 ib_unregister_mad_agent(port->mad_agent);
3286         }
3287         kfree(cm_dev);
3288 }
3289
3290 static int __init ib_cm_init(void)
3291 {
3292         int ret;
3293
3294         memset(&cm, 0, sizeof cm);
3295         INIT_LIST_HEAD(&cm.device_list);
3296         rwlock_init(&cm.device_lock);
3297         spin_lock_init(&cm.lock);
3298         cm.listen_service_table = RB_ROOT;
3299         cm.listen_service_id = __constant_be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
3300         cm.remote_id_table = RB_ROOT;
3301         cm.remote_qp_table = RB_ROOT;
3302         cm.remote_sidr_table = RB_ROOT;
3303         idr_init(&cm.local_id_table);
3304         idr_pre_get(&cm.local_id_table, GFP_KERNEL);
3305
3306         cm.wq = create_workqueue("ib_cm");
3307         if (!cm.wq)
3308                 return -ENOMEM;
3309
3310         ret = ib_register_client(&cm_client);
3311         if (ret)
3312                 goto error;
3313
3314         return 0;
3315 error:
3316         destroy_workqueue(cm.wq);
3317         return ret;
3318 }
3319
3320 static void __exit ib_cm_cleanup(void)
3321 {
3322         flush_workqueue(cm.wq);
3323         destroy_workqueue(cm.wq);
3324         ib_unregister_client(&cm_client);
3325 }
3326
3327 module_init(ib_cm_init);
3328 module_exit(ib_cm_cleanup);
3329