?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Committer: Michael Beasley <mike@snafu.setup>
[mikesnafu-overlay.git] / drivers / infiniband / core / cma.c
blobd81c156a22b41b85ce727bd585376bed2345e903
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 licensed under one of the following licenses:
8 *
9 * 1) under the terms of the "Common Public License 1.0" a copy of which is
10 * available from the Open Source Initiative, see
11 * http://www.opensource.org/licenses/cpl.php.
12 *
13 * 2) under the terms of the "The BSD License" a copy of which is
14 * available from the Open Source Initiative, see
15 * http://www.opensource.org/licenses/bsd-license.php.
16 *
17 * 3) under the terms of the "GNU General Public License (GPL) Version 2" a
18 * copy of which is available from the Open Source Initiative, see
19 * http://www.opensource.org/licenses/gpl-license.php.
20 *
21 * Licensee has the right to choose one of the above licenses.
22 *
23 * Redistributions of source code must retain the above copyright
24 * notice and one of the license notices.
25 *
26 * Redistributions in binary form must reproduce both the above copyright
27 * notice, one of the license notices in the documentation
28 * and/or other materials provided with the distribution.
29 *
30 */
31
32 #include <linux/completion.h>
33 #include <linux/in.h>
34 #include <linux/in6.h>
35 #include <linux/mutex.h>
36 #include <linux/random.h>
37 #include <linux/idr.h>
38 #include <linux/inetdevice.h>
39
40 #include <net/tcp.h>
41
42 #include <rdma/rdma_cm.h>
43 #include <rdma/rdma_cm_ib.h>
44 #include <rdma/ib_cache.h>
45 #include <rdma/ib_cm.h>
46 #include <rdma/ib_sa.h>
47 #include <rdma/iw_cm.h>
48
49 MODULE_AUTHOR("Sean Hefty");
50 MODULE_DESCRIPTION("Generic RDMA CM Agent");
51 MODULE_LICENSE("Dual BSD/GPL");
52
53 #define CMA_CM_RESPONSE_TIMEOUT 20
54 #define CMA_MAX_CM_RETRIES 15
55 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
56
57 static void cma_add_one(struct ib_device *device);
58 static void cma_remove_one(struct ib_device *device);
59
60 static struct ib_client cma_client = {
61 .name = "cma",
62 .add = cma_add_one,
63 .remove = cma_remove_one
64 };
65
66 static struct ib_sa_client sa_client;
67 static struct rdma_addr_client addr_client;
68 static LIST_HEAD(dev_list);
69 static LIST_HEAD(listen_any_list);
70 static DEFINE_MUTEX(lock);
71 static struct workqueue_struct *cma_wq;
72 static DEFINE_IDR(sdp_ps);
73 static DEFINE_IDR(tcp_ps);
74 static DEFINE_IDR(udp_ps);
75 static DEFINE_IDR(ipoib_ps);
76 static int next_port;
77
78 struct cma_device {
79 struct list_head list;
80 struct ib_device *device;
81 struct completion comp;
82 atomic_t refcount;
83 struct list_head id_list;
84 };
85
86 enum cma_state {
87 CMA_IDLE,
88 CMA_ADDR_QUERY,
89 CMA_ADDR_RESOLVED,
90 CMA_ROUTE_QUERY,
91 CMA_ROUTE_RESOLVED,
92 CMA_CONNECT,
93 CMA_DISCONNECT,
94 CMA_ADDR_BOUND,
95 CMA_LISTEN,
96 CMA_DEVICE_REMOVAL,
97 CMA_DESTROYING
98 };
99
100 struct rdma_bind_list {
101 struct idr *ps;
102 struct hlist_head owners;
103 unsigned short port;
104 };
105
106 /*
107 * Device removal can occur at anytime, so we need extra handling to
108 * serialize notifying the user of device removal with other callbacks.
109 * We do this by disabling removal notification while a callback is in process,
110 * and reporting it after the callback completes.
111 */
112 struct rdma_id_private {
113 struct rdma_cm_id id;
114
115 struct rdma_bind_list *bind_list;
116 struct hlist_node node;
117 struct list_head list; /* listen_any_list or cma_device.list */
118 struct list_head listen_list; /* per device listens */
119 struct cma_device *cma_dev;
120 struct list_head mc_list;
121
122 int internal_id;
123 enum cma_state state;
124 spinlock_t lock;
125 struct mutex qp_mutex;
126
127 struct completion comp;
128 atomic_t refcount;
129 wait_queue_head_t wait_remove;
130 atomic_t dev_remove;
131
132 int backlog;
133 int timeout_ms;
134 struct ib_sa_query *query;
135 int query_id;
136 union {
137 struct ib_cm_id *ib;
138 struct iw_cm_id *iw;
139 } cm_id;
140
141 u32 seq_num;
142 u32 qkey;
143 u32 qp_num;
144 u8 srq;
145 u8 tos;
146 };
147
148 struct cma_multicast {
149 struct rdma_id_private *id_priv;
150 union {
151 struct ib_sa_multicast *ib;
152 } multicast;
153 struct list_head list;
154 void *context;
155 struct sockaddr addr;
156 u8 pad[sizeof(struct sockaddr_in6) -
157 sizeof(struct sockaddr)];
158 };
159
160 struct cma_work {
161 struct work_struct work;
162 struct rdma_id_private *id;
163 enum cma_state old_state;
164 enum cma_state new_state;
165 struct rdma_cm_event event;
166 };
167
168 union cma_ip_addr {
169 struct in6_addr ip6;
170 struct {
171 __be32 pad[3];
172 __be32 addr;
173 } ip4;
174 };
175
176 struct cma_hdr {
177 u8 cma_version;
178 u8 ip_version; /* IP version: 7:4 */
179 __be16 port;
180 union cma_ip_addr src_addr;
181 union cma_ip_addr dst_addr;
182 };
183
184 struct sdp_hh {
185 u8 bsdh[16];
186 u8 sdp_version; /* Major version: 7:4 */
187 u8 ip_version; /* IP version: 7:4 */
188 u8 sdp_specific1[10];
189 __be16 port;
190 __be16 sdp_specific2;
191 union cma_ip_addr src_addr;
192 union cma_ip_addr dst_addr;
193 };
194
195 struct sdp_hah {
196 u8 bsdh[16];
197 u8 sdp_version;
198 };
199
200 #define CMA_VERSION 0x00
201 #define SDP_MAJ_VERSION 0x2
202
203 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
204 {
205 unsigned long flags;
206 int ret;
207
208 spin_lock_irqsave(&id_priv->lock, flags);
209 ret = (id_priv->state == comp);
210 spin_unlock_irqrestore(&id_priv->lock, flags);
211 return ret;
212 }
213
214 static int cma_comp_exch(struct rdma_id_private *id_priv,
215 enum cma_state comp, enum cma_state exch)
216 {
217 unsigned long flags;
218 int ret;
219
220 spin_lock_irqsave(&id_priv->lock, flags);
221 if ((ret = (id_priv->state == comp)))
222 id_priv->state = exch;
223 spin_unlock_irqrestore(&id_priv->lock, flags);
224 return ret;
225 }
226
227 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
228 enum cma_state exch)
229 {
230 unsigned long flags;
231 enum cma_state old;
232
233 spin_lock_irqsave(&id_priv->lock, flags);
234 old = id_priv->state;
235 id_priv->state = exch;
236 spin_unlock_irqrestore(&id_priv->lock, flags);
237 return old;
238 }
239
240 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
241 {
242 return hdr->ip_version >> 4;
243 }
244
245 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
246 {
247 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
248 }
249
250 static inline u8 sdp_get_majv(u8 sdp_version)
251 {
252 return sdp_version >> 4;
253 }
254
255 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
256 {
257 return hh->ip_version >> 4;
258 }
259
260 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
261 {
262 hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
263 }
264
265 static inline int cma_is_ud_ps(enum rdma_port_space ps)
266 {
267 return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB);
268 }
269
270 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
271 struct cma_device *cma_dev)
272 {
273 atomic_inc(&cma_dev->refcount);
274 id_priv->cma_dev = cma_dev;
275 id_priv->id.device = cma_dev->device;
276 list_add_tail(&id_priv->list, &cma_dev->id_list);
277 }
278
279 static inline void cma_deref_dev(struct cma_device *cma_dev)
280 {
281 if (atomic_dec_and_test(&cma_dev->refcount))
282 complete(&cma_dev->comp);
283 }
284
285 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
286 {
287 list_del(&id_priv->list);
288 cma_deref_dev(id_priv->cma_dev);
289 id_priv->cma_dev = NULL;
290 }
291
292 static int cma_set_qkey(struct ib_device *device, u8 port_num,
293 enum rdma_port_space ps,
294 struct rdma_dev_addr *dev_addr, u32 *qkey)
295 {
296 struct ib_sa_mcmember_rec rec;
297 int ret = 0;
298
299 switch (ps) {
300 case RDMA_PS_UDP:
301 *qkey = RDMA_UDP_QKEY;
302 break;
303 case RDMA_PS_IPOIB:
304 ib_addr_get_mgid(dev_addr, &rec.mgid);
305 ret = ib_sa_get_mcmember_rec(device, port_num, &rec.mgid, &rec);
306 *qkey = be32_to_cpu(rec.qkey);
307 break;
308 default:
309 break;
310 }
311 return ret;
312 }
313
314 static int cma_acquire_dev(struct rdma_id_private *id_priv)
315 {
316 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
317 struct cma_device *cma_dev;
318 union ib_gid gid;
319 int ret = -ENODEV;
320
321 switch (rdma_node_get_transport(dev_addr->dev_type)) {
322 case RDMA_TRANSPORT_IB:
323 ib_addr_get_sgid(dev_addr, &gid);
324 break;
325 case RDMA_TRANSPORT_IWARP:
326 iw_addr_get_sgid(dev_addr, &gid);
327 break;
328 default:
329 return -ENODEV;
330 }
331
332 list_for_each_entry(cma_dev, &dev_list, list) {
333 ret = ib_find_cached_gid(cma_dev->device, &gid,
334 &id_priv->id.port_num, NULL);
335 if (!ret) {
336 ret = cma_set_qkey(cma_dev->device,
337 id_priv->id.port_num,
338 id_priv->id.ps, dev_addr,
339 &id_priv->qkey);
340 if (!ret)
341 cma_attach_to_dev(id_priv, cma_dev);
342 break;
343 }
344 }
345 return ret;
346 }
347
348 static void cma_deref_id(struct rdma_id_private *id_priv)
349 {
350 if (atomic_dec_and_test(&id_priv->refcount))
351 complete(&id_priv->comp);
352 }
353
354 static int cma_disable_remove(struct rdma_id_private *id_priv,
355 enum cma_state state)
356 {
357 unsigned long flags;
358 int ret;
359
360 spin_lock_irqsave(&id_priv->lock, flags);
361 if (id_priv->state == state) {
362 atomic_inc(&id_priv->dev_remove);
363 ret = 0;
364 } else
365 ret = -EINVAL;
366 spin_unlock_irqrestore(&id_priv->lock, flags);
367 return ret;
368 }
369
370 static void cma_enable_remove(struct rdma_id_private *id_priv)
371 {
372 if (atomic_dec_and_test(&id_priv->dev_remove))
373 wake_up(&id_priv->wait_remove);
374 }
375
376 static int cma_has_cm_dev(struct rdma_id_private *id_priv)
377 {
378 return (id_priv->id.device && id_priv->cm_id.ib);
379 }
380
381 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
382 void *context, enum rdma_port_space ps)
383 {
384 struct rdma_id_private *id_priv;
385
386 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
387 if (!id_priv)
388 return ERR_PTR(-ENOMEM);
389
390 id_priv->state = CMA_IDLE;
391 id_priv->id.context = context;
392 id_priv->id.event_handler = event_handler;
393 id_priv->id.ps = ps;
394 spin_lock_init(&id_priv->lock);
395 mutex_init(&id_priv->qp_mutex);
396 init_completion(&id_priv->comp);
397 atomic_set(&id_priv->refcount, 1);
398 init_waitqueue_head(&id_priv->wait_remove);
399 atomic_set(&id_priv->dev_remove, 0);
400 INIT_LIST_HEAD(&id_priv->listen_list);
401 INIT_LIST_HEAD(&id_priv->mc_list);
402 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
403
404 return &id_priv->id;
405 }
406 EXPORT_SYMBOL(rdma_create_id);
407
408 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
409 {
410 struct ib_qp_attr qp_attr;
411 int qp_attr_mask, ret;
412
413 qp_attr.qp_state = IB_QPS_INIT;
414 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
415 if (ret)
416 return ret;
417
418 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
419 if (ret)
420 return ret;
421
422 qp_attr.qp_state = IB_QPS_RTR;
423 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
424 if (ret)
425 return ret;
426
427 qp_attr.qp_state = IB_QPS_RTS;
428 qp_attr.sq_psn = 0;
429 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
430
431 return ret;
432 }
433
434 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
435 {
436 struct ib_qp_attr qp_attr;
437 int qp_attr_mask, ret;
438
439 qp_attr.qp_state = IB_QPS_INIT;
440 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
441 if (ret)
442 return ret;
443
444 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
445 }
446
447 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
448 struct ib_qp_init_attr *qp_init_attr)
449 {
450 struct rdma_id_private *id_priv;
451 struct ib_qp *qp;
452 int ret;
453
454 id_priv = container_of(id, struct rdma_id_private, id);
455 if (id->device != pd->device)
456 return -EINVAL;
457
458 qp = ib_create_qp(pd, qp_init_attr);
459 if (IS_ERR(qp))
460 return PTR_ERR(qp);
461
462 if (cma_is_ud_ps(id_priv->id.ps))
463 ret = cma_init_ud_qp(id_priv, qp);
464 else
465 ret = cma_init_conn_qp(id_priv, qp);
466 if (ret)
467 goto err;
468
469 id->qp = qp;
470 id_priv->qp_num = qp->qp_num;
471 id_priv->srq = (qp->srq != NULL);
472 return 0;
473 err:
474 ib_destroy_qp(qp);
475 return ret;
476 }
477 EXPORT_SYMBOL(rdma_create_qp);
478
479 void rdma_destroy_qp(struct rdma_cm_id *id)
480 {
481 struct rdma_id_private *id_priv;
482
483 id_priv = container_of(id, struct rdma_id_private, id);
484 mutex_lock(&id_priv->qp_mutex);
485 ib_destroy_qp(id_priv->id.qp);
486 id_priv->id.qp = NULL;
487 mutex_unlock(&id_priv->qp_mutex);
488 }
489 EXPORT_SYMBOL(rdma_destroy_qp);
490
491 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
492 struct rdma_conn_param *conn_param)
493 {
494 struct ib_qp_attr qp_attr;
495 int qp_attr_mask, ret;
496
497 mutex_lock(&id_priv->qp_mutex);
498 if (!id_priv->id.qp) {
499 ret = 0;
500 goto out;
501 }
502
503 /* Need to update QP attributes from default values. */
504 qp_attr.qp_state = IB_QPS_INIT;
505 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
506 if (ret)
507 goto out;
508
509 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
510 if (ret)
511 goto out;
512
513 qp_attr.qp_state = IB_QPS_RTR;
514 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
515 if (ret)
516 goto out;
517
518 if (conn_param)
519 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
520 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
521 out:
522 mutex_unlock(&id_priv->qp_mutex);
523 return ret;
524 }
525
526 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
527 struct rdma_conn_param *conn_param)
528 {
529 struct ib_qp_attr qp_attr;
530 int qp_attr_mask, ret;
531
532 mutex_lock(&id_priv->qp_mutex);
533 if (!id_priv->id.qp) {
534 ret = 0;
535 goto out;
536 }
537
538 qp_attr.qp_state = IB_QPS_RTS;
539 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
540 if (ret)
541 goto out;
542
543 if (conn_param)
544 qp_attr.max_rd_atomic = conn_param->initiator_depth;
545 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
546 out:
547 mutex_unlock(&id_priv->qp_mutex);
548 return ret;
549 }
550
551 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
552 {
553 struct ib_qp_attr qp_attr;
554 int ret;
555
556 mutex_lock(&id_priv->qp_mutex);
557 if (!id_priv->id.qp) {
558 ret = 0;
559 goto out;
560 }
561
562 qp_attr.qp_state = IB_QPS_ERR;
563 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
564 out:
565 mutex_unlock(&id_priv->qp_mutex);
566 return ret;
567 }
568
569 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
570 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
571 {
572 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
573 int ret;
574
575 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
576 ib_addr_get_pkey(dev_addr),
577 &qp_attr->pkey_index);
578 if (ret)
579 return ret;
580
581 qp_attr->port_num = id_priv->id.port_num;
582 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
583
584 if (cma_is_ud_ps(id_priv->id.ps)) {
585 qp_attr->qkey = id_priv->qkey;
586 *qp_attr_mask |= IB_QP_QKEY;
587 } else {
588 qp_attr->qp_access_flags = 0;
589 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
590 }
591 return 0;
592 }
593
594 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
595 int *qp_attr_mask)
596 {
597 struct rdma_id_private *id_priv;
598 int ret = 0;
599
600 id_priv = container_of(id, struct rdma_id_private, id);
601 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
602 case RDMA_TRANSPORT_IB:
603 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps))
604 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
605 else
606 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
607 qp_attr_mask);
608 if (qp_attr->qp_state == IB_QPS_RTR)
609 qp_attr->rq_psn = id_priv->seq_num;
610 break;
611 case RDMA_TRANSPORT_IWARP:
612 if (!id_priv->cm_id.iw) {
613 qp_attr->qp_access_flags = 0;
614 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
615 } else
616 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
617 qp_attr_mask);
618 break;
619 default:
620 ret = -ENOSYS;
621 break;
622 }
623
624 return ret;
625 }
626 EXPORT_SYMBOL(rdma_init_qp_attr);
627
628 static inline int cma_zero_addr(struct sockaddr *addr)
629 {
630 struct in6_addr *ip6;
631
632 if (addr->sa_family == AF_INET)
633 return ipv4_is_zeronet(
634 ((struct sockaddr_in *)addr)->sin_addr.s_addr);
635 else {
636 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
637 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
638 ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
639 }
640 }
641
642 static inline int cma_loopback_addr(struct sockaddr *addr)
643 {
644 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
645 }
646
647 static inline int cma_any_addr(struct sockaddr *addr)
648 {
649 return cma_zero_addr(addr) || cma_loopback_addr(addr);
650 }
651
652 static inline __be16 cma_port(struct sockaddr *addr)
653 {
654 if (addr->sa_family == AF_INET)
655 return ((struct sockaddr_in *) addr)->sin_port;
656 else
657 return ((struct sockaddr_in6 *) addr)->sin6_port;
658 }
659
660 static inline int cma_any_port(struct sockaddr *addr)
661 {
662 return !cma_port(addr);
663 }
664
665 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
666 u8 *ip_ver, __be16 *port,
667 union cma_ip_addr **src, union cma_ip_addr **dst)
668 {
669 switch (ps) {
670 case RDMA_PS_SDP:
671 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
672 SDP_MAJ_VERSION)
673 return -EINVAL;
674
675 *ip_ver = sdp_get_ip_ver(hdr);
676 *port = ((struct sdp_hh *) hdr)->port;
677 *src = &((struct sdp_hh *) hdr)->src_addr;
678 *dst = &((struct sdp_hh *) hdr)->dst_addr;
679 break;
680 default:
681 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
682 return -EINVAL;
683
684 *ip_ver = cma_get_ip_ver(hdr);
685 *port = ((struct cma_hdr *) hdr)->port;
686 *src = &((struct cma_hdr *) hdr)->src_addr;
687 *dst = &((struct cma_hdr *) hdr)->dst_addr;
688 break;
689 }
690
691 if (*ip_ver != 4 && *ip_ver != 6)
692 return -EINVAL;
693 return 0;
694 }
695
696 static void cma_save_net_info(struct rdma_addr *addr,
697 struct rdma_addr *listen_addr,
698 u8 ip_ver, __be16 port,
699 union cma_ip_addr *src, union cma_ip_addr *dst)
700 {
701 struct sockaddr_in *listen4, *ip4;
702 struct sockaddr_in6 *listen6, *ip6;
703
704 switch (ip_ver) {
705 case 4:
706 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
707 ip4 = (struct sockaddr_in *) &addr->src_addr;
708 ip4->sin_family = listen4->sin_family;
709 ip4->sin_addr.s_addr = dst->ip4.addr;
710 ip4->sin_port = listen4->sin_port;
711
712 ip4 = (struct sockaddr_in *) &addr->dst_addr;
713 ip4->sin_family = listen4->sin_family;
714 ip4->sin_addr.s_addr = src->ip4.addr;
715 ip4->sin_port = port;
716 break;
717 case 6:
718 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
719 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
720 ip6->sin6_family = listen6->sin6_family;
721 ip6->sin6_addr = dst->ip6;
722 ip6->sin6_port = listen6->sin6_port;
723
724 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
725 ip6->sin6_family = listen6->sin6_family;
726 ip6->sin6_addr = src->ip6;
727 ip6->sin6_port = port;
728 break;
729 default:
730 break;
731 }
732 }
733
734 static inline int cma_user_data_offset(enum rdma_port_space ps)
735 {
736 switch (ps) {
737 case RDMA_PS_SDP:
738 return 0;
739 default:
740 return sizeof(struct cma_hdr);
741 }
742 }
743
744 static void cma_cancel_route(struct rdma_id_private *id_priv)
745 {
746 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
747 case RDMA_TRANSPORT_IB:
748 if (id_priv->query)
749 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
750 break;
751 default:
752 break;
753 }
754 }
755
756 static void cma_cancel_listens(struct rdma_id_private *id_priv)
757 {
758 struct rdma_id_private *dev_id_priv;
759
760 /*
761 * Remove from listen_any_list to prevent added devices from spawning
762 * additional listen requests.
763 */
764 mutex_lock(&lock);
765 list_del(&id_priv->list);
766
767 while (!list_empty(&id_priv->listen_list)) {
768 dev_id_priv = list_entry(id_priv->listen_list.next,
769 struct rdma_id_private, listen_list);
770 /* sync with device removal to avoid duplicate destruction */
771 list_del_init(&dev_id_priv->list);
772 list_del(&dev_id_priv->listen_list);
773 mutex_unlock(&lock);
774
775 rdma_destroy_id(&dev_id_priv->id);
776 mutex_lock(&lock);
777 }
778 mutex_unlock(&lock);
779 }
780
781 static void cma_cancel_operation(struct rdma_id_private *id_priv,
782 enum cma_state state)
783 {
784 switch (state) {
785 case CMA_ADDR_QUERY:
786 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
787 break;
788 case CMA_ROUTE_QUERY:
789 cma_cancel_route(id_priv);
790 break;
791 case CMA_LISTEN:
792 if (cma_any_addr(&id_priv->id.route.addr.src_addr) &&
793 !id_priv->cma_dev)
794 cma_cancel_listens(id_priv);
795 break;
796 default:
797 break;
798 }
799 }
800
801 static void cma_release_port(struct rdma_id_private *id_priv)
802 {
803 struct rdma_bind_list *bind_list = id_priv->bind_list;
804
805 if (!bind_list)
806 return;
807
808 mutex_lock(&lock);
809 hlist_del(&id_priv->node);
810 if (hlist_empty(&bind_list->owners)) {
811 idr_remove(bind_list->ps, bind_list->port);
812 kfree(bind_list);
813 }
814 mutex_unlock(&lock);
815 }
816
817 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
818 {
819 struct cma_multicast *mc;
820
821 while (!list_empty(&id_priv->mc_list)) {
822 mc = container_of(id_priv->mc_list.next,
823 struct cma_multicast, list);
824 list_del(&mc->list);
825 ib_sa_free_multicast(mc->multicast.ib);
826 kfree(mc);
827 }
828 }
829
830 void rdma_destroy_id(struct rdma_cm_id *id)
831 {
832 struct rdma_id_private *id_priv;
833 enum cma_state state;
834
835 id_priv = container_of(id, struct rdma_id_private, id);
836 state = cma_exch(id_priv, CMA_DESTROYING);
837 cma_cancel_operation(id_priv, state);
838
839 mutex_lock(&lock);
840 if (id_priv->cma_dev) {
841 mutex_unlock(&lock);
842 switch (rdma_node_get_transport(id->device->node_type)) {
843 case RDMA_TRANSPORT_IB:
844 if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
845 ib_destroy_cm_id(id_priv->cm_id.ib);
846 break;
847 case RDMA_TRANSPORT_IWARP:
848 if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
849 iw_destroy_cm_id(id_priv->cm_id.iw);
850 break;
851 default:
852 break;
853 }
854 cma_leave_mc_groups(id_priv);
855 mutex_lock(&lock);
856 cma_detach_from_dev(id_priv);
857 }
858 mutex_unlock(&lock);
859
860 cma_release_port(id_priv);
861 cma_deref_id(id_priv);
862 wait_for_completion(&id_priv->comp);
863
864 if (id_priv->internal_id)
865 cma_deref_id(id_priv->id.context);
866
867 kfree(id_priv->id.route.path_rec);
868 kfree(id_priv);
869 }
870 EXPORT_SYMBOL(rdma_destroy_id);
871
872 static int cma_rep_recv(struct rdma_id_private *id_priv)
873 {
874 int ret;
875
876 ret = cma_modify_qp_rtr(id_priv, NULL);
877 if (ret)
878 goto reject;
879
880 ret = cma_modify_qp_rts(id_priv, NULL);
881 if (ret)
882 goto reject;
883
884 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
885 if (ret)
886 goto reject;
887
888 return 0;
889 reject:
890 cma_modify_qp_err(id_priv);
891 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
892 NULL, 0, NULL, 0);
893 return ret;
894 }
895
896 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
897 {
898 if (id_priv->id.ps == RDMA_PS_SDP &&
899 sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
900 SDP_MAJ_VERSION)
901 return -EINVAL;
902
903 return 0;
904 }
905
906 static void cma_set_rep_event_data(struct rdma_cm_event *event,
907 struct ib_cm_rep_event_param *rep_data,
908 void *private_data)
909 {
910 event->param.conn.private_data = private_data;
911 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
912 event->param.conn.responder_resources = rep_data->responder_resources;
913 event->param.conn.initiator_depth = rep_data->initiator_depth;
914 event->param.conn.flow_control = rep_data->flow_control;
915 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
916 event->param.conn.srq = rep_data->srq;
917 event->param.conn.qp_num = rep_data->remote_qpn;
918 }
919
920 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
921 {
922 struct rdma_id_private *id_priv = cm_id->context;
923 struct rdma_cm_event event;
924 int ret = 0;
925
926 if (cma_disable_remove(id_priv, CMA_CONNECT))
927 return 0;
928
929 memset(&event, 0, sizeof event);
930 switch (ib_event->event) {
931 case IB_CM_REQ_ERROR:
932 case IB_CM_REP_ERROR:
933 event.event = RDMA_CM_EVENT_UNREACHABLE;
934 event.status = -ETIMEDOUT;
935 break;
936 case IB_CM_REP_RECEIVED:
937 event.status = cma_verify_rep(id_priv, ib_event->private_data);
938 if (event.status)
939 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
940 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
941 event.status = cma_rep_recv(id_priv);
942 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
943 RDMA_CM_EVENT_ESTABLISHED;
944 } else
945 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
946 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
947 ib_event->private_data);
948 break;
949 case IB_CM_RTU_RECEIVED:
950 case IB_CM_USER_ESTABLISHED:
951 event.event = RDMA_CM_EVENT_ESTABLISHED;
952 break;
953 case IB_CM_DREQ_ERROR:
954 event.status = -ETIMEDOUT; /* fall through */
955 case IB_CM_DREQ_RECEIVED:
956 case IB_CM_DREP_RECEIVED:
957 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
958 goto out;
959 event.event = RDMA_CM_EVENT_DISCONNECTED;
960 break;
961 case IB_CM_TIMEWAIT_EXIT:
962 case IB_CM_MRA_RECEIVED:
963 /* ignore event */
964 goto out;
965 case IB_CM_REJ_RECEIVED:
966 cma_modify_qp_err(id_priv);
967 event.status = ib_event->param.rej_rcvd.reason;
968 event.event = RDMA_CM_EVENT_REJECTED;
969 event.param.conn.private_data = ib_event->private_data;
970 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
971 break;
972 default:
973 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
974 ib_event->event);
975 goto out;
976 }
977
978 ret = id_priv->id.event_handler(&id_priv->id, &event);
979 if (ret) {
980 /* Destroy the CM ID by returning a non-zero value. */
981 id_priv->cm_id.ib = NULL;
982 cma_exch(id_priv, CMA_DESTROYING);
983 cma_enable_remove(id_priv);
984 rdma_destroy_id(&id_priv->id);
985 return ret;
986 }
987 out:
988 cma_enable_remove(id_priv);
989 return ret;
990 }
991
992 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
993 struct ib_cm_event *ib_event)
994 {
995 struct rdma_id_private *id_priv;
996 struct rdma_cm_id *id;
997 struct rdma_route *rt;
998 union cma_ip_addr *src, *dst;
999 __be16 port;
1000 u8 ip_ver;
1001
1002 if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1003 &ip_ver, &port, &src, &dst))
1004 goto err;
1005
1006 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1007 listen_id->ps);
1008 if (IS_ERR(id))
1009 goto err;
1010
1011 cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1012 ip_ver, port, src, dst);
1013
1014 rt = &id->route;
1015 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1016 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1017 GFP_KERNEL);
1018 if (!rt->path_rec)
1019 goto destroy_id;
1020
1021 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1022 if (rt->num_paths == 2)
1023 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1024
1025 ib_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1026 ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1027 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1028 rt->addr.dev_addr.dev_type = RDMA_NODE_IB_CA;
1029
1030 id_priv = container_of(id, struct rdma_id_private, id);
1031 id_priv->state = CMA_CONNECT;
1032 return id_priv;
1033
1034 destroy_id:
1035 rdma_destroy_id(id);
1036 err:
1037 return NULL;
1038 }
1039
1040 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1041 struct ib_cm_event *ib_event)
1042 {
1043 struct rdma_id_private *id_priv;
1044 struct rdma_cm_id *id;
1045 union cma_ip_addr *src, *dst;
1046 __be16 port;
1047 u8 ip_ver;
1048 int ret;
1049
1050 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1051 listen_id->ps);
1052 if (IS_ERR(id))
1053 return NULL;
1054
1055
1056 if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1057 &ip_ver, &port, &src, &dst))
1058 goto err;
1059
1060 cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1061 ip_ver, port, src, dst);
1062
1063 ret = rdma_translate_ip(&id->route.addr.src_addr,
1064 &id->route.addr.dev_addr);
1065 if (ret)
1066 goto err;
1067
1068 id_priv = container_of(id, struct rdma_id_private, id);
1069 id_priv->state = CMA_CONNECT;
1070 return id_priv;
1071 err:
1072 rdma_destroy_id(id);
1073 return NULL;
1074 }
1075
1076 static void cma_set_req_event_data(struct rdma_cm_event *event,
1077 struct ib_cm_req_event_param *req_data,
1078 void *private_data, int offset)
1079 {
1080 event->param.conn.private_data = private_data + offset;
1081 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1082 event->param.conn.responder_resources = req_data->responder_resources;
1083 event->param.conn.initiator_depth = req_data->initiator_depth;
1084 event->param.conn.flow_control = req_data->flow_control;
1085 event->param.conn.retry_count = req_data->retry_count;
1086 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1087 event->param.conn.srq = req_data->srq;
1088 event->param.conn.qp_num = req_data->remote_qpn;
1089 }
1090
1091 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1092 {
1093 struct rdma_id_private *listen_id, *conn_id;
1094 struct rdma_cm_event event;
1095 int offset, ret;
1096
1097 listen_id = cm_id->context;
1098 if (cma_disable_remove(listen_id, CMA_LISTEN))
1099 return -ECONNABORTED;
1100
1101 memset(&event, 0, sizeof event);
1102 offset = cma_user_data_offset(listen_id->id.ps);
1103 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1104 if (cma_is_ud_ps(listen_id->id.ps)) {
1105 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1106 event.param.ud.private_data = ib_event->private_data + offset;
1107 event.param.ud.private_data_len =
1108 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1109 } else {
1110 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1111 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1112 ib_event->private_data, offset);
1113 }
1114 if (!conn_id) {
1115 ret = -ENOMEM;
1116 goto out;
1117 }
1118
1119 atomic_inc(&conn_id->dev_remove);
1120 mutex_lock(&lock);
1121 ret = cma_acquire_dev(conn_id);
1122 mutex_unlock(&lock);
1123 if (ret)
1124 goto release_conn_id;
1125
1126 conn_id->cm_id.ib = cm_id;
1127 cm_id->context = conn_id;
1128 cm_id->cm_handler = cma_ib_handler;
1129
1130 ret = conn_id->id.event_handler(&conn_id->id, &event);
1131 if (!ret) {
1132 /*
1133 * Acquire mutex to prevent user executing rdma_destroy_id()
1134 * while we're accessing the cm_id.
1135 */
1136 mutex_lock(&lock);
1137 if (cma_comp(conn_id, CMA_CONNECT) &&
1138 !cma_is_ud_ps(conn_id->id.ps))
1139 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1140 mutex_unlock(&lock);
1141 cma_enable_remove(conn_id);
1142 goto out;
1143 }
1144
1145 /* Destroy the CM ID by returning a non-zero value. */
1146 conn_id->cm_id.ib = NULL;
1147
1148 release_conn_id:
1149 cma_exch(conn_id, CMA_DESTROYING);
1150 cma_enable_remove(conn_id);
1151 rdma_destroy_id(&conn_id->id);
1152
1153 out:
1154 cma_enable_remove(listen_id);
1155 return ret;
1156 }
1157
1158 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1159 {
1160 return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1161 }
1162
1163 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1164 struct ib_cm_compare_data *compare)
1165 {
1166 struct cma_hdr *cma_data, *cma_mask;
1167 struct sdp_hh *sdp_data, *sdp_mask;
1168 __be32 ip4_addr;
1169 struct in6_addr ip6_addr;
1170
1171 memset(compare, 0, sizeof *compare);
1172 cma_data = (void *) compare->data;
1173 cma_mask = (void *) compare->mask;
1174 sdp_data = (void *) compare->data;
1175 sdp_mask = (void *) compare->mask;
1176
1177 switch (addr->sa_family) {
1178 case AF_INET:
1179 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1180 if (ps == RDMA_PS_SDP) {
1181 sdp_set_ip_ver(sdp_data, 4);
1182 sdp_set_ip_ver(sdp_mask, 0xF);
1183 sdp_data->dst_addr.ip4.addr = ip4_addr;
1184 sdp_mask->dst_addr.ip4.addr = htonl(~0);
1185 } else {
1186 cma_set_ip_ver(cma_data, 4);
1187 cma_set_ip_ver(cma_mask, 0xF);
1188 cma_data->dst_addr.ip4.addr = ip4_addr;
1189 cma_mask->dst_addr.ip4.addr = htonl(~0);
1190 }
1191 break;
1192 case AF_INET6:
1193 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1194 if (ps == RDMA_PS_SDP) {
1195 sdp_set_ip_ver(sdp_data, 6);
1196 sdp_set_ip_ver(sdp_mask, 0xF);
1197 sdp_data->dst_addr.ip6 = ip6_addr;
1198 memset(&sdp_mask->dst_addr.ip6, 0xFF,
1199 sizeof sdp_mask->dst_addr.ip6);
1200 } else {
1201 cma_set_ip_ver(cma_data, 6);
1202 cma_set_ip_ver(cma_mask, 0xF);
1203 cma_data->dst_addr.ip6 = ip6_addr;
1204 memset(&cma_mask->dst_addr.ip6, 0xFF,
1205 sizeof cma_mask->dst_addr.ip6);
1206 }
1207 break;
1208 default:
1209 break;
1210 }
1211 }
1212
1213 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1214 {
1215 struct rdma_id_private *id_priv = iw_id->context;
1216 struct rdma_cm_event event;
1217 struct sockaddr_in *sin;
1218 int ret = 0;
1219
1220 if (cma_disable_remove(id_priv, CMA_CONNECT))
1221 return 0;
1222
1223 memset(&event, 0, sizeof event);
1224 switch (iw_event->event) {
1225 case IW_CM_EVENT_CLOSE:
1226 event.event = RDMA_CM_EVENT_DISCONNECTED;
1227 break;
1228 case IW_CM_EVENT_CONNECT_REPLY:
1229 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1230 *sin = iw_event->local_addr;
1231 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1232 *sin = iw_event->remote_addr;
1233 switch (iw_event->status) {
1234 case 0:
1235 event.event = RDMA_CM_EVENT_ESTABLISHED;
1236 break;
1237 case -ECONNRESET:
1238 case -ECONNREFUSED:
1239 event.event = RDMA_CM_EVENT_REJECTED;
1240 break;
1241 case -ETIMEDOUT:
1242 event.event = RDMA_CM_EVENT_UNREACHABLE;
1243 break;
1244 default:
1245 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1246 break;
1247 }
1248 break;
1249 case IW_CM_EVENT_ESTABLISHED:
1250 event.event = RDMA_CM_EVENT_ESTABLISHED;
1251 break;
1252 default:
1253 BUG_ON(1);
1254 }
1255
1256 event.status = iw_event->status;
1257 event.param.conn.private_data = iw_event->private_data;
1258 event.param.conn.private_data_len = iw_event->private_data_len;
1259 ret = id_priv->id.event_handler(&id_priv->id, &event);
1260 if (ret) {
1261 /* Destroy the CM ID by returning a non-zero value. */
1262 id_priv->cm_id.iw = NULL;
1263 cma_exch(id_priv, CMA_DESTROYING);
1264 cma_enable_remove(id_priv);
1265 rdma_destroy_id(&id_priv->id);
1266 return ret;
1267 }
1268
1269 cma_enable_remove(id_priv);
1270 return ret;
1271 }
1272
1273 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1274 struct iw_cm_event *iw_event)
1275 {
1276 struct rdma_cm_id *new_cm_id;
1277 struct rdma_id_private *listen_id, *conn_id;
1278 struct sockaddr_in *sin;
1279 struct net_device *dev = NULL;
1280 struct rdma_cm_event event;
1281 int ret;
1282 struct ib_device_attr attr;
1283
1284 listen_id = cm_id->context;
1285 if (cma_disable_remove(listen_id, CMA_LISTEN))
1286 return -ECONNABORTED;
1287
1288 /* Create a new RDMA id for the new IW CM ID */
1289 new_cm_id = rdma_create_id(listen_id->id.event_handler,
1290 listen_id->id.context,
1291 RDMA_PS_TCP);
1292 if (!new_cm_id) {
1293 ret = -ENOMEM;
1294 goto out;
1295 }
1296 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1297 atomic_inc(&conn_id->dev_remove);
1298 conn_id->state = CMA_CONNECT;
1299
1300 dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1301 if (!dev) {
1302 ret = -EADDRNOTAVAIL;
1303 cma_enable_remove(conn_id);
1304 rdma_destroy_id(new_cm_id);
1305 goto out;
1306 }
1307 ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1308 if (ret) {
1309 cma_enable_remove(conn_id);
1310 rdma_destroy_id(new_cm_id);
1311 goto out;
1312 }
1313
1314 mutex_lock(&lock);
1315 ret = cma_acquire_dev(conn_id);
1316 mutex_unlock(&lock);
1317 if (ret) {
1318 cma_enable_remove(conn_id);
1319 rdma_destroy_id(new_cm_id);
1320 goto out;
1321 }
1322
1323 conn_id->cm_id.iw = cm_id;
1324 cm_id->context = conn_id;
1325 cm_id->cm_handler = cma_iw_handler;
1326
1327 sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1328 *sin = iw_event->local_addr;
1329 sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1330 *sin = iw_event->remote_addr;
1331
1332 ret = ib_query_device(conn_id->id.device, &attr);
1333 if (ret) {
1334 cma_enable_remove(conn_id);
1335 rdma_destroy_id(new_cm_id);
1336 goto out;
1337 }
1338
1339 memset(&event, 0, sizeof event);
1340 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1341 event.param.conn.private_data = iw_event->private_data;
1342 event.param.conn.private_data_len = iw_event->private_data_len;
1343 event.param.conn.initiator_depth = attr.max_qp_init_rd_atom;
1344 event.param.conn.responder_resources = attr.max_qp_rd_atom;
1345 ret = conn_id->id.event_handler(&conn_id->id, &event);
1346 if (ret) {
1347 /* User wants to destroy the CM ID */
1348 conn_id->cm_id.iw = NULL;
1349 cma_exch(conn_id, CMA_DESTROYING);
1350 cma_enable_remove(conn_id);
1351 rdma_destroy_id(&conn_id->id);
1352 }
1353
1354 out:
1355 if (dev)
1356 dev_put(dev);
1357 cma_enable_remove(listen_id);
1358 return ret;
1359 }
1360
1361 static int cma_ib_listen(struct rdma_id_private *id_priv)
1362 {
1363 struct ib_cm_compare_data compare_data;
1364 struct sockaddr *addr;
1365 __be64 svc_id;
1366 int ret;
1367
1368 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1369 id_priv);
1370 if (IS_ERR(id_priv->cm_id.ib))
1371 return PTR_ERR(id_priv->cm_id.ib);
1372
1373 addr = &id_priv->id.route.addr.src_addr;
1374 svc_id = cma_get_service_id(id_priv->id.ps, addr);
1375 if (cma_any_addr(addr))
1376 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1377 else {
1378 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1379 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1380 }
1381
1382 if (ret) {
1383 ib_destroy_cm_id(id_priv->cm_id.ib);
1384 id_priv->cm_id.ib = NULL;
1385 }
1386
1387 return ret;
1388 }
1389
1390 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1391 {
1392 int ret;
1393 struct sockaddr_in *sin;
1394
1395 id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1396 iw_conn_req_handler,
1397 id_priv);
1398 if (IS_ERR(id_priv->cm_id.iw))
1399 return PTR_ERR(id_priv->cm_id.iw);
1400
1401 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1402 id_priv->cm_id.iw->local_addr = *sin;
1403
1404 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1405
1406 if (ret) {
1407 iw_destroy_cm_id(id_priv->cm_id.iw);
1408 id_priv->cm_id.iw = NULL;
1409 }
1410
1411 return ret;
1412 }
1413
1414 static int cma_listen_handler(struct rdma_cm_id *id,
1415 struct rdma_cm_event *event)
1416 {
1417 struct rdma_id_private *id_priv = id->context;
1418
1419 id->context = id_priv->id.context;
1420 id->event_handler = id_priv->id.event_handler;
1421 return id_priv->id.event_handler(id, event);
1422 }
1423
1424 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1425 struct cma_device *cma_dev)
1426 {
1427 struct rdma_id_private *dev_id_priv;
1428 struct rdma_cm_id *id;
1429 int ret;
1430
1431 id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1432 if (IS_ERR(id))
1433 return;
1434
1435 dev_id_priv = container_of(id, struct rdma_id_private, id);
1436
1437 dev_id_priv->state = CMA_ADDR_BOUND;
1438 memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1439 ip_addr_size(&id_priv->id.route.addr.src_addr));
1440
1441 cma_attach_to_dev(dev_id_priv, cma_dev);
1442 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1443 atomic_inc(&id_priv->refcount);
1444 dev_id_priv->internal_id = 1;
1445
1446 ret = rdma_listen(id, id_priv->backlog);
1447 if (ret)
1448 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1449 "listening on device %s", ret, cma_dev->device->name);
1450 }
1451
1452 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1453 {
1454 struct cma_device *cma_dev;
1455
1456 mutex_lock(&lock);
1457 list_add_tail(&id_priv->list, &listen_any_list);
1458 list_for_each_entry(cma_dev, &dev_list, list)
1459 cma_listen_on_dev(id_priv, cma_dev);
1460 mutex_unlock(&lock);
1461 }
1462
1463 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1464 {
1465 struct sockaddr_in addr_in;
1466
1467 memset(&addr_in, 0, sizeof addr_in);
1468 addr_in.sin_family = af;
1469 return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1470 }
1471
1472 int rdma_listen(struct rdma_cm_id *id, int backlog)
1473 {
1474 struct rdma_id_private *id_priv;
1475 int ret;
1476
1477 id_priv = container_of(id, struct rdma_id_private, id);
1478 if (id_priv->state == CMA_IDLE) {
1479 ret = cma_bind_any(id, AF_INET);
1480 if (ret)
1481 return ret;
1482 }
1483
1484 if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1485 return -EINVAL;
1486
1487 id_priv->backlog = backlog;
1488 if (id->device) {
1489 switch (rdma_node_get_transport(id->device->node_type)) {
1490 case RDMA_TRANSPORT_IB:
1491 ret = cma_ib_listen(id_priv);
1492 if (ret)
1493 goto err;
1494 break;
1495 case RDMA_TRANSPORT_IWARP:
1496 ret = cma_iw_listen(id_priv, backlog);
1497 if (ret)
1498 goto err;
1499 break;
1500 default:
1501 ret = -ENOSYS;
1502 goto err;
1503 }
1504 } else
1505 cma_listen_on_all(id_priv);
1506
1507 return 0;
1508 err:
1509 id_priv->backlog = 0;
1510 cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1511 return ret;
1512 }
1513 EXPORT_SYMBOL(rdma_listen);
1514
1515 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1516 {
1517 struct rdma_id_private *id_priv;
1518
1519 id_priv = container_of(id, struct rdma_id_private, id);
1520 id_priv->tos = (u8) tos;
1521 }
1522 EXPORT_SYMBOL(rdma_set_service_type);
1523
1524 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1525 void *context)
1526 {
1527 struct cma_work *work = context;
1528 struct rdma_route *route;
1529
1530 route = &work->id->id.route;
1531
1532 if (!status) {
1533 route->num_paths = 1;
1534 *route->path_rec = *path_rec;
1535 } else {
1536 work->old_state = CMA_ROUTE_QUERY;
1537 work->new_state = CMA_ADDR_RESOLVED;
1538 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1539 work->event.status = status;
1540 }
1541
1542 queue_work(cma_wq, &work->work);
1543 }
1544
1545 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1546 struct cma_work *work)
1547 {
1548 struct rdma_addr *addr = &id_priv->id.route.addr;
1549 struct ib_sa_path_rec path_rec;
1550 ib_sa_comp_mask comp_mask;
1551 struct sockaddr_in6 *sin6;
1552
1553 memset(&path_rec, 0, sizeof path_rec);
1554 ib_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1555 ib_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1556 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1557 path_rec.numb_path = 1;
1558 path_rec.reversible = 1;
1559 path_rec.service_id = cma_get_service_id(id_priv->id.ps, &addr->dst_addr);
1560
1561 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1562 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1563 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1564
1565 if (addr->src_addr.sa_family == AF_INET) {
1566 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1567 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1568 } else {
1569 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1570 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1571 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1572 }
1573
1574 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1575 id_priv->id.port_num, &path_rec,
1576 comp_mask, timeout_ms,
1577 GFP_KERNEL, cma_query_handler,
1578 work, &id_priv->query);
1579
1580 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1581 }
1582
1583 static void cma_work_handler(struct work_struct *_work)
1584 {
1585 struct cma_work *work = container_of(_work, struct cma_work, work);
1586 struct rdma_id_private *id_priv = work->id;
1587 int destroy = 0;
1588
1589 atomic_inc(&id_priv->dev_remove);
1590 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1591 goto out;
1592
1593 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1594 cma_exch(id_priv, CMA_DESTROYING);
1595 destroy = 1;
1596 }
1597 out:
1598 cma_enable_remove(id_priv);
1599 cma_deref_id(id_priv);
1600 if (destroy)
1601 rdma_destroy_id(&id_priv->id);
1602 kfree(work);
1603 }
1604
1605 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1606 {
1607 struct rdma_route *route = &id_priv->id.route;
1608 struct cma_work *work;
1609 int ret;
1610
1611 work = kzalloc(sizeof *work, GFP_KERNEL);
1612 if (!work)
1613 return -ENOMEM;
1614
1615 work->id = id_priv;
1616 INIT_WORK(&work->work, cma_work_handler);
1617 work->old_state = CMA_ROUTE_QUERY;
1618 work->new_state = CMA_ROUTE_RESOLVED;
1619 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1620
1621 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1622 if (!route->path_rec) {
1623 ret = -ENOMEM;
1624 goto err1;
1625 }
1626
1627 ret = cma_query_ib_route(id_priv, timeout_ms, work);
1628 if (ret)
1629 goto err2;
1630
1631 return 0;
1632 err2:
1633 kfree(route->path_rec);
1634 route->path_rec = NULL;
1635 err1:
1636 kfree(work);
1637 return ret;
1638 }
1639
1640 int rdma_set_ib_paths(struct rdma_cm_id *id,
1641 struct ib_sa_path_rec *path_rec, int num_paths)
1642 {
1643 struct rdma_id_private *id_priv;
1644 int ret;
1645
1646 id_priv = container_of(id, struct rdma_id_private, id);
1647 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1648 return -EINVAL;
1649
1650 id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1651 if (!id->route.path_rec) {
1652 ret = -ENOMEM;
1653 goto err;
1654 }
1655
1656 memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1657 return 0;
1658 err:
1659 cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1660 return ret;
1661 }
1662 EXPORT_SYMBOL(rdma_set_ib_paths);
1663
1664 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1665 {
1666 struct cma_work *work;
1667
1668 work = kzalloc(sizeof *work, GFP_KERNEL);
1669 if (!work)
1670 return -ENOMEM;
1671
1672 work->id = id_priv;
1673 INIT_WORK(&work->work, cma_work_handler);
1674 work->old_state = CMA_ROUTE_QUERY;
1675 work->new_state = CMA_ROUTE_RESOLVED;
1676 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1677 queue_work(cma_wq, &work->work);
1678 return 0;
1679 }
1680
1681 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1682 {
1683 struct rdma_id_private *id_priv;
1684 int ret;
1685
1686 id_priv = container_of(id, struct rdma_id_private, id);
1687 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1688 return -EINVAL;
1689
1690 atomic_inc(&id_priv->refcount);
1691 switch (rdma_node_get_transport(id->device->node_type)) {
1692 case RDMA_TRANSPORT_IB:
1693 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1694 break;
1695 case RDMA_TRANSPORT_IWARP:
1696 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1697 break;
1698 default:
1699 ret = -ENOSYS;
1700 break;
1701 }
1702 if (ret)
1703 goto err;
1704
1705 return 0;
1706 err:
1707 cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1708 cma_deref_id(id_priv);
1709 return ret;
1710 }
1711 EXPORT_SYMBOL(rdma_resolve_route);
1712
1713 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1714 {
1715 struct cma_device *cma_dev;
1716 struct ib_port_attr port_attr;
1717 union ib_gid gid;
1718 u16 pkey;
1719 int ret;
1720 u8 p;
1721
1722 mutex_lock(&lock);
1723 if (list_empty(&dev_list)) {
1724 ret = -ENODEV;
1725 goto out;
1726 }
1727 list_for_each_entry(cma_dev, &dev_list, list)
1728 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1729 if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1730 port_attr.state == IB_PORT_ACTIVE)
1731 goto port_found;
1732
1733 p = 1;
1734 cma_dev = list_entry(dev_list.next, struct cma_device, list);
1735
1736 port_found:
1737 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1738 if (ret)
1739 goto out;
1740
1741 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1742 if (ret)
1743 goto out;
1744
1745 ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1746 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1747 id_priv->id.port_num = p;
1748 cma_attach_to_dev(id_priv, cma_dev);
1749 out:
1750 mutex_unlock(&lock);
1751 return ret;
1752 }
1753
1754 static void addr_handler(int status, struct sockaddr *src_addr,
1755 struct rdma_dev_addr *dev_addr, void *context)
1756 {
1757 struct rdma_id_private *id_priv = context;
1758 struct rdma_cm_event event;
1759
1760 memset(&event, 0, sizeof event);
1761 atomic_inc(&id_priv->dev_remove);
1762
1763 /*
1764 * Grab mutex to block rdma_destroy_id() from removing the device while
1765 * we're trying to acquire it.
1766 */
1767 mutex_lock(&lock);
1768 if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1769 mutex_unlock(&lock);
1770 goto out;
1771 }
1772
1773 if (!status && !id_priv->cma_dev)
1774 status = cma_acquire_dev(id_priv);
1775 mutex_unlock(&lock);
1776
1777 if (status) {
1778 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1779 goto out;
1780 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1781 event.status = status;
1782 } else {
1783 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1784 ip_addr_size(src_addr));
1785 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1786 }
1787
1788 if (id_priv->id.event_handler(&id_priv->id, &event)) {
1789 cma_exch(id_priv, CMA_DESTROYING);
1790 cma_enable_remove(id_priv);
1791 cma_deref_id(id_priv);
1792 rdma_destroy_id(&id_priv->id);
1793 return;
1794 }
1795 out:
1796 cma_enable_remove(id_priv);
1797 cma_deref_id(id_priv);
1798 }
1799
1800 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1801 {
1802 struct cma_work *work;
1803 struct sockaddr_in *src_in, *dst_in;
1804 union ib_gid gid;
1805 int ret;
1806
1807 work = kzalloc(sizeof *work, GFP_KERNEL);
1808 if (!work)
1809 return -ENOMEM;
1810
1811 if (!id_priv->cma_dev) {
1812 ret = cma_bind_loopback(id_priv);
1813 if (ret)
1814 goto err;
1815 }
1816
1817 ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1818 ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1819
1820 if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) {
1821 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1822 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1823 src_in->sin_family = dst_in->sin_family;
1824 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1825 }
1826
1827 work->id = id_priv;
1828 INIT_WORK(&work->work, cma_work_handler);
1829 work->old_state = CMA_ADDR_QUERY;
1830 work->new_state = CMA_ADDR_RESOLVED;
1831 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1832 queue_work(cma_wq, &work->work);
1833 return 0;
1834 err:
1835 kfree(work);
1836 return ret;
1837 }
1838
1839 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1840 struct sockaddr *dst_addr)
1841 {
1842 if (src_addr && src_addr->sa_family)
1843 return rdma_bind_addr(id, src_addr);
1844 else
1845 return cma_bind_any(id, dst_addr->sa_family);
1846 }
1847
1848 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1849 struct sockaddr *dst_addr, int timeout_ms)
1850 {
1851 struct rdma_id_private *id_priv;
1852 int ret;
1853
1854 id_priv = container_of(id, struct rdma_id_private, id);
1855 if (id_priv->state == CMA_IDLE) {
1856 ret = cma_bind_addr(id, src_addr, dst_addr);
1857 if (ret)
1858 return ret;
1859 }
1860
1861 if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1862 return -EINVAL;
1863
1864 atomic_inc(&id_priv->refcount);
1865 memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1866 if (cma_any_addr(dst_addr))
1867 ret = cma_resolve_loopback(id_priv);
1868 else
1869 ret = rdma_resolve_ip(&addr_client, &id->route.addr.src_addr,
1870 dst_addr, &id->route.addr.dev_addr,
1871 timeout_ms, addr_handler, id_priv);
1872 if (ret)
1873 goto err;
1874
1875 return 0;
1876 err:
1877 cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1878 cma_deref_id(id_priv);
1879 return ret;
1880 }
1881 EXPORT_SYMBOL(rdma_resolve_addr);
1882
1883 static void cma_bind_port(struct rdma_bind_list *bind_list,
1884 struct rdma_id_private *id_priv)
1885 {
1886 struct sockaddr_in *sin;
1887
1888 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1889 sin->sin_port = htons(bind_list->port);
1890 id_priv->bind_list = bind_list;
1891 hlist_add_head(&id_priv->node, &bind_list->owners);
1892 }
1893
1894 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1895 unsigned short snum)
1896 {
1897 struct rdma_bind_list *bind_list;
1898 int port, ret;
1899
1900 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1901 if (!bind_list)
1902 return -ENOMEM;
1903
1904 do {
1905 ret = idr_get_new_above(ps, bind_list, snum, &port);
1906 } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1907
1908 if (ret)
1909 goto err1;
1910
1911 if (port != snum) {
1912 ret = -EADDRNOTAVAIL;
1913 goto err2;
1914 }
1915
1916 bind_list->ps = ps;
1917 bind_list->port = (unsigned short) port;
1918 cma_bind_port(bind_list, id_priv);
1919 return 0;
1920 err2:
1921 idr_remove(ps, port);
1922 err1:
1923 kfree(bind_list);
1924 return ret;
1925 }
1926
1927 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
1928 {
1929 struct rdma_bind_list *bind_list;
1930 int port, ret, low, high;
1931
1932 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1933 if (!bind_list)
1934 return -ENOMEM;
1935
1936 retry:
1937 /* FIXME: add proper port randomization per like inet_csk_get_port */
1938 do {
1939 ret = idr_get_new_above(ps, bind_list, next_port, &port);
1940 } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1941
1942 if (ret)
1943 goto err1;
1944
1945 inet_get_local_port_range(&low, &high);
1946 if (port > high) {
1947 if (next_port != low) {
1948 idr_remove(ps, port);
1949 next_port = low;
1950 goto retry;
1951 }
1952 ret = -EADDRNOTAVAIL;
1953 goto err2;
1954 }
1955
1956 if (port == high)
1957 next_port = low;
1958 else
1959 next_port = port + 1;
1960
1961 bind_list->ps = ps;
1962 bind_list->port = (unsigned short) port;
1963 cma_bind_port(bind_list, id_priv);
1964 return 0;
1965 err2:
1966 idr_remove(ps, port);
1967 err1:
1968 kfree(bind_list);
1969 return ret;
1970 }
1971
1972 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
1973 {
1974 struct rdma_id_private *cur_id;
1975 struct sockaddr_in *sin, *cur_sin;
1976 struct rdma_bind_list *bind_list;
1977 struct hlist_node *node;
1978 unsigned short snum;
1979
1980 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1981 snum = ntohs(sin->sin_port);
1982 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
1983 return -EACCES;
1984
1985 bind_list = idr_find(ps, snum);
1986 if (!bind_list)
1987 return cma_alloc_port(ps, id_priv, snum);
1988
1989 /*
1990 * We don't support binding to any address if anyone is bound to
1991 * a specific address on the same port.
1992 */
1993 if (cma_any_addr(&id_priv->id.route.addr.src_addr))
1994 return -EADDRNOTAVAIL;
1995
1996 hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
1997 if (cma_any_addr(&cur_id->id.route.addr.src_addr))
1998 return -EADDRNOTAVAIL;
1999
2000 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
2001 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
2002 return -EADDRINUSE;
2003 }
2004
2005 cma_bind_port(bind_list, id_priv);
2006 return 0;
2007 }
2008
2009 static int cma_get_port(struct rdma_id_private *id_priv)
2010 {
2011 struct idr *ps;
2012 int ret;
2013
2014 switch (id_priv->id.ps) {
2015 case RDMA_PS_SDP:
2016 ps = &sdp_ps;
2017 break;
2018 case RDMA_PS_TCP:
2019 ps = &tcp_ps;
2020 break;
2021 case RDMA_PS_UDP:
2022 ps = &udp_ps;
2023 break;
2024 case RDMA_PS_IPOIB:
2025 ps = &ipoib_ps;
2026 break;
2027 default:
2028 return -EPROTONOSUPPORT;
2029 }
2030
2031 mutex_lock(&lock);
2032 if (cma_any_port(&id_priv->id.route.addr.src_addr))
2033 ret = cma_alloc_any_port(ps, id_priv);
2034 else
2035 ret = cma_use_port(ps, id_priv);
2036 mutex_unlock(&lock);
2037
2038 return ret;
2039 }
2040
2041 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2042 {
2043 struct rdma_id_private *id_priv;
2044 int ret;
2045
2046 if (addr->sa_family != AF_INET)
2047 return -EAFNOSUPPORT;
2048
2049 id_priv = container_of(id, struct rdma_id_private, id);
2050 if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
2051 return -EINVAL;
2052
2053 if (!cma_any_addr(addr)) {
2054 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2055 if (ret)
2056 goto err1;
2057
2058 mutex_lock(&lock);
2059 ret = cma_acquire_dev(id_priv);
2060 mutex_unlock(&lock);
2061 if (ret)
2062 goto err1;
2063 }
2064
2065 memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2066 ret = cma_get_port(id_priv);
2067 if (ret)
2068 goto err2;
2069
2070 return 0;
2071 err2:
2072 if (!cma_any_addr(addr)) {
2073 mutex_lock(&lock);
2074 cma_detach_from_dev(id_priv);
2075 mutex_unlock(&lock);
2076 }
2077 err1:
2078 cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
2079 return ret;
2080 }
2081 EXPORT_SYMBOL(rdma_bind_addr);
2082
2083 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2084 struct rdma_route *route)
2085 {
2086 struct sockaddr_in *src4, *dst4;
2087 struct cma_hdr *cma_hdr;
2088 struct sdp_hh *sdp_hdr;
2089
2090 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2091 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2092
2093 switch (ps) {
2094 case RDMA_PS_SDP:
2095 sdp_hdr = hdr;
2096 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2097 return -EINVAL;
2098 sdp_set_ip_ver(sdp_hdr, 4);
2099 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2100 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2101 sdp_hdr->port = src4->sin_port;
2102 break;
2103 default:
2104 cma_hdr = hdr;
2105 cma_hdr->cma_version = CMA_VERSION;
2106 cma_set_ip_ver(cma_hdr, 4);
2107 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2108 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2109 cma_hdr->port = src4->sin_port;
2110 break;
2111 }
2112 return 0;
2113 }
2114
2115 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2116 struct ib_cm_event *ib_event)
2117 {
2118 struct rdma_id_private *id_priv = cm_id->context;
2119 struct rdma_cm_event event;
2120 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2121 int ret = 0;
2122
2123 if (cma_disable_remove(id_priv, CMA_CONNECT))
2124 return 0;
2125
2126 memset(&event, 0, sizeof event);
2127 switch (ib_event->event) {
2128 case IB_CM_SIDR_REQ_ERROR:
2129 event.event = RDMA_CM_EVENT_UNREACHABLE;
2130 event.status = -ETIMEDOUT;
2131 break;
2132 case IB_CM_SIDR_REP_RECEIVED:
2133 event.param.ud.private_data = ib_event->private_data;
2134 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2135 if (rep->status != IB_SIDR_SUCCESS) {
2136 event.event = RDMA_CM_EVENT_UNREACHABLE;
2137 event.status = ib_event->param.sidr_rep_rcvd.status;
2138 break;
2139 }
2140 if (id_priv->qkey != rep->qkey) {
2141 event.event = RDMA_CM_EVENT_UNREACHABLE;
2142 event.status = -EINVAL;
2143 break;
2144 }
2145 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2146 id_priv->id.route.path_rec,
2147 &event.param.ud.ah_attr);
2148 event.param.ud.qp_num = rep->qpn;
2149 event.param.ud.qkey = rep->qkey;
2150 event.event = RDMA_CM_EVENT_ESTABLISHED;
2151 event.status = 0;
2152 break;
2153 default:
2154 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
2155 ib_event->event);
2156 goto out;
2157 }
2158
2159 ret = id_priv->id.event_handler(&id_priv->id, &event);
2160 if (ret) {
2161 /* Destroy the CM ID by returning a non-zero value. */
2162 id_priv->cm_id.ib = NULL;
2163 cma_exch(id_priv, CMA_DESTROYING);
2164 cma_enable_remove(id_priv);
2165 rdma_destroy_id(&id_priv->id);
2166 return ret;
2167 }
2168 out:
2169 cma_enable_remove(id_priv);
2170 return ret;
2171 }
2172
2173 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2174 struct rdma_conn_param *conn_param)
2175 {
2176 struct ib_cm_sidr_req_param req;
2177 struct rdma_route *route;
2178 int ret;
2179
2180 req.private_data_len = sizeof(struct cma_hdr) +
2181 conn_param->private_data_len;
2182 req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2183 if (!req.private_data)
2184 return -ENOMEM;
2185
2186 if (conn_param->private_data && conn_param->private_data_len)
2187 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2188 conn_param->private_data, conn_param->private_data_len);
2189
2190 route = &id_priv->id.route;
2191 ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2192 if (ret)
2193 goto out;
2194
2195 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2196 cma_sidr_rep_handler, id_priv);
2197 if (IS_ERR(id_priv->cm_id.ib)) {
2198 ret = PTR_ERR(id_priv->cm_id.ib);
2199 goto out;
2200 }
2201
2202 req.path = route->path_rec;
2203 req.service_id = cma_get_service_id(id_priv->id.ps,
2204 &route->addr.dst_addr);
2205 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2206 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2207
2208 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2209 if (ret) {
2210 ib_destroy_cm_id(id_priv->cm_id.ib);
2211 id_priv->cm_id.ib = NULL;
2212 }
2213 out:
2214 kfree(req.private_data);
2215 return ret;
2216 }
2217
2218 static int cma_connect_ib(struct rdma_id_private *id_priv,
2219 struct rdma_conn_param *conn_param)
2220 {
2221 struct ib_cm_req_param req;
2222 struct rdma_route *route;
2223 void *private_data;
2224 int offset, ret;
2225
2226 memset(&req, 0, sizeof req);
2227 offset = cma_user_data_offset(id_priv->id.ps);
2228 req.private_data_len = offset + conn_param->private_data_len;
2229 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2230 if (!private_data)
2231 return -ENOMEM;
2232
2233 if (conn_param->private_data && conn_param->private_data_len)
2234 memcpy(private_data + offset, conn_param->private_data,
2235 conn_param->private_data_len);
2236
2237 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2238 id_priv);
2239 if (IS_ERR(id_priv->cm_id.ib)) {
2240 ret = PTR_ERR(id_priv->cm_id.ib);
2241 goto out;
2242 }
2243
2244 route = &id_priv->id.route;
2245 ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2246 if (ret)
2247 goto out;
2248 req.private_data = private_data;
2249
2250 req.primary_path = &route->path_rec[0];
2251 if (route->num_paths == 2)
2252 req.alternate_path = &route->path_rec[1];
2253
2254 req.service_id = cma_get_service_id(id_priv->id.ps,
2255 &route->addr.dst_addr);
2256 req.qp_num = id_priv->qp_num;
2257 req.qp_type = IB_QPT_RC;
2258 req.starting_psn = id_priv->seq_num;
2259 req.responder_resources = conn_param->responder_resources;
2260 req.initiator_depth = conn_param->initiator_depth;
2261 req.flow_control = conn_param->flow_control;
2262 req.retry_count = conn_param->retry_count;
2263 req.rnr_retry_count = conn_param->rnr_retry_count;
2264 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2265 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2266 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2267 req.srq = id_priv->srq ? 1 : 0;
2268
2269 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2270 out:
2271 if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2272 ib_destroy_cm_id(id_priv->cm_id.ib);
2273 id_priv->cm_id.ib = NULL;
2274 }
2275
2276 kfree(private_data);
2277 return ret;
2278 }
2279
2280 static int cma_connect_iw(struct rdma_id_private *id_priv,
2281 struct rdma_conn_param *conn_param)
2282 {
2283 struct iw_cm_id *cm_id;
2284 struct sockaddr_in* sin;
2285 int ret;
2286 struct iw_cm_conn_param iw_param;
2287
2288 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2289 if (IS_ERR(cm_id)) {
2290 ret = PTR_ERR(cm_id);
2291 goto out;
2292 }
2293
2294 id_priv->cm_id.iw = cm_id;
2295
2296 sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2297 cm_id->local_addr = *sin;
2298
2299 sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2300 cm_id->remote_addr = *sin;
2301
2302 ret = cma_modify_qp_rtr(id_priv, conn_param);
2303 if (ret)
2304 goto out;
2305
2306 iw_param.ord = conn_param->initiator_depth;
2307 iw_param.ird = conn_param->responder_resources;
2308 iw_param.private_data = conn_param->private_data;
2309 iw_param.private_data_len = conn_param->private_data_len;
2310 if (id_priv->id.qp)
2311 iw_param.qpn = id_priv->qp_num;
2312 else
2313 iw_param.qpn = conn_param->qp_num;
2314 ret = iw_cm_connect(cm_id, &iw_param);
2315 out:
2316 if (ret && !IS_ERR(cm_id)) {
2317 iw_destroy_cm_id(cm_id);
2318 id_priv->cm_id.iw = NULL;
2319 }
2320 return ret;
2321 }
2322
2323 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2324 {
2325 struct rdma_id_private *id_priv;
2326 int ret;
2327
2328 id_priv = container_of(id, struct rdma_id_private, id);
2329 if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2330 return -EINVAL;
2331
2332 if (!id->qp) {
2333 id_priv->qp_num = conn_param->qp_num;
2334 id_priv->srq = conn_param->srq;
2335 }
2336
2337 switch (rdma_node_get_transport(id->device->node_type)) {
2338 case RDMA_TRANSPORT_IB:
2339 if (cma_is_ud_ps(id->ps))
2340 ret = cma_resolve_ib_udp(id_priv, conn_param);
2341 else
2342 ret = cma_connect_ib(id_priv, conn_param);
2343 break;
2344 case RDMA_TRANSPORT_IWARP:
2345 ret = cma_connect_iw(id_priv, conn_param);
2346 break;
2347 default:
2348 ret = -ENOSYS;
2349 break;
2350 }
2351 if (ret)
2352 goto err;
2353
2354 return 0;
2355 err:
2356 cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2357 return ret;
2358 }
2359 EXPORT_SYMBOL(rdma_connect);
2360
2361 static int cma_accept_ib(struct rdma_id_private *id_priv,
2362 struct rdma_conn_param *conn_param)
2363 {
2364 struct ib_cm_rep_param rep;
2365 int ret;
2366
2367 ret = cma_modify_qp_rtr(id_priv, conn_param);
2368 if (ret)
2369 goto out;
2370
2371 ret = cma_modify_qp_rts(id_priv, conn_param);
2372 if (ret)
2373 goto out;
2374
2375 memset(&rep, 0, sizeof rep);
2376 rep.qp_num = id_priv->qp_num;
2377 rep.starting_psn = id_priv->seq_num;
2378 rep.private_data = conn_param->private_data;
2379 rep.private_data_len = conn_param->private_data_len;
2380 rep.responder_resources = conn_param->responder_resources;
2381 rep.initiator_depth = conn_param->initiator_depth;
2382 rep.failover_accepted = 0;
2383 rep.flow_control = conn_param->flow_control;
2384 rep.rnr_retry_count = conn_param->rnr_retry_count;
2385 rep.srq = id_priv->srq ? 1 : 0;
2386
2387 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2388 out:
2389 return ret;
2390 }
2391
2392 static int cma_accept_iw(struct rdma_id_private *id_priv,
2393 struct rdma_conn_param *conn_param)
2394 {
2395 struct iw_cm_conn_param iw_param;
2396 int ret;
2397
2398 ret = cma_modify_qp_rtr(id_priv, conn_param);
2399 if (ret)
2400 return ret;
2401
2402 iw_param.ord = conn_param->initiator_depth;
2403 iw_param.ird = conn_param->responder_resources;
2404 iw_param.private_data = conn_param->private_data;
2405 iw_param.private_data_len = conn_param->private_data_len;
2406 if (id_priv->id.qp) {
2407 iw_param.qpn = id_priv->qp_num;
2408 } else
2409 iw_param.qpn = conn_param->qp_num;
2410
2411 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2412 }
2413
2414 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2415 enum ib_cm_sidr_status status,
2416 const void *private_data, int private_data_len)
2417 {
2418 struct ib_cm_sidr_rep_param rep;
2419
2420 memset(&rep, 0, sizeof rep);
2421 rep.status = status;
2422 if (status == IB_SIDR_SUCCESS) {
2423 rep.qp_num = id_priv->qp_num;
2424 rep.qkey = id_priv->qkey;
2425 }
2426 rep.private_data = private_data;
2427 rep.private_data_len = private_data_len;
2428
2429 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2430 }
2431
2432 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2433 {
2434 struct rdma_id_private *id_priv;
2435 int ret;
2436
2437 id_priv = container_of(id, struct rdma_id_private, id);
2438 if (!cma_comp(id_priv, CMA_CONNECT))
2439 return -EINVAL;
2440
2441 if (!id->qp && conn_param) {
2442 id_priv->qp_num = conn_param->qp_num;
2443 id_priv->srq = conn_param->srq;
2444 }
2445
2446 switch (rdma_node_get_transport(id->device->node_type)) {
2447 case RDMA_TRANSPORT_IB:
2448 if (cma_is_ud_ps(id->ps))
2449 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2450 conn_param->private_data,
2451 conn_param->private_data_len);
2452 else if (conn_param)
2453 ret = cma_accept_ib(id_priv, conn_param);
2454 else
2455 ret = cma_rep_recv(id_priv);
2456 break;
2457 case RDMA_TRANSPORT_IWARP:
2458 ret = cma_accept_iw(id_priv, conn_param);
2459 break;
2460 default:
2461 ret = -ENOSYS;
2462 break;
2463 }
2464
2465 if (ret)
2466 goto reject;
2467
2468 return 0;
2469 reject:
2470 cma_modify_qp_err(id_priv);
2471 rdma_reject(id, NULL, 0);
2472 return ret;
2473 }
2474 EXPORT_SYMBOL(rdma_accept);
2475
2476 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2477 {
2478 struct rdma_id_private *id_priv;
2479 int ret;
2480
2481 id_priv = container_of(id, struct rdma_id_private, id);
2482 if (!cma_has_cm_dev(id_priv))
2483 return -EINVAL;
2484
2485 switch (id->device->node_type) {
2486 case RDMA_NODE_IB_CA:
2487 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2488 break;
2489 default:
2490 ret = 0;
2491 break;
2492 }
2493 return ret;
2494 }
2495 EXPORT_SYMBOL(rdma_notify);
2496
2497 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2498 u8 private_data_len)
2499 {
2500 struct rdma_id_private *id_priv;
2501 int ret;
2502
2503 id_priv = container_of(id, struct rdma_id_private, id);
2504 if (!cma_has_cm_dev(id_priv))
2505 return -EINVAL;
2506
2507 switch (rdma_node_get_transport(id->device->node_type)) {
2508 case RDMA_TRANSPORT_IB:
2509 if (cma_is_ud_ps(id->ps))
2510 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2511 private_data, private_data_len);
2512 else
2513 ret = ib_send_cm_rej(id_priv->cm_id.ib,
2514 IB_CM_REJ_CONSUMER_DEFINED, NULL,
2515 0, private_data, private_data_len);
2516 break;
2517 case RDMA_TRANSPORT_IWARP:
2518 ret = iw_cm_reject(id_priv->cm_id.iw,
2519 private_data, private_data_len);
2520 break;
2521 default:
2522 ret = -ENOSYS;
2523 break;
2524 }
2525 return ret;
2526 }
2527 EXPORT_SYMBOL(rdma_reject);
2528
2529 int rdma_disconnect(struct rdma_cm_id *id)
2530 {
2531 struct rdma_id_private *id_priv;
2532 int ret;
2533
2534 id_priv = container_of(id, struct rdma_id_private, id);
2535 if (!cma_has_cm_dev(id_priv))
2536 return -EINVAL;
2537
2538 switch (rdma_node_get_transport(id->device->node_type)) {
2539 case RDMA_TRANSPORT_IB:
2540 ret = cma_modify_qp_err(id_priv);
2541 if (ret)
2542 goto out;
2543 /* Initiate or respond to a disconnect. */
2544 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2545 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2546 break;
2547 case RDMA_TRANSPORT_IWARP:
2548 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2549 break;
2550 default:
2551 ret = -EINVAL;
2552 break;
2553 }
2554 out:
2555 return ret;
2556 }
2557 EXPORT_SYMBOL(rdma_disconnect);
2558
2559 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2560 {
2561 struct rdma_id_private *id_priv;
2562 struct cma_multicast *mc = multicast->context;
2563 struct rdma_cm_event event;
2564 int ret;
2565
2566 id_priv = mc->id_priv;
2567 if (cma_disable_remove(id_priv, CMA_ADDR_BOUND) &&
2568 cma_disable_remove(id_priv, CMA_ADDR_RESOLVED))
2569 return 0;
2570
2571 mutex_lock(&id_priv->qp_mutex);
2572 if (!status && id_priv->id.qp)
2573 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2574 multicast->rec.mlid);
2575 mutex_unlock(&id_priv->qp_mutex);
2576
2577 memset(&event, 0, sizeof event);
2578 event.status = status;
2579 event.param.ud.private_data = mc->context;
2580 if (!status) {
2581 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2582 ib_init_ah_from_mcmember(id_priv->id.device,
2583 id_priv->id.port_num, &multicast->rec,
2584 &event.param.ud.ah_attr);
2585 event.param.ud.qp_num = 0xFFFFFF;
2586 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2587 } else
2588 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2589
2590 ret = id_priv->id.event_handler(&id_priv->id, &event);
2591 if (ret) {
2592 cma_exch(id_priv, CMA_DESTROYING);
2593 cma_enable_remove(id_priv);
2594 rdma_destroy_id(&id_priv->id);
2595 return 0;
2596 }
2597
2598 cma_enable_remove(id_priv);
2599 return 0;
2600 }
2601
2602 static void cma_set_mgid(struct rdma_id_private *id_priv,
2603 struct sockaddr *addr, union ib_gid *mgid)
2604 {
2605 unsigned char mc_map[MAX_ADDR_LEN];
2606 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2607 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2608 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2609
2610 if (cma_any_addr(addr)) {
2611 memset(mgid, 0, sizeof *mgid);
2612 } else if ((addr->sa_family == AF_INET6) &&
2613 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFF10A01B) ==
2614 0xFF10A01B)) {
2615 /* IPv6 address is an SA assigned MGID. */
2616 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2617 } else {
2618 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2619 if (id_priv->id.ps == RDMA_PS_UDP)
2620 mc_map[7] = 0x01; /* Use RDMA CM signature */
2621 *mgid = *(union ib_gid *) (mc_map + 4);
2622 }
2623 }
2624
2625 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2626 struct cma_multicast *mc)
2627 {
2628 struct ib_sa_mcmember_rec rec;
2629 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2630 ib_sa_comp_mask comp_mask;
2631 int ret;
2632
2633 ib_addr_get_mgid(dev_addr, &rec.mgid);
2634 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2635 &rec.mgid, &rec);
2636 if (ret)
2637 return ret;
2638
2639 cma_set_mgid(id_priv, &mc->addr, &rec.mgid);
2640 if (id_priv->id.ps == RDMA_PS_UDP)
2641 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2642 ib_addr_get_sgid(dev_addr, &rec.port_gid);
2643 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2644 rec.join_state = 1;
2645
2646 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2647 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2648 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2649 IB_SA_MCMEMBER_REC_FLOW_LABEL |
2650 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2651
2652 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2653 id_priv->id.port_num, &rec,
2654 comp_mask, GFP_KERNEL,
2655 cma_ib_mc_handler, mc);
2656 if (IS_ERR(mc->multicast.ib))
2657 return PTR_ERR(mc->multicast.ib);
2658
2659 return 0;
2660 }
2661
2662 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
2663 void *context)
2664 {
2665 struct rdma_id_private *id_priv;
2666 struct cma_multicast *mc;
2667 int ret;
2668
2669 id_priv = container_of(id, struct rdma_id_private, id);
2670 if (!cma_comp(id_priv, CMA_ADDR_BOUND) &&
2671 !cma_comp(id_priv, CMA_ADDR_RESOLVED))
2672 return -EINVAL;
2673
2674 mc = kmalloc(sizeof *mc, GFP_KERNEL);
2675 if (!mc)
2676 return -ENOMEM;
2677
2678 memcpy(&mc->addr, addr, ip_addr_size(addr));
2679 mc->context = context;
2680 mc->id_priv = id_priv;
2681
2682 spin_lock(&id_priv->lock);
2683 list_add(&mc->list, &id_priv->mc_list);
2684 spin_unlock(&id_priv->lock);
2685
2686 switch (rdma_node_get_transport(id->device->node_type)) {
2687 case RDMA_TRANSPORT_IB:
2688 ret = cma_join_ib_multicast(id_priv, mc);
2689 break;
2690 default:
2691 ret = -ENOSYS;
2692 break;
2693 }
2694
2695 if (ret) {
2696 spin_lock_irq(&id_priv->lock);
2697 list_del(&mc->list);
2698 spin_unlock_irq(&id_priv->lock);
2699 kfree(mc);
2700 }
2701 return ret;
2702 }
2703 EXPORT_SYMBOL(rdma_join_multicast);
2704
2705 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
2706 {
2707 struct rdma_id_private *id_priv;
2708 struct cma_multicast *mc;
2709
2710 id_priv = container_of(id, struct rdma_id_private, id);
2711 spin_lock_irq(&id_priv->lock);
2712 list_for_each_entry(mc, &id_priv->mc_list, list) {
2713 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
2714 list_del(&mc->list);
2715 spin_unlock_irq(&id_priv->lock);
2716
2717 if (id->qp)
2718 ib_detach_mcast(id->qp,
2719 &mc->multicast.ib->rec.mgid,
2720 mc->multicast.ib->rec.mlid);
2721 ib_sa_free_multicast(mc->multicast.ib);
2722 kfree(mc);
2723 return;
2724 }
2725 }
2726 spin_unlock_irq(&id_priv->lock);
2727 }
2728 EXPORT_SYMBOL(rdma_leave_multicast);
2729
2730 static void cma_add_one(struct ib_device *device)
2731 {
2732 struct cma_device *cma_dev;
2733 struct rdma_id_private *id_priv;
2734
2735 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2736 if (!cma_dev)
2737 return;
2738
2739 cma_dev->device = device;
2740
2741 init_completion(&cma_dev->comp);
2742 atomic_set(&cma_dev->refcount, 1);
2743 INIT_LIST_HEAD(&cma_dev->id_list);
2744 ib_set_client_data(device, &cma_client, cma_dev);
2745
2746 mutex_lock(&lock);
2747 list_add_tail(&cma_dev->list, &dev_list);
2748 list_for_each_entry(id_priv, &listen_any_list, list)
2749 cma_listen_on_dev(id_priv, cma_dev);
2750 mutex_unlock(&lock);
2751 }
2752
2753 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2754 {
2755 struct rdma_cm_event event;
2756 enum cma_state state;
2757
2758 /* Record that we want to remove the device */
2759 state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2760 if (state == CMA_DESTROYING)
2761 return 0;
2762
2763 cma_cancel_operation(id_priv, state);
2764 wait_event(id_priv->wait_remove, !atomic_read(&id_priv->dev_remove));
2765
2766 /* Check for destruction from another callback. */
2767 if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2768 return 0;
2769
2770 memset(&event, 0, sizeof event);
2771 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
2772 return id_priv->id.event_handler(&id_priv->id, &event);
2773 }
2774
2775 static void cma_process_remove(struct cma_device *cma_dev)
2776 {
2777 struct rdma_id_private *id_priv;
2778 int ret;
2779
2780 mutex_lock(&lock);
2781 while (!list_empty(&cma_dev->id_list)) {
2782 id_priv = list_entry(cma_dev->id_list.next,
2783 struct rdma_id_private, list);
2784
2785 list_del(&id_priv->listen_list);
2786 list_del_init(&id_priv->list);
2787 atomic_inc(&id_priv->refcount);
2788 mutex_unlock(&lock);
2789
2790 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
2791 cma_deref_id(id_priv);
2792 if (ret)
2793 rdma_destroy_id(&id_priv->id);
2794
2795 mutex_lock(&lock);
2796 }
2797 mutex_unlock(&lock);
2798
2799 cma_deref_dev(cma_dev);
2800 wait_for_completion(&cma_dev->comp);
2801 }
2802
2803 static void cma_remove_one(struct ib_device *device)
2804 {
2805 struct cma_device *cma_dev;
2806
2807 cma_dev = ib_get_client_data(device, &cma_client);
2808 if (!cma_dev)
2809 return;
2810
2811 mutex_lock(&lock);
2812 list_del(&cma_dev->list);
2813 mutex_unlock(&lock);
2814
2815 cma_process_remove(cma_dev);
2816 kfree(cma_dev);
2817 }
2818
2819 static int cma_init(void)
2820 {
2821 int ret, low, high, remaining;
2822
2823 get_random_bytes(&next_port, sizeof next_port);
2824 inet_get_local_port_range(&low, &high);
2825 remaining = (high - low) + 1;
2826 next_port = ((unsigned int) next_port % remaining) + low;
2827
2828 cma_wq = create_singlethread_workqueue("rdma_cm");
2829 if (!cma_wq)
2830 return -ENOMEM;
2831
2832 ib_sa_register_client(&sa_client);
2833 rdma_addr_register_client(&addr_client);
2834
2835 ret = ib_register_client(&cma_client);
2836 if (ret)
2837 goto err;
2838 return 0;
2839
2840 err:
2841 rdma_addr_unregister_client(&addr_client);
2842 ib_sa_unregister_client(&sa_client);
2843 destroy_workqueue(cma_wq);
2844 return ret;
2845 }
2846
2847 static void cma_cleanup(void)
2848 {
2849 ib_unregister_client(&cma_client);
2850 rdma_addr_unregister_client(&addr_client);
2851 ib_sa_unregister_client(&sa_client);
2852 destroy_workqueue(cma_wq);
2853 idr_destroy(&sdp_ps);
2854 idr_destroy(&tcp_ps);
2855 idr_destroy(&udp_ps);
2856 idr_destroy(&ipoib_ps);
2857 }
2858
2859 module_init(cma_init);
2860 module_exit(cma_cleanup);
mah project overlay