Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland...
[sfrench/cifs-2.6.git] / 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 #include <linux/slab.h>
44
45 #include <net/tcp.h>
46 #include <net/ipv6.h>
47
48 #include <rdma/rdma_cm.h>
49 #include <rdma/rdma_cm_ib.h>
50 #include <rdma/ib_cache.h>
51 #include <rdma/ib_cm.h>
52 #include <rdma/ib_sa.h>
53 #include <rdma/iw_cm.h>
54
55 MODULE_AUTHOR("Sean Hefty");
56 MODULE_DESCRIPTION("Generic RDMA CM Agent");
57 MODULE_LICENSE("Dual BSD/GPL");
58
59 #define CMA_CM_RESPONSE_TIMEOUT 20
60 #define CMA_MAX_CM_RETRIES 15
61 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
62
63 static void cma_add_one(struct ib_device *device);
64 static void cma_remove_one(struct ib_device *device);
65
66 static struct ib_client cma_client = {
67         .name   = "cma",
68         .add    = cma_add_one,
69         .remove = cma_remove_one
70 };
71
72 static struct ib_sa_client sa_client;
73 static struct rdma_addr_client addr_client;
74 static LIST_HEAD(dev_list);
75 static LIST_HEAD(listen_any_list);
76 static DEFINE_MUTEX(lock);
77 static struct workqueue_struct *cma_wq;
78 static DEFINE_IDR(sdp_ps);
79 static DEFINE_IDR(tcp_ps);
80 static DEFINE_IDR(udp_ps);
81 static DEFINE_IDR(ipoib_ps);
82 static int next_port;
83
84 struct cma_device {
85         struct list_head        list;
86         struct ib_device        *device;
87         struct completion       comp;
88         atomic_t                refcount;
89         struct list_head        id_list;
90 };
91
92 enum cma_state {
93         CMA_IDLE,
94         CMA_ADDR_QUERY,
95         CMA_ADDR_RESOLVED,
96         CMA_ROUTE_QUERY,
97         CMA_ROUTE_RESOLVED,
98         CMA_CONNECT,
99         CMA_DISCONNECT,
100         CMA_ADDR_BOUND,
101         CMA_LISTEN,
102         CMA_DEVICE_REMOVAL,
103         CMA_DESTROYING
104 };
105
106 struct rdma_bind_list {
107         struct idr              *ps;
108         struct hlist_head       owners;
109         unsigned short          port;
110 };
111
112 /*
113  * Device removal can occur at anytime, so we need extra handling to
114  * serialize notifying the user of device removal with other callbacks.
115  * We do this by disabling removal notification while a callback is in process,
116  * and reporting it after the callback completes.
117  */
118 struct rdma_id_private {
119         struct rdma_cm_id       id;
120
121         struct rdma_bind_list   *bind_list;
122         struct hlist_node       node;
123         struct list_head        list; /* listen_any_list or cma_device.list */
124         struct list_head        listen_list; /* per device listens */
125         struct cma_device       *cma_dev;
126         struct list_head        mc_list;
127
128         int                     internal_id;
129         enum cma_state          state;
130         spinlock_t              lock;
131         struct mutex            qp_mutex;
132
133         struct completion       comp;
134         atomic_t                refcount;
135         struct mutex            handler_mutex;
136
137         int                     backlog;
138         int                     timeout_ms;
139         struct ib_sa_query      *query;
140         int                     query_id;
141         union {
142                 struct ib_cm_id *ib;
143                 struct iw_cm_id *iw;
144         } cm_id;
145
146         u32                     seq_num;
147         u32                     qkey;
148         u32                     qp_num;
149         u8                      srq;
150         u8                      tos;
151 };
152
153 struct cma_multicast {
154         struct rdma_id_private *id_priv;
155         union {
156                 struct ib_sa_multicast *ib;
157         } multicast;
158         struct list_head        list;
159         void                    *context;
160         struct sockaddr_storage addr;
161 };
162
163 struct cma_work {
164         struct work_struct      work;
165         struct rdma_id_private  *id;
166         enum cma_state          old_state;
167         enum cma_state          new_state;
168         struct rdma_cm_event    event;
169 };
170
171 struct cma_ndev_work {
172         struct work_struct      work;
173         struct rdma_id_private  *id;
174         struct rdma_cm_event    event;
175 };
176
177 union cma_ip_addr {
178         struct in6_addr ip6;
179         struct {
180                 __be32 pad[3];
181                 __be32 addr;
182         } ip4;
183 };
184
185 struct cma_hdr {
186         u8 cma_version;
187         u8 ip_version;  /* IP version: 7:4 */
188         __be16 port;
189         union cma_ip_addr src_addr;
190         union cma_ip_addr dst_addr;
191 };
192
193 struct sdp_hh {
194         u8 bsdh[16];
195         u8 sdp_version; /* Major version: 7:4 */
196         u8 ip_version;  /* IP version: 7:4 */
197         u8 sdp_specific1[10];
198         __be16 port;
199         __be16 sdp_specific2;
200         union cma_ip_addr src_addr;
201         union cma_ip_addr dst_addr;
202 };
203
204 struct sdp_hah {
205         u8 bsdh[16];
206         u8 sdp_version;
207 };
208
209 #define CMA_VERSION 0x00
210 #define SDP_MAJ_VERSION 0x2
211
212 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
213 {
214         unsigned long flags;
215         int ret;
216
217         spin_lock_irqsave(&id_priv->lock, flags);
218         ret = (id_priv->state == comp);
219         spin_unlock_irqrestore(&id_priv->lock, flags);
220         return ret;
221 }
222
223 static int cma_comp_exch(struct rdma_id_private *id_priv,
224                          enum cma_state comp, enum cma_state exch)
225 {
226         unsigned long flags;
227         int ret;
228
229         spin_lock_irqsave(&id_priv->lock, flags);
230         if ((ret = (id_priv->state == comp)))
231                 id_priv->state = exch;
232         spin_unlock_irqrestore(&id_priv->lock, flags);
233         return ret;
234 }
235
236 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
237                                enum cma_state exch)
238 {
239         unsigned long flags;
240         enum cma_state old;
241
242         spin_lock_irqsave(&id_priv->lock, flags);
243         old = id_priv->state;
244         id_priv->state = exch;
245         spin_unlock_irqrestore(&id_priv->lock, flags);
246         return old;
247 }
248
249 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
250 {
251         return hdr->ip_version >> 4;
252 }
253
254 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
255 {
256         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
257 }
258
259 static inline u8 sdp_get_majv(u8 sdp_version)
260 {
261         return sdp_version >> 4;
262 }
263
264 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
265 {
266         return hh->ip_version >> 4;
267 }
268
269 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
270 {
271         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
272 }
273
274 static inline int cma_is_ud_ps(enum rdma_port_space ps)
275 {
276         return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB);
277 }
278
279 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
280                               struct cma_device *cma_dev)
281 {
282         atomic_inc(&cma_dev->refcount);
283         id_priv->cma_dev = cma_dev;
284         id_priv->id.device = cma_dev->device;
285         list_add_tail(&id_priv->list, &cma_dev->id_list);
286 }
287
288 static inline void cma_deref_dev(struct cma_device *cma_dev)
289 {
290         if (atomic_dec_and_test(&cma_dev->refcount))
291                 complete(&cma_dev->comp);
292 }
293
294 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
295 {
296         list_del(&id_priv->list);
297         cma_deref_dev(id_priv->cma_dev);
298         id_priv->cma_dev = NULL;
299 }
300
301 static int cma_set_qkey(struct rdma_id_private *id_priv)
302 {
303         struct ib_sa_mcmember_rec rec;
304         int ret = 0;
305
306         if (id_priv->qkey)
307                 return 0;
308
309         switch (id_priv->id.ps) {
310         case RDMA_PS_UDP:
311                 id_priv->qkey = RDMA_UDP_QKEY;
312                 break;
313         case RDMA_PS_IPOIB:
314                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
315                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
316                                              id_priv->id.port_num, &rec.mgid,
317                                              &rec);
318                 if (!ret)
319                         id_priv->qkey = be32_to_cpu(rec.qkey);
320                 break;
321         default:
322                 break;
323         }
324         return ret;
325 }
326
327 static int cma_acquire_dev(struct rdma_id_private *id_priv)
328 {
329         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
330         struct cma_device *cma_dev;
331         union ib_gid gid;
332         int ret = -ENODEV;
333
334         rdma_addr_get_sgid(dev_addr, &gid);
335         list_for_each_entry(cma_dev, &dev_list, list) {
336                 ret = ib_find_cached_gid(cma_dev->device, &gid,
337                                          &id_priv->id.port_num, NULL);
338                 if (!ret) {
339                         cma_attach_to_dev(id_priv, cma_dev);
340                         break;
341                 }
342         }
343         return ret;
344 }
345
346 static void cma_deref_id(struct rdma_id_private *id_priv)
347 {
348         if (atomic_dec_and_test(&id_priv->refcount))
349                 complete(&id_priv->comp);
350 }
351
352 static int cma_disable_callback(struct rdma_id_private *id_priv,
353                               enum cma_state state)
354 {
355         mutex_lock(&id_priv->handler_mutex);
356         if (id_priv->state != state) {
357                 mutex_unlock(&id_priv->handler_mutex);
358                 return -EINVAL;
359         }
360         return 0;
361 }
362
363 static int cma_has_cm_dev(struct rdma_id_private *id_priv)
364 {
365         return (id_priv->id.device && id_priv->cm_id.ib);
366 }
367
368 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
369                                   void *context, enum rdma_port_space ps)
370 {
371         struct rdma_id_private *id_priv;
372
373         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
374         if (!id_priv)
375                 return ERR_PTR(-ENOMEM);
376
377         id_priv->state = CMA_IDLE;
378         id_priv->id.context = context;
379         id_priv->id.event_handler = event_handler;
380         id_priv->id.ps = ps;
381         spin_lock_init(&id_priv->lock);
382         mutex_init(&id_priv->qp_mutex);
383         init_completion(&id_priv->comp);
384         atomic_set(&id_priv->refcount, 1);
385         mutex_init(&id_priv->handler_mutex);
386         INIT_LIST_HEAD(&id_priv->listen_list);
387         INIT_LIST_HEAD(&id_priv->mc_list);
388         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
389
390         return &id_priv->id;
391 }
392 EXPORT_SYMBOL(rdma_create_id);
393
394 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
395 {
396         struct ib_qp_attr qp_attr;
397         int qp_attr_mask, ret;
398
399         qp_attr.qp_state = IB_QPS_INIT;
400         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
401         if (ret)
402                 return ret;
403
404         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
405         if (ret)
406                 return ret;
407
408         qp_attr.qp_state = IB_QPS_RTR;
409         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
410         if (ret)
411                 return ret;
412
413         qp_attr.qp_state = IB_QPS_RTS;
414         qp_attr.sq_psn = 0;
415         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
416
417         return ret;
418 }
419
420 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
421 {
422         struct ib_qp_attr qp_attr;
423         int qp_attr_mask, ret;
424
425         qp_attr.qp_state = IB_QPS_INIT;
426         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
427         if (ret)
428                 return ret;
429
430         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
431 }
432
433 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
434                    struct ib_qp_init_attr *qp_init_attr)
435 {
436         struct rdma_id_private *id_priv;
437         struct ib_qp *qp;
438         int ret;
439
440         id_priv = container_of(id, struct rdma_id_private, id);
441         if (id->device != pd->device)
442                 return -EINVAL;
443
444         qp = ib_create_qp(pd, qp_init_attr);
445         if (IS_ERR(qp))
446                 return PTR_ERR(qp);
447
448         if (cma_is_ud_ps(id_priv->id.ps))
449                 ret = cma_init_ud_qp(id_priv, qp);
450         else
451                 ret = cma_init_conn_qp(id_priv, qp);
452         if (ret)
453                 goto err;
454
455         id->qp = qp;
456         id_priv->qp_num = qp->qp_num;
457         id_priv->srq = (qp->srq != NULL);
458         return 0;
459 err:
460         ib_destroy_qp(qp);
461         return ret;
462 }
463 EXPORT_SYMBOL(rdma_create_qp);
464
465 void rdma_destroy_qp(struct rdma_cm_id *id)
466 {
467         struct rdma_id_private *id_priv;
468
469         id_priv = container_of(id, struct rdma_id_private, id);
470         mutex_lock(&id_priv->qp_mutex);
471         ib_destroy_qp(id_priv->id.qp);
472         id_priv->id.qp = NULL;
473         mutex_unlock(&id_priv->qp_mutex);
474 }
475 EXPORT_SYMBOL(rdma_destroy_qp);
476
477 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
478                              struct rdma_conn_param *conn_param)
479 {
480         struct ib_qp_attr qp_attr;
481         int qp_attr_mask, ret;
482
483         mutex_lock(&id_priv->qp_mutex);
484         if (!id_priv->id.qp) {
485                 ret = 0;
486                 goto out;
487         }
488
489         /* Need to update QP attributes from default values. */
490         qp_attr.qp_state = IB_QPS_INIT;
491         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
492         if (ret)
493                 goto out;
494
495         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
496         if (ret)
497                 goto out;
498
499         qp_attr.qp_state = IB_QPS_RTR;
500         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
501         if (ret)
502                 goto out;
503
504         if (conn_param)
505                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
506         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
507 out:
508         mutex_unlock(&id_priv->qp_mutex);
509         return ret;
510 }
511
512 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
513                              struct rdma_conn_param *conn_param)
514 {
515         struct ib_qp_attr qp_attr;
516         int qp_attr_mask, ret;
517
518         mutex_lock(&id_priv->qp_mutex);
519         if (!id_priv->id.qp) {
520                 ret = 0;
521                 goto out;
522         }
523
524         qp_attr.qp_state = IB_QPS_RTS;
525         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
526         if (ret)
527                 goto out;
528
529         if (conn_param)
530                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
531         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
532 out:
533         mutex_unlock(&id_priv->qp_mutex);
534         return ret;
535 }
536
537 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
538 {
539         struct ib_qp_attr qp_attr;
540         int ret;
541
542         mutex_lock(&id_priv->qp_mutex);
543         if (!id_priv->id.qp) {
544                 ret = 0;
545                 goto out;
546         }
547
548         qp_attr.qp_state = IB_QPS_ERR;
549         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
550 out:
551         mutex_unlock(&id_priv->qp_mutex);
552         return ret;
553 }
554
555 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
556                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
557 {
558         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
559         int ret;
560
561         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
562                                   ib_addr_get_pkey(dev_addr),
563                                   &qp_attr->pkey_index);
564         if (ret)
565                 return ret;
566
567         qp_attr->port_num = id_priv->id.port_num;
568         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
569
570         if (cma_is_ud_ps(id_priv->id.ps)) {
571                 ret = cma_set_qkey(id_priv);
572                 if (ret)
573                         return ret;
574
575                 qp_attr->qkey = id_priv->qkey;
576                 *qp_attr_mask |= IB_QP_QKEY;
577         } else {
578                 qp_attr->qp_access_flags = 0;
579                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
580         }
581         return 0;
582 }
583
584 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
585                        int *qp_attr_mask)
586 {
587         struct rdma_id_private *id_priv;
588         int ret = 0;
589
590         id_priv = container_of(id, struct rdma_id_private, id);
591         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
592         case RDMA_TRANSPORT_IB:
593                 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps))
594                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
595                 else
596                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
597                                                  qp_attr_mask);
598                 if (qp_attr->qp_state == IB_QPS_RTR)
599                         qp_attr->rq_psn = id_priv->seq_num;
600                 break;
601         case RDMA_TRANSPORT_IWARP:
602                 if (!id_priv->cm_id.iw) {
603                         qp_attr->qp_access_flags = 0;
604                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
605                 } else
606                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
607                                                  qp_attr_mask);
608                 break;
609         default:
610                 ret = -ENOSYS;
611                 break;
612         }
613
614         return ret;
615 }
616 EXPORT_SYMBOL(rdma_init_qp_attr);
617
618 static inline int cma_zero_addr(struct sockaddr *addr)
619 {
620         struct in6_addr *ip6;
621
622         if (addr->sa_family == AF_INET)
623                 return ipv4_is_zeronet(
624                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
625         else {
626                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
627                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
628                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
629         }
630 }
631
632 static inline int cma_loopback_addr(struct sockaddr *addr)
633 {
634         if (addr->sa_family == AF_INET)
635                 return ipv4_is_loopback(
636                         ((struct sockaddr_in *) addr)->sin_addr.s_addr);
637         else
638                 return ipv6_addr_loopback(
639                         &((struct sockaddr_in6 *) addr)->sin6_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         if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1027                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1028                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1029                 ib_addr_set_pkey(&rt->addr.dev_addr, rt->path_rec[0].pkey);
1030         } else {
1031                 ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1032                                         &rt->addr.dev_addr);
1033                 if (ret)
1034                         goto destroy_id;
1035         }
1036         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1037
1038         id_priv = container_of(id, struct rdma_id_private, id);
1039         id_priv->state = CMA_CONNECT;
1040         return id_priv;
1041
1042 destroy_id:
1043         rdma_destroy_id(id);
1044 err:
1045         return NULL;
1046 }
1047
1048 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1049                                               struct ib_cm_event *ib_event)
1050 {
1051         struct rdma_id_private *id_priv;
1052         struct rdma_cm_id *id;
1053         union cma_ip_addr *src, *dst;
1054         __be16 port;
1055         u8 ip_ver;
1056         int ret;
1057
1058         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1059                             listen_id->ps);
1060         if (IS_ERR(id))
1061                 return NULL;
1062
1063
1064         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1065                              &ip_ver, &port, &src, &dst))
1066                 goto err;
1067
1068         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1069                           ip_ver, port, src, dst);
1070
1071         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1072                 ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1073                                         &id->route.addr.dev_addr);
1074                 if (ret)
1075                         goto err;
1076         }
1077
1078         id_priv = container_of(id, struct rdma_id_private, id);
1079         id_priv->state = CMA_CONNECT;
1080         return id_priv;
1081 err:
1082         rdma_destroy_id(id);
1083         return NULL;
1084 }
1085
1086 static void cma_set_req_event_data(struct rdma_cm_event *event,
1087                                    struct ib_cm_req_event_param *req_data,
1088                                    void *private_data, int offset)
1089 {
1090         event->param.conn.private_data = private_data + offset;
1091         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1092         event->param.conn.responder_resources = req_data->responder_resources;
1093         event->param.conn.initiator_depth = req_data->initiator_depth;
1094         event->param.conn.flow_control = req_data->flow_control;
1095         event->param.conn.retry_count = req_data->retry_count;
1096         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1097         event->param.conn.srq = req_data->srq;
1098         event->param.conn.qp_num = req_data->remote_qpn;
1099 }
1100
1101 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1102 {
1103         struct rdma_id_private *listen_id, *conn_id;
1104         struct rdma_cm_event event;
1105         int offset, ret;
1106
1107         listen_id = cm_id->context;
1108         if (cma_disable_callback(listen_id, CMA_LISTEN))
1109                 return -ECONNABORTED;
1110
1111         memset(&event, 0, sizeof event);
1112         offset = cma_user_data_offset(listen_id->id.ps);
1113         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1114         if (cma_is_ud_ps(listen_id->id.ps)) {
1115                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1116                 event.param.ud.private_data = ib_event->private_data + offset;
1117                 event.param.ud.private_data_len =
1118                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1119         } else {
1120                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1121                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1122                                        ib_event->private_data, offset);
1123         }
1124         if (!conn_id) {
1125                 ret = -ENOMEM;
1126                 goto out;
1127         }
1128
1129         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1130         mutex_lock(&lock);
1131         ret = cma_acquire_dev(conn_id);
1132         mutex_unlock(&lock);
1133         if (ret)
1134                 goto release_conn_id;
1135
1136         conn_id->cm_id.ib = cm_id;
1137         cm_id->context = conn_id;
1138         cm_id->cm_handler = cma_ib_handler;
1139
1140         ret = conn_id->id.event_handler(&conn_id->id, &event);
1141         if (!ret) {
1142                 /*
1143                  * Acquire mutex to prevent user executing rdma_destroy_id()
1144                  * while we're accessing the cm_id.
1145                  */
1146                 mutex_lock(&lock);
1147                 if (cma_comp(conn_id, CMA_CONNECT) &&
1148                     !cma_is_ud_ps(conn_id->id.ps))
1149                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1150                 mutex_unlock(&lock);
1151                 mutex_unlock(&conn_id->handler_mutex);
1152                 goto out;
1153         }
1154
1155         /* Destroy the CM ID by returning a non-zero value. */
1156         conn_id->cm_id.ib = NULL;
1157
1158 release_conn_id:
1159         cma_exch(conn_id, CMA_DESTROYING);
1160         mutex_unlock(&conn_id->handler_mutex);
1161         rdma_destroy_id(&conn_id->id);
1162
1163 out:
1164         mutex_unlock(&listen_id->handler_mutex);
1165         return ret;
1166 }
1167
1168 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1169 {
1170         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1171 }
1172
1173 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1174                                  struct ib_cm_compare_data *compare)
1175 {
1176         struct cma_hdr *cma_data, *cma_mask;
1177         struct sdp_hh *sdp_data, *sdp_mask;
1178         __be32 ip4_addr;
1179         struct in6_addr ip6_addr;
1180
1181         memset(compare, 0, sizeof *compare);
1182         cma_data = (void *) compare->data;
1183         cma_mask = (void *) compare->mask;
1184         sdp_data = (void *) compare->data;
1185         sdp_mask = (void *) compare->mask;
1186
1187         switch (addr->sa_family) {
1188         case AF_INET:
1189                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1190                 if (ps == RDMA_PS_SDP) {
1191                         sdp_set_ip_ver(sdp_data, 4);
1192                         sdp_set_ip_ver(sdp_mask, 0xF);
1193                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1194                         sdp_mask->dst_addr.ip4.addr = htonl(~0);
1195                 } else {
1196                         cma_set_ip_ver(cma_data, 4);
1197                         cma_set_ip_ver(cma_mask, 0xF);
1198                         cma_data->dst_addr.ip4.addr = ip4_addr;
1199                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1200                 }
1201                 break;
1202         case AF_INET6:
1203                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1204                 if (ps == RDMA_PS_SDP) {
1205                         sdp_set_ip_ver(sdp_data, 6);
1206                         sdp_set_ip_ver(sdp_mask, 0xF);
1207                         sdp_data->dst_addr.ip6 = ip6_addr;
1208                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1209                                sizeof sdp_mask->dst_addr.ip6);
1210                 } else {
1211                         cma_set_ip_ver(cma_data, 6);
1212                         cma_set_ip_ver(cma_mask, 0xF);
1213                         cma_data->dst_addr.ip6 = ip6_addr;
1214                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1215                                sizeof cma_mask->dst_addr.ip6);
1216                 }
1217                 break;
1218         default:
1219                 break;
1220         }
1221 }
1222
1223 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1224 {
1225         struct rdma_id_private *id_priv = iw_id->context;
1226         struct rdma_cm_event event;
1227         struct sockaddr_in *sin;
1228         int ret = 0;
1229
1230         if (cma_disable_callback(id_priv, CMA_CONNECT))
1231                 return 0;
1232
1233         memset(&event, 0, sizeof event);
1234         switch (iw_event->event) {
1235         case IW_CM_EVENT_CLOSE:
1236                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1237                 break;
1238         case IW_CM_EVENT_CONNECT_REPLY:
1239                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1240                 *sin = iw_event->local_addr;
1241                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1242                 *sin = iw_event->remote_addr;
1243                 switch (iw_event->status) {
1244                 case 0:
1245                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1246                         break;
1247                 case -ECONNRESET:
1248                 case -ECONNREFUSED:
1249                         event.event = RDMA_CM_EVENT_REJECTED;
1250                         break;
1251                 case -ETIMEDOUT:
1252                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1253                         break;
1254                 default:
1255                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1256                         break;
1257                 }
1258                 break;
1259         case IW_CM_EVENT_ESTABLISHED:
1260                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1261                 break;
1262         default:
1263                 BUG_ON(1);
1264         }
1265
1266         event.status = iw_event->status;
1267         event.param.conn.private_data = iw_event->private_data;
1268         event.param.conn.private_data_len = iw_event->private_data_len;
1269         ret = id_priv->id.event_handler(&id_priv->id, &event);
1270         if (ret) {
1271                 /* Destroy the CM ID by returning a non-zero value. */
1272                 id_priv->cm_id.iw = NULL;
1273                 cma_exch(id_priv, CMA_DESTROYING);
1274                 mutex_unlock(&id_priv->handler_mutex);
1275                 rdma_destroy_id(&id_priv->id);
1276                 return ret;
1277         }
1278
1279         mutex_unlock(&id_priv->handler_mutex);
1280         return ret;
1281 }
1282
1283 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1284                                struct iw_cm_event *iw_event)
1285 {
1286         struct rdma_cm_id *new_cm_id;
1287         struct rdma_id_private *listen_id, *conn_id;
1288         struct sockaddr_in *sin;
1289         struct net_device *dev = NULL;
1290         struct rdma_cm_event event;
1291         int ret;
1292         struct ib_device_attr attr;
1293
1294         listen_id = cm_id->context;
1295         if (cma_disable_callback(listen_id, CMA_LISTEN))
1296                 return -ECONNABORTED;
1297
1298         /* Create a new RDMA id for the new IW CM ID */
1299         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1300                                    listen_id->id.context,
1301                                    RDMA_PS_TCP);
1302         if (IS_ERR(new_cm_id)) {
1303                 ret = -ENOMEM;
1304                 goto out;
1305         }
1306         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1307         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1308         conn_id->state = CMA_CONNECT;
1309
1310         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1311         if (!dev) {
1312                 ret = -EADDRNOTAVAIL;
1313                 mutex_unlock(&conn_id->handler_mutex);
1314                 rdma_destroy_id(new_cm_id);
1315                 goto out;
1316         }
1317         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1318         if (ret) {
1319                 mutex_unlock(&conn_id->handler_mutex);
1320                 rdma_destroy_id(new_cm_id);
1321                 goto out;
1322         }
1323
1324         mutex_lock(&lock);
1325         ret = cma_acquire_dev(conn_id);
1326         mutex_unlock(&lock);
1327         if (ret) {
1328                 mutex_unlock(&conn_id->handler_mutex);
1329                 rdma_destroy_id(new_cm_id);
1330                 goto out;
1331         }
1332
1333         conn_id->cm_id.iw = cm_id;
1334         cm_id->context = conn_id;
1335         cm_id->cm_handler = cma_iw_handler;
1336
1337         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1338         *sin = iw_event->local_addr;
1339         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1340         *sin = iw_event->remote_addr;
1341
1342         ret = ib_query_device(conn_id->id.device, &attr);
1343         if (ret) {
1344                 mutex_unlock(&conn_id->handler_mutex);
1345                 rdma_destroy_id(new_cm_id);
1346                 goto out;
1347         }
1348
1349         memset(&event, 0, sizeof event);
1350         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1351         event.param.conn.private_data = iw_event->private_data;
1352         event.param.conn.private_data_len = iw_event->private_data_len;
1353         event.param.conn.initiator_depth = attr.max_qp_init_rd_atom;
1354         event.param.conn.responder_resources = attr.max_qp_rd_atom;
1355         ret = conn_id->id.event_handler(&conn_id->id, &event);
1356         if (ret) {
1357                 /* User wants to destroy the CM ID */
1358                 conn_id->cm_id.iw = NULL;
1359                 cma_exch(conn_id, CMA_DESTROYING);
1360                 mutex_unlock(&conn_id->handler_mutex);
1361                 rdma_destroy_id(&conn_id->id);
1362                 goto out;
1363         }
1364
1365         mutex_unlock(&conn_id->handler_mutex);
1366
1367 out:
1368         if (dev)
1369                 dev_put(dev);
1370         mutex_unlock(&listen_id->handler_mutex);
1371         return ret;
1372 }
1373
1374 static int cma_ib_listen(struct rdma_id_private *id_priv)
1375 {
1376         struct ib_cm_compare_data compare_data;
1377         struct sockaddr *addr;
1378         __be64 svc_id;
1379         int ret;
1380
1381         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1382                                             id_priv);
1383         if (IS_ERR(id_priv->cm_id.ib))
1384                 return PTR_ERR(id_priv->cm_id.ib);
1385
1386         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1387         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1388         if (cma_any_addr(addr))
1389                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1390         else {
1391                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1392                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1393         }
1394
1395         if (ret) {
1396                 ib_destroy_cm_id(id_priv->cm_id.ib);
1397                 id_priv->cm_id.ib = NULL;
1398         }
1399
1400         return ret;
1401 }
1402
1403 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1404 {
1405         int ret;
1406         struct sockaddr_in *sin;
1407
1408         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1409                                             iw_conn_req_handler,
1410                                             id_priv);
1411         if (IS_ERR(id_priv->cm_id.iw))
1412                 return PTR_ERR(id_priv->cm_id.iw);
1413
1414         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1415         id_priv->cm_id.iw->local_addr = *sin;
1416
1417         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1418
1419         if (ret) {
1420                 iw_destroy_cm_id(id_priv->cm_id.iw);
1421                 id_priv->cm_id.iw = NULL;
1422         }
1423
1424         return ret;
1425 }
1426
1427 static int cma_listen_handler(struct rdma_cm_id *id,
1428                               struct rdma_cm_event *event)
1429 {
1430         struct rdma_id_private *id_priv = id->context;
1431
1432         id->context = id_priv->id.context;
1433         id->event_handler = id_priv->id.event_handler;
1434         return id_priv->id.event_handler(id, event);
1435 }
1436
1437 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1438                               struct cma_device *cma_dev)
1439 {
1440         struct rdma_id_private *dev_id_priv;
1441         struct rdma_cm_id *id;
1442         int ret;
1443
1444         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1445         if (IS_ERR(id))
1446                 return;
1447
1448         dev_id_priv = container_of(id, struct rdma_id_private, id);
1449
1450         dev_id_priv->state = CMA_ADDR_BOUND;
1451         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1452                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1453
1454         cma_attach_to_dev(dev_id_priv, cma_dev);
1455         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1456         atomic_inc(&id_priv->refcount);
1457         dev_id_priv->internal_id = 1;
1458
1459         ret = rdma_listen(id, id_priv->backlog);
1460         if (ret)
1461                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1462                        "listening on device %s\n", ret, cma_dev->device->name);
1463 }
1464
1465 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1466 {
1467         struct cma_device *cma_dev;
1468
1469         mutex_lock(&lock);
1470         list_add_tail(&id_priv->list, &listen_any_list);
1471         list_for_each_entry(cma_dev, &dev_list, list)
1472                 cma_listen_on_dev(id_priv, cma_dev);
1473         mutex_unlock(&lock);
1474 }
1475
1476 int rdma_listen(struct rdma_cm_id *id, int backlog)
1477 {
1478         struct rdma_id_private *id_priv;
1479         int ret;
1480
1481         id_priv = container_of(id, struct rdma_id_private, id);
1482         if (id_priv->state == CMA_IDLE) {
1483                 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
1484                 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
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         rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1560         rdma_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         id->route.num_paths = num_paths;
1688         return 0;
1689 err:
1690         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1691         return ret;
1692 }
1693 EXPORT_SYMBOL(rdma_set_ib_paths);
1694
1695 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1696 {
1697         struct cma_work *work;
1698
1699         work = kzalloc(sizeof *work, GFP_KERNEL);
1700         if (!work)
1701                 return -ENOMEM;
1702
1703         work->id = id_priv;
1704         INIT_WORK(&work->work, cma_work_handler);
1705         work->old_state = CMA_ROUTE_QUERY;
1706         work->new_state = CMA_ROUTE_RESOLVED;
1707         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1708         queue_work(cma_wq, &work->work);
1709         return 0;
1710 }
1711
1712 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1713 {
1714         struct rdma_id_private *id_priv;
1715         int ret;
1716
1717         id_priv = container_of(id, struct rdma_id_private, id);
1718         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1719                 return -EINVAL;
1720
1721         atomic_inc(&id_priv->refcount);
1722         switch (rdma_node_get_transport(id->device->node_type)) {
1723         case RDMA_TRANSPORT_IB:
1724                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1725                 break;
1726         case RDMA_TRANSPORT_IWARP:
1727                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1728                 break;
1729         default:
1730                 ret = -ENOSYS;
1731                 break;
1732         }
1733         if (ret)
1734                 goto err;
1735
1736         return 0;
1737 err:
1738         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1739         cma_deref_id(id_priv);
1740         return ret;
1741 }
1742 EXPORT_SYMBOL(rdma_resolve_route);
1743
1744 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1745 {
1746         struct cma_device *cma_dev;
1747         struct ib_port_attr port_attr;
1748         union ib_gid gid;
1749         u16 pkey;
1750         int ret;
1751         u8 p;
1752
1753         mutex_lock(&lock);
1754         if (list_empty(&dev_list)) {
1755                 ret = -ENODEV;
1756                 goto out;
1757         }
1758         list_for_each_entry(cma_dev, &dev_list, list)
1759                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1760                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1761                             port_attr.state == IB_PORT_ACTIVE)
1762                                 goto port_found;
1763
1764         p = 1;
1765         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1766
1767 port_found:
1768         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1769         if (ret)
1770                 goto out;
1771
1772         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1773         if (ret)
1774                 goto out;
1775
1776         id_priv->id.route.addr.dev_addr.dev_type =
1777                 (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB) ?
1778                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1779
1780         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1781         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1782         id_priv->id.port_num = p;
1783         cma_attach_to_dev(id_priv, cma_dev);
1784 out:
1785         mutex_unlock(&lock);
1786         return ret;
1787 }
1788
1789 static void addr_handler(int status, struct sockaddr *src_addr,
1790                          struct rdma_dev_addr *dev_addr, void *context)
1791 {
1792         struct rdma_id_private *id_priv = context;
1793         struct rdma_cm_event event;
1794
1795         memset(&event, 0, sizeof event);
1796         mutex_lock(&id_priv->handler_mutex);
1797
1798         /*
1799          * Grab mutex to block rdma_destroy_id() from removing the device while
1800          * we're trying to acquire it.
1801          */
1802         mutex_lock(&lock);
1803         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1804                 mutex_unlock(&lock);
1805                 goto out;
1806         }
1807
1808         if (!status && !id_priv->cma_dev)
1809                 status = cma_acquire_dev(id_priv);
1810         mutex_unlock(&lock);
1811
1812         if (status) {
1813                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1814                         goto out;
1815                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1816                 event.status = status;
1817         } else {
1818                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1819                        ip_addr_size(src_addr));
1820                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1821         }
1822
1823         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1824                 cma_exch(id_priv, CMA_DESTROYING);
1825                 mutex_unlock(&id_priv->handler_mutex);
1826                 cma_deref_id(id_priv);
1827                 rdma_destroy_id(&id_priv->id);
1828                 return;
1829         }
1830 out:
1831         mutex_unlock(&id_priv->handler_mutex);
1832         cma_deref_id(id_priv);
1833 }
1834
1835 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1836 {
1837         struct cma_work *work;
1838         struct sockaddr *src, *dst;
1839         union ib_gid gid;
1840         int ret;
1841
1842         work = kzalloc(sizeof *work, GFP_KERNEL);
1843         if (!work)
1844                 return -ENOMEM;
1845
1846         if (!id_priv->cma_dev) {
1847                 ret = cma_bind_loopback(id_priv);
1848                 if (ret)
1849                         goto err;
1850         }
1851
1852         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1853         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1854
1855         src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1856         if (cma_zero_addr(src)) {
1857                 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
1858                 if ((src->sa_family = dst->sa_family) == AF_INET) {
1859                         ((struct sockaddr_in *) src)->sin_addr.s_addr =
1860                                 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
1861                 } else {
1862                         ipv6_addr_copy(&((struct sockaddr_in6 *) src)->sin6_addr,
1863                                        &((struct sockaddr_in6 *) dst)->sin6_addr);
1864                 }
1865         }
1866
1867         work->id = id_priv;
1868         INIT_WORK(&work->work, cma_work_handler);
1869         work->old_state = CMA_ADDR_QUERY;
1870         work->new_state = CMA_ADDR_RESOLVED;
1871         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1872         queue_work(cma_wq, &work->work);
1873         return 0;
1874 err:
1875         kfree(work);
1876         return ret;
1877 }
1878
1879 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1880                          struct sockaddr *dst_addr)
1881 {
1882         if (!src_addr || !src_addr->sa_family) {
1883                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
1884                 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
1885                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
1886                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
1887                 }
1888         }
1889         return rdma_bind_addr(id, src_addr);
1890 }
1891
1892 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1893                       struct sockaddr *dst_addr, int timeout_ms)
1894 {
1895         struct rdma_id_private *id_priv;
1896         int ret;
1897
1898         id_priv = container_of(id, struct rdma_id_private, id);
1899         if (id_priv->state == CMA_IDLE) {
1900                 ret = cma_bind_addr(id, src_addr, dst_addr);
1901                 if (ret)
1902                         return ret;
1903         }
1904
1905         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1906                 return -EINVAL;
1907
1908         atomic_inc(&id_priv->refcount);
1909         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1910         if (cma_any_addr(dst_addr))
1911                 ret = cma_resolve_loopback(id_priv);
1912         else
1913                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
1914                                       dst_addr, &id->route.addr.dev_addr,
1915                                       timeout_ms, addr_handler, id_priv);
1916         if (ret)
1917                 goto err;
1918
1919         return 0;
1920 err:
1921         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1922         cma_deref_id(id_priv);
1923         return ret;
1924 }
1925 EXPORT_SYMBOL(rdma_resolve_addr);
1926
1927 static void cma_bind_port(struct rdma_bind_list *bind_list,
1928                           struct rdma_id_private *id_priv)
1929 {
1930         struct sockaddr_in *sin;
1931
1932         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1933         sin->sin_port = htons(bind_list->port);
1934         id_priv->bind_list = bind_list;
1935         hlist_add_head(&id_priv->node, &bind_list->owners);
1936 }
1937
1938 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1939                           unsigned short snum)
1940 {
1941         struct rdma_bind_list *bind_list;
1942         int port, ret;
1943
1944         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1945         if (!bind_list)
1946                 return -ENOMEM;
1947
1948         do {
1949                 ret = idr_get_new_above(ps, bind_list, snum, &port);
1950         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1951
1952         if (ret)
1953                 goto err1;
1954
1955         if (port != snum) {
1956                 ret = -EADDRNOTAVAIL;
1957                 goto err2;
1958         }
1959
1960         bind_list->ps = ps;
1961         bind_list->port = (unsigned short) port;
1962         cma_bind_port(bind_list, id_priv);
1963         return 0;
1964 err2:
1965         idr_remove(ps, port);
1966 err1:
1967         kfree(bind_list);
1968         return ret;
1969 }
1970
1971 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
1972 {
1973         struct rdma_bind_list *bind_list;
1974         int port, ret, low, high;
1975
1976         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1977         if (!bind_list)
1978                 return -ENOMEM;
1979
1980 retry:
1981         /* FIXME: add proper port randomization per like inet_csk_get_port */
1982         do {
1983                 ret = idr_get_new_above(ps, bind_list, next_port, &port);
1984         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1985
1986         if (ret)
1987                 goto err1;
1988
1989         inet_get_local_port_range(&low, &high);
1990         if (port > high) {
1991                 if (next_port != low) {
1992                         idr_remove(ps, port);
1993                         next_port = low;
1994                         goto retry;
1995                 }
1996                 ret = -EADDRNOTAVAIL;
1997                 goto err2;
1998         }
1999
2000         if (port == high)
2001                 next_port = low;
2002         else
2003                 next_port = port + 1;
2004
2005         bind_list->ps = ps;
2006         bind_list->port = (unsigned short) port;
2007         cma_bind_port(bind_list, id_priv);
2008         return 0;
2009 err2:
2010         idr_remove(ps, port);
2011 err1:
2012         kfree(bind_list);
2013         return ret;
2014 }
2015
2016 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2017 {
2018         struct rdma_id_private *cur_id;
2019         struct sockaddr_in *sin, *cur_sin;
2020         struct rdma_bind_list *bind_list;
2021         struct hlist_node *node;
2022         unsigned short snum;
2023
2024         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2025         snum = ntohs(sin->sin_port);
2026         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2027                 return -EACCES;
2028
2029         bind_list = idr_find(ps, snum);
2030         if (!bind_list)
2031                 return cma_alloc_port(ps, id_priv, snum);
2032
2033         /*
2034          * We don't support binding to any address if anyone is bound to
2035          * a specific address on the same port.
2036          */
2037         if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2038                 return -EADDRNOTAVAIL;
2039
2040         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2041                 if (cma_any_addr((struct sockaddr *) &cur_id->id.route.addr.src_addr))
2042                         return -EADDRNOTAVAIL;
2043
2044                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
2045                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
2046                         return -EADDRINUSE;
2047         }
2048
2049         cma_bind_port(bind_list, id_priv);
2050         return 0;
2051 }
2052
2053 static int cma_get_port(struct rdma_id_private *id_priv)
2054 {
2055         struct idr *ps;
2056         int ret;
2057
2058         switch (id_priv->id.ps) {
2059         case RDMA_PS_SDP:
2060                 ps = &sdp_ps;
2061                 break;
2062         case RDMA_PS_TCP:
2063                 ps = &tcp_ps;
2064                 break;
2065         case RDMA_PS_UDP:
2066                 ps = &udp_ps;
2067                 break;
2068         case RDMA_PS_IPOIB:
2069                 ps = &ipoib_ps;
2070                 break;
2071         default:
2072                 return -EPROTONOSUPPORT;
2073         }
2074
2075         mutex_lock(&lock);
2076         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2077                 ret = cma_alloc_any_port(ps, id_priv);
2078         else
2079                 ret = cma_use_port(ps, id_priv);
2080         mutex_unlock(&lock);
2081
2082         return ret;
2083 }
2084
2085 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2086                                struct sockaddr *addr)
2087 {
2088 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2089         struct sockaddr_in6 *sin6;
2090
2091         if (addr->sa_family != AF_INET6)
2092                 return 0;
2093
2094         sin6 = (struct sockaddr_in6 *) addr;
2095         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2096             !sin6->sin6_scope_id)
2097                         return -EINVAL;
2098
2099         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2100 #endif
2101         return 0;
2102 }
2103
2104 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2105 {
2106         struct rdma_id_private *id_priv;
2107         int ret;
2108
2109         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2110                 return -EAFNOSUPPORT;
2111
2112         id_priv = container_of(id, struct rdma_id_private, id);
2113         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
2114                 return -EINVAL;
2115
2116         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2117         if (ret)
2118                 goto err1;
2119
2120         if (!cma_any_addr(addr)) {
2121                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2122                 if (ret)
2123                         goto err1;
2124
2125                 mutex_lock(&lock);
2126                 ret = cma_acquire_dev(id_priv);
2127                 mutex_unlock(&lock);
2128                 if (ret)
2129                         goto err1;
2130         }
2131
2132         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2133         ret = cma_get_port(id_priv);
2134         if (ret)
2135                 goto err2;
2136
2137         return 0;
2138 err2:
2139         if (id_priv->cma_dev) {
2140                 mutex_lock(&lock);
2141                 cma_detach_from_dev(id_priv);
2142                 mutex_unlock(&lock);
2143         }
2144 err1:
2145         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
2146         return ret;
2147 }
2148 EXPORT_SYMBOL(rdma_bind_addr);
2149
2150 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2151                           struct rdma_route *route)
2152 {
2153         struct cma_hdr *cma_hdr;
2154         struct sdp_hh *sdp_hdr;
2155
2156         if (route->addr.src_addr.ss_family == AF_INET) {
2157                 struct sockaddr_in *src4, *dst4;
2158
2159                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2160                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2161
2162                 switch (ps) {
2163                 case RDMA_PS_SDP:
2164                         sdp_hdr = hdr;
2165                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2166                                 return -EINVAL;
2167                         sdp_set_ip_ver(sdp_hdr, 4);
2168                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2169                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2170                         sdp_hdr->port = src4->sin_port;
2171                         break;
2172                 default:
2173                         cma_hdr = hdr;
2174                         cma_hdr->cma_version = CMA_VERSION;
2175                         cma_set_ip_ver(cma_hdr, 4);
2176                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2177                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2178                         cma_hdr->port = src4->sin_port;
2179                         break;
2180                 }
2181         } else {
2182                 struct sockaddr_in6 *src6, *dst6;
2183
2184                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2185                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2186
2187                 switch (ps) {
2188                 case RDMA_PS_SDP:
2189                         sdp_hdr = hdr;
2190                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2191                                 return -EINVAL;
2192                         sdp_set_ip_ver(sdp_hdr, 6);
2193                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2194                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2195                         sdp_hdr->port = src6->sin6_port;
2196                         break;
2197                 default:
2198                         cma_hdr = hdr;
2199                         cma_hdr->cma_version = CMA_VERSION;
2200                         cma_set_ip_ver(cma_hdr, 6);
2201                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2202                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2203                         cma_hdr->port = src6->sin6_port;
2204                         break;
2205                 }
2206         }
2207         return 0;
2208 }
2209
2210 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2211                                 struct ib_cm_event *ib_event)
2212 {
2213         struct rdma_id_private *id_priv = cm_id->context;
2214         struct rdma_cm_event event;
2215         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2216         int ret = 0;
2217
2218         if (cma_disable_callback(id_priv, CMA_CONNECT))
2219                 return 0;
2220
2221         memset(&event, 0, sizeof event);
2222         switch (ib_event->event) {
2223         case IB_CM_SIDR_REQ_ERROR:
2224                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2225                 event.status = -ETIMEDOUT;
2226                 break;
2227         case IB_CM_SIDR_REP_RECEIVED:
2228                 event.param.ud.private_data = ib_event->private_data;
2229                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2230                 if (rep->status != IB_SIDR_SUCCESS) {
2231                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2232                         event.status = ib_event->param.sidr_rep_rcvd.status;
2233                         break;
2234                 }
2235                 ret = cma_set_qkey(id_priv);
2236                 if (ret) {
2237                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2238                         event.status = -EINVAL;
2239                         break;
2240                 }
2241                 if (id_priv->qkey != rep->qkey) {
2242                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2243                         event.status = -EINVAL;
2244                         break;
2245                 }
2246                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2247                                      id_priv->id.route.path_rec,
2248                                      &event.param.ud.ah_attr);
2249                 event.param.ud.qp_num = rep->qpn;
2250                 event.param.ud.qkey = rep->qkey;
2251                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2252                 event.status = 0;
2253                 break;
2254         default:
2255                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2256                        ib_event->event);
2257                 goto out;
2258         }
2259
2260         ret = id_priv->id.event_handler(&id_priv->id, &event);
2261         if (ret) {
2262                 /* Destroy the CM ID by returning a non-zero value. */
2263                 id_priv->cm_id.ib = NULL;
2264                 cma_exch(id_priv, CMA_DESTROYING);
2265                 mutex_unlock(&id_priv->handler_mutex);
2266                 rdma_destroy_id(&id_priv->id);
2267                 return ret;
2268         }
2269 out:
2270         mutex_unlock(&id_priv->handler_mutex);
2271         return ret;
2272 }
2273
2274 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2275                               struct rdma_conn_param *conn_param)
2276 {
2277         struct ib_cm_sidr_req_param req;
2278         struct rdma_route *route;
2279         int ret;
2280
2281         req.private_data_len = sizeof(struct cma_hdr) +
2282                                conn_param->private_data_len;
2283         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2284         if (!req.private_data)
2285                 return -ENOMEM;
2286
2287         if (conn_param->private_data && conn_param->private_data_len)
2288                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2289                        conn_param->private_data, conn_param->private_data_len);
2290
2291         route = &id_priv->id.route;
2292         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2293         if (ret)
2294                 goto out;
2295
2296         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2297                                             cma_sidr_rep_handler, id_priv);
2298         if (IS_ERR(id_priv->cm_id.ib)) {
2299                 ret = PTR_ERR(id_priv->cm_id.ib);
2300                 goto out;
2301         }
2302
2303         req.path = route->path_rec;
2304         req.service_id = cma_get_service_id(id_priv->id.ps,
2305                                             (struct sockaddr *) &route->addr.dst_addr);
2306         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2307         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2308
2309         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2310         if (ret) {
2311                 ib_destroy_cm_id(id_priv->cm_id.ib);
2312                 id_priv->cm_id.ib = NULL;
2313         }
2314 out:
2315         kfree(req.private_data);
2316         return ret;
2317 }
2318
2319 static int cma_connect_ib(struct rdma_id_private *id_priv,
2320                           struct rdma_conn_param *conn_param)
2321 {
2322         struct ib_cm_req_param req;
2323         struct rdma_route *route;
2324         void *private_data;
2325         int offset, ret;
2326
2327         memset(&req, 0, sizeof req);
2328         offset = cma_user_data_offset(id_priv->id.ps);
2329         req.private_data_len = offset + conn_param->private_data_len;
2330         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2331         if (!private_data)
2332                 return -ENOMEM;
2333
2334         if (conn_param->private_data && conn_param->private_data_len)
2335                 memcpy(private_data + offset, conn_param->private_data,
2336                        conn_param->private_data_len);
2337
2338         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2339                                             id_priv);
2340         if (IS_ERR(id_priv->cm_id.ib)) {
2341                 ret = PTR_ERR(id_priv->cm_id.ib);
2342                 goto out;
2343         }
2344
2345         route = &id_priv->id.route;
2346         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2347         if (ret)
2348                 goto out;
2349         req.private_data = private_data;
2350
2351         req.primary_path = &route->path_rec[0];
2352         if (route->num_paths == 2)
2353                 req.alternate_path = &route->path_rec[1];
2354
2355         req.service_id = cma_get_service_id(id_priv->id.ps,
2356                                             (struct sockaddr *) &route->addr.dst_addr);
2357         req.qp_num = id_priv->qp_num;
2358         req.qp_type = IB_QPT_RC;
2359         req.starting_psn = id_priv->seq_num;
2360         req.responder_resources = conn_param->responder_resources;
2361         req.initiator_depth = conn_param->initiator_depth;
2362         req.flow_control = conn_param->flow_control;
2363         req.retry_count = conn_param->retry_count;
2364         req.rnr_retry_count = conn_param->rnr_retry_count;
2365         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2366         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2367         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2368         req.srq = id_priv->srq ? 1 : 0;
2369
2370         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2371 out:
2372         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2373                 ib_destroy_cm_id(id_priv->cm_id.ib);
2374                 id_priv->cm_id.ib = NULL;
2375         }
2376
2377         kfree(private_data);
2378         return ret;
2379 }
2380
2381 static int cma_connect_iw(struct rdma_id_private *id_priv,
2382                           struct rdma_conn_param *conn_param)
2383 {
2384         struct iw_cm_id *cm_id;
2385         struct sockaddr_in* sin;
2386         int ret;
2387         struct iw_cm_conn_param iw_param;
2388
2389         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2390         if (IS_ERR(cm_id)) {
2391                 ret = PTR_ERR(cm_id);
2392                 goto out;
2393         }
2394
2395         id_priv->cm_id.iw = cm_id;
2396
2397         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2398         cm_id->local_addr = *sin;
2399
2400         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2401         cm_id->remote_addr = *sin;
2402
2403         ret = cma_modify_qp_rtr(id_priv, conn_param);
2404         if (ret)
2405                 goto out;
2406
2407         iw_param.ord = conn_param->initiator_depth;
2408         iw_param.ird = conn_param->responder_resources;
2409         iw_param.private_data = conn_param->private_data;
2410         iw_param.private_data_len = conn_param->private_data_len;
2411         if (id_priv->id.qp)
2412                 iw_param.qpn = id_priv->qp_num;
2413         else
2414                 iw_param.qpn = conn_param->qp_num;
2415         ret = iw_cm_connect(cm_id, &iw_param);
2416 out:
2417         if (ret && !IS_ERR(cm_id)) {
2418                 iw_destroy_cm_id(cm_id);
2419                 id_priv->cm_id.iw = NULL;
2420         }
2421         return ret;
2422 }
2423
2424 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2425 {
2426         struct rdma_id_private *id_priv;
2427         int ret;
2428
2429         id_priv = container_of(id, struct rdma_id_private, id);
2430         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2431                 return -EINVAL;
2432
2433         if (!id->qp) {
2434                 id_priv->qp_num = conn_param->qp_num;
2435                 id_priv->srq = conn_param->srq;
2436         }
2437
2438         switch (rdma_node_get_transport(id->device->node_type)) {
2439         case RDMA_TRANSPORT_IB:
2440                 if (cma_is_ud_ps(id->ps))
2441                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2442                 else
2443                         ret = cma_connect_ib(id_priv, conn_param);
2444                 break;
2445         case RDMA_TRANSPORT_IWARP:
2446                 ret = cma_connect_iw(id_priv, conn_param);
2447                 break;
2448         default:
2449                 ret = -ENOSYS;
2450                 break;
2451         }
2452         if (ret)
2453                 goto err;
2454
2455         return 0;
2456 err:
2457         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2458         return ret;
2459 }
2460 EXPORT_SYMBOL(rdma_connect);
2461
2462 static int cma_accept_ib(struct rdma_id_private *id_priv,
2463                          struct rdma_conn_param *conn_param)
2464 {
2465         struct ib_cm_rep_param rep;
2466         int ret;
2467
2468         ret = cma_modify_qp_rtr(id_priv, conn_param);
2469         if (ret)
2470                 goto out;
2471
2472         ret = cma_modify_qp_rts(id_priv, conn_param);
2473         if (ret)
2474                 goto out;
2475
2476         memset(&rep, 0, sizeof rep);
2477         rep.qp_num = id_priv->qp_num;
2478         rep.starting_psn = id_priv->seq_num;
2479         rep.private_data = conn_param->private_data;
2480         rep.private_data_len = conn_param->private_data_len;
2481         rep.responder_resources = conn_param->responder_resources;
2482         rep.initiator_depth = conn_param->initiator_depth;
2483         rep.failover_accepted = 0;
2484         rep.flow_control = conn_param->flow_control;
2485         rep.rnr_retry_count = conn_param->rnr_retry_count;
2486         rep.srq = id_priv->srq ? 1 : 0;
2487
2488         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2489 out:
2490         return ret;
2491 }
2492
2493 static int cma_accept_iw(struct rdma_id_private *id_priv,
2494                   struct rdma_conn_param *conn_param)
2495 {
2496         struct iw_cm_conn_param iw_param;
2497         int ret;
2498
2499         ret = cma_modify_qp_rtr(id_priv, conn_param);
2500         if (ret)
2501                 return ret;
2502
2503         iw_param.ord = conn_param->initiator_depth;
2504         iw_param.ird = conn_param->responder_resources;
2505         iw_param.private_data = conn_param->private_data;
2506         iw_param.private_data_len = conn_param->private_data_len;
2507         if (id_priv->id.qp) {
2508                 iw_param.qpn = id_priv->qp_num;
2509         } else
2510                 iw_param.qpn = conn_param->qp_num;
2511
2512         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2513 }
2514
2515 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2516                              enum ib_cm_sidr_status status,
2517                              const void *private_data, int private_data_len)
2518 {
2519         struct ib_cm_sidr_rep_param rep;
2520         int ret;
2521
2522         memset(&rep, 0, sizeof rep);
2523         rep.status = status;
2524         if (status == IB_SIDR_SUCCESS) {
2525                 ret = cma_set_qkey(id_priv);
2526                 if (ret)
2527                         return ret;
2528                 rep.qp_num = id_priv->qp_num;
2529                 rep.qkey = id_priv->qkey;
2530         }
2531         rep.private_data = private_data;
2532         rep.private_data_len = private_data_len;
2533
2534         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2535 }
2536
2537 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2538 {
2539         struct rdma_id_private *id_priv;
2540         int ret;
2541
2542         id_priv = container_of(id, struct rdma_id_private, id);
2543         if (!cma_comp(id_priv, CMA_CONNECT))
2544                 return -EINVAL;
2545
2546         if (!id->qp && conn_param) {
2547                 id_priv->qp_num = conn_param->qp_num;
2548                 id_priv->srq = conn_param->srq;
2549         }
2550
2551         switch (rdma_node_get_transport(id->device->node_type)) {
2552         case RDMA_TRANSPORT_IB:
2553                 if (cma_is_ud_ps(id->ps))
2554                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2555                                                 conn_param->private_data,
2556                                                 conn_param->private_data_len);
2557                 else if (conn_param)
2558                         ret = cma_accept_ib(id_priv, conn_param);
2559                 else
2560                         ret = cma_rep_recv(id_priv);
2561                 break;
2562         case RDMA_TRANSPORT_IWARP:
2563                 ret = cma_accept_iw(id_priv, conn_param);
2564                 break;
2565         default:
2566                 ret = -ENOSYS;
2567                 break;
2568         }
2569
2570         if (ret)
2571                 goto reject;
2572
2573         return 0;
2574 reject:
2575         cma_modify_qp_err(id_priv);
2576         rdma_reject(id, NULL, 0);
2577         return ret;
2578 }
2579 EXPORT_SYMBOL(rdma_accept);
2580
2581 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2582 {
2583         struct rdma_id_private *id_priv;
2584         int ret;
2585
2586         id_priv = container_of(id, struct rdma_id_private, id);
2587         if (!cma_has_cm_dev(id_priv))
2588                 return -EINVAL;
2589
2590         switch (id->device->node_type) {
2591         case RDMA_NODE_IB_CA:
2592                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2593                 break;
2594         default:
2595                 ret = 0;
2596                 break;
2597         }
2598         return ret;
2599 }
2600 EXPORT_SYMBOL(rdma_notify);
2601
2602 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2603                 u8 private_data_len)
2604 {
2605         struct rdma_id_private *id_priv;
2606         int ret;
2607
2608         id_priv = container_of(id, struct rdma_id_private, id);
2609         if (!cma_has_cm_dev(id_priv))
2610                 return -EINVAL;
2611
2612         switch (rdma_node_get_transport(id->device->node_type)) {
2613         case RDMA_TRANSPORT_IB:
2614                 if (cma_is_ud_ps(id->ps))
2615                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2616                                                 private_data, private_data_len);
2617                 else
2618                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2619                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2620                                              0, private_data, private_data_len);
2621                 break;
2622         case RDMA_TRANSPORT_IWARP:
2623                 ret = iw_cm_reject(id_priv->cm_id.iw,
2624                                    private_data, private_data_len);
2625                 break;
2626         default:
2627                 ret = -ENOSYS;
2628                 break;
2629         }
2630         return ret;
2631 }
2632 EXPORT_SYMBOL(rdma_reject);
2633
2634 int rdma_disconnect(struct rdma_cm_id *id)
2635 {
2636         struct rdma_id_private *id_priv;
2637         int ret;
2638
2639         id_priv = container_of(id, struct rdma_id_private, id);
2640         if (!cma_has_cm_dev(id_priv))
2641                 return -EINVAL;
2642
2643         switch (rdma_node_get_transport(id->device->node_type)) {
2644         case RDMA_TRANSPORT_IB:
2645                 ret = cma_modify_qp_err(id_priv);
2646                 if (ret)
2647                         goto out;
2648                 /* Initiate or respond to a disconnect. */
2649                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2650                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2651                 break;
2652         case RDMA_TRANSPORT_IWARP:
2653                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2654                 break;
2655         default:
2656                 ret = -EINVAL;
2657                 break;
2658         }
2659 out:
2660         return ret;
2661 }
2662 EXPORT_SYMBOL(rdma_disconnect);
2663
2664 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2665 {
2666         struct rdma_id_private *id_priv;
2667         struct cma_multicast *mc = multicast->context;
2668         struct rdma_cm_event event;
2669         int ret;
2670
2671         id_priv = mc->id_priv;
2672         if (cma_disable_callback(id_priv, CMA_ADDR_BOUND) &&
2673             cma_disable_callback(id_priv, CMA_ADDR_RESOLVED))
2674                 return 0;
2675
2676         mutex_lock(&id_priv->qp_mutex);
2677         if (!status && id_priv->id.qp)
2678                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2679                                          multicast->rec.mlid);
2680         mutex_unlock(&id_priv->qp_mutex);
2681
2682         memset(&event, 0, sizeof event);
2683         event.status = status;
2684         event.param.ud.private_data = mc->context;
2685         if (!status) {
2686                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2687                 ib_init_ah_from_mcmember(id_priv->id.device,
2688                                          id_priv->id.port_num, &multicast->rec,
2689                                          &event.param.ud.ah_attr);
2690                 event.param.ud.qp_num = 0xFFFFFF;
2691                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2692         } else
2693                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2694
2695         ret = id_priv->id.event_handler(&id_priv->id, &event);
2696         if (ret) {
2697                 cma_exch(id_priv, CMA_DESTROYING);
2698                 mutex_unlock(&id_priv->handler_mutex);
2699                 rdma_destroy_id(&id_priv->id);
2700                 return 0;
2701         }
2702
2703         mutex_unlock(&id_priv->handler_mutex);
2704         return 0;
2705 }
2706
2707 static void cma_set_mgid(struct rdma_id_private *id_priv,
2708                          struct sockaddr *addr, union ib_gid *mgid)
2709 {
2710         unsigned char mc_map[MAX_ADDR_LEN];
2711         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2712         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2713         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2714
2715         if (cma_any_addr(addr)) {
2716                 memset(mgid, 0, sizeof *mgid);
2717         } else if ((addr->sa_family == AF_INET6) &&
2718                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2719                                                                  0xFF10A01B)) {
2720                 /* IPv6 address is an SA assigned MGID. */
2721                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2722         } else if ((addr->sa_family == AF_INET6)) {
2723                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2724                 if (id_priv->id.ps == RDMA_PS_UDP)
2725                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2726                 *mgid = *(union ib_gid *) (mc_map + 4);
2727         } else {
2728                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2729                 if (id_priv->id.ps == RDMA_PS_UDP)
2730                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2731                 *mgid = *(union ib_gid *) (mc_map + 4);
2732         }
2733 }
2734
2735 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2736                                  struct cma_multicast *mc)
2737 {
2738         struct ib_sa_mcmember_rec rec;
2739         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2740         ib_sa_comp_mask comp_mask;
2741         int ret;
2742
2743         ib_addr_get_mgid(dev_addr, &rec.mgid);
2744         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2745                                      &rec.mgid, &rec);
2746         if (ret)
2747                 return ret;
2748
2749         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
2750         if (id_priv->id.ps == RDMA_PS_UDP)
2751                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2752         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
2753         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2754         rec.join_state = 1;
2755
2756         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2757                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2758                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2759                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
2760                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2761
2762         if (id_priv->id.ps == RDMA_PS_IPOIB)
2763                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
2764                              IB_SA_MCMEMBER_REC_RATE_SELECTOR;
2765
2766         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2767                                                 id_priv->id.port_num, &rec,
2768                                                 comp_mask, GFP_KERNEL,
2769                                                 cma_ib_mc_handler, mc);
2770         if (IS_ERR(mc->multicast.ib))
2771                 return PTR_ERR(mc->multicast.ib);
2772
2773         return 0;
2774 }
2775
2776 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
2777                         void *context)
2778 {
2779         struct rdma_id_private *id_priv;
2780         struct cma_multicast *mc;
2781         int ret;
2782
2783         id_priv = container_of(id, struct rdma_id_private, id);
2784         if (!cma_comp(id_priv, CMA_ADDR_BOUND) &&
2785             !cma_comp(id_priv, CMA_ADDR_RESOLVED))
2786                 return -EINVAL;
2787
2788         mc = kmalloc(sizeof *mc, GFP_KERNEL);
2789         if (!mc)
2790                 return -ENOMEM;
2791
2792         memcpy(&mc->addr, addr, ip_addr_size(addr));
2793         mc->context = context;
2794         mc->id_priv = id_priv;
2795
2796         spin_lock(&id_priv->lock);
2797         list_add(&mc->list, &id_priv->mc_list);
2798         spin_unlock(&id_priv->lock);
2799
2800         switch (rdma_node_get_transport(id->device->node_type)) {
2801         case RDMA_TRANSPORT_IB:
2802                 ret = cma_join_ib_multicast(id_priv, mc);
2803                 break;
2804         default:
2805                 ret = -ENOSYS;
2806                 break;
2807         }
2808
2809         if (ret) {
2810                 spin_lock_irq(&id_priv->lock);
2811                 list_del(&mc->list);
2812                 spin_unlock_irq(&id_priv->lock);
2813                 kfree(mc);
2814         }
2815         return ret;
2816 }
2817 EXPORT_SYMBOL(rdma_join_multicast);
2818
2819 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
2820 {
2821         struct rdma_id_private *id_priv;
2822         struct cma_multicast *mc;
2823
2824         id_priv = container_of(id, struct rdma_id_private, id);
2825         spin_lock_irq(&id_priv->lock);
2826         list_for_each_entry(mc, &id_priv->mc_list, list) {
2827                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
2828                         list_del(&mc->list);
2829                         spin_unlock_irq(&id_priv->lock);
2830
2831                         if (id->qp)
2832                                 ib_detach_mcast(id->qp,
2833                                                 &mc->multicast.ib->rec.mgid,
2834                                                 mc->multicast.ib->rec.mlid);
2835                         ib_sa_free_multicast(mc->multicast.ib);
2836                         kfree(mc);
2837                         return;
2838                 }
2839         }
2840         spin_unlock_irq(&id_priv->lock);
2841 }
2842 EXPORT_SYMBOL(rdma_leave_multicast);
2843
2844 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
2845 {
2846         struct rdma_dev_addr *dev_addr;
2847         struct cma_ndev_work *work;
2848
2849         dev_addr = &id_priv->id.route.addr.dev_addr;
2850
2851         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
2852             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
2853                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
2854                        ndev->name, &id_priv->id);
2855                 work = kzalloc(sizeof *work, GFP_KERNEL);
2856                 if (!work)
2857                         return -ENOMEM;
2858
2859                 INIT_WORK(&work->work, cma_ndev_work_handler);
2860                 work->id = id_priv;
2861                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
2862                 atomic_inc(&id_priv->refcount);
2863                 queue_work(cma_wq, &work->work);
2864         }
2865
2866         return 0;
2867 }
2868
2869 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
2870                                void *ctx)
2871 {
2872         struct net_device *ndev = (struct net_device *)ctx;
2873         struct cma_device *cma_dev;
2874         struct rdma_id_private *id_priv;
2875         int ret = NOTIFY_DONE;
2876
2877         if (dev_net(ndev) != &init_net)
2878                 return NOTIFY_DONE;
2879
2880         if (event != NETDEV_BONDING_FAILOVER)
2881                 return NOTIFY_DONE;
2882
2883         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
2884                 return NOTIFY_DONE;
2885
2886         mutex_lock(&lock);
2887         list_for_each_entry(cma_dev, &dev_list, list)
2888                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
2889                         ret = cma_netdev_change(ndev, id_priv);
2890                         if (ret)
2891                                 goto out;
2892                 }
2893
2894 out:
2895         mutex_unlock(&lock);
2896         return ret;
2897 }
2898
2899 static struct notifier_block cma_nb = {
2900         .notifier_call = cma_netdev_callback
2901 };
2902
2903 static void cma_add_one(struct ib_device *device)
2904 {
2905         struct cma_device *cma_dev;
2906         struct rdma_id_private *id_priv;
2907
2908         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2909         if (!cma_dev)
2910                 return;
2911
2912         cma_dev->device = device;
2913
2914         init_completion(&cma_dev->comp);
2915         atomic_set(&cma_dev->refcount, 1);
2916         INIT_LIST_HEAD(&cma_dev->id_list);
2917         ib_set_client_data(device, &cma_client, cma_dev);
2918
2919         mutex_lock(&lock);
2920         list_add_tail(&cma_dev->list, &dev_list);
2921         list_for_each_entry(id_priv, &listen_any_list, list)
2922                 cma_listen_on_dev(id_priv, cma_dev);
2923         mutex_unlock(&lock);
2924 }
2925
2926 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2927 {
2928         struct rdma_cm_event event;
2929         enum cma_state state;
2930         int ret = 0;
2931
2932         /* Record that we want to remove the device */
2933         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2934         if (state == CMA_DESTROYING)
2935                 return 0;
2936
2937         cma_cancel_operation(id_priv, state);
2938         mutex_lock(&id_priv->handler_mutex);
2939
2940         /* Check for destruction from another callback. */
2941         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2942                 goto out;
2943
2944         memset(&event, 0, sizeof event);
2945         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
2946         ret = id_priv->id.event_handler(&id_priv->id, &event);
2947 out:
2948         mutex_unlock(&id_priv->handler_mutex);
2949         return ret;
2950 }
2951
2952 static void cma_process_remove(struct cma_device *cma_dev)
2953 {
2954         struct rdma_id_private *id_priv;
2955         int ret;
2956
2957         mutex_lock(&lock);
2958         while (!list_empty(&cma_dev->id_list)) {
2959                 id_priv = list_entry(cma_dev->id_list.next,
2960                                      struct rdma_id_private, list);
2961
2962                 list_del(&id_priv->listen_list);
2963                 list_del_init(&id_priv->list);
2964                 atomic_inc(&id_priv->refcount);
2965                 mutex_unlock(&lock);
2966
2967                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
2968                 cma_deref_id(id_priv);
2969                 if (ret)
2970                         rdma_destroy_id(&id_priv->id);
2971
2972                 mutex_lock(&lock);
2973         }
2974         mutex_unlock(&lock);
2975
2976         cma_deref_dev(cma_dev);
2977         wait_for_completion(&cma_dev->comp);
2978 }
2979
2980 static void cma_remove_one(struct ib_device *device)
2981 {
2982         struct cma_device *cma_dev;
2983
2984         cma_dev = ib_get_client_data(device, &cma_client);
2985         if (!cma_dev)
2986                 return;
2987
2988         mutex_lock(&lock);
2989         list_del(&cma_dev->list);
2990         mutex_unlock(&lock);
2991
2992         cma_process_remove(cma_dev);
2993         kfree(cma_dev);
2994 }
2995
2996 static int __init cma_init(void)
2997 {
2998         int ret, low, high, remaining;
2999
3000         get_random_bytes(&next_port, sizeof next_port);
3001         inet_get_local_port_range(&low, &high);
3002         remaining = (high - low) + 1;
3003         next_port = ((unsigned int) next_port % remaining) + low;
3004
3005         cma_wq = create_singlethread_workqueue("rdma_cm");
3006         if (!cma_wq)
3007                 return -ENOMEM;
3008
3009         ib_sa_register_client(&sa_client);
3010         rdma_addr_register_client(&addr_client);
3011         register_netdevice_notifier(&cma_nb);
3012
3013         ret = ib_register_client(&cma_client);
3014         if (ret)
3015                 goto err;
3016         return 0;
3017
3018 err:
3019         unregister_netdevice_notifier(&cma_nb);
3020         rdma_addr_unregister_client(&addr_client);
3021         ib_sa_unregister_client(&sa_client);
3022         destroy_workqueue(cma_wq);
3023         return ret;
3024 }
3025
3026 static void __exit cma_cleanup(void)
3027 {
3028         ib_unregister_client(&cma_client);
3029         unregister_netdevice_notifier(&cma_nb);
3030         rdma_addr_unregister_client(&addr_client);
3031         ib_sa_unregister_client(&sa_client);
3032         destroy_workqueue(cma_wq);
3033         idr_destroy(&sdp_ps);
3034         idr_destroy(&tcp_ps);
3035         idr_destroy(&udp_ps);
3036         idr_destroy(&ipoib_ps);
3037 }
3038
3039 module_init(cma_init);
3040 module_exit(cma_cleanup);