Merge branch 'core-iommu-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[linux-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 #define CMA_IBOE_PACKET_LIFETIME 18
63
64 static void cma_add_one(struct ib_device *device);
65 static void cma_remove_one(struct ib_device *device);
66
67 static struct ib_client cma_client = {
68         .name   = "cma",
69         .add    = cma_add_one,
70         .remove = cma_remove_one
71 };
72
73 static struct ib_sa_client sa_client;
74 static struct rdma_addr_client addr_client;
75 static LIST_HEAD(dev_list);
76 static LIST_HEAD(listen_any_list);
77 static DEFINE_MUTEX(lock);
78 static struct workqueue_struct *cma_wq;
79 static DEFINE_IDR(sdp_ps);
80 static DEFINE_IDR(tcp_ps);
81 static DEFINE_IDR(udp_ps);
82 static DEFINE_IDR(ipoib_ps);
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         u8                      reuseaddr;
152 };
153
154 struct cma_multicast {
155         struct rdma_id_private *id_priv;
156         union {
157                 struct ib_sa_multicast *ib;
158         } multicast;
159         struct list_head        list;
160         void                    *context;
161         struct sockaddr_storage addr;
162         struct kref             mcref;
163 };
164
165 struct cma_work {
166         struct work_struct      work;
167         struct rdma_id_private  *id;
168         enum cma_state          old_state;
169         enum cma_state          new_state;
170         struct rdma_cm_event    event;
171 };
172
173 struct cma_ndev_work {
174         struct work_struct      work;
175         struct rdma_id_private  *id;
176         struct rdma_cm_event    event;
177 };
178
179 struct iboe_mcast_work {
180         struct work_struct       work;
181         struct rdma_id_private  *id;
182         struct cma_multicast    *mc;
183 };
184
185 union cma_ip_addr {
186         struct in6_addr ip6;
187         struct {
188                 __be32 pad[3];
189                 __be32 addr;
190         } ip4;
191 };
192
193 struct cma_hdr {
194         u8 cma_version;
195         u8 ip_version;  /* IP version: 7:4 */
196         __be16 port;
197         union cma_ip_addr src_addr;
198         union cma_ip_addr dst_addr;
199 };
200
201 struct sdp_hh {
202         u8 bsdh[16];
203         u8 sdp_version; /* Major version: 7:4 */
204         u8 ip_version;  /* IP version: 7:4 */
205         u8 sdp_specific1[10];
206         __be16 port;
207         __be16 sdp_specific2;
208         union cma_ip_addr src_addr;
209         union cma_ip_addr dst_addr;
210 };
211
212 struct sdp_hah {
213         u8 bsdh[16];
214         u8 sdp_version;
215 };
216
217 #define CMA_VERSION 0x00
218 #define SDP_MAJ_VERSION 0x2
219
220 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
221 {
222         unsigned long flags;
223         int ret;
224
225         spin_lock_irqsave(&id_priv->lock, flags);
226         ret = (id_priv->state == comp);
227         spin_unlock_irqrestore(&id_priv->lock, flags);
228         return ret;
229 }
230
231 static int cma_comp_exch(struct rdma_id_private *id_priv,
232                          enum cma_state comp, enum cma_state exch)
233 {
234         unsigned long flags;
235         int ret;
236
237         spin_lock_irqsave(&id_priv->lock, flags);
238         if ((ret = (id_priv->state == comp)))
239                 id_priv->state = exch;
240         spin_unlock_irqrestore(&id_priv->lock, flags);
241         return ret;
242 }
243
244 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
245                                enum cma_state exch)
246 {
247         unsigned long flags;
248         enum cma_state old;
249
250         spin_lock_irqsave(&id_priv->lock, flags);
251         old = id_priv->state;
252         id_priv->state = exch;
253         spin_unlock_irqrestore(&id_priv->lock, flags);
254         return old;
255 }
256
257 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
258 {
259         return hdr->ip_version >> 4;
260 }
261
262 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
263 {
264         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
265 }
266
267 static inline u8 sdp_get_majv(u8 sdp_version)
268 {
269         return sdp_version >> 4;
270 }
271
272 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
273 {
274         return hh->ip_version >> 4;
275 }
276
277 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
278 {
279         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
280 }
281
282 static inline int cma_is_ud_ps(enum rdma_port_space ps)
283 {
284         return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB);
285 }
286
287 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
288                               struct cma_device *cma_dev)
289 {
290         atomic_inc(&cma_dev->refcount);
291         id_priv->cma_dev = cma_dev;
292         id_priv->id.device = cma_dev->device;
293         id_priv->id.route.addr.dev_addr.transport =
294                 rdma_node_get_transport(cma_dev->device->node_type);
295         list_add_tail(&id_priv->list, &cma_dev->id_list);
296 }
297
298 static inline void cma_deref_dev(struct cma_device *cma_dev)
299 {
300         if (atomic_dec_and_test(&cma_dev->refcount))
301                 complete(&cma_dev->comp);
302 }
303
304 static inline void release_mc(struct kref *kref)
305 {
306         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
307
308         kfree(mc->multicast.ib);
309         kfree(mc);
310 }
311
312 static void cma_release_dev(struct rdma_id_private *id_priv)
313 {
314         mutex_lock(&lock);
315         list_del(&id_priv->list);
316         cma_deref_dev(id_priv->cma_dev);
317         id_priv->cma_dev = NULL;
318         mutex_unlock(&lock);
319 }
320
321 static int cma_set_qkey(struct rdma_id_private *id_priv)
322 {
323         struct ib_sa_mcmember_rec rec;
324         int ret = 0;
325
326         if (id_priv->qkey)
327                 return 0;
328
329         switch (id_priv->id.ps) {
330         case RDMA_PS_UDP:
331                 id_priv->qkey = RDMA_UDP_QKEY;
332                 break;
333         case RDMA_PS_IPOIB:
334                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
335                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
336                                              id_priv->id.port_num, &rec.mgid,
337                                              &rec);
338                 if (!ret)
339                         id_priv->qkey = be32_to_cpu(rec.qkey);
340                 break;
341         default:
342                 break;
343         }
344         return ret;
345 }
346
347 static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
348 {
349         int i;
350         int err;
351         struct ib_port_attr props;
352         union ib_gid tmp;
353
354         err = ib_query_port(device, port_num, &props);
355         if (err)
356                 return 1;
357
358         for (i = 0; i < props.gid_tbl_len; ++i) {
359                 err = ib_query_gid(device, port_num, i, &tmp);
360                 if (err)
361                         return 1;
362                 if (!memcmp(&tmp, gid, sizeof tmp))
363                         return 0;
364         }
365
366         return -EAGAIN;
367 }
368
369 static int cma_acquire_dev(struct rdma_id_private *id_priv)
370 {
371         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
372         struct cma_device *cma_dev;
373         union ib_gid gid, iboe_gid;
374         int ret = -ENODEV;
375         u8 port;
376         enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
377                 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
378
379         mutex_lock(&lock);
380         iboe_addr_get_sgid(dev_addr, &iboe_gid);
381         memcpy(&gid, dev_addr->src_dev_addr +
382                rdma_addr_gid_offset(dev_addr), sizeof gid);
383         list_for_each_entry(cma_dev, &dev_list, list) {
384                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
385                         if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
386                                 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
387                                     rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
388                                         ret = find_gid_port(cma_dev->device, &iboe_gid, port);
389                                 else
390                                         ret = find_gid_port(cma_dev->device, &gid, port);
391
392                                 if (!ret) {
393                                         id_priv->id.port_num = port;
394                                         goto out;
395                                 } else if (ret == 1)
396                                         break;
397                         }
398                 }
399         }
400
401 out:
402         if (!ret)
403                 cma_attach_to_dev(id_priv, cma_dev);
404
405         mutex_unlock(&lock);
406         return ret;
407 }
408
409 static void cma_deref_id(struct rdma_id_private *id_priv)
410 {
411         if (atomic_dec_and_test(&id_priv->refcount))
412                 complete(&id_priv->comp);
413 }
414
415 static int cma_disable_callback(struct rdma_id_private *id_priv,
416                               enum cma_state state)
417 {
418         mutex_lock(&id_priv->handler_mutex);
419         if (id_priv->state != state) {
420                 mutex_unlock(&id_priv->handler_mutex);
421                 return -EINVAL;
422         }
423         return 0;
424 }
425
426 static int cma_has_cm_dev(struct rdma_id_private *id_priv)
427 {
428         return (id_priv->id.device && id_priv->cm_id.ib);
429 }
430
431 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
432                                   void *context, enum rdma_port_space ps)
433 {
434         struct rdma_id_private *id_priv;
435
436         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
437         if (!id_priv)
438                 return ERR_PTR(-ENOMEM);
439
440         id_priv->state = CMA_IDLE;
441         id_priv->id.context = context;
442         id_priv->id.event_handler = event_handler;
443         id_priv->id.ps = ps;
444         spin_lock_init(&id_priv->lock);
445         mutex_init(&id_priv->qp_mutex);
446         init_completion(&id_priv->comp);
447         atomic_set(&id_priv->refcount, 1);
448         mutex_init(&id_priv->handler_mutex);
449         INIT_LIST_HEAD(&id_priv->listen_list);
450         INIT_LIST_HEAD(&id_priv->mc_list);
451         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
452
453         return &id_priv->id;
454 }
455 EXPORT_SYMBOL(rdma_create_id);
456
457 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
458 {
459         struct ib_qp_attr qp_attr;
460         int qp_attr_mask, ret;
461
462         qp_attr.qp_state = IB_QPS_INIT;
463         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
464         if (ret)
465                 return ret;
466
467         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
468         if (ret)
469                 return ret;
470
471         qp_attr.qp_state = IB_QPS_RTR;
472         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
473         if (ret)
474                 return ret;
475
476         qp_attr.qp_state = IB_QPS_RTS;
477         qp_attr.sq_psn = 0;
478         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
479
480         return ret;
481 }
482
483 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
484 {
485         struct ib_qp_attr qp_attr;
486         int qp_attr_mask, ret;
487
488         qp_attr.qp_state = IB_QPS_INIT;
489         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
490         if (ret)
491                 return ret;
492
493         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
494 }
495
496 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
497                    struct ib_qp_init_attr *qp_init_attr)
498 {
499         struct rdma_id_private *id_priv;
500         struct ib_qp *qp;
501         int ret;
502
503         id_priv = container_of(id, struct rdma_id_private, id);
504         if (id->device != pd->device)
505                 return -EINVAL;
506
507         qp = ib_create_qp(pd, qp_init_attr);
508         if (IS_ERR(qp))
509                 return PTR_ERR(qp);
510
511         if (cma_is_ud_ps(id_priv->id.ps))
512                 ret = cma_init_ud_qp(id_priv, qp);
513         else
514                 ret = cma_init_conn_qp(id_priv, qp);
515         if (ret)
516                 goto err;
517
518         id->qp = qp;
519         id_priv->qp_num = qp->qp_num;
520         id_priv->srq = (qp->srq != NULL);
521         return 0;
522 err:
523         ib_destroy_qp(qp);
524         return ret;
525 }
526 EXPORT_SYMBOL(rdma_create_qp);
527
528 void rdma_destroy_qp(struct rdma_cm_id *id)
529 {
530         struct rdma_id_private *id_priv;
531
532         id_priv = container_of(id, struct rdma_id_private, id);
533         mutex_lock(&id_priv->qp_mutex);
534         ib_destroy_qp(id_priv->id.qp);
535         id_priv->id.qp = NULL;
536         mutex_unlock(&id_priv->qp_mutex);
537 }
538 EXPORT_SYMBOL(rdma_destroy_qp);
539
540 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
541                              struct rdma_conn_param *conn_param)
542 {
543         struct ib_qp_attr qp_attr;
544         int qp_attr_mask, ret;
545
546         mutex_lock(&id_priv->qp_mutex);
547         if (!id_priv->id.qp) {
548                 ret = 0;
549                 goto out;
550         }
551
552         /* Need to update QP attributes from default values. */
553         qp_attr.qp_state = IB_QPS_INIT;
554         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
555         if (ret)
556                 goto out;
557
558         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
559         if (ret)
560                 goto out;
561
562         qp_attr.qp_state = IB_QPS_RTR;
563         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
564         if (ret)
565                 goto out;
566
567         if (conn_param)
568                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
569         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
570 out:
571         mutex_unlock(&id_priv->qp_mutex);
572         return ret;
573 }
574
575 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
576                              struct rdma_conn_param *conn_param)
577 {
578         struct ib_qp_attr qp_attr;
579         int qp_attr_mask, ret;
580
581         mutex_lock(&id_priv->qp_mutex);
582         if (!id_priv->id.qp) {
583                 ret = 0;
584                 goto out;
585         }
586
587         qp_attr.qp_state = IB_QPS_RTS;
588         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
589         if (ret)
590                 goto out;
591
592         if (conn_param)
593                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
594         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
595 out:
596         mutex_unlock(&id_priv->qp_mutex);
597         return ret;
598 }
599
600 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
601 {
602         struct ib_qp_attr qp_attr;
603         int ret;
604
605         mutex_lock(&id_priv->qp_mutex);
606         if (!id_priv->id.qp) {
607                 ret = 0;
608                 goto out;
609         }
610
611         qp_attr.qp_state = IB_QPS_ERR;
612         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
613 out:
614         mutex_unlock(&id_priv->qp_mutex);
615         return ret;
616 }
617
618 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
619                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
620 {
621         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
622         int ret;
623         u16 pkey;
624
625         if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
626             IB_LINK_LAYER_INFINIBAND)
627                 pkey = ib_addr_get_pkey(dev_addr);
628         else
629                 pkey = 0xffff;
630
631         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
632                                   pkey, &qp_attr->pkey_index);
633         if (ret)
634                 return ret;
635
636         qp_attr->port_num = id_priv->id.port_num;
637         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
638
639         if (cma_is_ud_ps(id_priv->id.ps)) {
640                 ret = cma_set_qkey(id_priv);
641                 if (ret)
642                         return ret;
643
644                 qp_attr->qkey = id_priv->qkey;
645                 *qp_attr_mask |= IB_QP_QKEY;
646         } else {
647                 qp_attr->qp_access_flags = 0;
648                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
649         }
650         return 0;
651 }
652
653 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
654                        int *qp_attr_mask)
655 {
656         struct rdma_id_private *id_priv;
657         int ret = 0;
658
659         id_priv = container_of(id, struct rdma_id_private, id);
660         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
661         case RDMA_TRANSPORT_IB:
662                 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps))
663                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
664                 else
665                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
666                                                  qp_attr_mask);
667                 if (qp_attr->qp_state == IB_QPS_RTR)
668                         qp_attr->rq_psn = id_priv->seq_num;
669                 break;
670         case RDMA_TRANSPORT_IWARP:
671                 if (!id_priv->cm_id.iw) {
672                         qp_attr->qp_access_flags = 0;
673                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
674                 } else
675                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
676                                                  qp_attr_mask);
677                 break;
678         default:
679                 ret = -ENOSYS;
680                 break;
681         }
682
683         return ret;
684 }
685 EXPORT_SYMBOL(rdma_init_qp_attr);
686
687 static inline int cma_zero_addr(struct sockaddr *addr)
688 {
689         struct in6_addr *ip6;
690
691         if (addr->sa_family == AF_INET)
692                 return ipv4_is_zeronet(
693                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
694         else {
695                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
696                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
697                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
698         }
699 }
700
701 static inline int cma_loopback_addr(struct sockaddr *addr)
702 {
703         if (addr->sa_family == AF_INET)
704                 return ipv4_is_loopback(
705                         ((struct sockaddr_in *) addr)->sin_addr.s_addr);
706         else
707                 return ipv6_addr_loopback(
708                         &((struct sockaddr_in6 *) addr)->sin6_addr);
709 }
710
711 static inline int cma_any_addr(struct sockaddr *addr)
712 {
713         return cma_zero_addr(addr) || cma_loopback_addr(addr);
714 }
715
716 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
717 {
718         if (src->sa_family != dst->sa_family)
719                 return -1;
720
721         switch (src->sa_family) {
722         case AF_INET:
723                 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
724                        ((struct sockaddr_in *) dst)->sin_addr.s_addr;
725         default:
726                 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
727                                      &((struct sockaddr_in6 *) dst)->sin6_addr);
728         }
729 }
730
731 static inline __be16 cma_port(struct sockaddr *addr)
732 {
733         if (addr->sa_family == AF_INET)
734                 return ((struct sockaddr_in *) addr)->sin_port;
735         else
736                 return ((struct sockaddr_in6 *) addr)->sin6_port;
737 }
738
739 static inline int cma_any_port(struct sockaddr *addr)
740 {
741         return !cma_port(addr);
742 }
743
744 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
745                             u8 *ip_ver, __be16 *port,
746                             union cma_ip_addr **src, union cma_ip_addr **dst)
747 {
748         switch (ps) {
749         case RDMA_PS_SDP:
750                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
751                     SDP_MAJ_VERSION)
752                         return -EINVAL;
753
754                 *ip_ver = sdp_get_ip_ver(hdr);
755                 *port   = ((struct sdp_hh *) hdr)->port;
756                 *src    = &((struct sdp_hh *) hdr)->src_addr;
757                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
758                 break;
759         default:
760                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
761                         return -EINVAL;
762
763                 *ip_ver = cma_get_ip_ver(hdr);
764                 *port   = ((struct cma_hdr *) hdr)->port;
765                 *src    = &((struct cma_hdr *) hdr)->src_addr;
766                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
767                 break;
768         }
769
770         if (*ip_ver != 4 && *ip_ver != 6)
771                 return -EINVAL;
772         return 0;
773 }
774
775 static void cma_save_net_info(struct rdma_addr *addr,
776                               struct rdma_addr *listen_addr,
777                               u8 ip_ver, __be16 port,
778                               union cma_ip_addr *src, union cma_ip_addr *dst)
779 {
780         struct sockaddr_in *listen4, *ip4;
781         struct sockaddr_in6 *listen6, *ip6;
782
783         switch (ip_ver) {
784         case 4:
785                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
786                 ip4 = (struct sockaddr_in *) &addr->src_addr;
787                 ip4->sin_family = listen4->sin_family;
788                 ip4->sin_addr.s_addr = dst->ip4.addr;
789                 ip4->sin_port = listen4->sin_port;
790
791                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
792                 ip4->sin_family = listen4->sin_family;
793                 ip4->sin_addr.s_addr = src->ip4.addr;
794                 ip4->sin_port = port;
795                 break;
796         case 6:
797                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
798                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
799                 ip6->sin6_family = listen6->sin6_family;
800                 ip6->sin6_addr = dst->ip6;
801                 ip6->sin6_port = listen6->sin6_port;
802
803                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
804                 ip6->sin6_family = listen6->sin6_family;
805                 ip6->sin6_addr = src->ip6;
806                 ip6->sin6_port = port;
807                 break;
808         default:
809                 break;
810         }
811 }
812
813 static inline int cma_user_data_offset(enum rdma_port_space ps)
814 {
815         switch (ps) {
816         case RDMA_PS_SDP:
817                 return 0;
818         default:
819                 return sizeof(struct cma_hdr);
820         }
821 }
822
823 static void cma_cancel_route(struct rdma_id_private *id_priv)
824 {
825         switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
826         case IB_LINK_LAYER_INFINIBAND:
827                 if (id_priv->query)
828                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
829                 break;
830         default:
831                 break;
832         }
833 }
834
835 static void cma_cancel_listens(struct rdma_id_private *id_priv)
836 {
837         struct rdma_id_private *dev_id_priv;
838
839         /*
840          * Remove from listen_any_list to prevent added devices from spawning
841          * additional listen requests.
842          */
843         mutex_lock(&lock);
844         list_del(&id_priv->list);
845
846         while (!list_empty(&id_priv->listen_list)) {
847                 dev_id_priv = list_entry(id_priv->listen_list.next,
848                                          struct rdma_id_private, listen_list);
849                 /* sync with device removal to avoid duplicate destruction */
850                 list_del_init(&dev_id_priv->list);
851                 list_del(&dev_id_priv->listen_list);
852                 mutex_unlock(&lock);
853
854                 rdma_destroy_id(&dev_id_priv->id);
855                 mutex_lock(&lock);
856         }
857         mutex_unlock(&lock);
858 }
859
860 static void cma_cancel_operation(struct rdma_id_private *id_priv,
861                                  enum cma_state state)
862 {
863         switch (state) {
864         case CMA_ADDR_QUERY:
865                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
866                 break;
867         case CMA_ROUTE_QUERY:
868                 cma_cancel_route(id_priv);
869                 break;
870         case CMA_LISTEN:
871                 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
872                                 && !id_priv->cma_dev)
873                         cma_cancel_listens(id_priv);
874                 break;
875         default:
876                 break;
877         }
878 }
879
880 static void cma_release_port(struct rdma_id_private *id_priv)
881 {
882         struct rdma_bind_list *bind_list = id_priv->bind_list;
883
884         if (!bind_list)
885                 return;
886
887         mutex_lock(&lock);
888         hlist_del(&id_priv->node);
889         if (hlist_empty(&bind_list->owners)) {
890                 idr_remove(bind_list->ps, bind_list->port);
891                 kfree(bind_list);
892         }
893         mutex_unlock(&lock);
894 }
895
896 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
897 {
898         struct cma_multicast *mc;
899
900         while (!list_empty(&id_priv->mc_list)) {
901                 mc = container_of(id_priv->mc_list.next,
902                                   struct cma_multicast, list);
903                 list_del(&mc->list);
904                 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
905                 case IB_LINK_LAYER_INFINIBAND:
906                         ib_sa_free_multicast(mc->multicast.ib);
907                         kfree(mc);
908                         break;
909                 case IB_LINK_LAYER_ETHERNET:
910                         kref_put(&mc->mcref, release_mc);
911                         break;
912                 default:
913                         break;
914                 }
915         }
916 }
917
918 void rdma_destroy_id(struct rdma_cm_id *id)
919 {
920         struct rdma_id_private *id_priv;
921         enum cma_state state;
922
923         id_priv = container_of(id, struct rdma_id_private, id);
924         state = cma_exch(id_priv, CMA_DESTROYING);
925         cma_cancel_operation(id_priv, state);
926
927         /*
928          * Wait for any active callback to finish.  New callbacks will find
929          * the id_priv state set to destroying and abort.
930          */
931         mutex_lock(&id_priv->handler_mutex);
932         mutex_unlock(&id_priv->handler_mutex);
933
934         if (id_priv->cma_dev) {
935                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
936                 case RDMA_TRANSPORT_IB:
937                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
938                                 ib_destroy_cm_id(id_priv->cm_id.ib);
939                         break;
940                 case RDMA_TRANSPORT_IWARP:
941                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
942                                 iw_destroy_cm_id(id_priv->cm_id.iw);
943                         break;
944                 default:
945                         break;
946                 }
947                 cma_leave_mc_groups(id_priv);
948                 cma_release_dev(id_priv);
949         }
950
951         cma_release_port(id_priv);
952         cma_deref_id(id_priv);
953         wait_for_completion(&id_priv->comp);
954
955         if (id_priv->internal_id)
956                 cma_deref_id(id_priv->id.context);
957
958         kfree(id_priv->id.route.path_rec);
959         kfree(id_priv);
960 }
961 EXPORT_SYMBOL(rdma_destroy_id);
962
963 static int cma_rep_recv(struct rdma_id_private *id_priv)
964 {
965         int ret;
966
967         ret = cma_modify_qp_rtr(id_priv, NULL);
968         if (ret)
969                 goto reject;
970
971         ret = cma_modify_qp_rts(id_priv, NULL);
972         if (ret)
973                 goto reject;
974
975         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
976         if (ret)
977                 goto reject;
978
979         return 0;
980 reject:
981         cma_modify_qp_err(id_priv);
982         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
983                        NULL, 0, NULL, 0);
984         return ret;
985 }
986
987 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
988 {
989         if (id_priv->id.ps == RDMA_PS_SDP &&
990             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
991             SDP_MAJ_VERSION)
992                 return -EINVAL;
993
994         return 0;
995 }
996
997 static void cma_set_rep_event_data(struct rdma_cm_event *event,
998                                    struct ib_cm_rep_event_param *rep_data,
999                                    void *private_data)
1000 {
1001         event->param.conn.private_data = private_data;
1002         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1003         event->param.conn.responder_resources = rep_data->responder_resources;
1004         event->param.conn.initiator_depth = rep_data->initiator_depth;
1005         event->param.conn.flow_control = rep_data->flow_control;
1006         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1007         event->param.conn.srq = rep_data->srq;
1008         event->param.conn.qp_num = rep_data->remote_qpn;
1009 }
1010
1011 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1012 {
1013         struct rdma_id_private *id_priv = cm_id->context;
1014         struct rdma_cm_event event;
1015         int ret = 0;
1016
1017         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1018                 cma_disable_callback(id_priv, CMA_CONNECT)) ||
1019             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1020                 cma_disable_callback(id_priv, CMA_DISCONNECT)))
1021                 return 0;
1022
1023         memset(&event, 0, sizeof event);
1024         switch (ib_event->event) {
1025         case IB_CM_REQ_ERROR:
1026         case IB_CM_REP_ERROR:
1027                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1028                 event.status = -ETIMEDOUT;
1029                 break;
1030         case IB_CM_REP_RECEIVED:
1031                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
1032                 if (event.status)
1033                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1034                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
1035                         event.status = cma_rep_recv(id_priv);
1036                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1037                                                      RDMA_CM_EVENT_ESTABLISHED;
1038                 } else
1039                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1040                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1041                                        ib_event->private_data);
1042                 break;
1043         case IB_CM_RTU_RECEIVED:
1044         case IB_CM_USER_ESTABLISHED:
1045                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1046                 break;
1047         case IB_CM_DREQ_ERROR:
1048                 event.status = -ETIMEDOUT; /* fall through */
1049         case IB_CM_DREQ_RECEIVED:
1050         case IB_CM_DREP_RECEIVED:
1051                 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
1052                         goto out;
1053                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1054                 break;
1055         case IB_CM_TIMEWAIT_EXIT:
1056                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1057                 break;
1058         case IB_CM_MRA_RECEIVED:
1059                 /* ignore event */
1060                 goto out;
1061         case IB_CM_REJ_RECEIVED:
1062                 cma_modify_qp_err(id_priv);
1063                 event.status = ib_event->param.rej_rcvd.reason;
1064                 event.event = RDMA_CM_EVENT_REJECTED;
1065                 event.param.conn.private_data = ib_event->private_data;
1066                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1067                 break;
1068         default:
1069                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1070                        ib_event->event);
1071                 goto out;
1072         }
1073
1074         ret = id_priv->id.event_handler(&id_priv->id, &event);
1075         if (ret) {
1076                 /* Destroy the CM ID by returning a non-zero value. */
1077                 id_priv->cm_id.ib = NULL;
1078                 cma_exch(id_priv, CMA_DESTROYING);
1079                 mutex_unlock(&id_priv->handler_mutex);
1080                 rdma_destroy_id(&id_priv->id);
1081                 return ret;
1082         }
1083 out:
1084         mutex_unlock(&id_priv->handler_mutex);
1085         return ret;
1086 }
1087
1088 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1089                                                struct ib_cm_event *ib_event)
1090 {
1091         struct rdma_id_private *id_priv;
1092         struct rdma_cm_id *id;
1093         struct rdma_route *rt;
1094         union cma_ip_addr *src, *dst;
1095         __be16 port;
1096         u8 ip_ver;
1097         int ret;
1098
1099         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1100                              &ip_ver, &port, &src, &dst))
1101                 goto err;
1102
1103         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1104                             listen_id->ps);
1105         if (IS_ERR(id))
1106                 goto err;
1107
1108         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1109                           ip_ver, port, src, dst);
1110
1111         rt = &id->route;
1112         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1113         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1114                                GFP_KERNEL);
1115         if (!rt->path_rec)
1116                 goto destroy_id;
1117
1118         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1119         if (rt->num_paths == 2)
1120                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1121
1122         if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1123                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1124                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1125                 ib_addr_set_pkey(&rt->addr.dev_addr, rt->path_rec[0].pkey);
1126         } else {
1127                 ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1128                                         &rt->addr.dev_addr);
1129                 if (ret)
1130                         goto destroy_id;
1131         }
1132         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1133
1134         id_priv = container_of(id, struct rdma_id_private, id);
1135         id_priv->state = CMA_CONNECT;
1136         return id_priv;
1137
1138 destroy_id:
1139         rdma_destroy_id(id);
1140 err:
1141         return NULL;
1142 }
1143
1144 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1145                                               struct ib_cm_event *ib_event)
1146 {
1147         struct rdma_id_private *id_priv;
1148         struct rdma_cm_id *id;
1149         union cma_ip_addr *src, *dst;
1150         __be16 port;
1151         u8 ip_ver;
1152         int ret;
1153
1154         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1155                             listen_id->ps);
1156         if (IS_ERR(id))
1157                 return NULL;
1158
1159
1160         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1161                              &ip_ver, &port, &src, &dst))
1162                 goto err;
1163
1164         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1165                           ip_ver, port, src, dst);
1166
1167         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1168                 ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1169                                         &id->route.addr.dev_addr);
1170                 if (ret)
1171                         goto err;
1172         }
1173
1174         id_priv = container_of(id, struct rdma_id_private, id);
1175         id_priv->state = CMA_CONNECT;
1176         return id_priv;
1177 err:
1178         rdma_destroy_id(id);
1179         return NULL;
1180 }
1181
1182 static void cma_set_req_event_data(struct rdma_cm_event *event,
1183                                    struct ib_cm_req_event_param *req_data,
1184                                    void *private_data, int offset)
1185 {
1186         event->param.conn.private_data = private_data + offset;
1187         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1188         event->param.conn.responder_resources = req_data->responder_resources;
1189         event->param.conn.initiator_depth = req_data->initiator_depth;
1190         event->param.conn.flow_control = req_data->flow_control;
1191         event->param.conn.retry_count = req_data->retry_count;
1192         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1193         event->param.conn.srq = req_data->srq;
1194         event->param.conn.qp_num = req_data->remote_qpn;
1195 }
1196
1197 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1198 {
1199         struct rdma_id_private *listen_id, *conn_id;
1200         struct rdma_cm_event event;
1201         int offset, ret;
1202
1203         listen_id = cm_id->context;
1204         if (cma_disable_callback(listen_id, CMA_LISTEN))
1205                 return -ECONNABORTED;
1206
1207         memset(&event, 0, sizeof event);
1208         offset = cma_user_data_offset(listen_id->id.ps);
1209         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1210         if (cma_is_ud_ps(listen_id->id.ps)) {
1211                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1212                 event.param.ud.private_data = ib_event->private_data + offset;
1213                 event.param.ud.private_data_len =
1214                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1215         } else {
1216                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1217                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1218                                        ib_event->private_data, offset);
1219         }
1220         if (!conn_id) {
1221                 ret = -ENOMEM;
1222                 goto out;
1223         }
1224
1225         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1226         ret = cma_acquire_dev(conn_id);
1227         if (ret)
1228                 goto release_conn_id;
1229
1230         conn_id->cm_id.ib = cm_id;
1231         cm_id->context = conn_id;
1232         cm_id->cm_handler = cma_ib_handler;
1233
1234         /*
1235          * Protect against the user destroying conn_id from another thread
1236          * until we're done accessing it.
1237          */
1238         atomic_inc(&conn_id->refcount);
1239         ret = conn_id->id.event_handler(&conn_id->id, &event);
1240         if (!ret) {
1241                 /*
1242                  * Acquire mutex to prevent user executing rdma_destroy_id()
1243                  * while we're accessing the cm_id.
1244                  */
1245                 mutex_lock(&lock);
1246                 if (cma_comp(conn_id, CMA_CONNECT) &&
1247                     !cma_is_ud_ps(conn_id->id.ps))
1248                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1249                 mutex_unlock(&lock);
1250                 mutex_unlock(&conn_id->handler_mutex);
1251                 cma_deref_id(conn_id);
1252                 goto out;
1253         }
1254         cma_deref_id(conn_id);
1255
1256         /* Destroy the CM ID by returning a non-zero value. */
1257         conn_id->cm_id.ib = NULL;
1258
1259 release_conn_id:
1260         cma_exch(conn_id, CMA_DESTROYING);
1261         mutex_unlock(&conn_id->handler_mutex);
1262         rdma_destroy_id(&conn_id->id);
1263
1264 out:
1265         mutex_unlock(&listen_id->handler_mutex);
1266         return ret;
1267 }
1268
1269 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1270 {
1271         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1272 }
1273
1274 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1275                                  struct ib_cm_compare_data *compare)
1276 {
1277         struct cma_hdr *cma_data, *cma_mask;
1278         struct sdp_hh *sdp_data, *sdp_mask;
1279         __be32 ip4_addr;
1280         struct in6_addr ip6_addr;
1281
1282         memset(compare, 0, sizeof *compare);
1283         cma_data = (void *) compare->data;
1284         cma_mask = (void *) compare->mask;
1285         sdp_data = (void *) compare->data;
1286         sdp_mask = (void *) compare->mask;
1287
1288         switch (addr->sa_family) {
1289         case AF_INET:
1290                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1291                 if (ps == RDMA_PS_SDP) {
1292                         sdp_set_ip_ver(sdp_data, 4);
1293                         sdp_set_ip_ver(sdp_mask, 0xF);
1294                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1295                         sdp_mask->dst_addr.ip4.addr = htonl(~0);
1296                 } else {
1297                         cma_set_ip_ver(cma_data, 4);
1298                         cma_set_ip_ver(cma_mask, 0xF);
1299                         cma_data->dst_addr.ip4.addr = ip4_addr;
1300                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1301                 }
1302                 break;
1303         case AF_INET6:
1304                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1305                 if (ps == RDMA_PS_SDP) {
1306                         sdp_set_ip_ver(sdp_data, 6);
1307                         sdp_set_ip_ver(sdp_mask, 0xF);
1308                         sdp_data->dst_addr.ip6 = ip6_addr;
1309                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1310                                sizeof sdp_mask->dst_addr.ip6);
1311                 } else {
1312                         cma_set_ip_ver(cma_data, 6);
1313                         cma_set_ip_ver(cma_mask, 0xF);
1314                         cma_data->dst_addr.ip6 = ip6_addr;
1315                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1316                                sizeof cma_mask->dst_addr.ip6);
1317                 }
1318                 break;
1319         default:
1320                 break;
1321         }
1322 }
1323
1324 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1325 {
1326         struct rdma_id_private *id_priv = iw_id->context;
1327         struct rdma_cm_event event;
1328         struct sockaddr_in *sin;
1329         int ret = 0;
1330
1331         if (cma_disable_callback(id_priv, CMA_CONNECT))
1332                 return 0;
1333
1334         memset(&event, 0, sizeof event);
1335         switch (iw_event->event) {
1336         case IW_CM_EVENT_CLOSE:
1337                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1338                 break;
1339         case IW_CM_EVENT_CONNECT_REPLY:
1340                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1341                 *sin = iw_event->local_addr;
1342                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1343                 *sin = iw_event->remote_addr;
1344                 switch (iw_event->status) {
1345                 case 0:
1346                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1347                         break;
1348                 case -ECONNRESET:
1349                 case -ECONNREFUSED:
1350                         event.event = RDMA_CM_EVENT_REJECTED;
1351                         break;
1352                 case -ETIMEDOUT:
1353                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1354                         break;
1355                 default:
1356                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1357                         break;
1358                 }
1359                 break;
1360         case IW_CM_EVENT_ESTABLISHED:
1361                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1362                 break;
1363         default:
1364                 BUG_ON(1);
1365         }
1366
1367         event.status = iw_event->status;
1368         event.param.conn.private_data = iw_event->private_data;
1369         event.param.conn.private_data_len = iw_event->private_data_len;
1370         ret = id_priv->id.event_handler(&id_priv->id, &event);
1371         if (ret) {
1372                 /* Destroy the CM ID by returning a non-zero value. */
1373                 id_priv->cm_id.iw = NULL;
1374                 cma_exch(id_priv, CMA_DESTROYING);
1375                 mutex_unlock(&id_priv->handler_mutex);
1376                 rdma_destroy_id(&id_priv->id);
1377                 return ret;
1378         }
1379
1380         mutex_unlock(&id_priv->handler_mutex);
1381         return ret;
1382 }
1383
1384 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1385                                struct iw_cm_event *iw_event)
1386 {
1387         struct rdma_cm_id *new_cm_id;
1388         struct rdma_id_private *listen_id, *conn_id;
1389         struct sockaddr_in *sin;
1390         struct net_device *dev = NULL;
1391         struct rdma_cm_event event;
1392         int ret;
1393         struct ib_device_attr attr;
1394
1395         listen_id = cm_id->context;
1396         if (cma_disable_callback(listen_id, CMA_LISTEN))
1397                 return -ECONNABORTED;
1398
1399         /* Create a new RDMA id for the new IW CM ID */
1400         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1401                                    listen_id->id.context,
1402                                    RDMA_PS_TCP);
1403         if (IS_ERR(new_cm_id)) {
1404                 ret = -ENOMEM;
1405                 goto out;
1406         }
1407         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1408         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1409         conn_id->state = CMA_CONNECT;
1410
1411         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1412         if (!dev) {
1413                 ret = -EADDRNOTAVAIL;
1414                 mutex_unlock(&conn_id->handler_mutex);
1415                 rdma_destroy_id(new_cm_id);
1416                 goto out;
1417         }
1418         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1419         if (ret) {
1420                 mutex_unlock(&conn_id->handler_mutex);
1421                 rdma_destroy_id(new_cm_id);
1422                 goto out;
1423         }
1424
1425         ret = cma_acquire_dev(conn_id);
1426         if (ret) {
1427                 mutex_unlock(&conn_id->handler_mutex);
1428                 rdma_destroy_id(new_cm_id);
1429                 goto out;
1430         }
1431
1432         conn_id->cm_id.iw = cm_id;
1433         cm_id->context = conn_id;
1434         cm_id->cm_handler = cma_iw_handler;
1435
1436         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1437         *sin = iw_event->local_addr;
1438         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1439         *sin = iw_event->remote_addr;
1440
1441         ret = ib_query_device(conn_id->id.device, &attr);
1442         if (ret) {
1443                 mutex_unlock(&conn_id->handler_mutex);
1444                 rdma_destroy_id(new_cm_id);
1445                 goto out;
1446         }
1447
1448         memset(&event, 0, sizeof event);
1449         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1450         event.param.conn.private_data = iw_event->private_data;
1451         event.param.conn.private_data_len = iw_event->private_data_len;
1452         event.param.conn.initiator_depth = attr.max_qp_init_rd_atom;
1453         event.param.conn.responder_resources = attr.max_qp_rd_atom;
1454
1455         /*
1456          * Protect against the user destroying conn_id from another thread
1457          * until we're done accessing it.
1458          */
1459         atomic_inc(&conn_id->refcount);
1460         ret = conn_id->id.event_handler(&conn_id->id, &event);
1461         if (ret) {
1462                 /* User wants to destroy the CM ID */
1463                 conn_id->cm_id.iw = NULL;
1464                 cma_exch(conn_id, CMA_DESTROYING);
1465                 mutex_unlock(&conn_id->handler_mutex);
1466                 cma_deref_id(conn_id);
1467                 rdma_destroy_id(&conn_id->id);
1468                 goto out;
1469         }
1470
1471         mutex_unlock(&conn_id->handler_mutex);
1472         cma_deref_id(conn_id);
1473
1474 out:
1475         if (dev)
1476                 dev_put(dev);
1477         mutex_unlock(&listen_id->handler_mutex);
1478         return ret;
1479 }
1480
1481 static int cma_ib_listen(struct rdma_id_private *id_priv)
1482 {
1483         struct ib_cm_compare_data compare_data;
1484         struct sockaddr *addr;
1485         __be64 svc_id;
1486         int ret;
1487
1488         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1489                                             id_priv);
1490         if (IS_ERR(id_priv->cm_id.ib))
1491                 return PTR_ERR(id_priv->cm_id.ib);
1492
1493         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1494         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1495         if (cma_any_addr(addr))
1496                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1497         else {
1498                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1499                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1500         }
1501
1502         if (ret) {
1503                 ib_destroy_cm_id(id_priv->cm_id.ib);
1504                 id_priv->cm_id.ib = NULL;
1505         }
1506
1507         return ret;
1508 }
1509
1510 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1511 {
1512         int ret;
1513         struct sockaddr_in *sin;
1514
1515         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1516                                             iw_conn_req_handler,
1517                                             id_priv);
1518         if (IS_ERR(id_priv->cm_id.iw))
1519                 return PTR_ERR(id_priv->cm_id.iw);
1520
1521         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1522         id_priv->cm_id.iw->local_addr = *sin;
1523
1524         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1525
1526         if (ret) {
1527                 iw_destroy_cm_id(id_priv->cm_id.iw);
1528                 id_priv->cm_id.iw = NULL;
1529         }
1530
1531         return ret;
1532 }
1533
1534 static int cma_listen_handler(struct rdma_cm_id *id,
1535                               struct rdma_cm_event *event)
1536 {
1537         struct rdma_id_private *id_priv = id->context;
1538
1539         id->context = id_priv->id.context;
1540         id->event_handler = id_priv->id.event_handler;
1541         return id_priv->id.event_handler(id, event);
1542 }
1543
1544 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1545                               struct cma_device *cma_dev)
1546 {
1547         struct rdma_id_private *dev_id_priv;
1548         struct rdma_cm_id *id;
1549         int ret;
1550
1551         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1552         if (IS_ERR(id))
1553                 return;
1554
1555         dev_id_priv = container_of(id, struct rdma_id_private, id);
1556
1557         dev_id_priv->state = CMA_ADDR_BOUND;
1558         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1559                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1560
1561         cma_attach_to_dev(dev_id_priv, cma_dev);
1562         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1563         atomic_inc(&id_priv->refcount);
1564         dev_id_priv->internal_id = 1;
1565
1566         ret = rdma_listen(id, id_priv->backlog);
1567         if (ret)
1568                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1569                        "listening on device %s\n", ret, cma_dev->device->name);
1570 }
1571
1572 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1573 {
1574         struct cma_device *cma_dev;
1575
1576         mutex_lock(&lock);
1577         list_add_tail(&id_priv->list, &listen_any_list);
1578         list_for_each_entry(cma_dev, &dev_list, list)
1579                 cma_listen_on_dev(id_priv, cma_dev);
1580         mutex_unlock(&lock);
1581 }
1582
1583 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1584 {
1585         struct rdma_id_private *id_priv;
1586
1587         id_priv = container_of(id, struct rdma_id_private, id);
1588         id_priv->tos = (u8) tos;
1589 }
1590 EXPORT_SYMBOL(rdma_set_service_type);
1591
1592 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1593                               void *context)
1594 {
1595         struct cma_work *work = context;
1596         struct rdma_route *route;
1597
1598         route = &work->id->id.route;
1599
1600         if (!status) {
1601                 route->num_paths = 1;
1602                 *route->path_rec = *path_rec;
1603         } else {
1604                 work->old_state = CMA_ROUTE_QUERY;
1605                 work->new_state = CMA_ADDR_RESOLVED;
1606                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1607                 work->event.status = status;
1608         }
1609
1610         queue_work(cma_wq, &work->work);
1611 }
1612
1613 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1614                               struct cma_work *work)
1615 {
1616         struct rdma_addr *addr = &id_priv->id.route.addr;
1617         struct ib_sa_path_rec path_rec;
1618         ib_sa_comp_mask comp_mask;
1619         struct sockaddr_in6 *sin6;
1620
1621         memset(&path_rec, 0, sizeof path_rec);
1622         rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1623         rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1624         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1625         path_rec.numb_path = 1;
1626         path_rec.reversible = 1;
1627         path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1628                                                         (struct sockaddr *) &addr->dst_addr);
1629
1630         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1631                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1632                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1633
1634         if (addr->src_addr.ss_family == AF_INET) {
1635                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1636                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1637         } else {
1638                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1639                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1640                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1641         }
1642
1643         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1644                                                id_priv->id.port_num, &path_rec,
1645                                                comp_mask, timeout_ms,
1646                                                GFP_KERNEL, cma_query_handler,
1647                                                work, &id_priv->query);
1648
1649         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1650 }
1651
1652 static void cma_work_handler(struct work_struct *_work)
1653 {
1654         struct cma_work *work = container_of(_work, struct cma_work, work);
1655         struct rdma_id_private *id_priv = work->id;
1656         int destroy = 0;
1657
1658         mutex_lock(&id_priv->handler_mutex);
1659         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1660                 goto out;
1661
1662         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1663                 cma_exch(id_priv, CMA_DESTROYING);
1664                 destroy = 1;
1665         }
1666 out:
1667         mutex_unlock(&id_priv->handler_mutex);
1668         cma_deref_id(id_priv);
1669         if (destroy)
1670                 rdma_destroy_id(&id_priv->id);
1671         kfree(work);
1672 }
1673
1674 static void cma_ndev_work_handler(struct work_struct *_work)
1675 {
1676         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1677         struct rdma_id_private *id_priv = work->id;
1678         int destroy = 0;
1679
1680         mutex_lock(&id_priv->handler_mutex);
1681         if (id_priv->state == CMA_DESTROYING ||
1682             id_priv->state == CMA_DEVICE_REMOVAL)
1683                 goto out;
1684
1685         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1686                 cma_exch(id_priv, CMA_DESTROYING);
1687                 destroy = 1;
1688         }
1689
1690 out:
1691         mutex_unlock(&id_priv->handler_mutex);
1692         cma_deref_id(id_priv);
1693         if (destroy)
1694                 rdma_destroy_id(&id_priv->id);
1695         kfree(work);
1696 }
1697
1698 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1699 {
1700         struct rdma_route *route = &id_priv->id.route;
1701         struct cma_work *work;
1702         int ret;
1703
1704         work = kzalloc(sizeof *work, GFP_KERNEL);
1705         if (!work)
1706                 return -ENOMEM;
1707
1708         work->id = id_priv;
1709         INIT_WORK(&work->work, cma_work_handler);
1710         work->old_state = CMA_ROUTE_QUERY;
1711         work->new_state = CMA_ROUTE_RESOLVED;
1712         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1713
1714         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1715         if (!route->path_rec) {
1716                 ret = -ENOMEM;
1717                 goto err1;
1718         }
1719
1720         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1721         if (ret)
1722                 goto err2;
1723
1724         return 0;
1725 err2:
1726         kfree(route->path_rec);
1727         route->path_rec = NULL;
1728 err1:
1729         kfree(work);
1730         return ret;
1731 }
1732
1733 int rdma_set_ib_paths(struct rdma_cm_id *id,
1734                       struct ib_sa_path_rec *path_rec, int num_paths)
1735 {
1736         struct rdma_id_private *id_priv;
1737         int ret;
1738
1739         id_priv = container_of(id, struct rdma_id_private, id);
1740         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1741                 return -EINVAL;
1742
1743         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1744                                      GFP_KERNEL);
1745         if (!id->route.path_rec) {
1746                 ret = -ENOMEM;
1747                 goto err;
1748         }
1749
1750         id->route.num_paths = num_paths;
1751         return 0;
1752 err:
1753         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1754         return ret;
1755 }
1756 EXPORT_SYMBOL(rdma_set_ib_paths);
1757
1758 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1759 {
1760         struct cma_work *work;
1761
1762         work = kzalloc(sizeof *work, GFP_KERNEL);
1763         if (!work)
1764                 return -ENOMEM;
1765
1766         work->id = id_priv;
1767         INIT_WORK(&work->work, cma_work_handler);
1768         work->old_state = CMA_ROUTE_QUERY;
1769         work->new_state = CMA_ROUTE_RESOLVED;
1770         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1771         queue_work(cma_wq, &work->work);
1772         return 0;
1773 }
1774
1775 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1776 {
1777         struct rdma_route *route = &id_priv->id.route;
1778         struct rdma_addr *addr = &route->addr;
1779         struct cma_work *work;
1780         int ret;
1781         struct sockaddr_in *src_addr = (struct sockaddr_in *)&route->addr.src_addr;
1782         struct sockaddr_in *dst_addr = (struct sockaddr_in *)&route->addr.dst_addr;
1783         struct net_device *ndev = NULL;
1784         u16 vid;
1785
1786         if (src_addr->sin_family != dst_addr->sin_family)
1787                 return -EINVAL;
1788
1789         work = kzalloc(sizeof *work, GFP_KERNEL);
1790         if (!work)
1791                 return -ENOMEM;
1792
1793         work->id = id_priv;
1794         INIT_WORK(&work->work, cma_work_handler);
1795
1796         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1797         if (!route->path_rec) {
1798                 ret = -ENOMEM;
1799                 goto err1;
1800         }
1801
1802         route->num_paths = 1;
1803
1804         if (addr->dev_addr.bound_dev_if)
1805                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1806         if (!ndev) {
1807                 ret = -ENODEV;
1808                 goto err2;
1809         }
1810
1811         vid = rdma_vlan_dev_vlan_id(ndev);
1812
1813         iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1814         iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1815
1816         route->path_rec->hop_limit = 1;
1817         route->path_rec->reversible = 1;
1818         route->path_rec->pkey = cpu_to_be16(0xffff);
1819         route->path_rec->mtu_selector = IB_SA_EQ;
1820         route->path_rec->sl = id_priv->tos >> 5;
1821
1822         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1823         route->path_rec->rate_selector = IB_SA_EQ;
1824         route->path_rec->rate = iboe_get_rate(ndev);
1825         dev_put(ndev);
1826         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1827         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1828         if (!route->path_rec->mtu) {
1829                 ret = -EINVAL;
1830                 goto err2;
1831         }
1832
1833         work->old_state = CMA_ROUTE_QUERY;
1834         work->new_state = CMA_ROUTE_RESOLVED;
1835         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1836         work->event.status = 0;
1837
1838         queue_work(cma_wq, &work->work);
1839
1840         return 0;
1841
1842 err2:
1843         kfree(route->path_rec);
1844         route->path_rec = NULL;
1845 err1:
1846         kfree(work);
1847         return ret;
1848 }
1849
1850 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1851 {
1852         struct rdma_id_private *id_priv;
1853         int ret;
1854
1855         id_priv = container_of(id, struct rdma_id_private, id);
1856         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1857                 return -EINVAL;
1858
1859         atomic_inc(&id_priv->refcount);
1860         switch (rdma_node_get_transport(id->device->node_type)) {
1861         case RDMA_TRANSPORT_IB:
1862                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1863                 case IB_LINK_LAYER_INFINIBAND:
1864                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1865                         break;
1866                 case IB_LINK_LAYER_ETHERNET:
1867                         ret = cma_resolve_iboe_route(id_priv);
1868                         break;
1869                 default:
1870                         ret = -ENOSYS;
1871                 }
1872                 break;
1873         case RDMA_TRANSPORT_IWARP:
1874                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1875                 break;
1876         default:
1877                 ret = -ENOSYS;
1878                 break;
1879         }
1880         if (ret)
1881                 goto err;
1882
1883         return 0;
1884 err:
1885         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1886         cma_deref_id(id_priv);
1887         return ret;
1888 }
1889 EXPORT_SYMBOL(rdma_resolve_route);
1890
1891 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1892 {
1893         struct cma_device *cma_dev;
1894         struct ib_port_attr port_attr;
1895         union ib_gid gid;
1896         u16 pkey;
1897         int ret;
1898         u8 p;
1899
1900         mutex_lock(&lock);
1901         if (list_empty(&dev_list)) {
1902                 ret = -ENODEV;
1903                 goto out;
1904         }
1905         list_for_each_entry(cma_dev, &dev_list, list)
1906                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1907                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1908                             port_attr.state == IB_PORT_ACTIVE)
1909                                 goto port_found;
1910
1911         p = 1;
1912         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1913
1914 port_found:
1915         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1916         if (ret)
1917                 goto out;
1918
1919         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1920         if (ret)
1921                 goto out;
1922
1923         id_priv->id.route.addr.dev_addr.dev_type =
1924                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
1925                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1926
1927         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1928         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1929         id_priv->id.port_num = p;
1930         cma_attach_to_dev(id_priv, cma_dev);
1931 out:
1932         mutex_unlock(&lock);
1933         return ret;
1934 }
1935
1936 static void addr_handler(int status, struct sockaddr *src_addr,
1937                          struct rdma_dev_addr *dev_addr, void *context)
1938 {
1939         struct rdma_id_private *id_priv = context;
1940         struct rdma_cm_event event;
1941
1942         memset(&event, 0, sizeof event);
1943         mutex_lock(&id_priv->handler_mutex);
1944         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED))
1945                 goto out;
1946
1947         if (!status && !id_priv->cma_dev)
1948                 status = cma_acquire_dev(id_priv);
1949
1950         if (status) {
1951                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1952                         goto out;
1953                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1954                 event.status = status;
1955         } else {
1956                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1957                        ip_addr_size(src_addr));
1958                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1959         }
1960
1961         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1962                 cma_exch(id_priv, CMA_DESTROYING);
1963                 mutex_unlock(&id_priv->handler_mutex);
1964                 cma_deref_id(id_priv);
1965                 rdma_destroy_id(&id_priv->id);
1966                 return;
1967         }
1968 out:
1969         mutex_unlock(&id_priv->handler_mutex);
1970         cma_deref_id(id_priv);
1971 }
1972
1973 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1974 {
1975         struct cma_work *work;
1976         struct sockaddr *src, *dst;
1977         union ib_gid gid;
1978         int ret;
1979
1980         work = kzalloc(sizeof *work, GFP_KERNEL);
1981         if (!work)
1982                 return -ENOMEM;
1983
1984         if (!id_priv->cma_dev) {
1985                 ret = cma_bind_loopback(id_priv);
1986                 if (ret)
1987                         goto err;
1988         }
1989
1990         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1991         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1992
1993         src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1994         if (cma_zero_addr(src)) {
1995                 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
1996                 if ((src->sa_family = dst->sa_family) == AF_INET) {
1997                         ((struct sockaddr_in *) src)->sin_addr.s_addr =
1998                                 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
1999                 } else {
2000                         ipv6_addr_copy(&((struct sockaddr_in6 *) src)->sin6_addr,
2001                                        &((struct sockaddr_in6 *) dst)->sin6_addr);
2002                 }
2003         }
2004
2005         work->id = id_priv;
2006         INIT_WORK(&work->work, cma_work_handler);
2007         work->old_state = CMA_ADDR_QUERY;
2008         work->new_state = CMA_ADDR_RESOLVED;
2009         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2010         queue_work(cma_wq, &work->work);
2011         return 0;
2012 err:
2013         kfree(work);
2014         return ret;
2015 }
2016
2017 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2018                          struct sockaddr *dst_addr)
2019 {
2020         if (!src_addr || !src_addr->sa_family) {
2021                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2022                 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
2023                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2024                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2025                 }
2026         }
2027         return rdma_bind_addr(id, src_addr);
2028 }
2029
2030 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2031                       struct sockaddr *dst_addr, int timeout_ms)
2032 {
2033         struct rdma_id_private *id_priv;
2034         int ret;
2035
2036         id_priv = container_of(id, struct rdma_id_private, id);
2037         if (id_priv->state == CMA_IDLE) {
2038                 ret = cma_bind_addr(id, src_addr, dst_addr);
2039                 if (ret)
2040                         return ret;
2041         }
2042
2043         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
2044                 return -EINVAL;
2045
2046         atomic_inc(&id_priv->refcount);
2047         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2048         if (cma_any_addr(dst_addr))
2049                 ret = cma_resolve_loopback(id_priv);
2050         else
2051                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2052                                       dst_addr, &id->route.addr.dev_addr,
2053                                       timeout_ms, addr_handler, id_priv);
2054         if (ret)
2055                 goto err;
2056
2057         return 0;
2058 err:
2059         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
2060         cma_deref_id(id_priv);
2061         return ret;
2062 }
2063 EXPORT_SYMBOL(rdma_resolve_addr);
2064
2065 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2066 {
2067         struct rdma_id_private *id_priv;
2068         unsigned long flags;
2069         int ret;
2070
2071         id_priv = container_of(id, struct rdma_id_private, id);
2072         spin_lock_irqsave(&id_priv->lock, flags);
2073         if (id_priv->state == CMA_IDLE) {
2074                 id_priv->reuseaddr = reuse;
2075                 ret = 0;
2076         } else {
2077                 ret = -EINVAL;
2078         }
2079         spin_unlock_irqrestore(&id_priv->lock, flags);
2080         return ret;
2081 }
2082 EXPORT_SYMBOL(rdma_set_reuseaddr);
2083
2084 static void cma_bind_port(struct rdma_bind_list *bind_list,
2085                           struct rdma_id_private *id_priv)
2086 {
2087         struct sockaddr_in *sin;
2088
2089         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2090         sin->sin_port = htons(bind_list->port);
2091         id_priv->bind_list = bind_list;
2092         hlist_add_head(&id_priv->node, &bind_list->owners);
2093 }
2094
2095 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2096                           unsigned short snum)
2097 {
2098         struct rdma_bind_list *bind_list;
2099         int port, ret;
2100
2101         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2102         if (!bind_list)
2103                 return -ENOMEM;
2104
2105         do {
2106                 ret = idr_get_new_above(ps, bind_list, snum, &port);
2107         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2108
2109         if (ret)
2110                 goto err1;
2111
2112         if (port != snum) {
2113                 ret = -EADDRNOTAVAIL;
2114                 goto err2;
2115         }
2116
2117         bind_list->ps = ps;
2118         bind_list->port = (unsigned short) port;
2119         cma_bind_port(bind_list, id_priv);
2120         return 0;
2121 err2:
2122         idr_remove(ps, port);
2123 err1:
2124         kfree(bind_list);
2125         return ret;
2126 }
2127
2128 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2129 {
2130         static unsigned int last_used_port;
2131         int low, high, remaining;
2132         unsigned int rover;
2133
2134         inet_get_local_port_range(&low, &high);
2135         remaining = (high - low) + 1;
2136         rover = net_random() % remaining + low;
2137 retry:
2138         if (last_used_port != rover &&
2139             !idr_find(ps, (unsigned short) rover)) {
2140                 int ret = cma_alloc_port(ps, id_priv, rover);
2141                 /*
2142                  * Remember previously used port number in order to avoid
2143                  * re-using same port immediately after it is closed.
2144                  */
2145                 if (!ret)
2146                         last_used_port = rover;
2147                 if (ret != -EADDRNOTAVAIL)
2148                         return ret;
2149         }
2150         if (--remaining) {
2151                 rover++;
2152                 if ((rover < low) || (rover > high))
2153                         rover = low;
2154                 goto retry;
2155         }
2156         return -EADDRNOTAVAIL;
2157 }
2158
2159 /*
2160  * Check that the requested port is available.  This is called when trying to
2161  * bind to a specific port, or when trying to listen on a bound port.  In
2162  * the latter case, the provided id_priv may already be on the bind_list, but
2163  * we still need to check that it's okay to start listening.
2164  */
2165 static int cma_check_port(struct rdma_bind_list *bind_list,
2166                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2167 {
2168         struct rdma_id_private *cur_id;
2169         struct sockaddr *addr, *cur_addr;
2170         struct hlist_node *node;
2171
2172         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2173         if (cma_any_addr(addr) && !reuseaddr)
2174                 return -EADDRNOTAVAIL;
2175
2176         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2177                 if (id_priv == cur_id)
2178                         continue;
2179
2180                 if ((cur_id->state == CMA_LISTEN) ||
2181                     !reuseaddr || !cur_id->reuseaddr) {
2182                         cur_addr = (struct sockaddr *) &cur_id->id.route.addr.src_addr;
2183                         if (cma_any_addr(cur_addr))
2184                                 return -EADDRNOTAVAIL;
2185
2186                         if (!cma_addr_cmp(addr, cur_addr))
2187                                 return -EADDRINUSE;
2188                 }
2189         }
2190         return 0;
2191 }
2192
2193 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2194 {
2195         struct rdma_bind_list *bind_list;
2196         unsigned short snum;
2197         int ret;
2198
2199         snum = ntohs(cma_port((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2200         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2201                 return -EACCES;
2202
2203         bind_list = idr_find(ps, snum);
2204         if (!bind_list) {
2205                 ret = cma_alloc_port(ps, id_priv, snum);
2206         } else {
2207                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2208                 if (!ret)
2209                         cma_bind_port(bind_list, id_priv);
2210         }
2211         return ret;
2212 }
2213
2214 static int cma_bind_listen(struct rdma_id_private *id_priv)
2215 {
2216         struct rdma_bind_list *bind_list = id_priv->bind_list;
2217         int ret = 0;
2218
2219         mutex_lock(&lock);
2220         if (bind_list->owners.first->next)
2221                 ret = cma_check_port(bind_list, id_priv, 0);
2222         mutex_unlock(&lock);
2223         return ret;
2224 }
2225
2226 static int cma_get_port(struct rdma_id_private *id_priv)
2227 {
2228         struct idr *ps;
2229         int ret;
2230
2231         switch (id_priv->id.ps) {
2232         case RDMA_PS_SDP:
2233                 ps = &sdp_ps;
2234                 break;
2235         case RDMA_PS_TCP:
2236                 ps = &tcp_ps;
2237                 break;
2238         case RDMA_PS_UDP:
2239                 ps = &udp_ps;
2240                 break;
2241         case RDMA_PS_IPOIB:
2242                 ps = &ipoib_ps;
2243                 break;
2244         default:
2245                 return -EPROTONOSUPPORT;
2246         }
2247
2248         mutex_lock(&lock);
2249         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2250                 ret = cma_alloc_any_port(ps, id_priv);
2251         else
2252                 ret = cma_use_port(ps, id_priv);
2253         mutex_unlock(&lock);
2254
2255         return ret;
2256 }
2257
2258 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2259                                struct sockaddr *addr)
2260 {
2261 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2262         struct sockaddr_in6 *sin6;
2263
2264         if (addr->sa_family != AF_INET6)
2265                 return 0;
2266
2267         sin6 = (struct sockaddr_in6 *) addr;
2268         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2269             !sin6->sin6_scope_id)
2270                         return -EINVAL;
2271
2272         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2273 #endif
2274         return 0;
2275 }
2276
2277 int rdma_listen(struct rdma_cm_id *id, int backlog)
2278 {
2279         struct rdma_id_private *id_priv;
2280         int ret;
2281
2282         id_priv = container_of(id, struct rdma_id_private, id);
2283         if (id_priv->state == CMA_IDLE) {
2284                 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
2285                 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
2286                 if (ret)
2287                         return ret;
2288         }
2289
2290         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
2291                 return -EINVAL;
2292
2293         if (id_priv->reuseaddr) {
2294                 ret = cma_bind_listen(id_priv);
2295                 if (ret)
2296                         goto err;
2297         }
2298
2299         id_priv->backlog = backlog;
2300         if (id->device) {
2301                 switch (rdma_node_get_transport(id->device->node_type)) {
2302                 case RDMA_TRANSPORT_IB:
2303                         ret = cma_ib_listen(id_priv);
2304                         if (ret)
2305                                 goto err;
2306                         break;
2307                 case RDMA_TRANSPORT_IWARP:
2308                         ret = cma_iw_listen(id_priv, backlog);
2309                         if (ret)
2310                                 goto err;
2311                         break;
2312                 default:
2313                         ret = -ENOSYS;
2314                         goto err;
2315                 }
2316         } else
2317                 cma_listen_on_all(id_priv);
2318
2319         return 0;
2320 err:
2321         id_priv->backlog = 0;
2322         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
2323         return ret;
2324 }
2325 EXPORT_SYMBOL(rdma_listen);
2326
2327 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2328 {
2329         struct rdma_id_private *id_priv;
2330         int ret;
2331
2332         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2333                 return -EAFNOSUPPORT;
2334
2335         id_priv = container_of(id, struct rdma_id_private, id);
2336         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
2337                 return -EINVAL;
2338
2339         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2340         if (ret)
2341                 goto err1;
2342
2343         if (!cma_any_addr(addr)) {
2344                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2345                 if (ret)
2346                         goto err1;
2347
2348                 ret = cma_acquire_dev(id_priv);
2349                 if (ret)
2350                         goto err1;
2351         }
2352
2353         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2354         ret = cma_get_port(id_priv);
2355         if (ret)
2356                 goto err2;
2357
2358         return 0;
2359 err2:
2360         if (id_priv->cma_dev)
2361                 cma_release_dev(id_priv);
2362 err1:
2363         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
2364         return ret;
2365 }
2366 EXPORT_SYMBOL(rdma_bind_addr);
2367
2368 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2369                           struct rdma_route *route)
2370 {
2371         struct cma_hdr *cma_hdr;
2372         struct sdp_hh *sdp_hdr;
2373
2374         if (route->addr.src_addr.ss_family == AF_INET) {
2375                 struct sockaddr_in *src4, *dst4;
2376
2377                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2378                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2379
2380                 switch (ps) {
2381                 case RDMA_PS_SDP:
2382                         sdp_hdr = hdr;
2383                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2384                                 return -EINVAL;
2385                         sdp_set_ip_ver(sdp_hdr, 4);
2386                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2387                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2388                         sdp_hdr->port = src4->sin_port;
2389                         break;
2390                 default:
2391                         cma_hdr = hdr;
2392                         cma_hdr->cma_version = CMA_VERSION;
2393                         cma_set_ip_ver(cma_hdr, 4);
2394                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2395                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2396                         cma_hdr->port = src4->sin_port;
2397                         break;
2398                 }
2399         } else {
2400                 struct sockaddr_in6 *src6, *dst6;
2401
2402                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2403                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2404
2405                 switch (ps) {
2406                 case RDMA_PS_SDP:
2407                         sdp_hdr = hdr;
2408                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2409                                 return -EINVAL;
2410                         sdp_set_ip_ver(sdp_hdr, 6);
2411                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2412                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2413                         sdp_hdr->port = src6->sin6_port;
2414                         break;
2415                 default:
2416                         cma_hdr = hdr;
2417                         cma_hdr->cma_version = CMA_VERSION;
2418                         cma_set_ip_ver(cma_hdr, 6);
2419                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2420                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2421                         cma_hdr->port = src6->sin6_port;
2422                         break;
2423                 }
2424         }
2425         return 0;
2426 }
2427
2428 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2429                                 struct ib_cm_event *ib_event)
2430 {
2431         struct rdma_id_private *id_priv = cm_id->context;
2432         struct rdma_cm_event event;
2433         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2434         int ret = 0;
2435
2436         if (cma_disable_callback(id_priv, CMA_CONNECT))
2437                 return 0;
2438
2439         memset(&event, 0, sizeof event);
2440         switch (ib_event->event) {
2441         case IB_CM_SIDR_REQ_ERROR:
2442                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2443                 event.status = -ETIMEDOUT;
2444                 break;
2445         case IB_CM_SIDR_REP_RECEIVED:
2446                 event.param.ud.private_data = ib_event->private_data;
2447                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2448                 if (rep->status != IB_SIDR_SUCCESS) {
2449                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2450                         event.status = ib_event->param.sidr_rep_rcvd.status;
2451                         break;
2452                 }
2453                 ret = cma_set_qkey(id_priv);
2454                 if (ret) {
2455                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2456                         event.status = -EINVAL;
2457                         break;
2458                 }
2459                 if (id_priv->qkey != rep->qkey) {
2460                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2461                         event.status = -EINVAL;
2462                         break;
2463                 }
2464                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2465                                      id_priv->id.route.path_rec,
2466                                      &event.param.ud.ah_attr);
2467                 event.param.ud.qp_num = rep->qpn;
2468                 event.param.ud.qkey = rep->qkey;
2469                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2470                 event.status = 0;
2471                 break;
2472         default:
2473                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2474                        ib_event->event);
2475                 goto out;
2476         }
2477
2478         ret = id_priv->id.event_handler(&id_priv->id, &event);
2479         if (ret) {
2480                 /* Destroy the CM ID by returning a non-zero value. */
2481                 id_priv->cm_id.ib = NULL;
2482                 cma_exch(id_priv, CMA_DESTROYING);
2483                 mutex_unlock(&id_priv->handler_mutex);
2484                 rdma_destroy_id(&id_priv->id);
2485                 return ret;
2486         }
2487 out:
2488         mutex_unlock(&id_priv->handler_mutex);
2489         return ret;
2490 }
2491
2492 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2493                               struct rdma_conn_param *conn_param)
2494 {
2495         struct ib_cm_sidr_req_param req;
2496         struct rdma_route *route;
2497         int ret;
2498
2499         req.private_data_len = sizeof(struct cma_hdr) +
2500                                conn_param->private_data_len;
2501         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2502         if (!req.private_data)
2503                 return -ENOMEM;
2504
2505         if (conn_param->private_data && conn_param->private_data_len)
2506                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2507                        conn_param->private_data, conn_param->private_data_len);
2508
2509         route = &id_priv->id.route;
2510         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2511         if (ret)
2512                 goto out;
2513
2514         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2515                                             cma_sidr_rep_handler, id_priv);
2516         if (IS_ERR(id_priv->cm_id.ib)) {
2517                 ret = PTR_ERR(id_priv->cm_id.ib);
2518                 goto out;
2519         }
2520
2521         req.path = route->path_rec;
2522         req.service_id = cma_get_service_id(id_priv->id.ps,
2523                                             (struct sockaddr *) &route->addr.dst_addr);
2524         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2525         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2526
2527         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2528         if (ret) {
2529                 ib_destroy_cm_id(id_priv->cm_id.ib);
2530                 id_priv->cm_id.ib = NULL;
2531         }
2532 out:
2533         kfree(req.private_data);
2534         return ret;
2535 }
2536
2537 static int cma_connect_ib(struct rdma_id_private *id_priv,
2538                           struct rdma_conn_param *conn_param)
2539 {
2540         struct ib_cm_req_param req;
2541         struct rdma_route *route;
2542         void *private_data;
2543         int offset, ret;
2544
2545         memset(&req, 0, sizeof req);
2546         offset = cma_user_data_offset(id_priv->id.ps);
2547         req.private_data_len = offset + conn_param->private_data_len;
2548         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2549         if (!private_data)
2550                 return -ENOMEM;
2551
2552         if (conn_param->private_data && conn_param->private_data_len)
2553                 memcpy(private_data + offset, conn_param->private_data,
2554                        conn_param->private_data_len);
2555
2556         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2557                                             id_priv);
2558         if (IS_ERR(id_priv->cm_id.ib)) {
2559                 ret = PTR_ERR(id_priv->cm_id.ib);
2560                 goto out;
2561         }
2562
2563         route = &id_priv->id.route;
2564         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2565         if (ret)
2566                 goto out;
2567         req.private_data = private_data;
2568
2569         req.primary_path = &route->path_rec[0];
2570         if (route->num_paths == 2)
2571                 req.alternate_path = &route->path_rec[1];
2572
2573         req.service_id = cma_get_service_id(id_priv->id.ps,
2574                                             (struct sockaddr *) &route->addr.dst_addr);
2575         req.qp_num = id_priv->qp_num;
2576         req.qp_type = IB_QPT_RC;
2577         req.starting_psn = id_priv->seq_num;
2578         req.responder_resources = conn_param->responder_resources;
2579         req.initiator_depth = conn_param->initiator_depth;
2580         req.flow_control = conn_param->flow_control;
2581         req.retry_count = conn_param->retry_count;
2582         req.rnr_retry_count = conn_param->rnr_retry_count;
2583         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2584         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2585         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2586         req.srq = id_priv->srq ? 1 : 0;
2587
2588         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2589 out:
2590         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2591                 ib_destroy_cm_id(id_priv->cm_id.ib);
2592                 id_priv->cm_id.ib = NULL;
2593         }
2594
2595         kfree(private_data);
2596         return ret;
2597 }
2598
2599 static int cma_connect_iw(struct rdma_id_private *id_priv,
2600                           struct rdma_conn_param *conn_param)
2601 {
2602         struct iw_cm_id *cm_id;
2603         struct sockaddr_in* sin;
2604         int ret;
2605         struct iw_cm_conn_param iw_param;
2606
2607         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2608         if (IS_ERR(cm_id)) {
2609                 ret = PTR_ERR(cm_id);
2610                 goto out;
2611         }
2612
2613         id_priv->cm_id.iw = cm_id;
2614
2615         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2616         cm_id->local_addr = *sin;
2617
2618         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2619         cm_id->remote_addr = *sin;
2620
2621         ret = cma_modify_qp_rtr(id_priv, conn_param);
2622         if (ret)
2623                 goto out;
2624
2625         iw_param.ord = conn_param->initiator_depth;
2626         iw_param.ird = conn_param->responder_resources;
2627         iw_param.private_data = conn_param->private_data;
2628         iw_param.private_data_len = conn_param->private_data_len;
2629         if (id_priv->id.qp)
2630                 iw_param.qpn = id_priv->qp_num;
2631         else
2632                 iw_param.qpn = conn_param->qp_num;
2633         ret = iw_cm_connect(cm_id, &iw_param);
2634 out:
2635         if (ret && !IS_ERR(cm_id)) {
2636                 iw_destroy_cm_id(cm_id);
2637                 id_priv->cm_id.iw = NULL;
2638         }
2639         return ret;
2640 }
2641
2642 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2643 {
2644         struct rdma_id_private *id_priv;
2645         int ret;
2646
2647         id_priv = container_of(id, struct rdma_id_private, id);
2648         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2649                 return -EINVAL;
2650
2651         if (!id->qp) {
2652                 id_priv->qp_num = conn_param->qp_num;
2653                 id_priv->srq = conn_param->srq;
2654         }
2655
2656         switch (rdma_node_get_transport(id->device->node_type)) {
2657         case RDMA_TRANSPORT_IB:
2658                 if (cma_is_ud_ps(id->ps))
2659                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2660                 else
2661                         ret = cma_connect_ib(id_priv, conn_param);
2662                 break;
2663         case RDMA_TRANSPORT_IWARP:
2664                 ret = cma_connect_iw(id_priv, conn_param);
2665                 break;
2666         default:
2667                 ret = -ENOSYS;
2668                 break;
2669         }
2670         if (ret)
2671                 goto err;
2672
2673         return 0;
2674 err:
2675         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2676         return ret;
2677 }
2678 EXPORT_SYMBOL(rdma_connect);
2679
2680 static int cma_accept_ib(struct rdma_id_private *id_priv,
2681                          struct rdma_conn_param *conn_param)
2682 {
2683         struct ib_cm_rep_param rep;
2684         int ret;
2685
2686         ret = cma_modify_qp_rtr(id_priv, conn_param);
2687         if (ret)
2688                 goto out;
2689
2690         ret = cma_modify_qp_rts(id_priv, conn_param);
2691         if (ret)
2692                 goto out;
2693
2694         memset(&rep, 0, sizeof rep);
2695         rep.qp_num = id_priv->qp_num;
2696         rep.starting_psn = id_priv->seq_num;
2697         rep.private_data = conn_param->private_data;
2698         rep.private_data_len = conn_param->private_data_len;
2699         rep.responder_resources = conn_param->responder_resources;
2700         rep.initiator_depth = conn_param->initiator_depth;
2701         rep.failover_accepted = 0;
2702         rep.flow_control = conn_param->flow_control;
2703         rep.rnr_retry_count = conn_param->rnr_retry_count;
2704         rep.srq = id_priv->srq ? 1 : 0;
2705
2706         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2707 out:
2708         return ret;
2709 }
2710
2711 static int cma_accept_iw(struct rdma_id_private *id_priv,
2712                   struct rdma_conn_param *conn_param)
2713 {
2714         struct iw_cm_conn_param iw_param;
2715         int ret;
2716
2717         ret = cma_modify_qp_rtr(id_priv, conn_param);
2718         if (ret)
2719                 return ret;
2720
2721         iw_param.ord = conn_param->initiator_depth;
2722         iw_param.ird = conn_param->responder_resources;
2723         iw_param.private_data = conn_param->private_data;
2724         iw_param.private_data_len = conn_param->private_data_len;
2725         if (id_priv->id.qp) {
2726                 iw_param.qpn = id_priv->qp_num;
2727         } else
2728                 iw_param.qpn = conn_param->qp_num;
2729
2730         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2731 }
2732
2733 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2734                              enum ib_cm_sidr_status status,
2735                              const void *private_data, int private_data_len)
2736 {
2737         struct ib_cm_sidr_rep_param rep;
2738         int ret;
2739
2740         memset(&rep, 0, sizeof rep);
2741         rep.status = status;
2742         if (status == IB_SIDR_SUCCESS) {
2743                 ret = cma_set_qkey(id_priv);
2744                 if (ret)
2745                         return ret;
2746                 rep.qp_num = id_priv->qp_num;
2747                 rep.qkey = id_priv->qkey;
2748         }
2749         rep.private_data = private_data;
2750         rep.private_data_len = private_data_len;
2751
2752         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2753 }
2754
2755 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2756 {
2757         struct rdma_id_private *id_priv;
2758         int ret;
2759
2760         id_priv = container_of(id, struct rdma_id_private, id);
2761         if (!cma_comp(id_priv, CMA_CONNECT))
2762                 return -EINVAL;
2763
2764         if (!id->qp && conn_param) {
2765                 id_priv->qp_num = conn_param->qp_num;
2766                 id_priv->srq = conn_param->srq;
2767         }
2768
2769         switch (rdma_node_get_transport(id->device->node_type)) {
2770         case RDMA_TRANSPORT_IB:
2771                 if (cma_is_ud_ps(id->ps))
2772                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2773                                                 conn_param->private_data,
2774                                                 conn_param->private_data_len);
2775                 else if (conn_param)
2776                         ret = cma_accept_ib(id_priv, conn_param);
2777                 else
2778                         ret = cma_rep_recv(id_priv);
2779                 break;
2780         case RDMA_TRANSPORT_IWARP:
2781                 ret = cma_accept_iw(id_priv, conn_param);
2782                 break;
2783         default:
2784                 ret = -ENOSYS;
2785                 break;
2786         }
2787
2788         if (ret)
2789                 goto reject;
2790
2791         return 0;
2792 reject:
2793         cma_modify_qp_err(id_priv);
2794         rdma_reject(id, NULL, 0);
2795         return ret;
2796 }
2797 EXPORT_SYMBOL(rdma_accept);
2798
2799 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2800 {
2801         struct rdma_id_private *id_priv;
2802         int ret;
2803
2804         id_priv = container_of(id, struct rdma_id_private, id);
2805         if (!cma_has_cm_dev(id_priv))
2806                 return -EINVAL;
2807
2808         switch (id->device->node_type) {
2809         case RDMA_NODE_IB_CA:
2810                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2811                 break;
2812         default:
2813                 ret = 0;
2814                 break;
2815         }
2816         return ret;
2817 }
2818 EXPORT_SYMBOL(rdma_notify);
2819
2820 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2821                 u8 private_data_len)
2822 {
2823         struct rdma_id_private *id_priv;
2824         int ret;
2825
2826         id_priv = container_of(id, struct rdma_id_private, id);
2827         if (!cma_has_cm_dev(id_priv))
2828                 return -EINVAL;
2829
2830         switch (rdma_node_get_transport(id->device->node_type)) {
2831         case RDMA_TRANSPORT_IB:
2832                 if (cma_is_ud_ps(id->ps))
2833                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2834                                                 private_data, private_data_len);
2835                 else
2836                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2837                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2838                                              0, private_data, private_data_len);
2839                 break;
2840         case RDMA_TRANSPORT_IWARP:
2841                 ret = iw_cm_reject(id_priv->cm_id.iw,
2842                                    private_data, private_data_len);
2843                 break;
2844         default:
2845                 ret = -ENOSYS;
2846                 break;
2847         }
2848         return ret;
2849 }
2850 EXPORT_SYMBOL(rdma_reject);
2851
2852 int rdma_disconnect(struct rdma_cm_id *id)
2853 {
2854         struct rdma_id_private *id_priv;
2855         int ret;
2856
2857         id_priv = container_of(id, struct rdma_id_private, id);
2858         if (!cma_has_cm_dev(id_priv))
2859                 return -EINVAL;
2860
2861         switch (rdma_node_get_transport(id->device->node_type)) {
2862         case RDMA_TRANSPORT_IB:
2863                 ret = cma_modify_qp_err(id_priv);
2864                 if (ret)
2865                         goto out;
2866                 /* Initiate or respond to a disconnect. */
2867                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2868                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2869                 break;
2870         case RDMA_TRANSPORT_IWARP:
2871                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2872                 break;
2873         default:
2874                 ret = -EINVAL;
2875                 break;
2876         }
2877 out:
2878         return ret;
2879 }
2880 EXPORT_SYMBOL(rdma_disconnect);
2881
2882 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2883 {
2884         struct rdma_id_private *id_priv;
2885         struct cma_multicast *mc = multicast->context;
2886         struct rdma_cm_event event;
2887         int ret;
2888
2889         id_priv = mc->id_priv;
2890         if (cma_disable_callback(id_priv, CMA_ADDR_BOUND) &&
2891             cma_disable_callback(id_priv, CMA_ADDR_RESOLVED))
2892                 return 0;
2893
2894         mutex_lock(&id_priv->qp_mutex);
2895         if (!status && id_priv->id.qp)
2896                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2897                                          multicast->rec.mlid);
2898         mutex_unlock(&id_priv->qp_mutex);
2899
2900         memset(&event, 0, sizeof event);
2901         event.status = status;
2902         event.param.ud.private_data = mc->context;
2903         if (!status) {
2904                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2905                 ib_init_ah_from_mcmember(id_priv->id.device,
2906                                          id_priv->id.port_num, &multicast->rec,
2907                                          &event.param.ud.ah_attr);
2908                 event.param.ud.qp_num = 0xFFFFFF;
2909                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2910         } else
2911                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2912
2913         ret = id_priv->id.event_handler(&id_priv->id, &event);
2914         if (ret) {
2915                 cma_exch(id_priv, CMA_DESTROYING);
2916                 mutex_unlock(&id_priv->handler_mutex);
2917                 rdma_destroy_id(&id_priv->id);
2918                 return 0;
2919         }
2920
2921         mutex_unlock(&id_priv->handler_mutex);
2922         return 0;
2923 }
2924
2925 static void cma_set_mgid(struct rdma_id_private *id_priv,
2926                          struct sockaddr *addr, union ib_gid *mgid)
2927 {
2928         unsigned char mc_map[MAX_ADDR_LEN];
2929         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2930         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2931         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2932
2933         if (cma_any_addr(addr)) {
2934                 memset(mgid, 0, sizeof *mgid);
2935         } else if ((addr->sa_family == AF_INET6) &&
2936                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2937                                                                  0xFF10A01B)) {
2938                 /* IPv6 address is an SA assigned MGID. */
2939                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2940         } else if ((addr->sa_family == AF_INET6)) {
2941                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2942                 if (id_priv->id.ps == RDMA_PS_UDP)
2943                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2944                 *mgid = *(union ib_gid *) (mc_map + 4);
2945         } else {
2946                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2947                 if (id_priv->id.ps == RDMA_PS_UDP)
2948                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2949                 *mgid = *(union ib_gid *) (mc_map + 4);
2950         }
2951 }
2952
2953 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2954                                  struct cma_multicast *mc)
2955 {
2956         struct ib_sa_mcmember_rec rec;
2957         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2958         ib_sa_comp_mask comp_mask;
2959         int ret;
2960
2961         ib_addr_get_mgid(dev_addr, &rec.mgid);
2962         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2963                                      &rec.mgid, &rec);
2964         if (ret)
2965                 return ret;
2966
2967         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
2968         if (id_priv->id.ps == RDMA_PS_UDP)
2969                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2970         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
2971         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2972         rec.join_state = 1;
2973
2974         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2975                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2976                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2977                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
2978                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2979
2980         if (id_priv->id.ps == RDMA_PS_IPOIB)
2981                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
2982                              IB_SA_MCMEMBER_REC_RATE_SELECTOR;
2983
2984         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2985                                                 id_priv->id.port_num, &rec,
2986                                                 comp_mask, GFP_KERNEL,
2987                                                 cma_ib_mc_handler, mc);
2988         if (IS_ERR(mc->multicast.ib))
2989                 return PTR_ERR(mc->multicast.ib);
2990
2991         return 0;
2992 }
2993
2994 static void iboe_mcast_work_handler(struct work_struct *work)
2995 {
2996         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
2997         struct cma_multicast *mc = mw->mc;
2998         struct ib_sa_multicast *m = mc->multicast.ib;
2999
3000         mc->multicast.ib->context = mc;
3001         cma_ib_mc_handler(0, m);
3002         kref_put(&mc->mcref, release_mc);
3003         kfree(mw);
3004 }
3005
3006 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3007 {
3008         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3009         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3010
3011         if (cma_any_addr(addr)) {
3012                 memset(mgid, 0, sizeof *mgid);
3013         } else if (addr->sa_family == AF_INET6) {
3014                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3015         } else {
3016                 mgid->raw[0] = 0xff;
3017                 mgid->raw[1] = 0x0e;
3018                 mgid->raw[2] = 0;
3019                 mgid->raw[3] = 0;
3020                 mgid->raw[4] = 0;
3021                 mgid->raw[5] = 0;
3022                 mgid->raw[6] = 0;
3023                 mgid->raw[7] = 0;
3024                 mgid->raw[8] = 0;
3025                 mgid->raw[9] = 0;
3026                 mgid->raw[10] = 0xff;
3027                 mgid->raw[11] = 0xff;
3028                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3029         }
3030 }
3031
3032 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3033                                    struct cma_multicast *mc)
3034 {
3035         struct iboe_mcast_work *work;
3036         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3037         int err;
3038         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3039         struct net_device *ndev = NULL;
3040
3041         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3042                 return -EINVAL;
3043
3044         work = kzalloc(sizeof *work, GFP_KERNEL);
3045         if (!work)
3046                 return -ENOMEM;
3047
3048         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3049         if (!mc->multicast.ib) {
3050                 err = -ENOMEM;
3051                 goto out1;
3052         }
3053
3054         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3055
3056         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3057         if (id_priv->id.ps == RDMA_PS_UDP)
3058                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3059
3060         if (dev_addr->bound_dev_if)
3061                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3062         if (!ndev) {
3063                 err = -ENODEV;
3064                 goto out2;
3065         }
3066         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3067         mc->multicast.ib->rec.hop_limit = 1;
3068         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3069         dev_put(ndev);
3070         if (!mc->multicast.ib->rec.mtu) {
3071                 err = -EINVAL;
3072                 goto out2;
3073         }
3074         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3075         work->id = id_priv;
3076         work->mc = mc;
3077         INIT_WORK(&work->work, iboe_mcast_work_handler);
3078         kref_get(&mc->mcref);
3079         queue_work(cma_wq, &work->work);
3080
3081         return 0;
3082
3083 out2:
3084         kfree(mc->multicast.ib);
3085 out1:
3086         kfree(work);
3087         return err;
3088 }
3089
3090 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3091                         void *context)
3092 {
3093         struct rdma_id_private *id_priv;
3094         struct cma_multicast *mc;
3095         int ret;
3096
3097         id_priv = container_of(id, struct rdma_id_private, id);
3098         if (!cma_comp(id_priv, CMA_ADDR_BOUND) &&
3099             !cma_comp(id_priv, CMA_ADDR_RESOLVED))
3100                 return -EINVAL;
3101
3102         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3103         if (!mc)
3104                 return -ENOMEM;
3105
3106         memcpy(&mc->addr, addr, ip_addr_size(addr));
3107         mc->context = context;
3108         mc->id_priv = id_priv;
3109
3110         spin_lock(&id_priv->lock);
3111         list_add(&mc->list, &id_priv->mc_list);
3112         spin_unlock(&id_priv->lock);
3113
3114         switch (rdma_node_get_transport(id->device->node_type)) {
3115         case RDMA_TRANSPORT_IB:
3116                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3117                 case IB_LINK_LAYER_INFINIBAND:
3118                         ret = cma_join_ib_multicast(id_priv, mc);
3119                         break;
3120                 case IB_LINK_LAYER_ETHERNET:
3121                         kref_init(&mc->mcref);
3122                         ret = cma_iboe_join_multicast(id_priv, mc);
3123                         break;
3124                 default:
3125                         ret = -EINVAL;
3126                 }
3127                 break;
3128         default:
3129                 ret = -ENOSYS;
3130                 break;
3131         }
3132
3133         if (ret) {
3134                 spin_lock_irq(&id_priv->lock);
3135                 list_del(&mc->list);
3136                 spin_unlock_irq(&id_priv->lock);
3137                 kfree(mc);
3138         }
3139         return ret;
3140 }
3141 EXPORT_SYMBOL(rdma_join_multicast);
3142
3143 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3144 {
3145         struct rdma_id_private *id_priv;
3146         struct cma_multicast *mc;
3147
3148         id_priv = container_of(id, struct rdma_id_private, id);
3149         spin_lock_irq(&id_priv->lock);
3150         list_for_each_entry(mc, &id_priv->mc_list, list) {
3151                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3152                         list_del(&mc->list);
3153                         spin_unlock_irq(&id_priv->lock);
3154
3155                         if (id->qp)
3156                                 ib_detach_mcast(id->qp,
3157                                                 &mc->multicast.ib->rec.mgid,
3158                                                 mc->multicast.ib->rec.mlid);
3159                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3160                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3161                                 case IB_LINK_LAYER_INFINIBAND:
3162                                         ib_sa_free_multicast(mc->multicast.ib);
3163                                         kfree(mc);
3164                                         break;
3165                                 case IB_LINK_LAYER_ETHERNET:
3166                                         kref_put(&mc->mcref, release_mc);
3167                                         break;
3168                                 default:
3169                                         break;
3170                                 }
3171                         }
3172                         return;
3173                 }
3174         }
3175         spin_unlock_irq(&id_priv->lock);
3176 }
3177 EXPORT_SYMBOL(rdma_leave_multicast);
3178
3179 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3180 {
3181         struct rdma_dev_addr *dev_addr;
3182         struct cma_ndev_work *work;
3183
3184         dev_addr = &id_priv->id.route.addr.dev_addr;
3185
3186         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3187             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3188                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3189                        ndev->name, &id_priv->id);
3190                 work = kzalloc(sizeof *work, GFP_KERNEL);
3191                 if (!work)
3192                         return -ENOMEM;
3193
3194                 INIT_WORK(&work->work, cma_ndev_work_handler);
3195                 work->id = id_priv;
3196                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3197                 atomic_inc(&id_priv->refcount);
3198                 queue_work(cma_wq, &work->work);
3199         }
3200
3201         return 0;
3202 }
3203
3204 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3205                                void *ctx)
3206 {
3207         struct net_device *ndev = (struct net_device *)ctx;
3208         struct cma_device *cma_dev;
3209         struct rdma_id_private *id_priv;
3210         int ret = NOTIFY_DONE;
3211
3212         if (dev_net(ndev) != &init_net)
3213                 return NOTIFY_DONE;
3214
3215         if (event != NETDEV_BONDING_FAILOVER)
3216                 return NOTIFY_DONE;
3217
3218         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3219                 return NOTIFY_DONE;
3220
3221         mutex_lock(&lock);
3222         list_for_each_entry(cma_dev, &dev_list, list)
3223                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3224                         ret = cma_netdev_change(ndev, id_priv);
3225                         if (ret)
3226                                 goto out;
3227                 }
3228
3229 out:
3230         mutex_unlock(&lock);
3231         return ret;
3232 }
3233
3234 static struct notifier_block cma_nb = {
3235         .notifier_call = cma_netdev_callback
3236 };
3237
3238 static void cma_add_one(struct ib_device *device)
3239 {
3240         struct cma_device *cma_dev;
3241         struct rdma_id_private *id_priv;
3242
3243         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3244         if (!cma_dev)
3245                 return;
3246
3247         cma_dev->device = device;
3248
3249         init_completion(&cma_dev->comp);
3250         atomic_set(&cma_dev->refcount, 1);
3251         INIT_LIST_HEAD(&cma_dev->id_list);
3252         ib_set_client_data(device, &cma_client, cma_dev);
3253
3254         mutex_lock(&lock);
3255         list_add_tail(&cma_dev->list, &dev_list);
3256         list_for_each_entry(id_priv, &listen_any_list, list)
3257                 cma_listen_on_dev(id_priv, cma_dev);
3258         mutex_unlock(&lock);
3259 }
3260
3261 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3262 {
3263         struct rdma_cm_event event;
3264         enum cma_state state;
3265         int ret = 0;
3266
3267         /* Record that we want to remove the device */
3268         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
3269         if (state == CMA_DESTROYING)
3270                 return 0;
3271
3272         cma_cancel_operation(id_priv, state);
3273         mutex_lock(&id_priv->handler_mutex);
3274
3275         /* Check for destruction from another callback. */
3276         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
3277                 goto out;
3278
3279         memset(&event, 0, sizeof event);
3280         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3281         ret = id_priv->id.event_handler(&id_priv->id, &event);
3282 out:
3283         mutex_unlock(&id_priv->handler_mutex);
3284         return ret;
3285 }
3286
3287 static void cma_process_remove(struct cma_device *cma_dev)
3288 {
3289         struct rdma_id_private *id_priv;
3290         int ret;
3291
3292         mutex_lock(&lock);
3293         while (!list_empty(&cma_dev->id_list)) {
3294                 id_priv = list_entry(cma_dev->id_list.next,
3295                                      struct rdma_id_private, list);
3296
3297                 list_del(&id_priv->listen_list);
3298                 list_del_init(&id_priv->list);
3299                 atomic_inc(&id_priv->refcount);
3300                 mutex_unlock(&lock);
3301
3302                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3303                 cma_deref_id(id_priv);
3304                 if (ret)
3305                         rdma_destroy_id(&id_priv->id);
3306
3307                 mutex_lock(&lock);
3308         }
3309         mutex_unlock(&lock);
3310
3311         cma_deref_dev(cma_dev);
3312         wait_for_completion(&cma_dev->comp);
3313 }
3314
3315 static void cma_remove_one(struct ib_device *device)
3316 {
3317         struct cma_device *cma_dev;
3318
3319         cma_dev = ib_get_client_data(device, &cma_client);
3320         if (!cma_dev)
3321                 return;
3322
3323         mutex_lock(&lock);
3324         list_del(&cma_dev->list);
3325         mutex_unlock(&lock);
3326
3327         cma_process_remove(cma_dev);
3328         kfree(cma_dev);
3329 }
3330
3331 static int __init cma_init(void)
3332 {
3333         int ret;
3334
3335         cma_wq = create_singlethread_workqueue("rdma_cm");
3336         if (!cma_wq)
3337                 return -ENOMEM;
3338
3339         ib_sa_register_client(&sa_client);
3340         rdma_addr_register_client(&addr_client);
3341         register_netdevice_notifier(&cma_nb);
3342
3343         ret = ib_register_client(&cma_client);
3344         if (ret)
3345                 goto err;
3346         return 0;
3347
3348 err:
3349         unregister_netdevice_notifier(&cma_nb);
3350         rdma_addr_unregister_client(&addr_client);
3351         ib_sa_unregister_client(&sa_client);
3352         destroy_workqueue(cma_wq);
3353         return ret;
3354 }
3355
3356 static void __exit cma_cleanup(void)
3357 {
3358         ib_unregister_client(&cma_client);
3359         unregister_netdevice_notifier(&cma_nb);
3360         rdma_addr_unregister_client(&addr_client);
3361         ib_sa_unregister_client(&sa_client);
3362         destroy_workqueue(cma_wq);
3363         idr_destroy(&sdp_ps);
3364         idr_destroy(&tcp_ps);
3365         idr_destroy(&udp_ps);
3366         idr_destroy(&ipoib_ps);
3367 }
3368
3369 module_init(cma_init);
3370 module_exit(cma_cleanup);