Merge branches 'x86/prototypes', 'x86/x2apic' and 'x86/debug' into x86/core
[linux-2.6] / drivers / infiniband / core / cma.c
1 /*
2  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
3  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
4  * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
5  * Copyright (c) 2005-2006 Intel Corporation.  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
36 #include <linux/completion.h>
37 #include <linux/in.h>
38 #include <linux/in6.h>
39 #include <linux/mutex.h>
40 #include <linux/random.h>
41 #include <linux/idr.h>
42 #include <linux/inetdevice.h>
43
44 #include <net/tcp.h>
45
46 #include <rdma/rdma_cm.h>
47 #include <rdma/rdma_cm_ib.h>
48 #include <rdma/ib_cache.h>
49 #include <rdma/ib_cm.h>
50 #include <rdma/ib_sa.h>
51 #include <rdma/iw_cm.h>
52
53 MODULE_AUTHOR("Sean Hefty");
54 MODULE_DESCRIPTION("Generic RDMA CM Agent");
55 MODULE_LICENSE("Dual BSD/GPL");
56
57 #define CMA_CM_RESPONSE_TIMEOUT 20
58 #define CMA_MAX_CM_RETRIES 15
59 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
60
61 static void cma_add_one(struct ib_device *device);
62 static void cma_remove_one(struct ib_device *device);
63
64 static struct ib_client cma_client = {
65         .name   = "cma",
66         .add    = cma_add_one,
67         .remove = cma_remove_one
68 };
69
70 static struct ib_sa_client sa_client;
71 static struct rdma_addr_client addr_client;
72 static LIST_HEAD(dev_list);
73 static LIST_HEAD(listen_any_list);
74 static DEFINE_MUTEX(lock);
75 static struct workqueue_struct *cma_wq;
76 static DEFINE_IDR(sdp_ps);
77 static DEFINE_IDR(tcp_ps);
78 static DEFINE_IDR(udp_ps);
79 static DEFINE_IDR(ipoib_ps);
80 static int next_port;
81
82 struct cma_device {
83         struct list_head        list;
84         struct ib_device        *device;
85         struct completion       comp;
86         atomic_t                refcount;
87         struct list_head        id_list;
88 };
89
90 enum cma_state {
91         CMA_IDLE,
92         CMA_ADDR_QUERY,
93         CMA_ADDR_RESOLVED,
94         CMA_ROUTE_QUERY,
95         CMA_ROUTE_RESOLVED,
96         CMA_CONNECT,
97         CMA_DISCONNECT,
98         CMA_ADDR_BOUND,
99         CMA_LISTEN,
100         CMA_DEVICE_REMOVAL,
101         CMA_DESTROYING
102 };
103
104 struct rdma_bind_list {
105         struct idr              *ps;
106         struct hlist_head       owners;
107         unsigned short          port;
108 };
109
110 /*
111  * Device removal can occur at anytime, so we need extra handling to
112  * serialize notifying the user of device removal with other callbacks.
113  * We do this by disabling removal notification while a callback is in process,
114  * and reporting it after the callback completes.
115  */
116 struct rdma_id_private {
117         struct rdma_cm_id       id;
118
119         struct rdma_bind_list   *bind_list;
120         struct hlist_node       node;
121         struct list_head        list; /* listen_any_list or cma_device.list */
122         struct list_head        listen_list; /* per device listens */
123         struct cma_device       *cma_dev;
124         struct list_head        mc_list;
125
126         int                     internal_id;
127         enum cma_state          state;
128         spinlock_t              lock;
129         struct mutex            qp_mutex;
130
131         struct completion       comp;
132         atomic_t                refcount;
133         struct mutex            handler_mutex;
134
135         int                     backlog;
136         int                     timeout_ms;
137         struct ib_sa_query      *query;
138         int                     query_id;
139         union {
140                 struct ib_cm_id *ib;
141                 struct iw_cm_id *iw;
142         } cm_id;
143
144         u32                     seq_num;
145         u32                     qkey;
146         u32                     qp_num;
147         u8                      srq;
148         u8                      tos;
149 };
150
151 struct cma_multicast {
152         struct rdma_id_private *id_priv;
153         union {
154                 struct ib_sa_multicast *ib;
155         } multicast;
156         struct list_head        list;
157         void                    *context;
158         struct sockaddr_storage addr;
159 };
160
161 struct cma_work {
162         struct work_struct      work;
163         struct rdma_id_private  *id;
164         enum cma_state          old_state;
165         enum cma_state          new_state;
166         struct rdma_cm_event    event;
167 };
168
169 struct cma_ndev_work {
170         struct work_struct      work;
171         struct rdma_id_private  *id;
172         struct rdma_cm_event    event;
173 };
174
175 union cma_ip_addr {
176         struct in6_addr ip6;
177         struct {
178                 __be32 pad[3];
179                 __be32 addr;
180         } ip4;
181 };
182
183 struct cma_hdr {
184         u8 cma_version;
185         u8 ip_version;  /* IP version: 7:4 */
186         __be16 port;
187         union cma_ip_addr src_addr;
188         union cma_ip_addr dst_addr;
189 };
190
191 struct sdp_hh {
192         u8 bsdh[16];
193         u8 sdp_version; /* Major version: 7:4 */
194         u8 ip_version;  /* IP version: 7:4 */
195         u8 sdp_specific1[10];
196         __be16 port;
197         __be16 sdp_specific2;
198         union cma_ip_addr src_addr;
199         union cma_ip_addr dst_addr;
200 };
201
202 struct sdp_hah {
203         u8 bsdh[16];
204         u8 sdp_version;
205 };
206
207 #define CMA_VERSION 0x00
208 #define SDP_MAJ_VERSION 0x2
209
210 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
211 {
212         unsigned long flags;
213         int ret;
214
215         spin_lock_irqsave(&id_priv->lock, flags);
216         ret = (id_priv->state == comp);
217         spin_unlock_irqrestore(&id_priv->lock, flags);
218         return ret;
219 }
220
221 static int cma_comp_exch(struct rdma_id_private *id_priv,
222                          enum cma_state comp, enum cma_state exch)
223 {
224         unsigned long flags;
225         int ret;
226
227         spin_lock_irqsave(&id_priv->lock, flags);
228         if ((ret = (id_priv->state == comp)))
229                 id_priv->state = exch;
230         spin_unlock_irqrestore(&id_priv->lock, flags);
231         return ret;
232 }
233
234 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
235                                enum cma_state exch)
236 {
237         unsigned long flags;
238         enum cma_state old;
239
240         spin_lock_irqsave(&id_priv->lock, flags);
241         old = id_priv->state;
242         id_priv->state = exch;
243         spin_unlock_irqrestore(&id_priv->lock, flags);
244         return old;
245 }
246
247 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
248 {
249         return hdr->ip_version >> 4;
250 }
251
252 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
253 {
254         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
255 }
256
257 static inline u8 sdp_get_majv(u8 sdp_version)
258 {
259         return sdp_version >> 4;
260 }
261
262 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
263 {
264         return hh->ip_version >> 4;
265 }
266
267 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
268 {
269         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
270 }
271
272 static inline int cma_is_ud_ps(enum rdma_port_space ps)
273 {
274         return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB);
275 }
276
277 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
278                               struct cma_device *cma_dev)
279 {
280         atomic_inc(&cma_dev->refcount);
281         id_priv->cma_dev = cma_dev;
282         id_priv->id.device = cma_dev->device;
283         list_add_tail(&id_priv->list, &cma_dev->id_list);
284 }
285
286 static inline void cma_deref_dev(struct cma_device *cma_dev)
287 {
288         if (atomic_dec_and_test(&cma_dev->refcount))
289                 complete(&cma_dev->comp);
290 }
291
292 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
293 {
294         list_del(&id_priv->list);
295         cma_deref_dev(id_priv->cma_dev);
296         id_priv->cma_dev = NULL;
297 }
298
299 static int cma_set_qkey(struct ib_device *device, u8 port_num,
300                         enum rdma_port_space ps,
301                         struct rdma_dev_addr *dev_addr, u32 *qkey)
302 {
303         struct ib_sa_mcmember_rec rec;
304         int ret = 0;
305
306         switch (ps) {
307         case RDMA_PS_UDP:
308                 *qkey = RDMA_UDP_QKEY;
309                 break;
310         case RDMA_PS_IPOIB:
311                 ib_addr_get_mgid(dev_addr, &rec.mgid);
312                 ret = ib_sa_get_mcmember_rec(device, port_num, &rec.mgid, &rec);
313                 *qkey = be32_to_cpu(rec.qkey);
314                 break;
315         default:
316                 break;
317         }
318         return ret;
319 }
320
321 static int cma_acquire_dev(struct rdma_id_private *id_priv)
322 {
323         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
324         struct cma_device *cma_dev;
325         union ib_gid gid;
326         int ret = -ENODEV;
327
328         switch (rdma_node_get_transport(dev_addr->dev_type)) {
329         case RDMA_TRANSPORT_IB:
330                 ib_addr_get_sgid(dev_addr, &gid);
331                 break;
332         case RDMA_TRANSPORT_IWARP:
333                 iw_addr_get_sgid(dev_addr, &gid);
334                 break;
335         default:
336                 return -ENODEV;
337         }
338
339         list_for_each_entry(cma_dev, &dev_list, list) {
340                 ret = ib_find_cached_gid(cma_dev->device, &gid,
341                                          &id_priv->id.port_num, NULL);
342                 if (!ret) {
343                         ret = cma_set_qkey(cma_dev->device,
344                                            id_priv->id.port_num,
345                                            id_priv->id.ps, dev_addr,
346                                            &id_priv->qkey);
347                         if (!ret)
348                                 cma_attach_to_dev(id_priv, cma_dev);
349                         break;
350                 }
351         }
352         return ret;
353 }
354
355 static void cma_deref_id(struct rdma_id_private *id_priv)
356 {
357         if (atomic_dec_and_test(&id_priv->refcount))
358                 complete(&id_priv->comp);
359 }
360
361 static int cma_disable_callback(struct rdma_id_private *id_priv,
362                               enum cma_state state)
363 {
364         mutex_lock(&id_priv->handler_mutex);
365         if (id_priv->state != state) {
366                 mutex_unlock(&id_priv->handler_mutex);
367                 return -EINVAL;
368         }
369         return 0;
370 }
371
372 static int cma_has_cm_dev(struct rdma_id_private *id_priv)
373 {
374         return (id_priv->id.device && id_priv->cm_id.ib);
375 }
376
377 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
378                                   void *context, enum rdma_port_space ps)
379 {
380         struct rdma_id_private *id_priv;
381
382         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
383         if (!id_priv)
384                 return ERR_PTR(-ENOMEM);
385
386         id_priv->state = CMA_IDLE;
387         id_priv->id.context = context;
388         id_priv->id.event_handler = event_handler;
389         id_priv->id.ps = ps;
390         spin_lock_init(&id_priv->lock);
391         mutex_init(&id_priv->qp_mutex);
392         init_completion(&id_priv->comp);
393         atomic_set(&id_priv->refcount, 1);
394         mutex_init(&id_priv->handler_mutex);
395         INIT_LIST_HEAD(&id_priv->listen_list);
396         INIT_LIST_HEAD(&id_priv->mc_list);
397         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
398
399         return &id_priv->id;
400 }
401 EXPORT_SYMBOL(rdma_create_id);
402
403 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
404 {
405         struct ib_qp_attr qp_attr;
406         int qp_attr_mask, ret;
407
408         qp_attr.qp_state = IB_QPS_INIT;
409         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
410         if (ret)
411                 return ret;
412
413         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
414         if (ret)
415                 return ret;
416
417         qp_attr.qp_state = IB_QPS_RTR;
418         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
419         if (ret)
420                 return ret;
421
422         qp_attr.qp_state = IB_QPS_RTS;
423         qp_attr.sq_psn = 0;
424         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
425
426         return ret;
427 }
428
429 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
430 {
431         struct ib_qp_attr qp_attr;
432         int qp_attr_mask, ret;
433
434         qp_attr.qp_state = IB_QPS_INIT;
435         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
436         if (ret)
437                 return ret;
438
439         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
440 }
441
442 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
443                    struct ib_qp_init_attr *qp_init_attr)
444 {
445         struct rdma_id_private *id_priv;
446         struct ib_qp *qp;
447         int ret;
448
449         id_priv = container_of(id, struct rdma_id_private, id);
450         if (id->device != pd->device)
451                 return -EINVAL;
452
453         qp = ib_create_qp(pd, qp_init_attr);
454         if (IS_ERR(qp))
455                 return PTR_ERR(qp);
456
457         if (cma_is_ud_ps(id_priv->id.ps))
458                 ret = cma_init_ud_qp(id_priv, qp);
459         else
460                 ret = cma_init_conn_qp(id_priv, qp);
461         if (ret)
462                 goto err;
463
464         id->qp = qp;
465         id_priv->qp_num = qp->qp_num;
466         id_priv->srq = (qp->srq != NULL);
467         return 0;
468 err:
469         ib_destroy_qp(qp);
470         return ret;
471 }
472 EXPORT_SYMBOL(rdma_create_qp);
473
474 void rdma_destroy_qp(struct rdma_cm_id *id)
475 {
476         struct rdma_id_private *id_priv;
477
478         id_priv = container_of(id, struct rdma_id_private, id);
479         mutex_lock(&id_priv->qp_mutex);
480         ib_destroy_qp(id_priv->id.qp);
481         id_priv->id.qp = NULL;
482         mutex_unlock(&id_priv->qp_mutex);
483 }
484 EXPORT_SYMBOL(rdma_destroy_qp);
485
486 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
487                              struct rdma_conn_param *conn_param)
488 {
489         struct ib_qp_attr qp_attr;
490         int qp_attr_mask, ret;
491
492         mutex_lock(&id_priv->qp_mutex);
493         if (!id_priv->id.qp) {
494                 ret = 0;
495                 goto out;
496         }
497
498         /* Need to update QP attributes from default values. */
499         qp_attr.qp_state = IB_QPS_INIT;
500         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
501         if (ret)
502                 goto out;
503
504         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
505         if (ret)
506                 goto out;
507
508         qp_attr.qp_state = IB_QPS_RTR;
509         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
510         if (ret)
511                 goto out;
512
513         if (conn_param)
514                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
515         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
516 out:
517         mutex_unlock(&id_priv->qp_mutex);
518         return ret;
519 }
520
521 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
522                              struct rdma_conn_param *conn_param)
523 {
524         struct ib_qp_attr qp_attr;
525         int qp_attr_mask, ret;
526
527         mutex_lock(&id_priv->qp_mutex);
528         if (!id_priv->id.qp) {
529                 ret = 0;
530                 goto out;
531         }
532
533         qp_attr.qp_state = IB_QPS_RTS;
534         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
535         if (ret)
536                 goto out;
537
538         if (conn_param)
539                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
540         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
541 out:
542         mutex_unlock(&id_priv->qp_mutex);
543         return ret;
544 }
545
546 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
547 {
548         struct ib_qp_attr qp_attr;
549         int ret;
550
551         mutex_lock(&id_priv->qp_mutex);
552         if (!id_priv->id.qp) {
553                 ret = 0;
554                 goto out;
555         }
556
557         qp_attr.qp_state = IB_QPS_ERR;
558         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
559 out:
560         mutex_unlock(&id_priv->qp_mutex);
561         return ret;
562 }
563
564 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
565                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
566 {
567         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
568         int ret;
569
570         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
571                                   ib_addr_get_pkey(dev_addr),
572                                   &qp_attr->pkey_index);
573         if (ret)
574                 return ret;
575
576         qp_attr->port_num = id_priv->id.port_num;
577         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
578
579         if (cma_is_ud_ps(id_priv->id.ps)) {
580                 qp_attr->qkey = id_priv->qkey;
581                 *qp_attr_mask |= IB_QP_QKEY;
582         } else {
583                 qp_attr->qp_access_flags = 0;
584                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
585         }
586         return 0;
587 }
588
589 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
590                        int *qp_attr_mask)
591 {
592         struct rdma_id_private *id_priv;
593         int ret = 0;
594
595         id_priv = container_of(id, struct rdma_id_private, id);
596         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
597         case RDMA_TRANSPORT_IB:
598                 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps))
599                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
600                 else
601                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
602                                                  qp_attr_mask);
603                 if (qp_attr->qp_state == IB_QPS_RTR)
604                         qp_attr->rq_psn = id_priv->seq_num;
605                 break;
606         case RDMA_TRANSPORT_IWARP:
607                 if (!id_priv->cm_id.iw) {
608                         qp_attr->qp_access_flags = 0;
609                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
610                 } else
611                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
612                                                  qp_attr_mask);
613                 break;
614         default:
615                 ret = -ENOSYS;
616                 break;
617         }
618
619         return ret;
620 }
621 EXPORT_SYMBOL(rdma_init_qp_attr);
622
623 static inline int cma_zero_addr(struct sockaddr *addr)
624 {
625         struct in6_addr *ip6;
626
627         if (addr->sa_family == AF_INET)
628                 return ipv4_is_zeronet(
629                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
630         else {
631                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
632                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
633                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
634         }
635 }
636
637 static inline int cma_loopback_addr(struct sockaddr *addr)
638 {
639         return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
640 }
641
642 static inline int cma_any_addr(struct sockaddr *addr)
643 {
644         return cma_zero_addr(addr) || cma_loopback_addr(addr);
645 }
646
647 static inline __be16 cma_port(struct sockaddr *addr)
648 {
649         if (addr->sa_family == AF_INET)
650                 return ((struct sockaddr_in *) addr)->sin_port;
651         else
652                 return ((struct sockaddr_in6 *) addr)->sin6_port;
653 }
654
655 static inline int cma_any_port(struct sockaddr *addr)
656 {
657         return !cma_port(addr);
658 }
659
660 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
661                             u8 *ip_ver, __be16 *port,
662                             union cma_ip_addr **src, union cma_ip_addr **dst)
663 {
664         switch (ps) {
665         case RDMA_PS_SDP:
666                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
667                     SDP_MAJ_VERSION)
668                         return -EINVAL;
669
670                 *ip_ver = sdp_get_ip_ver(hdr);
671                 *port   = ((struct sdp_hh *) hdr)->port;
672                 *src    = &((struct sdp_hh *) hdr)->src_addr;
673                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
674                 break;
675         default:
676                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
677                         return -EINVAL;
678
679                 *ip_ver = cma_get_ip_ver(hdr);
680                 *port   = ((struct cma_hdr *) hdr)->port;
681                 *src    = &((struct cma_hdr *) hdr)->src_addr;
682                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
683                 break;
684         }
685
686         if (*ip_ver != 4 && *ip_ver != 6)
687                 return -EINVAL;
688         return 0;
689 }
690
691 static void cma_save_net_info(struct rdma_addr *addr,
692                               struct rdma_addr *listen_addr,
693                               u8 ip_ver, __be16 port,
694                               union cma_ip_addr *src, union cma_ip_addr *dst)
695 {
696         struct sockaddr_in *listen4, *ip4;
697         struct sockaddr_in6 *listen6, *ip6;
698
699         switch (ip_ver) {
700         case 4:
701                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
702                 ip4 = (struct sockaddr_in *) &addr->src_addr;
703                 ip4->sin_family = listen4->sin_family;
704                 ip4->sin_addr.s_addr = dst->ip4.addr;
705                 ip4->sin_port = listen4->sin_port;
706
707                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
708                 ip4->sin_family = listen4->sin_family;
709                 ip4->sin_addr.s_addr = src->ip4.addr;
710                 ip4->sin_port = port;
711                 break;
712         case 6:
713                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
714                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
715                 ip6->sin6_family = listen6->sin6_family;
716                 ip6->sin6_addr = dst->ip6;
717                 ip6->sin6_port = listen6->sin6_port;
718
719                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
720                 ip6->sin6_family = listen6->sin6_family;
721                 ip6->sin6_addr = src->ip6;
722                 ip6->sin6_port = port;
723                 break;
724         default:
725                 break;
726         }
727 }
728
729 static inline int cma_user_data_offset(enum rdma_port_space ps)
730 {
731         switch (ps) {
732         case RDMA_PS_SDP:
733                 return 0;
734         default:
735                 return sizeof(struct cma_hdr);
736         }
737 }
738
739 static void cma_cancel_route(struct rdma_id_private *id_priv)
740 {
741         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
742         case RDMA_TRANSPORT_IB:
743                 if (id_priv->query)
744                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
745                 break;
746         default:
747                 break;
748         }
749 }
750
751 static void cma_cancel_listens(struct rdma_id_private *id_priv)
752 {
753         struct rdma_id_private *dev_id_priv;
754
755         /*
756          * Remove from listen_any_list to prevent added devices from spawning
757          * additional listen requests.
758          */
759         mutex_lock(&lock);
760         list_del(&id_priv->list);
761
762         while (!list_empty(&id_priv->listen_list)) {
763                 dev_id_priv = list_entry(id_priv->listen_list.next,
764                                          struct rdma_id_private, listen_list);
765                 /* sync with device removal to avoid duplicate destruction */
766                 list_del_init(&dev_id_priv->list);
767                 list_del(&dev_id_priv->listen_list);
768                 mutex_unlock(&lock);
769
770                 rdma_destroy_id(&dev_id_priv->id);
771                 mutex_lock(&lock);
772         }
773         mutex_unlock(&lock);
774 }
775
776 static void cma_cancel_operation(struct rdma_id_private *id_priv,
777                                  enum cma_state state)
778 {
779         switch (state) {
780         case CMA_ADDR_QUERY:
781                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
782                 break;
783         case CMA_ROUTE_QUERY:
784                 cma_cancel_route(id_priv);
785                 break;
786         case CMA_LISTEN:
787                 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
788                                 && !id_priv->cma_dev)
789                         cma_cancel_listens(id_priv);
790                 break;
791         default:
792                 break;
793         }
794 }
795
796 static void cma_release_port(struct rdma_id_private *id_priv)
797 {
798         struct rdma_bind_list *bind_list = id_priv->bind_list;
799
800         if (!bind_list)
801                 return;
802
803         mutex_lock(&lock);
804         hlist_del(&id_priv->node);
805         if (hlist_empty(&bind_list->owners)) {
806                 idr_remove(bind_list->ps, bind_list->port);
807                 kfree(bind_list);
808         }
809         mutex_unlock(&lock);
810 }
811
812 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
813 {
814         struct cma_multicast *mc;
815
816         while (!list_empty(&id_priv->mc_list)) {
817                 mc = container_of(id_priv->mc_list.next,
818                                   struct cma_multicast, list);
819                 list_del(&mc->list);
820                 ib_sa_free_multicast(mc->multicast.ib);
821                 kfree(mc);
822         }
823 }
824
825 void rdma_destroy_id(struct rdma_cm_id *id)
826 {
827         struct rdma_id_private *id_priv;
828         enum cma_state state;
829
830         id_priv = container_of(id, struct rdma_id_private, id);
831         state = cma_exch(id_priv, CMA_DESTROYING);
832         cma_cancel_operation(id_priv, state);
833
834         mutex_lock(&lock);
835         if (id_priv->cma_dev) {
836                 mutex_unlock(&lock);
837                 switch (rdma_node_get_transport(id->device->node_type)) {
838                 case RDMA_TRANSPORT_IB:
839                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
840                                 ib_destroy_cm_id(id_priv->cm_id.ib);
841                         break;
842                 case RDMA_TRANSPORT_IWARP:
843                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
844                                 iw_destroy_cm_id(id_priv->cm_id.iw);
845                         break;
846                 default:
847                         break;
848                 }
849                 cma_leave_mc_groups(id_priv);
850                 mutex_lock(&lock);
851                 cma_detach_from_dev(id_priv);
852         }
853         mutex_unlock(&lock);
854
855         cma_release_port(id_priv);
856         cma_deref_id(id_priv);
857         wait_for_completion(&id_priv->comp);
858
859         if (id_priv->internal_id)
860                 cma_deref_id(id_priv->id.context);
861
862         kfree(id_priv->id.route.path_rec);
863         kfree(id_priv);
864 }
865 EXPORT_SYMBOL(rdma_destroy_id);
866
867 static int cma_rep_recv(struct rdma_id_private *id_priv)
868 {
869         int ret;
870
871         ret = cma_modify_qp_rtr(id_priv, NULL);
872         if (ret)
873                 goto reject;
874
875         ret = cma_modify_qp_rts(id_priv, NULL);
876         if (ret)
877                 goto reject;
878
879         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
880         if (ret)
881                 goto reject;
882
883         return 0;
884 reject:
885         cma_modify_qp_err(id_priv);
886         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
887                        NULL, 0, NULL, 0);
888         return ret;
889 }
890
891 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
892 {
893         if (id_priv->id.ps == RDMA_PS_SDP &&
894             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
895             SDP_MAJ_VERSION)
896                 return -EINVAL;
897
898         return 0;
899 }
900
901 static void cma_set_rep_event_data(struct rdma_cm_event *event,
902                                    struct ib_cm_rep_event_param *rep_data,
903                                    void *private_data)
904 {
905         event->param.conn.private_data = private_data;
906         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
907         event->param.conn.responder_resources = rep_data->responder_resources;
908         event->param.conn.initiator_depth = rep_data->initiator_depth;
909         event->param.conn.flow_control = rep_data->flow_control;
910         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
911         event->param.conn.srq = rep_data->srq;
912         event->param.conn.qp_num = rep_data->remote_qpn;
913 }
914
915 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
916 {
917         struct rdma_id_private *id_priv = cm_id->context;
918         struct rdma_cm_event event;
919         int ret = 0;
920
921         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
922                 cma_disable_callback(id_priv, CMA_CONNECT)) ||
923             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
924                 cma_disable_callback(id_priv, CMA_DISCONNECT)))
925                 return 0;
926
927         memset(&event, 0, sizeof event);
928         switch (ib_event->event) {
929         case IB_CM_REQ_ERROR:
930         case IB_CM_REP_ERROR:
931                 event.event = RDMA_CM_EVENT_UNREACHABLE;
932                 event.status = -ETIMEDOUT;
933                 break;
934         case IB_CM_REP_RECEIVED:
935                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
936                 if (event.status)
937                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
938                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
939                         event.status = cma_rep_recv(id_priv);
940                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
941                                                      RDMA_CM_EVENT_ESTABLISHED;
942                 } else
943                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
944                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
945                                        ib_event->private_data);
946                 break;
947         case IB_CM_RTU_RECEIVED:
948         case IB_CM_USER_ESTABLISHED:
949                 event.event = RDMA_CM_EVENT_ESTABLISHED;
950                 break;
951         case IB_CM_DREQ_ERROR:
952                 event.status = -ETIMEDOUT; /* fall through */
953         case IB_CM_DREQ_RECEIVED:
954         case IB_CM_DREP_RECEIVED:
955                 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
956                         goto out;
957                 event.event = RDMA_CM_EVENT_DISCONNECTED;
958                 break;
959         case IB_CM_TIMEWAIT_EXIT:
960                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
961                 break;
962         case IB_CM_MRA_RECEIVED:
963                 /* ignore event */
964                 goto out;
965         case IB_CM_REJ_RECEIVED:
966                 cma_modify_qp_err(id_priv);
967                 event.status = ib_event->param.rej_rcvd.reason;
968                 event.event = RDMA_CM_EVENT_REJECTED;
969                 event.param.conn.private_data = ib_event->private_data;
970                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
971                 break;
972         default:
973                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
974                        ib_event->event);
975                 goto out;
976         }
977
978         ret = id_priv->id.event_handler(&id_priv->id, &event);
979         if (ret) {
980                 /* Destroy the CM ID by returning a non-zero value. */
981                 id_priv->cm_id.ib = NULL;
982                 cma_exch(id_priv, CMA_DESTROYING);
983                 mutex_unlock(&id_priv->handler_mutex);
984                 rdma_destroy_id(&id_priv->id);
985                 return ret;
986         }
987 out:
988         mutex_unlock(&id_priv->handler_mutex);
989         return ret;
990 }
991
992 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
993                                                struct ib_cm_event *ib_event)
994 {
995         struct rdma_id_private *id_priv;
996         struct rdma_cm_id *id;
997         struct rdma_route *rt;
998         union cma_ip_addr *src, *dst;
999         __be16 port;
1000         u8 ip_ver;
1001         int ret;
1002
1003         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1004                              &ip_ver, &port, &src, &dst))
1005                 goto err;
1006
1007         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1008                             listen_id->ps);
1009         if (IS_ERR(id))
1010                 goto err;
1011
1012         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1013                           ip_ver, port, src, dst);
1014
1015         rt = &id->route;
1016         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1017         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1018                                GFP_KERNEL);
1019         if (!rt->path_rec)
1020                 goto destroy_id;
1021
1022         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1023         if (rt->num_paths == 2)
1024                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1025
1026         ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1027         ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1028                                 &id->route.addr.dev_addr);
1029         if (ret)
1030                 goto destroy_id;
1031
1032         id_priv = container_of(id, struct rdma_id_private, id);
1033         id_priv->state = CMA_CONNECT;
1034         return id_priv;
1035
1036 destroy_id:
1037         rdma_destroy_id(id);
1038 err:
1039         return NULL;
1040 }
1041
1042 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1043                                               struct ib_cm_event *ib_event)
1044 {
1045         struct rdma_id_private *id_priv;
1046         struct rdma_cm_id *id;
1047         union cma_ip_addr *src, *dst;
1048         __be16 port;
1049         u8 ip_ver;
1050         int ret;
1051
1052         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1053                             listen_id->ps);
1054         if (IS_ERR(id))
1055                 return NULL;
1056
1057
1058         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1059                              &ip_ver, &port, &src, &dst))
1060                 goto err;
1061
1062         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1063                           ip_ver, port, src, dst);
1064
1065         ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1066                                 &id->route.addr.dev_addr);
1067         if (ret)
1068                 goto err;
1069
1070         id_priv = container_of(id, struct rdma_id_private, id);
1071         id_priv->state = CMA_CONNECT;
1072         return id_priv;
1073 err:
1074         rdma_destroy_id(id);
1075         return NULL;
1076 }
1077
1078 static void cma_set_req_event_data(struct rdma_cm_event *event,
1079                                    struct ib_cm_req_event_param *req_data,
1080                                    void *private_data, int offset)
1081 {
1082         event->param.conn.private_data = private_data + offset;
1083         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1084         event->param.conn.responder_resources = req_data->responder_resources;
1085         event->param.conn.initiator_depth = req_data->initiator_depth;
1086         event->param.conn.flow_control = req_data->flow_control;
1087         event->param.conn.retry_count = req_data->retry_count;
1088         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1089         event->param.conn.srq = req_data->srq;
1090         event->param.conn.qp_num = req_data->remote_qpn;
1091 }
1092
1093 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1094 {
1095         struct rdma_id_private *listen_id, *conn_id;
1096         struct rdma_cm_event event;
1097         int offset, ret;
1098
1099         listen_id = cm_id->context;
1100         if (cma_disable_callback(listen_id, CMA_LISTEN))
1101                 return -ECONNABORTED;
1102
1103         memset(&event, 0, sizeof event);
1104         offset = cma_user_data_offset(listen_id->id.ps);
1105         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1106         if (cma_is_ud_ps(listen_id->id.ps)) {
1107                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1108                 event.param.ud.private_data = ib_event->private_data + offset;
1109                 event.param.ud.private_data_len =
1110                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1111         } else {
1112                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1113                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1114                                        ib_event->private_data, offset);
1115         }
1116         if (!conn_id) {
1117                 ret = -ENOMEM;
1118                 goto out;
1119         }
1120
1121         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1122         mutex_lock(&lock);
1123         ret = cma_acquire_dev(conn_id);
1124         mutex_unlock(&lock);
1125         if (ret)
1126                 goto release_conn_id;
1127
1128         conn_id->cm_id.ib = cm_id;
1129         cm_id->context = conn_id;
1130         cm_id->cm_handler = cma_ib_handler;
1131
1132         ret = conn_id->id.event_handler(&conn_id->id, &event);
1133         if (!ret) {
1134                 /*
1135                  * Acquire mutex to prevent user executing rdma_destroy_id()
1136                  * while we're accessing the cm_id.
1137                  */
1138                 mutex_lock(&lock);
1139                 if (cma_comp(conn_id, CMA_CONNECT) &&
1140                     !cma_is_ud_ps(conn_id->id.ps))
1141                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1142                 mutex_unlock(&lock);
1143                 mutex_unlock(&conn_id->handler_mutex);
1144                 goto out;
1145         }
1146
1147         /* Destroy the CM ID by returning a non-zero value. */
1148         conn_id->cm_id.ib = NULL;
1149
1150 release_conn_id:
1151         cma_exch(conn_id, CMA_DESTROYING);
1152         mutex_unlock(&conn_id->handler_mutex);
1153         rdma_destroy_id(&conn_id->id);
1154
1155 out:
1156         mutex_unlock(&listen_id->handler_mutex);
1157         return ret;
1158 }
1159
1160 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1161 {
1162         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1163 }
1164
1165 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1166                                  struct ib_cm_compare_data *compare)
1167 {
1168         struct cma_hdr *cma_data, *cma_mask;
1169         struct sdp_hh *sdp_data, *sdp_mask;
1170         __be32 ip4_addr;
1171         struct in6_addr ip6_addr;
1172
1173         memset(compare, 0, sizeof *compare);
1174         cma_data = (void *) compare->data;
1175         cma_mask = (void *) compare->mask;
1176         sdp_data = (void *) compare->data;
1177         sdp_mask = (void *) compare->mask;
1178
1179         switch (addr->sa_family) {
1180         case AF_INET:
1181                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1182                 if (ps == RDMA_PS_SDP) {
1183                         sdp_set_ip_ver(sdp_data, 4);
1184                         sdp_set_ip_ver(sdp_mask, 0xF);
1185                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1186                         sdp_mask->dst_addr.ip4.addr = htonl(~0);
1187                 } else {
1188                         cma_set_ip_ver(cma_data, 4);
1189                         cma_set_ip_ver(cma_mask, 0xF);
1190                         cma_data->dst_addr.ip4.addr = ip4_addr;
1191                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1192                 }
1193                 break;
1194         case AF_INET6:
1195                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1196                 if (ps == RDMA_PS_SDP) {
1197                         sdp_set_ip_ver(sdp_data, 6);
1198                         sdp_set_ip_ver(sdp_mask, 0xF);
1199                         sdp_data->dst_addr.ip6 = ip6_addr;
1200                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1201                                sizeof sdp_mask->dst_addr.ip6);
1202                 } else {
1203                         cma_set_ip_ver(cma_data, 6);
1204                         cma_set_ip_ver(cma_mask, 0xF);
1205                         cma_data->dst_addr.ip6 = ip6_addr;
1206                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1207                                sizeof cma_mask->dst_addr.ip6);
1208                 }
1209                 break;
1210         default:
1211                 break;
1212         }
1213 }
1214
1215 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1216 {
1217         struct rdma_id_private *id_priv = iw_id->context;
1218         struct rdma_cm_event event;
1219         struct sockaddr_in *sin;
1220         int ret = 0;
1221
1222         if (cma_disable_callback(id_priv, CMA_CONNECT))
1223                 return 0;
1224
1225         memset(&event, 0, sizeof event);
1226         switch (iw_event->event) {
1227         case IW_CM_EVENT_CLOSE:
1228                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1229                 break;
1230         case IW_CM_EVENT_CONNECT_REPLY:
1231                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1232                 *sin = iw_event->local_addr;
1233                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1234                 *sin = iw_event->remote_addr;
1235                 switch (iw_event->status) {
1236                 case 0:
1237                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1238                         break;
1239                 case -ECONNRESET:
1240                 case -ECONNREFUSED:
1241                         event.event = RDMA_CM_EVENT_REJECTED;
1242                         break;
1243                 case -ETIMEDOUT:
1244                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1245                         break;
1246                 default:
1247                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1248                         break;
1249                 }
1250                 break;
1251         case IW_CM_EVENT_ESTABLISHED:
1252                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1253                 break;
1254         default:
1255                 BUG_ON(1);
1256         }
1257
1258         event.status = iw_event->status;
1259         event.param.conn.private_data = iw_event->private_data;
1260         event.param.conn.private_data_len = iw_event->private_data_len;
1261         ret = id_priv->id.event_handler(&id_priv->id, &event);
1262         if (ret) {
1263                 /* Destroy the CM ID by returning a non-zero value. */
1264                 id_priv->cm_id.iw = NULL;
1265                 cma_exch(id_priv, CMA_DESTROYING);
1266                 mutex_unlock(&id_priv->handler_mutex);
1267                 rdma_destroy_id(&id_priv->id);
1268                 return ret;
1269         }
1270
1271         mutex_unlock(&id_priv->handler_mutex);
1272         return ret;
1273 }
1274
1275 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1276                                struct iw_cm_event *iw_event)
1277 {
1278         struct rdma_cm_id *new_cm_id;
1279         struct rdma_id_private *listen_id, *conn_id;
1280         struct sockaddr_in *sin;
1281         struct net_device *dev = NULL;
1282         struct rdma_cm_event event;
1283         int ret;
1284         struct ib_device_attr attr;
1285
1286         listen_id = cm_id->context;
1287         if (cma_disable_callback(listen_id, CMA_LISTEN))
1288                 return -ECONNABORTED;
1289
1290         /* Create a new RDMA id for the new IW CM ID */
1291         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1292                                    listen_id->id.context,
1293                                    RDMA_PS_TCP);
1294         if (IS_ERR(new_cm_id)) {
1295                 ret = -ENOMEM;
1296                 goto out;
1297         }
1298         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1299         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1300         conn_id->state = CMA_CONNECT;
1301
1302         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1303         if (!dev) {
1304                 ret = -EADDRNOTAVAIL;
1305                 mutex_unlock(&conn_id->handler_mutex);
1306                 rdma_destroy_id(new_cm_id);
1307                 goto out;
1308         }
1309         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1310         if (ret) {
1311                 mutex_unlock(&conn_id->handler_mutex);
1312                 rdma_destroy_id(new_cm_id);
1313                 goto out;
1314         }
1315
1316         mutex_lock(&lock);
1317         ret = cma_acquire_dev(conn_id);
1318         mutex_unlock(&lock);
1319         if (ret) {
1320                 mutex_unlock(&conn_id->handler_mutex);
1321                 rdma_destroy_id(new_cm_id);
1322                 goto out;
1323         }
1324
1325         conn_id->cm_id.iw = cm_id;
1326         cm_id->context = conn_id;
1327         cm_id->cm_handler = cma_iw_handler;
1328
1329         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1330         *sin = iw_event->local_addr;
1331         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1332         *sin = iw_event->remote_addr;
1333
1334         ret = ib_query_device(conn_id->id.device, &attr);
1335         if (ret) {
1336                 mutex_unlock(&conn_id->handler_mutex);
1337                 rdma_destroy_id(new_cm_id);
1338                 goto out;
1339         }
1340
1341         memset(&event, 0, sizeof event);
1342         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1343         event.param.conn.private_data = iw_event->private_data;
1344         event.param.conn.private_data_len = iw_event->private_data_len;
1345         event.param.conn.initiator_depth = attr.max_qp_init_rd_atom;
1346         event.param.conn.responder_resources = attr.max_qp_rd_atom;
1347         ret = conn_id->id.event_handler(&conn_id->id, &event);
1348         if (ret) {
1349                 /* User wants to destroy the CM ID */
1350                 conn_id->cm_id.iw = NULL;
1351                 cma_exch(conn_id, CMA_DESTROYING);
1352                 mutex_unlock(&conn_id->handler_mutex);
1353                 rdma_destroy_id(&conn_id->id);
1354                 goto out;
1355         }
1356
1357         mutex_unlock(&conn_id->handler_mutex);
1358
1359 out:
1360         if (dev)
1361                 dev_put(dev);
1362         mutex_unlock(&listen_id->handler_mutex);
1363         return ret;
1364 }
1365
1366 static int cma_ib_listen(struct rdma_id_private *id_priv)
1367 {
1368         struct ib_cm_compare_data compare_data;
1369         struct sockaddr *addr;
1370         __be64 svc_id;
1371         int ret;
1372
1373         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1374                                             id_priv);
1375         if (IS_ERR(id_priv->cm_id.ib))
1376                 return PTR_ERR(id_priv->cm_id.ib);
1377
1378         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1379         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1380         if (cma_any_addr(addr))
1381                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1382         else {
1383                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1384                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1385         }
1386
1387         if (ret) {
1388                 ib_destroy_cm_id(id_priv->cm_id.ib);
1389                 id_priv->cm_id.ib = NULL;
1390         }
1391
1392         return ret;
1393 }
1394
1395 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1396 {
1397         int ret;
1398         struct sockaddr_in *sin;
1399
1400         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1401                                             iw_conn_req_handler,
1402                                             id_priv);
1403         if (IS_ERR(id_priv->cm_id.iw))
1404                 return PTR_ERR(id_priv->cm_id.iw);
1405
1406         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1407         id_priv->cm_id.iw->local_addr = *sin;
1408
1409         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1410
1411         if (ret) {
1412                 iw_destroy_cm_id(id_priv->cm_id.iw);
1413                 id_priv->cm_id.iw = NULL;
1414         }
1415
1416         return ret;
1417 }
1418
1419 static int cma_listen_handler(struct rdma_cm_id *id,
1420                               struct rdma_cm_event *event)
1421 {
1422         struct rdma_id_private *id_priv = id->context;
1423
1424         id->context = id_priv->id.context;
1425         id->event_handler = id_priv->id.event_handler;
1426         return id_priv->id.event_handler(id, event);
1427 }
1428
1429 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1430                               struct cma_device *cma_dev)
1431 {
1432         struct rdma_id_private *dev_id_priv;
1433         struct rdma_cm_id *id;
1434         int ret;
1435
1436         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1437         if (IS_ERR(id))
1438                 return;
1439
1440         dev_id_priv = container_of(id, struct rdma_id_private, id);
1441
1442         dev_id_priv->state = CMA_ADDR_BOUND;
1443         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1444                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1445
1446         cma_attach_to_dev(dev_id_priv, cma_dev);
1447         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1448         atomic_inc(&id_priv->refcount);
1449         dev_id_priv->internal_id = 1;
1450
1451         ret = rdma_listen(id, id_priv->backlog);
1452         if (ret)
1453                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1454                        "listening on device %s\n", ret, cma_dev->device->name);
1455 }
1456
1457 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1458 {
1459         struct cma_device *cma_dev;
1460
1461         mutex_lock(&lock);
1462         list_add_tail(&id_priv->list, &listen_any_list);
1463         list_for_each_entry(cma_dev, &dev_list, list)
1464                 cma_listen_on_dev(id_priv, cma_dev);
1465         mutex_unlock(&lock);
1466 }
1467
1468 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1469 {
1470         struct sockaddr_in addr_in;
1471
1472         memset(&addr_in, 0, sizeof addr_in);
1473         addr_in.sin_family = af;
1474         return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1475 }
1476
1477 int rdma_listen(struct rdma_cm_id *id, int backlog)
1478 {
1479         struct rdma_id_private *id_priv;
1480         int ret;
1481
1482         id_priv = container_of(id, struct rdma_id_private, id);
1483         if (id_priv->state == CMA_IDLE) {
1484                 ret = cma_bind_any(id, AF_INET);
1485                 if (ret)
1486                         return ret;
1487         }
1488
1489         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1490                 return -EINVAL;
1491
1492         id_priv->backlog = backlog;
1493         if (id->device) {
1494                 switch (rdma_node_get_transport(id->device->node_type)) {
1495                 case RDMA_TRANSPORT_IB:
1496                         ret = cma_ib_listen(id_priv);
1497                         if (ret)
1498                                 goto err;
1499                         break;
1500                 case RDMA_TRANSPORT_IWARP:
1501                         ret = cma_iw_listen(id_priv, backlog);
1502                         if (ret)
1503                                 goto err;
1504                         break;
1505                 default:
1506                         ret = -ENOSYS;
1507                         goto err;
1508                 }
1509         } else
1510                 cma_listen_on_all(id_priv);
1511
1512         return 0;
1513 err:
1514         id_priv->backlog = 0;
1515         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1516         return ret;
1517 }
1518 EXPORT_SYMBOL(rdma_listen);
1519
1520 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1521 {
1522         struct rdma_id_private *id_priv;
1523
1524         id_priv = container_of(id, struct rdma_id_private, id);
1525         id_priv->tos = (u8) tos;
1526 }
1527 EXPORT_SYMBOL(rdma_set_service_type);
1528
1529 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1530                               void *context)
1531 {
1532         struct cma_work *work = context;
1533         struct rdma_route *route;
1534
1535         route = &work->id->id.route;
1536
1537         if (!status) {
1538                 route->num_paths = 1;
1539                 *route->path_rec = *path_rec;
1540         } else {
1541                 work->old_state = CMA_ROUTE_QUERY;
1542                 work->new_state = CMA_ADDR_RESOLVED;
1543                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1544                 work->event.status = status;
1545         }
1546
1547         queue_work(cma_wq, &work->work);
1548 }
1549
1550 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1551                               struct cma_work *work)
1552 {
1553         struct rdma_addr *addr = &id_priv->id.route.addr;
1554         struct ib_sa_path_rec path_rec;
1555         ib_sa_comp_mask comp_mask;
1556         struct sockaddr_in6 *sin6;
1557
1558         memset(&path_rec, 0, sizeof path_rec);
1559         ib_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1560         ib_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1561         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1562         path_rec.numb_path = 1;
1563         path_rec.reversible = 1;
1564         path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1565                                                         (struct sockaddr *) &addr->dst_addr);
1566
1567         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1568                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1569                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1570
1571         if (addr->src_addr.ss_family == AF_INET) {
1572                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1573                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1574         } else {
1575                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1576                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1577                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1578         }
1579
1580         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1581                                                id_priv->id.port_num, &path_rec,
1582                                                comp_mask, timeout_ms,
1583                                                GFP_KERNEL, cma_query_handler,
1584                                                work, &id_priv->query);
1585
1586         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1587 }
1588
1589 static void cma_work_handler(struct work_struct *_work)
1590 {
1591         struct cma_work *work = container_of(_work, struct cma_work, work);
1592         struct rdma_id_private *id_priv = work->id;
1593         int destroy = 0;
1594
1595         mutex_lock(&id_priv->handler_mutex);
1596         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1597                 goto out;
1598
1599         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1600                 cma_exch(id_priv, CMA_DESTROYING);
1601                 destroy = 1;
1602         }
1603 out:
1604         mutex_unlock(&id_priv->handler_mutex);
1605         cma_deref_id(id_priv);
1606         if (destroy)
1607                 rdma_destroy_id(&id_priv->id);
1608         kfree(work);
1609 }
1610
1611 static void cma_ndev_work_handler(struct work_struct *_work)
1612 {
1613         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1614         struct rdma_id_private *id_priv = work->id;
1615         int destroy = 0;
1616
1617         mutex_lock(&id_priv->handler_mutex);
1618         if (id_priv->state == CMA_DESTROYING ||
1619             id_priv->state == CMA_DEVICE_REMOVAL)
1620                 goto out;
1621
1622         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1623                 cma_exch(id_priv, CMA_DESTROYING);
1624                 destroy = 1;
1625         }
1626
1627 out:
1628         mutex_unlock(&id_priv->handler_mutex);
1629         cma_deref_id(id_priv);
1630         if (destroy)
1631                 rdma_destroy_id(&id_priv->id);
1632         kfree(work);
1633 }
1634
1635 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1636 {
1637         struct rdma_route *route = &id_priv->id.route;
1638         struct cma_work *work;
1639         int ret;
1640
1641         work = kzalloc(sizeof *work, GFP_KERNEL);
1642         if (!work)
1643                 return -ENOMEM;
1644
1645         work->id = id_priv;
1646         INIT_WORK(&work->work, cma_work_handler);
1647         work->old_state = CMA_ROUTE_QUERY;
1648         work->new_state = CMA_ROUTE_RESOLVED;
1649         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1650
1651         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1652         if (!route->path_rec) {
1653                 ret = -ENOMEM;
1654                 goto err1;
1655         }
1656
1657         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1658         if (ret)
1659                 goto err2;
1660
1661         return 0;
1662 err2:
1663         kfree(route->path_rec);
1664         route->path_rec = NULL;
1665 err1:
1666         kfree(work);
1667         return ret;
1668 }
1669
1670 int rdma_set_ib_paths(struct rdma_cm_id *id,
1671                       struct ib_sa_path_rec *path_rec, int num_paths)
1672 {
1673         struct rdma_id_private *id_priv;
1674         int ret;
1675
1676         id_priv = container_of(id, struct rdma_id_private, id);
1677         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1678                 return -EINVAL;
1679
1680         id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1681         if (!id->route.path_rec) {
1682                 ret = -ENOMEM;
1683                 goto err;
1684         }
1685
1686         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1687         return 0;
1688 err:
1689         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1690         return ret;
1691 }
1692 EXPORT_SYMBOL(rdma_set_ib_paths);
1693
1694 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1695 {
1696         struct cma_work *work;
1697
1698         work = kzalloc(sizeof *work, GFP_KERNEL);
1699         if (!work)
1700                 return -ENOMEM;
1701
1702         work->id = id_priv;
1703         INIT_WORK(&work->work, cma_work_handler);
1704         work->old_state = CMA_ROUTE_QUERY;
1705         work->new_state = CMA_ROUTE_RESOLVED;
1706         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1707         queue_work(cma_wq, &work->work);
1708         return 0;
1709 }
1710
1711 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1712 {
1713         struct rdma_id_private *id_priv;
1714         int ret;
1715
1716         id_priv = container_of(id, struct rdma_id_private, id);
1717         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1718                 return -EINVAL;
1719
1720         atomic_inc(&id_priv->refcount);
1721         switch (rdma_node_get_transport(id->device->node_type)) {
1722         case RDMA_TRANSPORT_IB:
1723                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1724                 break;
1725         case RDMA_TRANSPORT_IWARP:
1726                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1727                 break;
1728         default:
1729                 ret = -ENOSYS;
1730                 break;
1731         }
1732         if (ret)
1733                 goto err;
1734
1735         return 0;
1736 err:
1737         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1738         cma_deref_id(id_priv);
1739         return ret;
1740 }
1741 EXPORT_SYMBOL(rdma_resolve_route);
1742
1743 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1744 {
1745         struct cma_device *cma_dev;
1746         struct ib_port_attr port_attr;
1747         union ib_gid gid;
1748         u16 pkey;
1749         int ret;
1750         u8 p;
1751
1752         mutex_lock(&lock);
1753         if (list_empty(&dev_list)) {
1754                 ret = -ENODEV;
1755                 goto out;
1756         }
1757         list_for_each_entry(cma_dev, &dev_list, list)
1758                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1759                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1760                             port_attr.state == IB_PORT_ACTIVE)
1761                                 goto port_found;
1762
1763         p = 1;
1764         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1765
1766 port_found:
1767         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1768         if (ret)
1769                 goto out;
1770
1771         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1772         if (ret)
1773                 goto out;
1774
1775         ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1776         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1777         id_priv->id.port_num = p;
1778         cma_attach_to_dev(id_priv, cma_dev);
1779 out:
1780         mutex_unlock(&lock);
1781         return ret;
1782 }
1783
1784 static void addr_handler(int status, struct sockaddr *src_addr,
1785                          struct rdma_dev_addr *dev_addr, void *context)
1786 {
1787         struct rdma_id_private *id_priv = context;
1788         struct rdma_cm_event event;
1789
1790         memset(&event, 0, sizeof event);
1791         mutex_lock(&id_priv->handler_mutex);
1792
1793         /*
1794          * Grab mutex to block rdma_destroy_id() from removing the device while
1795          * we're trying to acquire it.
1796          */
1797         mutex_lock(&lock);
1798         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1799                 mutex_unlock(&lock);
1800                 goto out;
1801         }
1802
1803         if (!status && !id_priv->cma_dev)
1804                 status = cma_acquire_dev(id_priv);
1805         mutex_unlock(&lock);
1806
1807         if (status) {
1808                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1809                         goto out;
1810                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1811                 event.status = status;
1812         } else {
1813                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1814                        ip_addr_size(src_addr));
1815                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1816         }
1817
1818         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1819                 cma_exch(id_priv, CMA_DESTROYING);
1820                 mutex_unlock(&id_priv->handler_mutex);
1821                 cma_deref_id(id_priv);
1822                 rdma_destroy_id(&id_priv->id);
1823                 return;
1824         }
1825 out:
1826         mutex_unlock(&id_priv->handler_mutex);
1827         cma_deref_id(id_priv);
1828 }
1829
1830 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1831 {
1832         struct cma_work *work;
1833         struct sockaddr_in *src_in, *dst_in;
1834         union ib_gid gid;
1835         int ret;
1836
1837         work = kzalloc(sizeof *work, GFP_KERNEL);
1838         if (!work)
1839                 return -ENOMEM;
1840
1841         if (!id_priv->cma_dev) {
1842                 ret = cma_bind_loopback(id_priv);
1843                 if (ret)
1844                         goto err;
1845         }
1846
1847         ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1848         ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1849
1850         if (cma_zero_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)) {
1851                 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1852                 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1853                 src_in->sin_family = dst_in->sin_family;
1854                 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1855         }
1856
1857         work->id = id_priv;
1858         INIT_WORK(&work->work, cma_work_handler);
1859         work->old_state = CMA_ADDR_QUERY;
1860         work->new_state = CMA_ADDR_RESOLVED;
1861         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1862         queue_work(cma_wq, &work->work);
1863         return 0;
1864 err:
1865         kfree(work);
1866         return ret;
1867 }
1868
1869 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1870                          struct sockaddr *dst_addr)
1871 {
1872         if (src_addr && src_addr->sa_family)
1873                 return rdma_bind_addr(id, src_addr);
1874         else
1875                 return cma_bind_any(id, dst_addr->sa_family);
1876 }
1877
1878 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1879                       struct sockaddr *dst_addr, int timeout_ms)
1880 {
1881         struct rdma_id_private *id_priv;
1882         int ret;
1883
1884         id_priv = container_of(id, struct rdma_id_private, id);
1885         if (id_priv->state == CMA_IDLE) {
1886                 ret = cma_bind_addr(id, src_addr, dst_addr);
1887                 if (ret)
1888                         return ret;
1889         }
1890
1891         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1892                 return -EINVAL;
1893
1894         atomic_inc(&id_priv->refcount);
1895         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1896         if (cma_any_addr(dst_addr))
1897                 ret = cma_resolve_loopback(id_priv);
1898         else
1899                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
1900                                       dst_addr, &id->route.addr.dev_addr,
1901                                       timeout_ms, addr_handler, id_priv);
1902         if (ret)
1903                 goto err;
1904
1905         return 0;
1906 err:
1907         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1908         cma_deref_id(id_priv);
1909         return ret;
1910 }
1911 EXPORT_SYMBOL(rdma_resolve_addr);
1912
1913 static void cma_bind_port(struct rdma_bind_list *bind_list,
1914                           struct rdma_id_private *id_priv)
1915 {
1916         struct sockaddr_in *sin;
1917
1918         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1919         sin->sin_port = htons(bind_list->port);
1920         id_priv->bind_list = bind_list;
1921         hlist_add_head(&id_priv->node, &bind_list->owners);
1922 }
1923
1924 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1925                           unsigned short snum)
1926 {
1927         struct rdma_bind_list *bind_list;
1928         int port, ret;
1929
1930         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1931         if (!bind_list)
1932                 return -ENOMEM;
1933
1934         do {
1935                 ret = idr_get_new_above(ps, bind_list, snum, &port);
1936         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1937
1938         if (ret)
1939                 goto err1;
1940
1941         if (port != snum) {
1942                 ret = -EADDRNOTAVAIL;
1943                 goto err2;
1944         }
1945
1946         bind_list->ps = ps;
1947         bind_list->port = (unsigned short) port;
1948         cma_bind_port(bind_list, id_priv);
1949         return 0;
1950 err2:
1951         idr_remove(ps, port);
1952 err1:
1953         kfree(bind_list);
1954         return ret;
1955 }
1956
1957 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
1958 {
1959         struct rdma_bind_list *bind_list;
1960         int port, ret, low, high;
1961
1962         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1963         if (!bind_list)
1964                 return -ENOMEM;
1965
1966 retry:
1967         /* FIXME: add proper port randomization per like inet_csk_get_port */
1968         do {
1969                 ret = idr_get_new_above(ps, bind_list, next_port, &port);
1970         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1971
1972         if (ret)
1973                 goto err1;
1974
1975         inet_get_local_port_range(&low, &high);
1976         if (port > high) {
1977                 if (next_port != low) {
1978                         idr_remove(ps, port);
1979                         next_port = low;
1980                         goto retry;
1981                 }
1982                 ret = -EADDRNOTAVAIL;
1983                 goto err2;
1984         }
1985
1986         if (port == high)
1987                 next_port = low;
1988         else
1989                 next_port = port + 1;
1990
1991         bind_list->ps = ps;
1992         bind_list->port = (unsigned short) port;
1993         cma_bind_port(bind_list, id_priv);
1994         return 0;
1995 err2:
1996         idr_remove(ps, port);
1997 err1:
1998         kfree(bind_list);
1999         return ret;
2000 }
2001
2002 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2003 {
2004         struct rdma_id_private *cur_id;
2005         struct sockaddr_in *sin, *cur_sin;
2006         struct rdma_bind_list *bind_list;
2007         struct hlist_node *node;
2008         unsigned short snum;
2009
2010         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2011         snum = ntohs(sin->sin_port);
2012         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2013                 return -EACCES;
2014
2015         bind_list = idr_find(ps, snum);
2016         if (!bind_list)
2017                 return cma_alloc_port(ps, id_priv, snum);
2018
2019         /*
2020          * We don't support binding to any address if anyone is bound to
2021          * a specific address on the same port.
2022          */
2023         if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2024                 return -EADDRNOTAVAIL;
2025
2026         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2027                 if (cma_any_addr((struct sockaddr *) &cur_id->id.route.addr.src_addr))
2028                         return -EADDRNOTAVAIL;
2029
2030                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
2031                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
2032                         return -EADDRINUSE;
2033         }
2034
2035         cma_bind_port(bind_list, id_priv);
2036         return 0;
2037 }
2038
2039 static int cma_get_port(struct rdma_id_private *id_priv)
2040 {
2041         struct idr *ps;
2042         int ret;
2043
2044         switch (id_priv->id.ps) {
2045         case RDMA_PS_SDP:
2046                 ps = &sdp_ps;
2047                 break;
2048         case RDMA_PS_TCP:
2049                 ps = &tcp_ps;
2050                 break;
2051         case RDMA_PS_UDP:
2052                 ps = &udp_ps;
2053                 break;
2054         case RDMA_PS_IPOIB:
2055                 ps = &ipoib_ps;
2056                 break;
2057         default:
2058                 return -EPROTONOSUPPORT;
2059         }
2060
2061         mutex_lock(&lock);
2062         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2063                 ret = cma_alloc_any_port(ps, id_priv);
2064         else
2065                 ret = cma_use_port(ps, id_priv);
2066         mutex_unlock(&lock);
2067
2068         return ret;
2069 }
2070
2071 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2072 {
2073         struct rdma_id_private *id_priv;
2074         int ret;
2075
2076         if (addr->sa_family != AF_INET)
2077                 return -EAFNOSUPPORT;
2078
2079         id_priv = container_of(id, struct rdma_id_private, id);
2080         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
2081                 return -EINVAL;
2082
2083         if (!cma_any_addr(addr)) {
2084                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2085                 if (ret)
2086                         goto err1;
2087
2088                 mutex_lock(&lock);
2089                 ret = cma_acquire_dev(id_priv);
2090                 mutex_unlock(&lock);
2091                 if (ret)
2092                         goto err1;
2093         }
2094
2095         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2096         ret = cma_get_port(id_priv);
2097         if (ret)
2098                 goto err2;
2099
2100         return 0;
2101 err2:
2102         if (!cma_any_addr(addr)) {
2103                 mutex_lock(&lock);
2104                 cma_detach_from_dev(id_priv);
2105                 mutex_unlock(&lock);
2106         }
2107 err1:
2108         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
2109         return ret;
2110 }
2111 EXPORT_SYMBOL(rdma_bind_addr);
2112
2113 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2114                           struct rdma_route *route)
2115 {
2116         struct sockaddr_in *src4, *dst4;
2117         struct cma_hdr *cma_hdr;
2118         struct sdp_hh *sdp_hdr;
2119
2120         src4 = (struct sockaddr_in *) &route->addr.src_addr;
2121         dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2122
2123         switch (ps) {
2124         case RDMA_PS_SDP:
2125                 sdp_hdr = hdr;
2126                 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2127                         return -EINVAL;
2128                 sdp_set_ip_ver(sdp_hdr, 4);
2129                 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2130                 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2131                 sdp_hdr->port = src4->sin_port;
2132                 break;
2133         default:
2134                 cma_hdr = hdr;
2135                 cma_hdr->cma_version = CMA_VERSION;
2136                 cma_set_ip_ver(cma_hdr, 4);
2137                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2138                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2139                 cma_hdr->port = src4->sin_port;
2140                 break;
2141         }
2142         return 0;
2143 }
2144
2145 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2146                                 struct ib_cm_event *ib_event)
2147 {
2148         struct rdma_id_private *id_priv = cm_id->context;
2149         struct rdma_cm_event event;
2150         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2151         int ret = 0;
2152
2153         if (cma_disable_callback(id_priv, CMA_CONNECT))
2154                 return 0;
2155
2156         memset(&event, 0, sizeof event);
2157         switch (ib_event->event) {
2158         case IB_CM_SIDR_REQ_ERROR:
2159                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2160                 event.status = -ETIMEDOUT;
2161                 break;
2162         case IB_CM_SIDR_REP_RECEIVED:
2163                 event.param.ud.private_data = ib_event->private_data;
2164                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2165                 if (rep->status != IB_SIDR_SUCCESS) {
2166                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2167                         event.status = ib_event->param.sidr_rep_rcvd.status;
2168                         break;
2169                 }
2170                 if (id_priv->qkey != rep->qkey) {
2171                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2172                         event.status = -EINVAL;
2173                         break;
2174                 }
2175                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2176                                      id_priv->id.route.path_rec,
2177                                      &event.param.ud.ah_attr);
2178                 event.param.ud.qp_num = rep->qpn;
2179                 event.param.ud.qkey = rep->qkey;
2180                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2181                 event.status = 0;
2182                 break;
2183         default:
2184                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2185                        ib_event->event);
2186                 goto out;
2187         }
2188
2189         ret = id_priv->id.event_handler(&id_priv->id, &event);
2190         if (ret) {
2191                 /* Destroy the CM ID by returning a non-zero value. */
2192                 id_priv->cm_id.ib = NULL;
2193                 cma_exch(id_priv, CMA_DESTROYING);
2194                 mutex_unlock(&id_priv->handler_mutex);
2195                 rdma_destroy_id(&id_priv->id);
2196                 return ret;
2197         }
2198 out:
2199         mutex_unlock(&id_priv->handler_mutex);
2200         return ret;
2201 }
2202
2203 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2204                               struct rdma_conn_param *conn_param)
2205 {
2206         struct ib_cm_sidr_req_param req;
2207         struct rdma_route *route;
2208         int ret;
2209
2210         req.private_data_len = sizeof(struct cma_hdr) +
2211                                conn_param->private_data_len;
2212         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2213         if (!req.private_data)
2214                 return -ENOMEM;
2215
2216         if (conn_param->private_data && conn_param->private_data_len)
2217                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2218                        conn_param->private_data, conn_param->private_data_len);
2219
2220         route = &id_priv->id.route;
2221         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2222         if (ret)
2223                 goto out;
2224
2225         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2226                                             cma_sidr_rep_handler, id_priv);
2227         if (IS_ERR(id_priv->cm_id.ib)) {
2228                 ret = PTR_ERR(id_priv->cm_id.ib);
2229                 goto out;
2230         }
2231
2232         req.path = route->path_rec;
2233         req.service_id = cma_get_service_id(id_priv->id.ps,
2234                                             (struct sockaddr *) &route->addr.dst_addr);
2235         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2236         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2237
2238         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2239         if (ret) {
2240                 ib_destroy_cm_id(id_priv->cm_id.ib);
2241                 id_priv->cm_id.ib = NULL;
2242         }
2243 out:
2244         kfree(req.private_data);
2245         return ret;
2246 }
2247
2248 static int cma_connect_ib(struct rdma_id_private *id_priv,
2249                           struct rdma_conn_param *conn_param)
2250 {
2251         struct ib_cm_req_param req;
2252         struct rdma_route *route;
2253         void *private_data;
2254         int offset, ret;
2255
2256         memset(&req, 0, sizeof req);
2257         offset = cma_user_data_offset(id_priv->id.ps);
2258         req.private_data_len = offset + conn_param->private_data_len;
2259         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2260         if (!private_data)
2261                 return -ENOMEM;
2262
2263         if (conn_param->private_data && conn_param->private_data_len)
2264                 memcpy(private_data + offset, conn_param->private_data,
2265                        conn_param->private_data_len);
2266
2267         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2268                                             id_priv);
2269         if (IS_ERR(id_priv->cm_id.ib)) {
2270                 ret = PTR_ERR(id_priv->cm_id.ib);
2271                 goto out;
2272         }
2273
2274         route = &id_priv->id.route;
2275         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2276         if (ret)
2277                 goto out;
2278         req.private_data = private_data;
2279
2280         req.primary_path = &route->path_rec[0];
2281         if (route->num_paths == 2)
2282                 req.alternate_path = &route->path_rec[1];
2283
2284         req.service_id = cma_get_service_id(id_priv->id.ps,
2285                                             (struct sockaddr *) &route->addr.dst_addr);
2286         req.qp_num = id_priv->qp_num;
2287         req.qp_type = IB_QPT_RC;
2288         req.starting_psn = id_priv->seq_num;
2289         req.responder_resources = conn_param->responder_resources;
2290         req.initiator_depth = conn_param->initiator_depth;
2291         req.flow_control = conn_param->flow_control;
2292         req.retry_count = conn_param->retry_count;
2293         req.rnr_retry_count = conn_param->rnr_retry_count;
2294         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2295         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2296         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2297         req.srq = id_priv->srq ? 1 : 0;
2298
2299         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2300 out:
2301         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2302                 ib_destroy_cm_id(id_priv->cm_id.ib);
2303                 id_priv->cm_id.ib = NULL;
2304         }
2305
2306         kfree(private_data);
2307         return ret;
2308 }
2309
2310 static int cma_connect_iw(struct rdma_id_private *id_priv,
2311                           struct rdma_conn_param *conn_param)
2312 {
2313         struct iw_cm_id *cm_id;
2314         struct sockaddr_in* sin;
2315         int ret;
2316         struct iw_cm_conn_param iw_param;
2317
2318         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2319         if (IS_ERR(cm_id)) {
2320                 ret = PTR_ERR(cm_id);
2321                 goto out;
2322         }
2323
2324         id_priv->cm_id.iw = cm_id;
2325
2326         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2327         cm_id->local_addr = *sin;
2328
2329         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2330         cm_id->remote_addr = *sin;
2331
2332         ret = cma_modify_qp_rtr(id_priv, conn_param);
2333         if (ret)
2334                 goto out;
2335
2336         iw_param.ord = conn_param->initiator_depth;
2337         iw_param.ird = conn_param->responder_resources;
2338         iw_param.private_data = conn_param->private_data;
2339         iw_param.private_data_len = conn_param->private_data_len;
2340         if (id_priv->id.qp)
2341                 iw_param.qpn = id_priv->qp_num;
2342         else
2343                 iw_param.qpn = conn_param->qp_num;
2344         ret = iw_cm_connect(cm_id, &iw_param);
2345 out:
2346         if (ret && !IS_ERR(cm_id)) {
2347                 iw_destroy_cm_id(cm_id);
2348                 id_priv->cm_id.iw = NULL;
2349         }
2350         return ret;
2351 }
2352
2353 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2354 {
2355         struct rdma_id_private *id_priv;
2356         int ret;
2357
2358         id_priv = container_of(id, struct rdma_id_private, id);
2359         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2360                 return -EINVAL;
2361
2362         if (!id->qp) {
2363                 id_priv->qp_num = conn_param->qp_num;
2364                 id_priv->srq = conn_param->srq;
2365         }
2366
2367         switch (rdma_node_get_transport(id->device->node_type)) {
2368         case RDMA_TRANSPORT_IB:
2369                 if (cma_is_ud_ps(id->ps))
2370                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2371                 else
2372                         ret = cma_connect_ib(id_priv, conn_param);
2373                 break;
2374         case RDMA_TRANSPORT_IWARP:
2375                 ret = cma_connect_iw(id_priv, conn_param);
2376                 break;
2377         default:
2378                 ret = -ENOSYS;
2379                 break;
2380         }
2381         if (ret)
2382                 goto err;
2383
2384         return 0;
2385 err:
2386         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2387         return ret;
2388 }
2389 EXPORT_SYMBOL(rdma_connect);
2390
2391 static int cma_accept_ib(struct rdma_id_private *id_priv,
2392                          struct rdma_conn_param *conn_param)
2393 {
2394         struct ib_cm_rep_param rep;
2395         int ret;
2396
2397         ret = cma_modify_qp_rtr(id_priv, conn_param);
2398         if (ret)
2399                 goto out;
2400
2401         ret = cma_modify_qp_rts(id_priv, conn_param);
2402         if (ret)
2403                 goto out;
2404
2405         memset(&rep, 0, sizeof rep);
2406         rep.qp_num = id_priv->qp_num;
2407         rep.starting_psn = id_priv->seq_num;
2408         rep.private_data = conn_param->private_data;
2409         rep.private_data_len = conn_param->private_data_len;
2410         rep.responder_resources = conn_param->responder_resources;
2411         rep.initiator_depth = conn_param->initiator_depth;
2412         rep.failover_accepted = 0;
2413         rep.flow_control = conn_param->flow_control;
2414         rep.rnr_retry_count = conn_param->rnr_retry_count;
2415         rep.srq = id_priv->srq ? 1 : 0;
2416
2417         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2418 out:
2419         return ret;
2420 }
2421
2422 static int cma_accept_iw(struct rdma_id_private *id_priv,
2423                   struct rdma_conn_param *conn_param)
2424 {
2425         struct iw_cm_conn_param iw_param;
2426         int ret;
2427
2428         ret = cma_modify_qp_rtr(id_priv, conn_param);
2429         if (ret)
2430                 return ret;
2431
2432         iw_param.ord = conn_param->initiator_depth;
2433         iw_param.ird = conn_param->responder_resources;
2434         iw_param.private_data = conn_param->private_data;
2435         iw_param.private_data_len = conn_param->private_data_len;
2436         if (id_priv->id.qp) {
2437                 iw_param.qpn = id_priv->qp_num;
2438         } else
2439                 iw_param.qpn = conn_param->qp_num;
2440
2441         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2442 }
2443
2444 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2445                              enum ib_cm_sidr_status status,
2446                              const void *private_data, int private_data_len)
2447 {
2448         struct ib_cm_sidr_rep_param rep;
2449
2450         memset(&rep, 0, sizeof rep);
2451         rep.status = status;
2452         if (status == IB_SIDR_SUCCESS) {
2453                 rep.qp_num = id_priv->qp_num;
2454                 rep.qkey = id_priv->qkey;
2455         }
2456         rep.private_data = private_data;
2457         rep.private_data_len = private_data_len;
2458
2459         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2460 }
2461
2462 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2463 {
2464         struct rdma_id_private *id_priv;
2465         int ret;
2466
2467         id_priv = container_of(id, struct rdma_id_private, id);
2468         if (!cma_comp(id_priv, CMA_CONNECT))
2469                 return -EINVAL;
2470
2471         if (!id->qp && conn_param) {
2472                 id_priv->qp_num = conn_param->qp_num;
2473                 id_priv->srq = conn_param->srq;
2474         }
2475
2476         switch (rdma_node_get_transport(id->device->node_type)) {
2477         case RDMA_TRANSPORT_IB:
2478                 if (cma_is_ud_ps(id->ps))
2479                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2480                                                 conn_param->private_data,
2481                                                 conn_param->private_data_len);
2482                 else if (conn_param)
2483                         ret = cma_accept_ib(id_priv, conn_param);
2484                 else
2485                         ret = cma_rep_recv(id_priv);
2486                 break;
2487         case RDMA_TRANSPORT_IWARP:
2488                 ret = cma_accept_iw(id_priv, conn_param);
2489                 break;
2490         default:
2491                 ret = -ENOSYS;
2492                 break;
2493         }
2494
2495         if (ret)
2496                 goto reject;
2497
2498         return 0;
2499 reject:
2500         cma_modify_qp_err(id_priv);
2501         rdma_reject(id, NULL, 0);
2502         return ret;
2503 }
2504 EXPORT_SYMBOL(rdma_accept);
2505
2506 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2507 {
2508         struct rdma_id_private *id_priv;
2509         int ret;
2510
2511         id_priv = container_of(id, struct rdma_id_private, id);
2512         if (!cma_has_cm_dev(id_priv))
2513                 return -EINVAL;
2514
2515         switch (id->device->node_type) {
2516         case RDMA_NODE_IB_CA:
2517                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2518                 break;
2519         default:
2520                 ret = 0;
2521                 break;
2522         }
2523         return ret;
2524 }
2525 EXPORT_SYMBOL(rdma_notify);
2526
2527 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2528                 u8 private_data_len)
2529 {
2530         struct rdma_id_private *id_priv;
2531         int ret;
2532
2533         id_priv = container_of(id, struct rdma_id_private, id);
2534         if (!cma_has_cm_dev(id_priv))
2535                 return -EINVAL;
2536
2537         switch (rdma_node_get_transport(id->device->node_type)) {
2538         case RDMA_TRANSPORT_IB:
2539                 if (cma_is_ud_ps(id->ps))
2540                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2541                                                 private_data, private_data_len);
2542                 else
2543                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2544                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2545                                              0, private_data, private_data_len);
2546                 break;
2547         case RDMA_TRANSPORT_IWARP:
2548                 ret = iw_cm_reject(id_priv->cm_id.iw,
2549                                    private_data, private_data_len);
2550                 break;
2551         default:
2552                 ret = -ENOSYS;
2553                 break;
2554         }
2555         return ret;
2556 }
2557 EXPORT_SYMBOL(rdma_reject);
2558
2559 int rdma_disconnect(struct rdma_cm_id *id)
2560 {
2561         struct rdma_id_private *id_priv;
2562         int ret;
2563
2564         id_priv = container_of(id, struct rdma_id_private, id);
2565         if (!cma_has_cm_dev(id_priv))
2566                 return -EINVAL;
2567
2568         switch (rdma_node_get_transport(id->device->node_type)) {
2569         case RDMA_TRANSPORT_IB:
2570                 ret = cma_modify_qp_err(id_priv);
2571                 if (ret)
2572                         goto out;
2573                 /* Initiate or respond to a disconnect. */
2574                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2575                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2576                 break;
2577         case RDMA_TRANSPORT_IWARP:
2578                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2579                 break;
2580         default:
2581                 ret = -EINVAL;
2582                 break;
2583         }
2584 out:
2585         return ret;
2586 }
2587 EXPORT_SYMBOL(rdma_disconnect);
2588
2589 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2590 {
2591         struct rdma_id_private *id_priv;
2592         struct cma_multicast *mc = multicast->context;
2593         struct rdma_cm_event event;
2594         int ret;
2595
2596         id_priv = mc->id_priv;
2597         if (cma_disable_callback(id_priv, CMA_ADDR_BOUND) &&
2598             cma_disable_callback(id_priv, CMA_ADDR_RESOLVED))
2599                 return 0;
2600
2601         mutex_lock(&id_priv->qp_mutex);
2602         if (!status && id_priv->id.qp)
2603                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2604                                          multicast->rec.mlid);
2605         mutex_unlock(&id_priv->qp_mutex);
2606
2607         memset(&event, 0, sizeof event);
2608         event.status = status;
2609         event.param.ud.private_data = mc->context;
2610         if (!status) {
2611                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2612                 ib_init_ah_from_mcmember(id_priv->id.device,
2613                                          id_priv->id.port_num, &multicast->rec,
2614                                          &event.param.ud.ah_attr);
2615                 event.param.ud.qp_num = 0xFFFFFF;
2616                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2617         } else
2618                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2619
2620         ret = id_priv->id.event_handler(&id_priv->id, &event);
2621         if (ret) {
2622                 cma_exch(id_priv, CMA_DESTROYING);
2623                 mutex_unlock(&id_priv->handler_mutex);
2624                 rdma_destroy_id(&id_priv->id);
2625                 return 0;
2626         }
2627
2628         mutex_unlock(&id_priv->handler_mutex);
2629         return 0;
2630 }
2631
2632 static void cma_set_mgid(struct rdma_id_private *id_priv,
2633                          struct sockaddr *addr, union ib_gid *mgid)
2634 {
2635         unsigned char mc_map[MAX_ADDR_LEN];
2636         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2637         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2638         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2639
2640         if (cma_any_addr(addr)) {
2641                 memset(mgid, 0, sizeof *mgid);
2642         } else if ((addr->sa_family == AF_INET6) &&
2643                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFF10A01B) ==
2644                                                                  0xFF10A01B)) {
2645                 /* IPv6 address is an SA assigned MGID. */
2646                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2647         } else {
2648                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2649                 if (id_priv->id.ps == RDMA_PS_UDP)
2650                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2651                 *mgid = *(union ib_gid *) (mc_map + 4);
2652         }
2653 }
2654
2655 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2656                                  struct cma_multicast *mc)
2657 {
2658         struct ib_sa_mcmember_rec rec;
2659         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2660         ib_sa_comp_mask comp_mask;
2661         int ret;
2662
2663         ib_addr_get_mgid(dev_addr, &rec.mgid);
2664         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2665                                      &rec.mgid, &rec);
2666         if (ret)
2667                 return ret;
2668
2669         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
2670         if (id_priv->id.ps == RDMA_PS_UDP)
2671                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2672         ib_addr_get_sgid(dev_addr, &rec.port_gid);
2673         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2674         rec.join_state = 1;
2675
2676         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2677                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2678                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2679                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
2680                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2681
2682         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2683                                                 id_priv->id.port_num, &rec,
2684                                                 comp_mask, GFP_KERNEL,
2685                                                 cma_ib_mc_handler, mc);
2686         if (IS_ERR(mc->multicast.ib))
2687                 return PTR_ERR(mc->multicast.ib);
2688
2689         return 0;
2690 }
2691
2692 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
2693                         void *context)
2694 {
2695         struct rdma_id_private *id_priv;
2696         struct cma_multicast *mc;
2697         int ret;
2698
2699         id_priv = container_of(id, struct rdma_id_private, id);
2700         if (!cma_comp(id_priv, CMA_ADDR_BOUND) &&
2701             !cma_comp(id_priv, CMA_ADDR_RESOLVED))
2702                 return -EINVAL;
2703
2704         mc = kmalloc(sizeof *mc, GFP_KERNEL);
2705         if (!mc)
2706                 return -ENOMEM;
2707
2708         memcpy(&mc->addr, addr, ip_addr_size(addr));
2709         mc->context = context;
2710         mc->id_priv = id_priv;
2711
2712         spin_lock(&id_priv->lock);
2713         list_add(&mc->list, &id_priv->mc_list);
2714         spin_unlock(&id_priv->lock);
2715
2716         switch (rdma_node_get_transport(id->device->node_type)) {
2717         case RDMA_TRANSPORT_IB:
2718                 ret = cma_join_ib_multicast(id_priv, mc);
2719                 break;
2720         default:
2721                 ret = -ENOSYS;
2722                 break;
2723         }
2724
2725         if (ret) {
2726                 spin_lock_irq(&id_priv->lock);
2727                 list_del(&mc->list);
2728                 spin_unlock_irq(&id_priv->lock);
2729                 kfree(mc);
2730         }
2731         return ret;
2732 }
2733 EXPORT_SYMBOL(rdma_join_multicast);
2734
2735 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
2736 {
2737         struct rdma_id_private *id_priv;
2738         struct cma_multicast *mc;
2739
2740         id_priv = container_of(id, struct rdma_id_private, id);
2741         spin_lock_irq(&id_priv->lock);
2742         list_for_each_entry(mc, &id_priv->mc_list, list) {
2743                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
2744                         list_del(&mc->list);
2745                         spin_unlock_irq(&id_priv->lock);
2746
2747                         if (id->qp)
2748                                 ib_detach_mcast(id->qp,
2749                                                 &mc->multicast.ib->rec.mgid,
2750                                                 mc->multicast.ib->rec.mlid);
2751                         ib_sa_free_multicast(mc->multicast.ib);
2752                         kfree(mc);
2753                         return;
2754                 }
2755         }
2756         spin_unlock_irq(&id_priv->lock);
2757 }
2758 EXPORT_SYMBOL(rdma_leave_multicast);
2759
2760 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
2761 {
2762         struct rdma_dev_addr *dev_addr;
2763         struct cma_ndev_work *work;
2764
2765         dev_addr = &id_priv->id.route.addr.dev_addr;
2766
2767         if ((dev_addr->src_dev == ndev) &&
2768             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
2769                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
2770                        ndev->name, &id_priv->id);
2771                 work = kzalloc(sizeof *work, GFP_KERNEL);
2772                 if (!work)
2773                         return -ENOMEM;
2774
2775                 INIT_WORK(&work->work, cma_ndev_work_handler);
2776                 work->id = id_priv;
2777                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
2778                 atomic_inc(&id_priv->refcount);
2779                 queue_work(cma_wq, &work->work);
2780         }
2781
2782         return 0;
2783 }
2784
2785 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
2786                                void *ctx)
2787 {
2788         struct net_device *ndev = (struct net_device *)ctx;
2789         struct cma_device *cma_dev;
2790         struct rdma_id_private *id_priv;
2791         int ret = NOTIFY_DONE;
2792
2793         if (dev_net(ndev) != &init_net)
2794                 return NOTIFY_DONE;
2795
2796         if (event != NETDEV_BONDING_FAILOVER)
2797                 return NOTIFY_DONE;
2798
2799         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
2800                 return NOTIFY_DONE;
2801
2802         mutex_lock(&lock);
2803         list_for_each_entry(cma_dev, &dev_list, list)
2804                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
2805                         ret = cma_netdev_change(ndev, id_priv);
2806                         if (ret)
2807                                 goto out;
2808                 }
2809
2810 out:
2811         mutex_unlock(&lock);
2812         return ret;
2813 }
2814
2815 static struct notifier_block cma_nb = {
2816         .notifier_call = cma_netdev_callback
2817 };
2818
2819 static void cma_add_one(struct ib_device *device)
2820 {
2821         struct cma_device *cma_dev;
2822         struct rdma_id_private *id_priv;
2823
2824         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2825         if (!cma_dev)
2826                 return;
2827
2828         cma_dev->device = device;
2829
2830         init_completion(&cma_dev->comp);
2831         atomic_set(&cma_dev->refcount, 1);
2832         INIT_LIST_HEAD(&cma_dev->id_list);
2833         ib_set_client_data(device, &cma_client, cma_dev);
2834
2835         mutex_lock(&lock);
2836         list_add_tail(&cma_dev->list, &dev_list);
2837         list_for_each_entry(id_priv, &listen_any_list, list)
2838                 cma_listen_on_dev(id_priv, cma_dev);
2839         mutex_unlock(&lock);
2840 }
2841
2842 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2843 {
2844         struct rdma_cm_event event;
2845         enum cma_state state;
2846         int ret = 0;
2847
2848         /* Record that we want to remove the device */
2849         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2850         if (state == CMA_DESTROYING)
2851                 return 0;
2852
2853         cma_cancel_operation(id_priv, state);
2854         mutex_lock(&id_priv->handler_mutex);
2855
2856         /* Check for destruction from another callback. */
2857         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2858                 goto out;
2859
2860         memset(&event, 0, sizeof event);
2861         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
2862         ret = id_priv->id.event_handler(&id_priv->id, &event);
2863 out:
2864         mutex_unlock(&id_priv->handler_mutex);
2865         return ret;
2866 }
2867
2868 static void cma_process_remove(struct cma_device *cma_dev)
2869 {
2870         struct rdma_id_private *id_priv;
2871         int ret;
2872
2873         mutex_lock(&lock);
2874         while (!list_empty(&cma_dev->id_list)) {
2875                 id_priv = list_entry(cma_dev->id_list.next,
2876                                      struct rdma_id_private, list);
2877
2878                 list_del(&id_priv->listen_list);
2879                 list_del_init(&id_priv->list);
2880                 atomic_inc(&id_priv->refcount);
2881                 mutex_unlock(&lock);
2882
2883                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
2884                 cma_deref_id(id_priv);
2885                 if (ret)
2886                         rdma_destroy_id(&id_priv->id);
2887
2888                 mutex_lock(&lock);
2889         }
2890         mutex_unlock(&lock);
2891
2892         cma_deref_dev(cma_dev);
2893         wait_for_completion(&cma_dev->comp);
2894 }
2895
2896 static void cma_remove_one(struct ib_device *device)
2897 {
2898         struct cma_device *cma_dev;
2899
2900         cma_dev = ib_get_client_data(device, &cma_client);
2901         if (!cma_dev)
2902                 return;
2903
2904         mutex_lock(&lock);
2905         list_del(&cma_dev->list);
2906         mutex_unlock(&lock);
2907
2908         cma_process_remove(cma_dev);
2909         kfree(cma_dev);
2910 }
2911
2912 static int cma_init(void)
2913 {
2914         int ret, low, high, remaining;
2915
2916         get_random_bytes(&next_port, sizeof next_port);
2917         inet_get_local_port_range(&low, &high);
2918         remaining = (high - low) + 1;
2919         next_port = ((unsigned int) next_port % remaining) + low;
2920
2921         cma_wq = create_singlethread_workqueue("rdma_cm");
2922         if (!cma_wq)
2923                 return -ENOMEM;
2924
2925         ib_sa_register_client(&sa_client);
2926         rdma_addr_register_client(&addr_client);
2927         register_netdevice_notifier(&cma_nb);
2928
2929         ret = ib_register_client(&cma_client);
2930         if (ret)
2931                 goto err;
2932         return 0;
2933
2934 err:
2935         unregister_netdevice_notifier(&cma_nb);
2936         rdma_addr_unregister_client(&addr_client);
2937         ib_sa_unregister_client(&sa_client);
2938         destroy_workqueue(cma_wq);
2939         return ret;
2940 }
2941
2942 static void cma_cleanup(void)
2943 {
2944         ib_unregister_client(&cma_client);
2945         unregister_netdevice_notifier(&cma_nb);
2946         rdma_addr_unregister_client(&addr_client);
2947         ib_sa_unregister_client(&sa_client);
2948         destroy_workqueue(cma_wq);
2949         idr_destroy(&sdp_ps);
2950         idr_destroy(&tcp_ps);
2951         idr_destroy(&udp_ps);
2952         idr_destroy(&ipoib_ps);
2953 }
2954
2955 module_init(cma_init);
2956 module_exit(cma_cleanup);