xref: /openbmc/linux/drivers/infiniband/core/cma.c (revision 261a9af6)
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/rdma_netlink.h>
51 #include <rdma/ib_cache.h>
52 #include <rdma/ib_cm.h>
53 #include <rdma/ib_sa.h>
54 #include <rdma/iw_cm.h>
55 
56 MODULE_AUTHOR("Sean Hefty");
57 MODULE_DESCRIPTION("Generic RDMA CM Agent");
58 MODULE_LICENSE("Dual BSD/GPL");
59 
60 #define CMA_CM_RESPONSE_TIMEOUT 20
61 #define CMA_MAX_CM_RETRIES 15
62 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
63 #define CMA_IBOE_PACKET_LIFETIME 18
64 
65 static void cma_add_one(struct ib_device *device);
66 static void cma_remove_one(struct ib_device *device);
67 
68 static struct ib_client cma_client = {
69 	.name   = "cma",
70 	.add    = cma_add_one,
71 	.remove = cma_remove_one
72 };
73 
74 static struct ib_sa_client sa_client;
75 static struct rdma_addr_client addr_client;
76 static LIST_HEAD(dev_list);
77 static LIST_HEAD(listen_any_list);
78 static DEFINE_MUTEX(lock);
79 static struct workqueue_struct *cma_wq;
80 static DEFINE_IDR(sdp_ps);
81 static DEFINE_IDR(tcp_ps);
82 static DEFINE_IDR(udp_ps);
83 static DEFINE_IDR(ipoib_ps);
84 
85 struct cma_device {
86 	struct list_head	list;
87 	struct ib_device	*device;
88 	struct completion	comp;
89 	atomic_t		refcount;
90 	struct list_head	id_list;
91 };
92 
93 struct rdma_bind_list {
94 	struct idr		*ps;
95 	struct hlist_head	owners;
96 	unsigned short		port;
97 };
98 
99 /*
100  * Device removal can occur at anytime, so we need extra handling to
101  * serialize notifying the user of device removal with other callbacks.
102  * We do this by disabling removal notification while a callback is in process,
103  * and reporting it after the callback completes.
104  */
105 struct rdma_id_private {
106 	struct rdma_cm_id	id;
107 
108 	struct rdma_bind_list	*bind_list;
109 	struct hlist_node	node;
110 	struct list_head	list; /* listen_any_list or cma_device.list */
111 	struct list_head	listen_list; /* per device listens */
112 	struct cma_device	*cma_dev;
113 	struct list_head	mc_list;
114 
115 	int			internal_id;
116 	enum rdma_cm_state	state;
117 	spinlock_t		lock;
118 	struct mutex		qp_mutex;
119 
120 	struct completion	comp;
121 	atomic_t		refcount;
122 	struct mutex		handler_mutex;
123 
124 	int			backlog;
125 	int			timeout_ms;
126 	struct ib_sa_query	*query;
127 	int			query_id;
128 	union {
129 		struct ib_cm_id	*ib;
130 		struct iw_cm_id	*iw;
131 	} cm_id;
132 
133 	u32			seq_num;
134 	u32			qkey;
135 	u32			qp_num;
136 	pid_t			owner;
137 	u8			srq;
138 	u8			tos;
139 	u8			reuseaddr;
140 };
141 
142 struct cma_multicast {
143 	struct rdma_id_private *id_priv;
144 	union {
145 		struct ib_sa_multicast *ib;
146 	} multicast;
147 	struct list_head	list;
148 	void			*context;
149 	struct sockaddr_storage	addr;
150 	struct kref		mcref;
151 };
152 
153 struct cma_work {
154 	struct work_struct	work;
155 	struct rdma_id_private	*id;
156 	enum rdma_cm_state	old_state;
157 	enum rdma_cm_state	new_state;
158 	struct rdma_cm_event	event;
159 };
160 
161 struct cma_ndev_work {
162 	struct work_struct	work;
163 	struct rdma_id_private	*id;
164 	struct rdma_cm_event	event;
165 };
166 
167 struct iboe_mcast_work {
168 	struct work_struct	 work;
169 	struct rdma_id_private	*id;
170 	struct cma_multicast	*mc;
171 };
172 
173 union cma_ip_addr {
174 	struct in6_addr ip6;
175 	struct {
176 		__be32 pad[3];
177 		__be32 addr;
178 	} ip4;
179 };
180 
181 struct cma_hdr {
182 	u8 cma_version;
183 	u8 ip_version;	/* IP version: 7:4 */
184 	__be16 port;
185 	union cma_ip_addr src_addr;
186 	union cma_ip_addr dst_addr;
187 };
188 
189 struct sdp_hh {
190 	u8 bsdh[16];
191 	u8 sdp_version; /* Major version: 7:4 */
192 	u8 ip_version;	/* IP version: 7:4 */
193 	u8 sdp_specific1[10];
194 	__be16 port;
195 	__be16 sdp_specific2;
196 	union cma_ip_addr src_addr;
197 	union cma_ip_addr dst_addr;
198 };
199 
200 struct sdp_hah {
201 	u8 bsdh[16];
202 	u8 sdp_version;
203 };
204 
205 #define CMA_VERSION 0x00
206 #define SDP_MAJ_VERSION 0x2
207 
208 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
209 {
210 	unsigned long flags;
211 	int ret;
212 
213 	spin_lock_irqsave(&id_priv->lock, flags);
214 	ret = (id_priv->state == comp);
215 	spin_unlock_irqrestore(&id_priv->lock, flags);
216 	return ret;
217 }
218 
219 static int cma_comp_exch(struct rdma_id_private *id_priv,
220 			 enum rdma_cm_state comp, enum rdma_cm_state exch)
221 {
222 	unsigned long flags;
223 	int ret;
224 
225 	spin_lock_irqsave(&id_priv->lock, flags);
226 	if ((ret = (id_priv->state == comp)))
227 		id_priv->state = exch;
228 	spin_unlock_irqrestore(&id_priv->lock, flags);
229 	return ret;
230 }
231 
232 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
233 				   enum rdma_cm_state exch)
234 {
235 	unsigned long flags;
236 	enum rdma_cm_state old;
237 
238 	spin_lock_irqsave(&id_priv->lock, flags);
239 	old = id_priv->state;
240 	id_priv->state = exch;
241 	spin_unlock_irqrestore(&id_priv->lock, flags);
242 	return old;
243 }
244 
245 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
246 {
247 	return hdr->ip_version >> 4;
248 }
249 
250 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
251 {
252 	hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
253 }
254 
255 static inline u8 sdp_get_majv(u8 sdp_version)
256 {
257 	return sdp_version >> 4;
258 }
259 
260 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
261 {
262 	return hh->ip_version >> 4;
263 }
264 
265 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
266 {
267 	hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
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 	id_priv->id.route.addr.dev_addr.transport =
277 		rdma_node_get_transport(cma_dev->device->node_type);
278 	list_add_tail(&id_priv->list, &cma_dev->id_list);
279 }
280 
281 static inline void cma_deref_dev(struct cma_device *cma_dev)
282 {
283 	if (atomic_dec_and_test(&cma_dev->refcount))
284 		complete(&cma_dev->comp);
285 }
286 
287 static inline void release_mc(struct kref *kref)
288 {
289 	struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
290 
291 	kfree(mc->multicast.ib);
292 	kfree(mc);
293 }
294 
295 static void cma_release_dev(struct rdma_id_private *id_priv)
296 {
297 	mutex_lock(&lock);
298 	list_del(&id_priv->list);
299 	cma_deref_dev(id_priv->cma_dev);
300 	id_priv->cma_dev = NULL;
301 	mutex_unlock(&lock);
302 }
303 
304 static int cma_set_qkey(struct rdma_id_private *id_priv)
305 {
306 	struct ib_sa_mcmember_rec rec;
307 	int ret = 0;
308 
309 	if (id_priv->qkey)
310 		return 0;
311 
312 	switch (id_priv->id.ps) {
313 	case RDMA_PS_UDP:
314 		id_priv->qkey = RDMA_UDP_QKEY;
315 		break;
316 	case RDMA_PS_IPOIB:
317 		ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
318 		ret = ib_sa_get_mcmember_rec(id_priv->id.device,
319 					     id_priv->id.port_num, &rec.mgid,
320 					     &rec);
321 		if (!ret)
322 			id_priv->qkey = be32_to_cpu(rec.qkey);
323 		break;
324 	default:
325 		break;
326 	}
327 	return ret;
328 }
329 
330 static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
331 {
332 	int i;
333 	int err;
334 	struct ib_port_attr props;
335 	union ib_gid tmp;
336 
337 	err = ib_query_port(device, port_num, &props);
338 	if (err)
339 		return 1;
340 
341 	for (i = 0; i < props.gid_tbl_len; ++i) {
342 		err = ib_query_gid(device, port_num, i, &tmp);
343 		if (err)
344 			return 1;
345 		if (!memcmp(&tmp, gid, sizeof tmp))
346 			return 0;
347 	}
348 
349 	return -EAGAIN;
350 }
351 
352 static int cma_acquire_dev(struct rdma_id_private *id_priv)
353 {
354 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
355 	struct cma_device *cma_dev;
356 	union ib_gid gid, iboe_gid;
357 	int ret = -ENODEV;
358 	u8 port;
359 	enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
360 		IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
361 
362 	mutex_lock(&lock);
363 	iboe_addr_get_sgid(dev_addr, &iboe_gid);
364 	memcpy(&gid, dev_addr->src_dev_addr +
365 	       rdma_addr_gid_offset(dev_addr), sizeof gid);
366 	list_for_each_entry(cma_dev, &dev_list, list) {
367 		for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
368 			if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
369 				if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
370 				    rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
371 					ret = find_gid_port(cma_dev->device, &iboe_gid, port);
372 				else
373 					ret = find_gid_port(cma_dev->device, &gid, port);
374 
375 				if (!ret) {
376 					id_priv->id.port_num = port;
377 					goto out;
378 				} else if (ret == 1)
379 					break;
380 			}
381 		}
382 	}
383 
384 out:
385 	if (!ret)
386 		cma_attach_to_dev(id_priv, cma_dev);
387 
388 	mutex_unlock(&lock);
389 	return ret;
390 }
391 
392 static void cma_deref_id(struct rdma_id_private *id_priv)
393 {
394 	if (atomic_dec_and_test(&id_priv->refcount))
395 		complete(&id_priv->comp);
396 }
397 
398 static int cma_disable_callback(struct rdma_id_private *id_priv,
399 				enum rdma_cm_state state)
400 {
401 	mutex_lock(&id_priv->handler_mutex);
402 	if (id_priv->state != state) {
403 		mutex_unlock(&id_priv->handler_mutex);
404 		return -EINVAL;
405 	}
406 	return 0;
407 }
408 
409 static int cma_has_cm_dev(struct rdma_id_private *id_priv)
410 {
411 	return (id_priv->id.device && id_priv->cm_id.ib);
412 }
413 
414 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
415 				  void *context, enum rdma_port_space ps,
416 				  enum ib_qp_type qp_type)
417 {
418 	struct rdma_id_private *id_priv;
419 
420 	id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
421 	if (!id_priv)
422 		return ERR_PTR(-ENOMEM);
423 
424 	id_priv->owner = task_pid_nr(current);
425 	id_priv->state = RDMA_CM_IDLE;
426 	id_priv->id.context = context;
427 	id_priv->id.event_handler = event_handler;
428 	id_priv->id.ps = ps;
429 	id_priv->id.qp_type = qp_type;
430 	spin_lock_init(&id_priv->lock);
431 	mutex_init(&id_priv->qp_mutex);
432 	init_completion(&id_priv->comp);
433 	atomic_set(&id_priv->refcount, 1);
434 	mutex_init(&id_priv->handler_mutex);
435 	INIT_LIST_HEAD(&id_priv->listen_list);
436 	INIT_LIST_HEAD(&id_priv->mc_list);
437 	get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
438 
439 	return &id_priv->id;
440 }
441 EXPORT_SYMBOL(rdma_create_id);
442 
443 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
444 {
445 	struct ib_qp_attr qp_attr;
446 	int qp_attr_mask, ret;
447 
448 	qp_attr.qp_state = IB_QPS_INIT;
449 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
450 	if (ret)
451 		return ret;
452 
453 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
454 	if (ret)
455 		return ret;
456 
457 	qp_attr.qp_state = IB_QPS_RTR;
458 	ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
459 	if (ret)
460 		return ret;
461 
462 	qp_attr.qp_state = IB_QPS_RTS;
463 	qp_attr.sq_psn = 0;
464 	ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
465 
466 	return ret;
467 }
468 
469 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
470 {
471 	struct ib_qp_attr qp_attr;
472 	int qp_attr_mask, ret;
473 
474 	qp_attr.qp_state = IB_QPS_INIT;
475 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
476 	if (ret)
477 		return ret;
478 
479 	return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
480 }
481 
482 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
483 		   struct ib_qp_init_attr *qp_init_attr)
484 {
485 	struct rdma_id_private *id_priv;
486 	struct ib_qp *qp;
487 	int ret;
488 
489 	id_priv = container_of(id, struct rdma_id_private, id);
490 	if (id->device != pd->device)
491 		return -EINVAL;
492 
493 	qp = ib_create_qp(pd, qp_init_attr);
494 	if (IS_ERR(qp))
495 		return PTR_ERR(qp);
496 
497 	if (id->qp_type == IB_QPT_UD)
498 		ret = cma_init_ud_qp(id_priv, qp);
499 	else
500 		ret = cma_init_conn_qp(id_priv, qp);
501 	if (ret)
502 		goto err;
503 
504 	id->qp = qp;
505 	id_priv->qp_num = qp->qp_num;
506 	id_priv->srq = (qp->srq != NULL);
507 	return 0;
508 err:
509 	ib_destroy_qp(qp);
510 	return ret;
511 }
512 EXPORT_SYMBOL(rdma_create_qp);
513 
514 void rdma_destroy_qp(struct rdma_cm_id *id)
515 {
516 	struct rdma_id_private *id_priv;
517 
518 	id_priv = container_of(id, struct rdma_id_private, id);
519 	mutex_lock(&id_priv->qp_mutex);
520 	ib_destroy_qp(id_priv->id.qp);
521 	id_priv->id.qp = NULL;
522 	mutex_unlock(&id_priv->qp_mutex);
523 }
524 EXPORT_SYMBOL(rdma_destroy_qp);
525 
526 static int cma_modify_qp_rtr(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 	/* Need to update QP attributes from default values. */
539 	qp_attr.qp_state = IB_QPS_INIT;
540 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
541 	if (ret)
542 		goto out;
543 
544 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
545 	if (ret)
546 		goto out;
547 
548 	qp_attr.qp_state = IB_QPS_RTR;
549 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
550 	if (ret)
551 		goto out;
552 
553 	if (conn_param)
554 		qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
555 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
556 out:
557 	mutex_unlock(&id_priv->qp_mutex);
558 	return ret;
559 }
560 
561 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
562 			     struct rdma_conn_param *conn_param)
563 {
564 	struct ib_qp_attr qp_attr;
565 	int qp_attr_mask, ret;
566 
567 	mutex_lock(&id_priv->qp_mutex);
568 	if (!id_priv->id.qp) {
569 		ret = 0;
570 		goto out;
571 	}
572 
573 	qp_attr.qp_state = IB_QPS_RTS;
574 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
575 	if (ret)
576 		goto out;
577 
578 	if (conn_param)
579 		qp_attr.max_rd_atomic = conn_param->initiator_depth;
580 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
581 out:
582 	mutex_unlock(&id_priv->qp_mutex);
583 	return ret;
584 }
585 
586 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
587 {
588 	struct ib_qp_attr qp_attr;
589 	int ret;
590 
591 	mutex_lock(&id_priv->qp_mutex);
592 	if (!id_priv->id.qp) {
593 		ret = 0;
594 		goto out;
595 	}
596 
597 	qp_attr.qp_state = IB_QPS_ERR;
598 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
599 out:
600 	mutex_unlock(&id_priv->qp_mutex);
601 	return ret;
602 }
603 
604 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
605 			       struct ib_qp_attr *qp_attr, int *qp_attr_mask)
606 {
607 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
608 	int ret;
609 	u16 pkey;
610 
611 	if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
612 	    IB_LINK_LAYER_INFINIBAND)
613 		pkey = ib_addr_get_pkey(dev_addr);
614 	else
615 		pkey = 0xffff;
616 
617 	ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
618 				  pkey, &qp_attr->pkey_index);
619 	if (ret)
620 		return ret;
621 
622 	qp_attr->port_num = id_priv->id.port_num;
623 	*qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
624 
625 	if (id_priv->id.qp_type == IB_QPT_UD) {
626 		ret = cma_set_qkey(id_priv);
627 		if (ret)
628 			return ret;
629 
630 		qp_attr->qkey = id_priv->qkey;
631 		*qp_attr_mask |= IB_QP_QKEY;
632 	} else {
633 		qp_attr->qp_access_flags = 0;
634 		*qp_attr_mask |= IB_QP_ACCESS_FLAGS;
635 	}
636 	return 0;
637 }
638 
639 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
640 		       int *qp_attr_mask)
641 {
642 	struct rdma_id_private *id_priv;
643 	int ret = 0;
644 
645 	id_priv = container_of(id, struct rdma_id_private, id);
646 	switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
647 	case RDMA_TRANSPORT_IB:
648 		if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
649 			ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
650 		else
651 			ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
652 						 qp_attr_mask);
653 		if (qp_attr->qp_state == IB_QPS_RTR)
654 			qp_attr->rq_psn = id_priv->seq_num;
655 		break;
656 	case RDMA_TRANSPORT_IWARP:
657 		if (!id_priv->cm_id.iw) {
658 			qp_attr->qp_access_flags = 0;
659 			*qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
660 		} else
661 			ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
662 						 qp_attr_mask);
663 		break;
664 	default:
665 		ret = -ENOSYS;
666 		break;
667 	}
668 
669 	return ret;
670 }
671 EXPORT_SYMBOL(rdma_init_qp_attr);
672 
673 static inline int cma_zero_addr(struct sockaddr *addr)
674 {
675 	struct in6_addr *ip6;
676 
677 	if (addr->sa_family == AF_INET)
678 		return ipv4_is_zeronet(
679 			((struct sockaddr_in *)addr)->sin_addr.s_addr);
680 	else {
681 		ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
682 		return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
683 			ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
684 	}
685 }
686 
687 static inline int cma_loopback_addr(struct sockaddr *addr)
688 {
689 	if (addr->sa_family == AF_INET)
690 		return ipv4_is_loopback(
691 			((struct sockaddr_in *) addr)->sin_addr.s_addr);
692 	else
693 		return ipv6_addr_loopback(
694 			&((struct sockaddr_in6 *) addr)->sin6_addr);
695 }
696 
697 static inline int cma_any_addr(struct sockaddr *addr)
698 {
699 	return cma_zero_addr(addr) || cma_loopback_addr(addr);
700 }
701 
702 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
703 {
704 	if (src->sa_family != dst->sa_family)
705 		return -1;
706 
707 	switch (src->sa_family) {
708 	case AF_INET:
709 		return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
710 		       ((struct sockaddr_in *) dst)->sin_addr.s_addr;
711 	default:
712 		return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
713 				     &((struct sockaddr_in6 *) dst)->sin6_addr);
714 	}
715 }
716 
717 static inline __be16 cma_port(struct sockaddr *addr)
718 {
719 	if (addr->sa_family == AF_INET)
720 		return ((struct sockaddr_in *) addr)->sin_port;
721 	else
722 		return ((struct sockaddr_in6 *) addr)->sin6_port;
723 }
724 
725 static inline int cma_any_port(struct sockaddr *addr)
726 {
727 	return !cma_port(addr);
728 }
729 
730 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
731 			    u8 *ip_ver, __be16 *port,
732 			    union cma_ip_addr **src, union cma_ip_addr **dst)
733 {
734 	switch (ps) {
735 	case RDMA_PS_SDP:
736 		if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
737 		    SDP_MAJ_VERSION)
738 			return -EINVAL;
739 
740 		*ip_ver	= sdp_get_ip_ver(hdr);
741 		*port	= ((struct sdp_hh *) hdr)->port;
742 		*src	= &((struct sdp_hh *) hdr)->src_addr;
743 		*dst	= &((struct sdp_hh *) hdr)->dst_addr;
744 		break;
745 	default:
746 		if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
747 			return -EINVAL;
748 
749 		*ip_ver	= cma_get_ip_ver(hdr);
750 		*port	= ((struct cma_hdr *) hdr)->port;
751 		*src	= &((struct cma_hdr *) hdr)->src_addr;
752 		*dst	= &((struct cma_hdr *) hdr)->dst_addr;
753 		break;
754 	}
755 
756 	if (*ip_ver != 4 && *ip_ver != 6)
757 		return -EINVAL;
758 	return 0;
759 }
760 
761 static void cma_save_net_info(struct rdma_addr *addr,
762 			      struct rdma_addr *listen_addr,
763 			      u8 ip_ver, __be16 port,
764 			      union cma_ip_addr *src, union cma_ip_addr *dst)
765 {
766 	struct sockaddr_in *listen4, *ip4;
767 	struct sockaddr_in6 *listen6, *ip6;
768 
769 	switch (ip_ver) {
770 	case 4:
771 		listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
772 		ip4 = (struct sockaddr_in *) &addr->src_addr;
773 		ip4->sin_family = listen4->sin_family;
774 		ip4->sin_addr.s_addr = dst->ip4.addr;
775 		ip4->sin_port = listen4->sin_port;
776 
777 		ip4 = (struct sockaddr_in *) &addr->dst_addr;
778 		ip4->sin_family = listen4->sin_family;
779 		ip4->sin_addr.s_addr = src->ip4.addr;
780 		ip4->sin_port = port;
781 		break;
782 	case 6:
783 		listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
784 		ip6 = (struct sockaddr_in6 *) &addr->src_addr;
785 		ip6->sin6_family = listen6->sin6_family;
786 		ip6->sin6_addr = dst->ip6;
787 		ip6->sin6_port = listen6->sin6_port;
788 
789 		ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
790 		ip6->sin6_family = listen6->sin6_family;
791 		ip6->sin6_addr = src->ip6;
792 		ip6->sin6_port = port;
793 		break;
794 	default:
795 		break;
796 	}
797 }
798 
799 static inline int cma_user_data_offset(enum rdma_port_space ps)
800 {
801 	switch (ps) {
802 	case RDMA_PS_SDP:
803 		return 0;
804 	default:
805 		return sizeof(struct cma_hdr);
806 	}
807 }
808 
809 static void cma_cancel_route(struct rdma_id_private *id_priv)
810 {
811 	switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
812 	case IB_LINK_LAYER_INFINIBAND:
813 		if (id_priv->query)
814 			ib_sa_cancel_query(id_priv->query_id, id_priv->query);
815 		break;
816 	default:
817 		break;
818 	}
819 }
820 
821 static void cma_cancel_listens(struct rdma_id_private *id_priv)
822 {
823 	struct rdma_id_private *dev_id_priv;
824 
825 	/*
826 	 * Remove from listen_any_list to prevent added devices from spawning
827 	 * additional listen requests.
828 	 */
829 	mutex_lock(&lock);
830 	list_del(&id_priv->list);
831 
832 	while (!list_empty(&id_priv->listen_list)) {
833 		dev_id_priv = list_entry(id_priv->listen_list.next,
834 					 struct rdma_id_private, listen_list);
835 		/* sync with device removal to avoid duplicate destruction */
836 		list_del_init(&dev_id_priv->list);
837 		list_del(&dev_id_priv->listen_list);
838 		mutex_unlock(&lock);
839 
840 		rdma_destroy_id(&dev_id_priv->id);
841 		mutex_lock(&lock);
842 	}
843 	mutex_unlock(&lock);
844 }
845 
846 static void cma_cancel_operation(struct rdma_id_private *id_priv,
847 				 enum rdma_cm_state state)
848 {
849 	switch (state) {
850 	case RDMA_CM_ADDR_QUERY:
851 		rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
852 		break;
853 	case RDMA_CM_ROUTE_QUERY:
854 		cma_cancel_route(id_priv);
855 		break;
856 	case RDMA_CM_LISTEN:
857 		if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
858 				&& !id_priv->cma_dev)
859 			cma_cancel_listens(id_priv);
860 		break;
861 	default:
862 		break;
863 	}
864 }
865 
866 static void cma_release_port(struct rdma_id_private *id_priv)
867 {
868 	struct rdma_bind_list *bind_list = id_priv->bind_list;
869 
870 	if (!bind_list)
871 		return;
872 
873 	mutex_lock(&lock);
874 	hlist_del(&id_priv->node);
875 	if (hlist_empty(&bind_list->owners)) {
876 		idr_remove(bind_list->ps, bind_list->port);
877 		kfree(bind_list);
878 	}
879 	mutex_unlock(&lock);
880 }
881 
882 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
883 {
884 	struct cma_multicast *mc;
885 
886 	while (!list_empty(&id_priv->mc_list)) {
887 		mc = container_of(id_priv->mc_list.next,
888 				  struct cma_multicast, list);
889 		list_del(&mc->list);
890 		switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
891 		case IB_LINK_LAYER_INFINIBAND:
892 			ib_sa_free_multicast(mc->multicast.ib);
893 			kfree(mc);
894 			break;
895 		case IB_LINK_LAYER_ETHERNET:
896 			kref_put(&mc->mcref, release_mc);
897 			break;
898 		default:
899 			break;
900 		}
901 	}
902 }
903 
904 void rdma_destroy_id(struct rdma_cm_id *id)
905 {
906 	struct rdma_id_private *id_priv;
907 	enum rdma_cm_state state;
908 
909 	id_priv = container_of(id, struct rdma_id_private, id);
910 	state = cma_exch(id_priv, RDMA_CM_DESTROYING);
911 	cma_cancel_operation(id_priv, state);
912 
913 	/*
914 	 * Wait for any active callback to finish.  New callbacks will find
915 	 * the id_priv state set to destroying and abort.
916 	 */
917 	mutex_lock(&id_priv->handler_mutex);
918 	mutex_unlock(&id_priv->handler_mutex);
919 
920 	if (id_priv->cma_dev) {
921 		switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
922 		case RDMA_TRANSPORT_IB:
923 			if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
924 				ib_destroy_cm_id(id_priv->cm_id.ib);
925 			break;
926 		case RDMA_TRANSPORT_IWARP:
927 			if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
928 				iw_destroy_cm_id(id_priv->cm_id.iw);
929 			break;
930 		default:
931 			break;
932 		}
933 		cma_leave_mc_groups(id_priv);
934 		cma_release_dev(id_priv);
935 	}
936 
937 	cma_release_port(id_priv);
938 	cma_deref_id(id_priv);
939 	wait_for_completion(&id_priv->comp);
940 
941 	if (id_priv->internal_id)
942 		cma_deref_id(id_priv->id.context);
943 
944 	kfree(id_priv->id.route.path_rec);
945 	kfree(id_priv);
946 }
947 EXPORT_SYMBOL(rdma_destroy_id);
948 
949 static int cma_rep_recv(struct rdma_id_private *id_priv)
950 {
951 	int ret;
952 
953 	ret = cma_modify_qp_rtr(id_priv, NULL);
954 	if (ret)
955 		goto reject;
956 
957 	ret = cma_modify_qp_rts(id_priv, NULL);
958 	if (ret)
959 		goto reject;
960 
961 	ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
962 	if (ret)
963 		goto reject;
964 
965 	return 0;
966 reject:
967 	cma_modify_qp_err(id_priv);
968 	ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
969 		       NULL, 0, NULL, 0);
970 	return ret;
971 }
972 
973 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
974 {
975 	if (id_priv->id.ps == RDMA_PS_SDP &&
976 	    sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
977 	    SDP_MAJ_VERSION)
978 		return -EINVAL;
979 
980 	return 0;
981 }
982 
983 static void cma_set_rep_event_data(struct rdma_cm_event *event,
984 				   struct ib_cm_rep_event_param *rep_data,
985 				   void *private_data)
986 {
987 	event->param.conn.private_data = private_data;
988 	event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
989 	event->param.conn.responder_resources = rep_data->responder_resources;
990 	event->param.conn.initiator_depth = rep_data->initiator_depth;
991 	event->param.conn.flow_control = rep_data->flow_control;
992 	event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
993 	event->param.conn.srq = rep_data->srq;
994 	event->param.conn.qp_num = rep_data->remote_qpn;
995 }
996 
997 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
998 {
999 	struct rdma_id_private *id_priv = cm_id->context;
1000 	struct rdma_cm_event event;
1001 	int ret = 0;
1002 
1003 	if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1004 		cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1005 	    (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1006 		cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1007 		return 0;
1008 
1009 	memset(&event, 0, sizeof event);
1010 	switch (ib_event->event) {
1011 	case IB_CM_REQ_ERROR:
1012 	case IB_CM_REP_ERROR:
1013 		event.event = RDMA_CM_EVENT_UNREACHABLE;
1014 		event.status = -ETIMEDOUT;
1015 		break;
1016 	case IB_CM_REP_RECEIVED:
1017 		event.status = cma_verify_rep(id_priv, ib_event->private_data);
1018 		if (event.status)
1019 			event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1020 		else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
1021 			event.status = cma_rep_recv(id_priv);
1022 			event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1023 						     RDMA_CM_EVENT_ESTABLISHED;
1024 		} else
1025 			event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1026 		cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1027 				       ib_event->private_data);
1028 		break;
1029 	case IB_CM_RTU_RECEIVED:
1030 	case IB_CM_USER_ESTABLISHED:
1031 		event.event = RDMA_CM_EVENT_ESTABLISHED;
1032 		break;
1033 	case IB_CM_DREQ_ERROR:
1034 		event.status = -ETIMEDOUT; /* fall through */
1035 	case IB_CM_DREQ_RECEIVED:
1036 	case IB_CM_DREP_RECEIVED:
1037 		if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1038 				   RDMA_CM_DISCONNECT))
1039 			goto out;
1040 		event.event = RDMA_CM_EVENT_DISCONNECTED;
1041 		break;
1042 	case IB_CM_TIMEWAIT_EXIT:
1043 		event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1044 		break;
1045 	case IB_CM_MRA_RECEIVED:
1046 		/* ignore event */
1047 		goto out;
1048 	case IB_CM_REJ_RECEIVED:
1049 		cma_modify_qp_err(id_priv);
1050 		event.status = ib_event->param.rej_rcvd.reason;
1051 		event.event = RDMA_CM_EVENT_REJECTED;
1052 		event.param.conn.private_data = ib_event->private_data;
1053 		event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1054 		break;
1055 	default:
1056 		printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1057 		       ib_event->event);
1058 		goto out;
1059 	}
1060 
1061 	ret = id_priv->id.event_handler(&id_priv->id, &event);
1062 	if (ret) {
1063 		/* Destroy the CM ID by returning a non-zero value. */
1064 		id_priv->cm_id.ib = NULL;
1065 		cma_exch(id_priv, RDMA_CM_DESTROYING);
1066 		mutex_unlock(&id_priv->handler_mutex);
1067 		rdma_destroy_id(&id_priv->id);
1068 		return ret;
1069 	}
1070 out:
1071 	mutex_unlock(&id_priv->handler_mutex);
1072 	return ret;
1073 }
1074 
1075 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1076 					       struct ib_cm_event *ib_event)
1077 {
1078 	struct rdma_id_private *id_priv;
1079 	struct rdma_cm_id *id;
1080 	struct rdma_route *rt;
1081 	union cma_ip_addr *src, *dst;
1082 	__be16 port;
1083 	u8 ip_ver;
1084 	int ret;
1085 
1086 	if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1087 			     &ip_ver, &port, &src, &dst))
1088 		goto err;
1089 
1090 	id = rdma_create_id(listen_id->event_handler, listen_id->context,
1091 			    listen_id->ps, ib_event->param.req_rcvd.qp_type);
1092 	if (IS_ERR(id))
1093 		goto err;
1094 
1095 	cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1096 			  ip_ver, port, src, dst);
1097 
1098 	rt = &id->route;
1099 	rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1100 	rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1101 			       GFP_KERNEL);
1102 	if (!rt->path_rec)
1103 		goto destroy_id;
1104 
1105 	rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1106 	if (rt->num_paths == 2)
1107 		rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1108 
1109 	if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1110 		rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1111 		rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1112 		ib_addr_set_pkey(&rt->addr.dev_addr, rt->path_rec[0].pkey);
1113 	} else {
1114 		ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1115 					&rt->addr.dev_addr);
1116 		if (ret)
1117 			goto destroy_id;
1118 	}
1119 	rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1120 
1121 	id_priv = container_of(id, struct rdma_id_private, id);
1122 	id_priv->state = RDMA_CM_CONNECT;
1123 	return id_priv;
1124 
1125 destroy_id:
1126 	rdma_destroy_id(id);
1127 err:
1128 	return NULL;
1129 }
1130 
1131 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1132 					      struct ib_cm_event *ib_event)
1133 {
1134 	struct rdma_id_private *id_priv;
1135 	struct rdma_cm_id *id;
1136 	union cma_ip_addr *src, *dst;
1137 	__be16 port;
1138 	u8 ip_ver;
1139 	int ret;
1140 
1141 	id = rdma_create_id(listen_id->event_handler, listen_id->context,
1142 			    listen_id->ps, IB_QPT_UD);
1143 	if (IS_ERR(id))
1144 		return NULL;
1145 
1146 
1147 	if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1148 			     &ip_ver, &port, &src, &dst))
1149 		goto err;
1150 
1151 	cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1152 			  ip_ver, port, src, dst);
1153 
1154 	if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1155 		ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1156 					&id->route.addr.dev_addr);
1157 		if (ret)
1158 			goto err;
1159 	}
1160 
1161 	id_priv = container_of(id, struct rdma_id_private, id);
1162 	id_priv->state = RDMA_CM_CONNECT;
1163 	return id_priv;
1164 err:
1165 	rdma_destroy_id(id);
1166 	return NULL;
1167 }
1168 
1169 static void cma_set_req_event_data(struct rdma_cm_event *event,
1170 				   struct ib_cm_req_event_param *req_data,
1171 				   void *private_data, int offset)
1172 {
1173 	event->param.conn.private_data = private_data + offset;
1174 	event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1175 	event->param.conn.responder_resources = req_data->responder_resources;
1176 	event->param.conn.initiator_depth = req_data->initiator_depth;
1177 	event->param.conn.flow_control = req_data->flow_control;
1178 	event->param.conn.retry_count = req_data->retry_count;
1179 	event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1180 	event->param.conn.srq = req_data->srq;
1181 	event->param.conn.qp_num = req_data->remote_qpn;
1182 }
1183 
1184 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1185 {
1186 	struct rdma_id_private *listen_id, *conn_id;
1187 	struct rdma_cm_event event;
1188 	int offset, ret;
1189 
1190 	listen_id = cm_id->context;
1191 	if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1192 		return -ECONNABORTED;
1193 
1194 	memset(&event, 0, sizeof event);
1195 	offset = cma_user_data_offset(listen_id->id.ps);
1196 	event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1197 	if (listen_id->id.qp_type == IB_QPT_UD) {
1198 		conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1199 		event.param.ud.private_data = ib_event->private_data + offset;
1200 		event.param.ud.private_data_len =
1201 				IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1202 	} else {
1203 		conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1204 		cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1205 				       ib_event->private_data, offset);
1206 	}
1207 	if (!conn_id) {
1208 		ret = -ENOMEM;
1209 		goto out;
1210 	}
1211 
1212 	mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1213 	ret = cma_acquire_dev(conn_id);
1214 	if (ret)
1215 		goto release_conn_id;
1216 
1217 	conn_id->cm_id.ib = cm_id;
1218 	cm_id->context = conn_id;
1219 	cm_id->cm_handler = cma_ib_handler;
1220 
1221 	/*
1222 	 * Protect against the user destroying conn_id from another thread
1223 	 * until we're done accessing it.
1224 	 */
1225 	atomic_inc(&conn_id->refcount);
1226 	ret = conn_id->id.event_handler(&conn_id->id, &event);
1227 	if (!ret) {
1228 		/*
1229 		 * Acquire mutex to prevent user executing rdma_destroy_id()
1230 		 * while we're accessing the cm_id.
1231 		 */
1232 		mutex_lock(&lock);
1233 		if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD))
1234 			ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1235 		mutex_unlock(&lock);
1236 		mutex_unlock(&conn_id->handler_mutex);
1237 		cma_deref_id(conn_id);
1238 		goto out;
1239 	}
1240 	cma_deref_id(conn_id);
1241 
1242 	/* Destroy the CM ID by returning a non-zero value. */
1243 	conn_id->cm_id.ib = NULL;
1244 
1245 release_conn_id:
1246 	cma_exch(conn_id, RDMA_CM_DESTROYING);
1247 	mutex_unlock(&conn_id->handler_mutex);
1248 	rdma_destroy_id(&conn_id->id);
1249 
1250 out:
1251 	mutex_unlock(&listen_id->handler_mutex);
1252 	return ret;
1253 }
1254 
1255 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1256 {
1257 	return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1258 }
1259 
1260 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1261 				 struct ib_cm_compare_data *compare)
1262 {
1263 	struct cma_hdr *cma_data, *cma_mask;
1264 	struct sdp_hh *sdp_data, *sdp_mask;
1265 	__be32 ip4_addr;
1266 	struct in6_addr ip6_addr;
1267 
1268 	memset(compare, 0, sizeof *compare);
1269 	cma_data = (void *) compare->data;
1270 	cma_mask = (void *) compare->mask;
1271 	sdp_data = (void *) compare->data;
1272 	sdp_mask = (void *) compare->mask;
1273 
1274 	switch (addr->sa_family) {
1275 	case AF_INET:
1276 		ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1277 		if (ps == RDMA_PS_SDP) {
1278 			sdp_set_ip_ver(sdp_data, 4);
1279 			sdp_set_ip_ver(sdp_mask, 0xF);
1280 			sdp_data->dst_addr.ip4.addr = ip4_addr;
1281 			sdp_mask->dst_addr.ip4.addr = htonl(~0);
1282 		} else {
1283 			cma_set_ip_ver(cma_data, 4);
1284 			cma_set_ip_ver(cma_mask, 0xF);
1285 			cma_data->dst_addr.ip4.addr = ip4_addr;
1286 			cma_mask->dst_addr.ip4.addr = htonl(~0);
1287 		}
1288 		break;
1289 	case AF_INET6:
1290 		ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1291 		if (ps == RDMA_PS_SDP) {
1292 			sdp_set_ip_ver(sdp_data, 6);
1293 			sdp_set_ip_ver(sdp_mask, 0xF);
1294 			sdp_data->dst_addr.ip6 = ip6_addr;
1295 			memset(&sdp_mask->dst_addr.ip6, 0xFF,
1296 			       sizeof sdp_mask->dst_addr.ip6);
1297 		} else {
1298 			cma_set_ip_ver(cma_data, 6);
1299 			cma_set_ip_ver(cma_mask, 0xF);
1300 			cma_data->dst_addr.ip6 = ip6_addr;
1301 			memset(&cma_mask->dst_addr.ip6, 0xFF,
1302 			       sizeof cma_mask->dst_addr.ip6);
1303 		}
1304 		break;
1305 	default:
1306 		break;
1307 	}
1308 }
1309 
1310 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1311 {
1312 	struct rdma_id_private *id_priv = iw_id->context;
1313 	struct rdma_cm_event event;
1314 	struct sockaddr_in *sin;
1315 	int ret = 0;
1316 
1317 	if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1318 		return 0;
1319 
1320 	memset(&event, 0, sizeof event);
1321 	switch (iw_event->event) {
1322 	case IW_CM_EVENT_CLOSE:
1323 		event.event = RDMA_CM_EVENT_DISCONNECTED;
1324 		break;
1325 	case IW_CM_EVENT_CONNECT_REPLY:
1326 		sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1327 		*sin = iw_event->local_addr;
1328 		sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1329 		*sin = iw_event->remote_addr;
1330 		switch (iw_event->status) {
1331 		case 0:
1332 			event.event = RDMA_CM_EVENT_ESTABLISHED;
1333 			break;
1334 		case -ECONNRESET:
1335 		case -ECONNREFUSED:
1336 			event.event = RDMA_CM_EVENT_REJECTED;
1337 			break;
1338 		case -ETIMEDOUT:
1339 			event.event = RDMA_CM_EVENT_UNREACHABLE;
1340 			break;
1341 		default:
1342 			event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1343 			break;
1344 		}
1345 		break;
1346 	case IW_CM_EVENT_ESTABLISHED:
1347 		event.event = RDMA_CM_EVENT_ESTABLISHED;
1348 		break;
1349 	default:
1350 		BUG_ON(1);
1351 	}
1352 
1353 	event.status = iw_event->status;
1354 	event.param.conn.private_data = iw_event->private_data;
1355 	event.param.conn.private_data_len = iw_event->private_data_len;
1356 	ret = id_priv->id.event_handler(&id_priv->id, &event);
1357 	if (ret) {
1358 		/* Destroy the CM ID by returning a non-zero value. */
1359 		id_priv->cm_id.iw = NULL;
1360 		cma_exch(id_priv, RDMA_CM_DESTROYING);
1361 		mutex_unlock(&id_priv->handler_mutex);
1362 		rdma_destroy_id(&id_priv->id);
1363 		return ret;
1364 	}
1365 
1366 	mutex_unlock(&id_priv->handler_mutex);
1367 	return ret;
1368 }
1369 
1370 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1371 			       struct iw_cm_event *iw_event)
1372 {
1373 	struct rdma_cm_id *new_cm_id;
1374 	struct rdma_id_private *listen_id, *conn_id;
1375 	struct sockaddr_in *sin;
1376 	struct net_device *dev = NULL;
1377 	struct rdma_cm_event event;
1378 	int ret;
1379 	struct ib_device_attr attr;
1380 
1381 	listen_id = cm_id->context;
1382 	if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1383 		return -ECONNABORTED;
1384 
1385 	/* Create a new RDMA id for the new IW CM ID */
1386 	new_cm_id = rdma_create_id(listen_id->id.event_handler,
1387 				   listen_id->id.context,
1388 				   RDMA_PS_TCP, IB_QPT_RC);
1389 	if (IS_ERR(new_cm_id)) {
1390 		ret = -ENOMEM;
1391 		goto out;
1392 	}
1393 	conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1394 	mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1395 	conn_id->state = RDMA_CM_CONNECT;
1396 
1397 	dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1398 	if (!dev) {
1399 		ret = -EADDRNOTAVAIL;
1400 		mutex_unlock(&conn_id->handler_mutex);
1401 		rdma_destroy_id(new_cm_id);
1402 		goto out;
1403 	}
1404 	ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1405 	if (ret) {
1406 		mutex_unlock(&conn_id->handler_mutex);
1407 		rdma_destroy_id(new_cm_id);
1408 		goto out;
1409 	}
1410 
1411 	ret = cma_acquire_dev(conn_id);
1412 	if (ret) {
1413 		mutex_unlock(&conn_id->handler_mutex);
1414 		rdma_destroy_id(new_cm_id);
1415 		goto out;
1416 	}
1417 
1418 	conn_id->cm_id.iw = cm_id;
1419 	cm_id->context = conn_id;
1420 	cm_id->cm_handler = cma_iw_handler;
1421 
1422 	sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1423 	*sin = iw_event->local_addr;
1424 	sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1425 	*sin = iw_event->remote_addr;
1426 
1427 	ret = ib_query_device(conn_id->id.device, &attr);
1428 	if (ret) {
1429 		mutex_unlock(&conn_id->handler_mutex);
1430 		rdma_destroy_id(new_cm_id);
1431 		goto out;
1432 	}
1433 
1434 	memset(&event, 0, sizeof event);
1435 	event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1436 	event.param.conn.private_data = iw_event->private_data;
1437 	event.param.conn.private_data_len = iw_event->private_data_len;
1438 	event.param.conn.initiator_depth = attr.max_qp_init_rd_atom;
1439 	event.param.conn.responder_resources = attr.max_qp_rd_atom;
1440 
1441 	/*
1442 	 * Protect against the user destroying conn_id from another thread
1443 	 * until we're done accessing it.
1444 	 */
1445 	atomic_inc(&conn_id->refcount);
1446 	ret = conn_id->id.event_handler(&conn_id->id, &event);
1447 	if (ret) {
1448 		/* User wants to destroy the CM ID */
1449 		conn_id->cm_id.iw = NULL;
1450 		cma_exch(conn_id, RDMA_CM_DESTROYING);
1451 		mutex_unlock(&conn_id->handler_mutex);
1452 		cma_deref_id(conn_id);
1453 		rdma_destroy_id(&conn_id->id);
1454 		goto out;
1455 	}
1456 
1457 	mutex_unlock(&conn_id->handler_mutex);
1458 	cma_deref_id(conn_id);
1459 
1460 out:
1461 	if (dev)
1462 		dev_put(dev);
1463 	mutex_unlock(&listen_id->handler_mutex);
1464 	return ret;
1465 }
1466 
1467 static int cma_ib_listen(struct rdma_id_private *id_priv)
1468 {
1469 	struct ib_cm_compare_data compare_data;
1470 	struct sockaddr *addr;
1471 	__be64 svc_id;
1472 	int ret;
1473 
1474 	id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1475 					    id_priv);
1476 	if (IS_ERR(id_priv->cm_id.ib))
1477 		return PTR_ERR(id_priv->cm_id.ib);
1478 
1479 	addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1480 	svc_id = cma_get_service_id(id_priv->id.ps, addr);
1481 	if (cma_any_addr(addr))
1482 		ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1483 	else {
1484 		cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1485 		ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1486 	}
1487 
1488 	if (ret) {
1489 		ib_destroy_cm_id(id_priv->cm_id.ib);
1490 		id_priv->cm_id.ib = NULL;
1491 	}
1492 
1493 	return ret;
1494 }
1495 
1496 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1497 {
1498 	int ret;
1499 	struct sockaddr_in *sin;
1500 
1501 	id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1502 					    iw_conn_req_handler,
1503 					    id_priv);
1504 	if (IS_ERR(id_priv->cm_id.iw))
1505 		return PTR_ERR(id_priv->cm_id.iw);
1506 
1507 	sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1508 	id_priv->cm_id.iw->local_addr = *sin;
1509 
1510 	ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1511 
1512 	if (ret) {
1513 		iw_destroy_cm_id(id_priv->cm_id.iw);
1514 		id_priv->cm_id.iw = NULL;
1515 	}
1516 
1517 	return ret;
1518 }
1519 
1520 static int cma_listen_handler(struct rdma_cm_id *id,
1521 			      struct rdma_cm_event *event)
1522 {
1523 	struct rdma_id_private *id_priv = id->context;
1524 
1525 	id->context = id_priv->id.context;
1526 	id->event_handler = id_priv->id.event_handler;
1527 	return id_priv->id.event_handler(id, event);
1528 }
1529 
1530 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1531 			      struct cma_device *cma_dev)
1532 {
1533 	struct rdma_id_private *dev_id_priv;
1534 	struct rdma_cm_id *id;
1535 	int ret;
1536 
1537 	id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1538 			    id_priv->id.qp_type);
1539 	if (IS_ERR(id))
1540 		return;
1541 
1542 	dev_id_priv = container_of(id, struct rdma_id_private, id);
1543 
1544 	dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1545 	memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1546 	       ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1547 
1548 	cma_attach_to_dev(dev_id_priv, cma_dev);
1549 	list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1550 	atomic_inc(&id_priv->refcount);
1551 	dev_id_priv->internal_id = 1;
1552 
1553 	ret = rdma_listen(id, id_priv->backlog);
1554 	if (ret)
1555 		printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1556 		       "listening on device %s\n", ret, cma_dev->device->name);
1557 }
1558 
1559 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1560 {
1561 	struct cma_device *cma_dev;
1562 
1563 	mutex_lock(&lock);
1564 	list_add_tail(&id_priv->list, &listen_any_list);
1565 	list_for_each_entry(cma_dev, &dev_list, list)
1566 		cma_listen_on_dev(id_priv, cma_dev);
1567 	mutex_unlock(&lock);
1568 }
1569 
1570 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1571 {
1572 	struct rdma_id_private *id_priv;
1573 
1574 	id_priv = container_of(id, struct rdma_id_private, id);
1575 	id_priv->tos = (u8) tos;
1576 }
1577 EXPORT_SYMBOL(rdma_set_service_type);
1578 
1579 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1580 			      void *context)
1581 {
1582 	struct cma_work *work = context;
1583 	struct rdma_route *route;
1584 
1585 	route = &work->id->id.route;
1586 
1587 	if (!status) {
1588 		route->num_paths = 1;
1589 		*route->path_rec = *path_rec;
1590 	} else {
1591 		work->old_state = RDMA_CM_ROUTE_QUERY;
1592 		work->new_state = RDMA_CM_ADDR_RESOLVED;
1593 		work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1594 		work->event.status = status;
1595 	}
1596 
1597 	queue_work(cma_wq, &work->work);
1598 }
1599 
1600 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1601 			      struct cma_work *work)
1602 {
1603 	struct rdma_addr *addr = &id_priv->id.route.addr;
1604 	struct ib_sa_path_rec path_rec;
1605 	ib_sa_comp_mask comp_mask;
1606 	struct sockaddr_in6 *sin6;
1607 
1608 	memset(&path_rec, 0, sizeof path_rec);
1609 	rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1610 	rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1611 	path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1612 	path_rec.numb_path = 1;
1613 	path_rec.reversible = 1;
1614 	path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1615 							(struct sockaddr *) &addr->dst_addr);
1616 
1617 	comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1618 		    IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1619 		    IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1620 
1621 	if (addr->src_addr.ss_family == AF_INET) {
1622 		path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1623 		comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1624 	} else {
1625 		sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1626 		path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1627 		comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1628 	}
1629 
1630 	id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1631 					       id_priv->id.port_num, &path_rec,
1632 					       comp_mask, timeout_ms,
1633 					       GFP_KERNEL, cma_query_handler,
1634 					       work, &id_priv->query);
1635 
1636 	return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1637 }
1638 
1639 static void cma_work_handler(struct work_struct *_work)
1640 {
1641 	struct cma_work *work = container_of(_work, struct cma_work, work);
1642 	struct rdma_id_private *id_priv = work->id;
1643 	int destroy = 0;
1644 
1645 	mutex_lock(&id_priv->handler_mutex);
1646 	if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1647 		goto out;
1648 
1649 	if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1650 		cma_exch(id_priv, RDMA_CM_DESTROYING);
1651 		destroy = 1;
1652 	}
1653 out:
1654 	mutex_unlock(&id_priv->handler_mutex);
1655 	cma_deref_id(id_priv);
1656 	if (destroy)
1657 		rdma_destroy_id(&id_priv->id);
1658 	kfree(work);
1659 }
1660 
1661 static void cma_ndev_work_handler(struct work_struct *_work)
1662 {
1663 	struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1664 	struct rdma_id_private *id_priv = work->id;
1665 	int destroy = 0;
1666 
1667 	mutex_lock(&id_priv->handler_mutex);
1668 	if (id_priv->state == RDMA_CM_DESTROYING ||
1669 	    id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1670 		goto out;
1671 
1672 	if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1673 		cma_exch(id_priv, RDMA_CM_DESTROYING);
1674 		destroy = 1;
1675 	}
1676 
1677 out:
1678 	mutex_unlock(&id_priv->handler_mutex);
1679 	cma_deref_id(id_priv);
1680 	if (destroy)
1681 		rdma_destroy_id(&id_priv->id);
1682 	kfree(work);
1683 }
1684 
1685 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1686 {
1687 	struct rdma_route *route = &id_priv->id.route;
1688 	struct cma_work *work;
1689 	int ret;
1690 
1691 	work = kzalloc(sizeof *work, GFP_KERNEL);
1692 	if (!work)
1693 		return -ENOMEM;
1694 
1695 	work->id = id_priv;
1696 	INIT_WORK(&work->work, cma_work_handler);
1697 	work->old_state = RDMA_CM_ROUTE_QUERY;
1698 	work->new_state = RDMA_CM_ROUTE_RESOLVED;
1699 	work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1700 
1701 	route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1702 	if (!route->path_rec) {
1703 		ret = -ENOMEM;
1704 		goto err1;
1705 	}
1706 
1707 	ret = cma_query_ib_route(id_priv, timeout_ms, work);
1708 	if (ret)
1709 		goto err2;
1710 
1711 	return 0;
1712 err2:
1713 	kfree(route->path_rec);
1714 	route->path_rec = NULL;
1715 err1:
1716 	kfree(work);
1717 	return ret;
1718 }
1719 
1720 int rdma_set_ib_paths(struct rdma_cm_id *id,
1721 		      struct ib_sa_path_rec *path_rec, int num_paths)
1722 {
1723 	struct rdma_id_private *id_priv;
1724 	int ret;
1725 
1726 	id_priv = container_of(id, struct rdma_id_private, id);
1727 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1728 			   RDMA_CM_ROUTE_RESOLVED))
1729 		return -EINVAL;
1730 
1731 	id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1732 				     GFP_KERNEL);
1733 	if (!id->route.path_rec) {
1734 		ret = -ENOMEM;
1735 		goto err;
1736 	}
1737 
1738 	id->route.num_paths = num_paths;
1739 	return 0;
1740 err:
1741 	cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1742 	return ret;
1743 }
1744 EXPORT_SYMBOL(rdma_set_ib_paths);
1745 
1746 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1747 {
1748 	struct cma_work *work;
1749 
1750 	work = kzalloc(sizeof *work, GFP_KERNEL);
1751 	if (!work)
1752 		return -ENOMEM;
1753 
1754 	work->id = id_priv;
1755 	INIT_WORK(&work->work, cma_work_handler);
1756 	work->old_state = RDMA_CM_ROUTE_QUERY;
1757 	work->new_state = RDMA_CM_ROUTE_RESOLVED;
1758 	work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1759 	queue_work(cma_wq, &work->work);
1760 	return 0;
1761 }
1762 
1763 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1764 {
1765 	struct rdma_route *route = &id_priv->id.route;
1766 	struct rdma_addr *addr = &route->addr;
1767 	struct cma_work *work;
1768 	int ret;
1769 	struct sockaddr_in *src_addr = (struct sockaddr_in *)&route->addr.src_addr;
1770 	struct sockaddr_in *dst_addr = (struct sockaddr_in *)&route->addr.dst_addr;
1771 	struct net_device *ndev = NULL;
1772 	u16 vid;
1773 
1774 	if (src_addr->sin_family != dst_addr->sin_family)
1775 		return -EINVAL;
1776 
1777 	work = kzalloc(sizeof *work, GFP_KERNEL);
1778 	if (!work)
1779 		return -ENOMEM;
1780 
1781 	work->id = id_priv;
1782 	INIT_WORK(&work->work, cma_work_handler);
1783 
1784 	route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1785 	if (!route->path_rec) {
1786 		ret = -ENOMEM;
1787 		goto err1;
1788 	}
1789 
1790 	route->num_paths = 1;
1791 
1792 	if (addr->dev_addr.bound_dev_if)
1793 		ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1794 	if (!ndev) {
1795 		ret = -ENODEV;
1796 		goto err2;
1797 	}
1798 
1799 	vid = rdma_vlan_dev_vlan_id(ndev);
1800 
1801 	iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1802 	iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1803 
1804 	route->path_rec->hop_limit = 1;
1805 	route->path_rec->reversible = 1;
1806 	route->path_rec->pkey = cpu_to_be16(0xffff);
1807 	route->path_rec->mtu_selector = IB_SA_EQ;
1808 	route->path_rec->sl = id_priv->tos >> 5;
1809 
1810 	route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1811 	route->path_rec->rate_selector = IB_SA_EQ;
1812 	route->path_rec->rate = iboe_get_rate(ndev);
1813 	dev_put(ndev);
1814 	route->path_rec->packet_life_time_selector = IB_SA_EQ;
1815 	route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1816 	if (!route->path_rec->mtu) {
1817 		ret = -EINVAL;
1818 		goto err2;
1819 	}
1820 
1821 	work->old_state = RDMA_CM_ROUTE_QUERY;
1822 	work->new_state = RDMA_CM_ROUTE_RESOLVED;
1823 	work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1824 	work->event.status = 0;
1825 
1826 	queue_work(cma_wq, &work->work);
1827 
1828 	return 0;
1829 
1830 err2:
1831 	kfree(route->path_rec);
1832 	route->path_rec = NULL;
1833 err1:
1834 	kfree(work);
1835 	return ret;
1836 }
1837 
1838 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1839 {
1840 	struct rdma_id_private *id_priv;
1841 	int ret;
1842 
1843 	id_priv = container_of(id, struct rdma_id_private, id);
1844 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1845 		return -EINVAL;
1846 
1847 	atomic_inc(&id_priv->refcount);
1848 	switch (rdma_node_get_transport(id->device->node_type)) {
1849 	case RDMA_TRANSPORT_IB:
1850 		switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1851 		case IB_LINK_LAYER_INFINIBAND:
1852 			ret = cma_resolve_ib_route(id_priv, timeout_ms);
1853 			break;
1854 		case IB_LINK_LAYER_ETHERNET:
1855 			ret = cma_resolve_iboe_route(id_priv);
1856 			break;
1857 		default:
1858 			ret = -ENOSYS;
1859 		}
1860 		break;
1861 	case RDMA_TRANSPORT_IWARP:
1862 		ret = cma_resolve_iw_route(id_priv, timeout_ms);
1863 		break;
1864 	default:
1865 		ret = -ENOSYS;
1866 		break;
1867 	}
1868 	if (ret)
1869 		goto err;
1870 
1871 	return 0;
1872 err:
1873 	cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1874 	cma_deref_id(id_priv);
1875 	return ret;
1876 }
1877 EXPORT_SYMBOL(rdma_resolve_route);
1878 
1879 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1880 {
1881 	struct cma_device *cma_dev;
1882 	struct ib_port_attr port_attr;
1883 	union ib_gid gid;
1884 	u16 pkey;
1885 	int ret;
1886 	u8 p;
1887 
1888 	mutex_lock(&lock);
1889 	if (list_empty(&dev_list)) {
1890 		ret = -ENODEV;
1891 		goto out;
1892 	}
1893 	list_for_each_entry(cma_dev, &dev_list, list)
1894 		for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1895 			if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1896 			    port_attr.state == IB_PORT_ACTIVE)
1897 				goto port_found;
1898 
1899 	p = 1;
1900 	cma_dev = list_entry(dev_list.next, struct cma_device, list);
1901 
1902 port_found:
1903 	ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1904 	if (ret)
1905 		goto out;
1906 
1907 	ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1908 	if (ret)
1909 		goto out;
1910 
1911 	id_priv->id.route.addr.dev_addr.dev_type =
1912 		(rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
1913 		ARPHRD_INFINIBAND : ARPHRD_ETHER;
1914 
1915 	rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1916 	ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1917 	id_priv->id.port_num = p;
1918 	cma_attach_to_dev(id_priv, cma_dev);
1919 out:
1920 	mutex_unlock(&lock);
1921 	return ret;
1922 }
1923 
1924 static void addr_handler(int status, struct sockaddr *src_addr,
1925 			 struct rdma_dev_addr *dev_addr, void *context)
1926 {
1927 	struct rdma_id_private *id_priv = context;
1928 	struct rdma_cm_event event;
1929 
1930 	memset(&event, 0, sizeof event);
1931 	mutex_lock(&id_priv->handler_mutex);
1932 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
1933 			   RDMA_CM_ADDR_RESOLVED))
1934 		goto out;
1935 
1936 	if (!status && !id_priv->cma_dev)
1937 		status = cma_acquire_dev(id_priv);
1938 
1939 	if (status) {
1940 		if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1941 				   RDMA_CM_ADDR_BOUND))
1942 			goto out;
1943 		event.event = RDMA_CM_EVENT_ADDR_ERROR;
1944 		event.status = status;
1945 	} else {
1946 		memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1947 		       ip_addr_size(src_addr));
1948 		event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1949 	}
1950 
1951 	if (id_priv->id.event_handler(&id_priv->id, &event)) {
1952 		cma_exch(id_priv, RDMA_CM_DESTROYING);
1953 		mutex_unlock(&id_priv->handler_mutex);
1954 		cma_deref_id(id_priv);
1955 		rdma_destroy_id(&id_priv->id);
1956 		return;
1957 	}
1958 out:
1959 	mutex_unlock(&id_priv->handler_mutex);
1960 	cma_deref_id(id_priv);
1961 }
1962 
1963 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1964 {
1965 	struct cma_work *work;
1966 	struct sockaddr *src, *dst;
1967 	union ib_gid gid;
1968 	int ret;
1969 
1970 	work = kzalloc(sizeof *work, GFP_KERNEL);
1971 	if (!work)
1972 		return -ENOMEM;
1973 
1974 	if (!id_priv->cma_dev) {
1975 		ret = cma_bind_loopback(id_priv);
1976 		if (ret)
1977 			goto err;
1978 	}
1979 
1980 	rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1981 	rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1982 
1983 	src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1984 	if (cma_zero_addr(src)) {
1985 		dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
1986 		if ((src->sa_family = dst->sa_family) == AF_INET) {
1987 			((struct sockaddr_in *) src)->sin_addr.s_addr =
1988 				((struct sockaddr_in *) dst)->sin_addr.s_addr;
1989 		} else {
1990 			ipv6_addr_copy(&((struct sockaddr_in6 *) src)->sin6_addr,
1991 				       &((struct sockaddr_in6 *) dst)->sin6_addr);
1992 		}
1993 	}
1994 
1995 	work->id = id_priv;
1996 	INIT_WORK(&work->work, cma_work_handler);
1997 	work->old_state = RDMA_CM_ADDR_QUERY;
1998 	work->new_state = RDMA_CM_ADDR_RESOLVED;
1999 	work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2000 	queue_work(cma_wq, &work->work);
2001 	return 0;
2002 err:
2003 	kfree(work);
2004 	return ret;
2005 }
2006 
2007 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2008 			 struct sockaddr *dst_addr)
2009 {
2010 	if (!src_addr || !src_addr->sa_family) {
2011 		src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2012 		if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
2013 			((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2014 				((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2015 		}
2016 	}
2017 	return rdma_bind_addr(id, src_addr);
2018 }
2019 
2020 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2021 		      struct sockaddr *dst_addr, int timeout_ms)
2022 {
2023 	struct rdma_id_private *id_priv;
2024 	int ret;
2025 
2026 	id_priv = container_of(id, struct rdma_id_private, id);
2027 	if (id_priv->state == RDMA_CM_IDLE) {
2028 		ret = cma_bind_addr(id, src_addr, dst_addr);
2029 		if (ret)
2030 			return ret;
2031 	}
2032 
2033 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2034 		return -EINVAL;
2035 
2036 	atomic_inc(&id_priv->refcount);
2037 	memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2038 	if (cma_any_addr(dst_addr))
2039 		ret = cma_resolve_loopback(id_priv);
2040 	else
2041 		ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2042 				      dst_addr, &id->route.addr.dev_addr,
2043 				      timeout_ms, addr_handler, id_priv);
2044 	if (ret)
2045 		goto err;
2046 
2047 	return 0;
2048 err:
2049 	cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2050 	cma_deref_id(id_priv);
2051 	return ret;
2052 }
2053 EXPORT_SYMBOL(rdma_resolve_addr);
2054 
2055 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2056 {
2057 	struct rdma_id_private *id_priv;
2058 	unsigned long flags;
2059 	int ret;
2060 
2061 	id_priv = container_of(id, struct rdma_id_private, id);
2062 	spin_lock_irqsave(&id_priv->lock, flags);
2063 	if (id_priv->state == RDMA_CM_IDLE) {
2064 		id_priv->reuseaddr = reuse;
2065 		ret = 0;
2066 	} else {
2067 		ret = -EINVAL;
2068 	}
2069 	spin_unlock_irqrestore(&id_priv->lock, flags);
2070 	return ret;
2071 }
2072 EXPORT_SYMBOL(rdma_set_reuseaddr);
2073 
2074 static void cma_bind_port(struct rdma_bind_list *bind_list,
2075 			  struct rdma_id_private *id_priv)
2076 {
2077 	struct sockaddr_in *sin;
2078 
2079 	sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2080 	sin->sin_port = htons(bind_list->port);
2081 	id_priv->bind_list = bind_list;
2082 	hlist_add_head(&id_priv->node, &bind_list->owners);
2083 }
2084 
2085 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2086 			  unsigned short snum)
2087 {
2088 	struct rdma_bind_list *bind_list;
2089 	int port, ret;
2090 
2091 	bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2092 	if (!bind_list)
2093 		return -ENOMEM;
2094 
2095 	do {
2096 		ret = idr_get_new_above(ps, bind_list, snum, &port);
2097 	} while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2098 
2099 	if (ret)
2100 		goto err1;
2101 
2102 	if (port != snum) {
2103 		ret = -EADDRNOTAVAIL;
2104 		goto err2;
2105 	}
2106 
2107 	bind_list->ps = ps;
2108 	bind_list->port = (unsigned short) port;
2109 	cma_bind_port(bind_list, id_priv);
2110 	return 0;
2111 err2:
2112 	idr_remove(ps, port);
2113 err1:
2114 	kfree(bind_list);
2115 	return ret;
2116 }
2117 
2118 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2119 {
2120 	static unsigned int last_used_port;
2121 	int low, high, remaining;
2122 	unsigned int rover;
2123 
2124 	inet_get_local_port_range(&low, &high);
2125 	remaining = (high - low) + 1;
2126 	rover = net_random() % remaining + low;
2127 retry:
2128 	if (last_used_port != rover &&
2129 	    !idr_find(ps, (unsigned short) rover)) {
2130 		int ret = cma_alloc_port(ps, id_priv, rover);
2131 		/*
2132 		 * Remember previously used port number in order to avoid
2133 		 * re-using same port immediately after it is closed.
2134 		 */
2135 		if (!ret)
2136 			last_used_port = rover;
2137 		if (ret != -EADDRNOTAVAIL)
2138 			return ret;
2139 	}
2140 	if (--remaining) {
2141 		rover++;
2142 		if ((rover < low) || (rover > high))
2143 			rover = low;
2144 		goto retry;
2145 	}
2146 	return -EADDRNOTAVAIL;
2147 }
2148 
2149 /*
2150  * Check that the requested port is available.  This is called when trying to
2151  * bind to a specific port, or when trying to listen on a bound port.  In
2152  * the latter case, the provided id_priv may already be on the bind_list, but
2153  * we still need to check that it's okay to start listening.
2154  */
2155 static int cma_check_port(struct rdma_bind_list *bind_list,
2156 			  struct rdma_id_private *id_priv, uint8_t reuseaddr)
2157 {
2158 	struct rdma_id_private *cur_id;
2159 	struct sockaddr *addr, *cur_addr;
2160 	struct hlist_node *node;
2161 
2162 	addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2163 	if (cma_any_addr(addr) && !reuseaddr)
2164 		return -EADDRNOTAVAIL;
2165 
2166 	hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2167 		if (id_priv == cur_id)
2168 			continue;
2169 
2170 		if ((cur_id->state == RDMA_CM_LISTEN) ||
2171 		    !reuseaddr || !cur_id->reuseaddr) {
2172 			cur_addr = (struct sockaddr *) &cur_id->id.route.addr.src_addr;
2173 			if (cma_any_addr(cur_addr))
2174 				return -EADDRNOTAVAIL;
2175 
2176 			if (!cma_addr_cmp(addr, cur_addr))
2177 				return -EADDRINUSE;
2178 		}
2179 	}
2180 	return 0;
2181 }
2182 
2183 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2184 {
2185 	struct rdma_bind_list *bind_list;
2186 	unsigned short snum;
2187 	int ret;
2188 
2189 	snum = ntohs(cma_port((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2190 	if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2191 		return -EACCES;
2192 
2193 	bind_list = idr_find(ps, snum);
2194 	if (!bind_list) {
2195 		ret = cma_alloc_port(ps, id_priv, snum);
2196 	} else {
2197 		ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2198 		if (!ret)
2199 			cma_bind_port(bind_list, id_priv);
2200 	}
2201 	return ret;
2202 }
2203 
2204 static int cma_bind_listen(struct rdma_id_private *id_priv)
2205 {
2206 	struct rdma_bind_list *bind_list = id_priv->bind_list;
2207 	int ret = 0;
2208 
2209 	mutex_lock(&lock);
2210 	if (bind_list->owners.first->next)
2211 		ret = cma_check_port(bind_list, id_priv, 0);
2212 	mutex_unlock(&lock);
2213 	return ret;
2214 }
2215 
2216 static int cma_get_port(struct rdma_id_private *id_priv)
2217 {
2218 	struct idr *ps;
2219 	int ret;
2220 
2221 	switch (id_priv->id.ps) {
2222 	case RDMA_PS_SDP:
2223 		ps = &sdp_ps;
2224 		break;
2225 	case RDMA_PS_TCP:
2226 		ps = &tcp_ps;
2227 		break;
2228 	case RDMA_PS_UDP:
2229 		ps = &udp_ps;
2230 		break;
2231 	case RDMA_PS_IPOIB:
2232 		ps = &ipoib_ps;
2233 		break;
2234 	default:
2235 		return -EPROTONOSUPPORT;
2236 	}
2237 
2238 	mutex_lock(&lock);
2239 	if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2240 		ret = cma_alloc_any_port(ps, id_priv);
2241 	else
2242 		ret = cma_use_port(ps, id_priv);
2243 	mutex_unlock(&lock);
2244 
2245 	return ret;
2246 }
2247 
2248 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2249 			       struct sockaddr *addr)
2250 {
2251 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2252 	struct sockaddr_in6 *sin6;
2253 
2254 	if (addr->sa_family != AF_INET6)
2255 		return 0;
2256 
2257 	sin6 = (struct sockaddr_in6 *) addr;
2258 	if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2259 	    !sin6->sin6_scope_id)
2260 			return -EINVAL;
2261 
2262 	dev_addr->bound_dev_if = sin6->sin6_scope_id;
2263 #endif
2264 	return 0;
2265 }
2266 
2267 int rdma_listen(struct rdma_cm_id *id, int backlog)
2268 {
2269 	struct rdma_id_private *id_priv;
2270 	int ret;
2271 
2272 	id_priv = container_of(id, struct rdma_id_private, id);
2273 	if (id_priv->state == RDMA_CM_IDLE) {
2274 		((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
2275 		ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
2276 		if (ret)
2277 			return ret;
2278 	}
2279 
2280 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2281 		return -EINVAL;
2282 
2283 	if (id_priv->reuseaddr) {
2284 		ret = cma_bind_listen(id_priv);
2285 		if (ret)
2286 			goto err;
2287 	}
2288 
2289 	id_priv->backlog = backlog;
2290 	if (id->device) {
2291 		switch (rdma_node_get_transport(id->device->node_type)) {
2292 		case RDMA_TRANSPORT_IB:
2293 			ret = cma_ib_listen(id_priv);
2294 			if (ret)
2295 				goto err;
2296 			break;
2297 		case RDMA_TRANSPORT_IWARP:
2298 			ret = cma_iw_listen(id_priv, backlog);
2299 			if (ret)
2300 				goto err;
2301 			break;
2302 		default:
2303 			ret = -ENOSYS;
2304 			goto err;
2305 		}
2306 	} else
2307 		cma_listen_on_all(id_priv);
2308 
2309 	return 0;
2310 err:
2311 	id_priv->backlog = 0;
2312 	cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2313 	return ret;
2314 }
2315 EXPORT_SYMBOL(rdma_listen);
2316 
2317 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2318 {
2319 	struct rdma_id_private *id_priv;
2320 	int ret;
2321 
2322 	if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2323 		return -EAFNOSUPPORT;
2324 
2325 	id_priv = container_of(id, struct rdma_id_private, id);
2326 	if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2327 		return -EINVAL;
2328 
2329 	ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2330 	if (ret)
2331 		goto err1;
2332 
2333 	if (!cma_any_addr(addr)) {
2334 		ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2335 		if (ret)
2336 			goto err1;
2337 
2338 		ret = cma_acquire_dev(id_priv);
2339 		if (ret)
2340 			goto err1;
2341 	}
2342 
2343 	memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2344 	ret = cma_get_port(id_priv);
2345 	if (ret)
2346 		goto err2;
2347 
2348 	return 0;
2349 err2:
2350 	if (id_priv->cma_dev)
2351 		cma_release_dev(id_priv);
2352 err1:
2353 	cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2354 	return ret;
2355 }
2356 EXPORT_SYMBOL(rdma_bind_addr);
2357 
2358 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2359 			  struct rdma_route *route)
2360 {
2361 	struct cma_hdr *cma_hdr;
2362 	struct sdp_hh *sdp_hdr;
2363 
2364 	if (route->addr.src_addr.ss_family == AF_INET) {
2365 		struct sockaddr_in *src4, *dst4;
2366 
2367 		src4 = (struct sockaddr_in *) &route->addr.src_addr;
2368 		dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2369 
2370 		switch (ps) {
2371 		case RDMA_PS_SDP:
2372 			sdp_hdr = hdr;
2373 			if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2374 				return -EINVAL;
2375 			sdp_set_ip_ver(sdp_hdr, 4);
2376 			sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2377 			sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2378 			sdp_hdr->port = src4->sin_port;
2379 			break;
2380 		default:
2381 			cma_hdr = hdr;
2382 			cma_hdr->cma_version = CMA_VERSION;
2383 			cma_set_ip_ver(cma_hdr, 4);
2384 			cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2385 			cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2386 			cma_hdr->port = src4->sin_port;
2387 			break;
2388 		}
2389 	} else {
2390 		struct sockaddr_in6 *src6, *dst6;
2391 
2392 		src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2393 		dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2394 
2395 		switch (ps) {
2396 		case RDMA_PS_SDP:
2397 			sdp_hdr = hdr;
2398 			if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2399 				return -EINVAL;
2400 			sdp_set_ip_ver(sdp_hdr, 6);
2401 			sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2402 			sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2403 			sdp_hdr->port = src6->sin6_port;
2404 			break;
2405 		default:
2406 			cma_hdr = hdr;
2407 			cma_hdr->cma_version = CMA_VERSION;
2408 			cma_set_ip_ver(cma_hdr, 6);
2409 			cma_hdr->src_addr.ip6 = src6->sin6_addr;
2410 			cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2411 			cma_hdr->port = src6->sin6_port;
2412 			break;
2413 		}
2414 	}
2415 	return 0;
2416 }
2417 
2418 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2419 				struct ib_cm_event *ib_event)
2420 {
2421 	struct rdma_id_private *id_priv = cm_id->context;
2422 	struct rdma_cm_event event;
2423 	struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2424 	int ret = 0;
2425 
2426 	if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2427 		return 0;
2428 
2429 	memset(&event, 0, sizeof event);
2430 	switch (ib_event->event) {
2431 	case IB_CM_SIDR_REQ_ERROR:
2432 		event.event = RDMA_CM_EVENT_UNREACHABLE;
2433 		event.status = -ETIMEDOUT;
2434 		break;
2435 	case IB_CM_SIDR_REP_RECEIVED:
2436 		event.param.ud.private_data = ib_event->private_data;
2437 		event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2438 		if (rep->status != IB_SIDR_SUCCESS) {
2439 			event.event = RDMA_CM_EVENT_UNREACHABLE;
2440 			event.status = ib_event->param.sidr_rep_rcvd.status;
2441 			break;
2442 		}
2443 		ret = cma_set_qkey(id_priv);
2444 		if (ret) {
2445 			event.event = RDMA_CM_EVENT_ADDR_ERROR;
2446 			event.status = -EINVAL;
2447 			break;
2448 		}
2449 		if (id_priv->qkey != rep->qkey) {
2450 			event.event = RDMA_CM_EVENT_UNREACHABLE;
2451 			event.status = -EINVAL;
2452 			break;
2453 		}
2454 		ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2455 				     id_priv->id.route.path_rec,
2456 				     &event.param.ud.ah_attr);
2457 		event.param.ud.qp_num = rep->qpn;
2458 		event.param.ud.qkey = rep->qkey;
2459 		event.event = RDMA_CM_EVENT_ESTABLISHED;
2460 		event.status = 0;
2461 		break;
2462 	default:
2463 		printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2464 		       ib_event->event);
2465 		goto out;
2466 	}
2467 
2468 	ret = id_priv->id.event_handler(&id_priv->id, &event);
2469 	if (ret) {
2470 		/* Destroy the CM ID by returning a non-zero value. */
2471 		id_priv->cm_id.ib = NULL;
2472 		cma_exch(id_priv, RDMA_CM_DESTROYING);
2473 		mutex_unlock(&id_priv->handler_mutex);
2474 		rdma_destroy_id(&id_priv->id);
2475 		return ret;
2476 	}
2477 out:
2478 	mutex_unlock(&id_priv->handler_mutex);
2479 	return ret;
2480 }
2481 
2482 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2483 			      struct rdma_conn_param *conn_param)
2484 {
2485 	struct ib_cm_sidr_req_param req;
2486 	struct rdma_route *route;
2487 	int ret;
2488 
2489 	req.private_data_len = sizeof(struct cma_hdr) +
2490 			       conn_param->private_data_len;
2491 	req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2492 	if (!req.private_data)
2493 		return -ENOMEM;
2494 
2495 	if (conn_param->private_data && conn_param->private_data_len)
2496 		memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2497 		       conn_param->private_data, conn_param->private_data_len);
2498 
2499 	route = &id_priv->id.route;
2500 	ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2501 	if (ret)
2502 		goto out;
2503 
2504 	id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2505 					    cma_sidr_rep_handler, id_priv);
2506 	if (IS_ERR(id_priv->cm_id.ib)) {
2507 		ret = PTR_ERR(id_priv->cm_id.ib);
2508 		goto out;
2509 	}
2510 
2511 	req.path = route->path_rec;
2512 	req.service_id = cma_get_service_id(id_priv->id.ps,
2513 					    (struct sockaddr *) &route->addr.dst_addr);
2514 	req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2515 	req.max_cm_retries = CMA_MAX_CM_RETRIES;
2516 
2517 	ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2518 	if (ret) {
2519 		ib_destroy_cm_id(id_priv->cm_id.ib);
2520 		id_priv->cm_id.ib = NULL;
2521 	}
2522 out:
2523 	kfree(req.private_data);
2524 	return ret;
2525 }
2526 
2527 static int cma_connect_ib(struct rdma_id_private *id_priv,
2528 			  struct rdma_conn_param *conn_param)
2529 {
2530 	struct ib_cm_req_param req;
2531 	struct rdma_route *route;
2532 	void *private_data;
2533 	int offset, ret;
2534 
2535 	memset(&req, 0, sizeof req);
2536 	offset = cma_user_data_offset(id_priv->id.ps);
2537 	req.private_data_len = offset + conn_param->private_data_len;
2538 	private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2539 	if (!private_data)
2540 		return -ENOMEM;
2541 
2542 	if (conn_param->private_data && conn_param->private_data_len)
2543 		memcpy(private_data + offset, conn_param->private_data,
2544 		       conn_param->private_data_len);
2545 
2546 	id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2547 					    id_priv);
2548 	if (IS_ERR(id_priv->cm_id.ib)) {
2549 		ret = PTR_ERR(id_priv->cm_id.ib);
2550 		goto out;
2551 	}
2552 
2553 	route = &id_priv->id.route;
2554 	ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2555 	if (ret)
2556 		goto out;
2557 	req.private_data = private_data;
2558 
2559 	req.primary_path = &route->path_rec[0];
2560 	if (route->num_paths == 2)
2561 		req.alternate_path = &route->path_rec[1];
2562 
2563 	req.service_id = cma_get_service_id(id_priv->id.ps,
2564 					    (struct sockaddr *) &route->addr.dst_addr);
2565 	req.qp_num = id_priv->qp_num;
2566 	req.qp_type = IB_QPT_RC;
2567 	req.starting_psn = id_priv->seq_num;
2568 	req.responder_resources = conn_param->responder_resources;
2569 	req.initiator_depth = conn_param->initiator_depth;
2570 	req.flow_control = conn_param->flow_control;
2571 	req.retry_count = conn_param->retry_count;
2572 	req.rnr_retry_count = conn_param->rnr_retry_count;
2573 	req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2574 	req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2575 	req.max_cm_retries = CMA_MAX_CM_RETRIES;
2576 	req.srq = id_priv->srq ? 1 : 0;
2577 
2578 	ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2579 out:
2580 	if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2581 		ib_destroy_cm_id(id_priv->cm_id.ib);
2582 		id_priv->cm_id.ib = NULL;
2583 	}
2584 
2585 	kfree(private_data);
2586 	return ret;
2587 }
2588 
2589 static int cma_connect_iw(struct rdma_id_private *id_priv,
2590 			  struct rdma_conn_param *conn_param)
2591 {
2592 	struct iw_cm_id *cm_id;
2593 	struct sockaddr_in* sin;
2594 	int ret;
2595 	struct iw_cm_conn_param iw_param;
2596 
2597 	cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2598 	if (IS_ERR(cm_id)) {
2599 		ret = PTR_ERR(cm_id);
2600 		goto out;
2601 	}
2602 
2603 	id_priv->cm_id.iw = cm_id;
2604 
2605 	sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2606 	cm_id->local_addr = *sin;
2607 
2608 	sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2609 	cm_id->remote_addr = *sin;
2610 
2611 	ret = cma_modify_qp_rtr(id_priv, conn_param);
2612 	if (ret)
2613 		goto out;
2614 
2615 	iw_param.ord = conn_param->initiator_depth;
2616 	iw_param.ird = conn_param->responder_resources;
2617 	iw_param.private_data = conn_param->private_data;
2618 	iw_param.private_data_len = conn_param->private_data_len;
2619 	if (id_priv->id.qp)
2620 		iw_param.qpn = id_priv->qp_num;
2621 	else
2622 		iw_param.qpn = conn_param->qp_num;
2623 	ret = iw_cm_connect(cm_id, &iw_param);
2624 out:
2625 	if (ret && !IS_ERR(cm_id)) {
2626 		iw_destroy_cm_id(cm_id);
2627 		id_priv->cm_id.iw = NULL;
2628 	}
2629 	return ret;
2630 }
2631 
2632 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2633 {
2634 	struct rdma_id_private *id_priv;
2635 	int ret;
2636 
2637 	id_priv = container_of(id, struct rdma_id_private, id);
2638 	if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2639 		return -EINVAL;
2640 
2641 	if (!id->qp) {
2642 		id_priv->qp_num = conn_param->qp_num;
2643 		id_priv->srq = conn_param->srq;
2644 	}
2645 
2646 	switch (rdma_node_get_transport(id->device->node_type)) {
2647 	case RDMA_TRANSPORT_IB:
2648 		if (id->qp_type == IB_QPT_UD)
2649 			ret = cma_resolve_ib_udp(id_priv, conn_param);
2650 		else
2651 			ret = cma_connect_ib(id_priv, conn_param);
2652 		break;
2653 	case RDMA_TRANSPORT_IWARP:
2654 		ret = cma_connect_iw(id_priv, conn_param);
2655 		break;
2656 	default:
2657 		ret = -ENOSYS;
2658 		break;
2659 	}
2660 	if (ret)
2661 		goto err;
2662 
2663 	return 0;
2664 err:
2665 	cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2666 	return ret;
2667 }
2668 EXPORT_SYMBOL(rdma_connect);
2669 
2670 static int cma_accept_ib(struct rdma_id_private *id_priv,
2671 			 struct rdma_conn_param *conn_param)
2672 {
2673 	struct ib_cm_rep_param rep;
2674 	int ret;
2675 
2676 	ret = cma_modify_qp_rtr(id_priv, conn_param);
2677 	if (ret)
2678 		goto out;
2679 
2680 	ret = cma_modify_qp_rts(id_priv, conn_param);
2681 	if (ret)
2682 		goto out;
2683 
2684 	memset(&rep, 0, sizeof rep);
2685 	rep.qp_num = id_priv->qp_num;
2686 	rep.starting_psn = id_priv->seq_num;
2687 	rep.private_data = conn_param->private_data;
2688 	rep.private_data_len = conn_param->private_data_len;
2689 	rep.responder_resources = conn_param->responder_resources;
2690 	rep.initiator_depth = conn_param->initiator_depth;
2691 	rep.failover_accepted = 0;
2692 	rep.flow_control = conn_param->flow_control;
2693 	rep.rnr_retry_count = conn_param->rnr_retry_count;
2694 	rep.srq = id_priv->srq ? 1 : 0;
2695 
2696 	ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2697 out:
2698 	return ret;
2699 }
2700 
2701 static int cma_accept_iw(struct rdma_id_private *id_priv,
2702 		  struct rdma_conn_param *conn_param)
2703 {
2704 	struct iw_cm_conn_param iw_param;
2705 	int ret;
2706 
2707 	ret = cma_modify_qp_rtr(id_priv, conn_param);
2708 	if (ret)
2709 		return ret;
2710 
2711 	iw_param.ord = conn_param->initiator_depth;
2712 	iw_param.ird = conn_param->responder_resources;
2713 	iw_param.private_data = conn_param->private_data;
2714 	iw_param.private_data_len = conn_param->private_data_len;
2715 	if (id_priv->id.qp) {
2716 		iw_param.qpn = id_priv->qp_num;
2717 	} else
2718 		iw_param.qpn = conn_param->qp_num;
2719 
2720 	return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2721 }
2722 
2723 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2724 			     enum ib_cm_sidr_status status,
2725 			     const void *private_data, int private_data_len)
2726 {
2727 	struct ib_cm_sidr_rep_param rep;
2728 	int ret;
2729 
2730 	memset(&rep, 0, sizeof rep);
2731 	rep.status = status;
2732 	if (status == IB_SIDR_SUCCESS) {
2733 		ret = cma_set_qkey(id_priv);
2734 		if (ret)
2735 			return ret;
2736 		rep.qp_num = id_priv->qp_num;
2737 		rep.qkey = id_priv->qkey;
2738 	}
2739 	rep.private_data = private_data;
2740 	rep.private_data_len = private_data_len;
2741 
2742 	return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2743 }
2744 
2745 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2746 {
2747 	struct rdma_id_private *id_priv;
2748 	int ret;
2749 
2750 	id_priv = container_of(id, struct rdma_id_private, id);
2751 
2752 	id_priv->owner = task_pid_nr(current);
2753 
2754 	if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2755 		return -EINVAL;
2756 
2757 	if (!id->qp && conn_param) {
2758 		id_priv->qp_num = conn_param->qp_num;
2759 		id_priv->srq = conn_param->srq;
2760 	}
2761 
2762 	switch (rdma_node_get_transport(id->device->node_type)) {
2763 	case RDMA_TRANSPORT_IB:
2764 		if (id->qp_type == IB_QPT_UD)
2765 			ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2766 						conn_param->private_data,
2767 						conn_param->private_data_len);
2768 		else if (conn_param)
2769 			ret = cma_accept_ib(id_priv, conn_param);
2770 		else
2771 			ret = cma_rep_recv(id_priv);
2772 		break;
2773 	case RDMA_TRANSPORT_IWARP:
2774 		ret = cma_accept_iw(id_priv, conn_param);
2775 		break;
2776 	default:
2777 		ret = -ENOSYS;
2778 		break;
2779 	}
2780 
2781 	if (ret)
2782 		goto reject;
2783 
2784 	return 0;
2785 reject:
2786 	cma_modify_qp_err(id_priv);
2787 	rdma_reject(id, NULL, 0);
2788 	return ret;
2789 }
2790 EXPORT_SYMBOL(rdma_accept);
2791 
2792 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2793 {
2794 	struct rdma_id_private *id_priv;
2795 	int ret;
2796 
2797 	id_priv = container_of(id, struct rdma_id_private, id);
2798 	if (!cma_has_cm_dev(id_priv))
2799 		return -EINVAL;
2800 
2801 	switch (id->device->node_type) {
2802 	case RDMA_NODE_IB_CA:
2803 		ret = ib_cm_notify(id_priv->cm_id.ib, event);
2804 		break;
2805 	default:
2806 		ret = 0;
2807 		break;
2808 	}
2809 	return ret;
2810 }
2811 EXPORT_SYMBOL(rdma_notify);
2812 
2813 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2814 		u8 private_data_len)
2815 {
2816 	struct rdma_id_private *id_priv;
2817 	int ret;
2818 
2819 	id_priv = container_of(id, struct rdma_id_private, id);
2820 	if (!cma_has_cm_dev(id_priv))
2821 		return -EINVAL;
2822 
2823 	switch (rdma_node_get_transport(id->device->node_type)) {
2824 	case RDMA_TRANSPORT_IB:
2825 		if (id->qp_type == IB_QPT_UD)
2826 			ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2827 						private_data, private_data_len);
2828 		else
2829 			ret = ib_send_cm_rej(id_priv->cm_id.ib,
2830 					     IB_CM_REJ_CONSUMER_DEFINED, NULL,
2831 					     0, private_data, private_data_len);
2832 		break;
2833 	case RDMA_TRANSPORT_IWARP:
2834 		ret = iw_cm_reject(id_priv->cm_id.iw,
2835 				   private_data, private_data_len);
2836 		break;
2837 	default:
2838 		ret = -ENOSYS;
2839 		break;
2840 	}
2841 	return ret;
2842 }
2843 EXPORT_SYMBOL(rdma_reject);
2844 
2845 int rdma_disconnect(struct rdma_cm_id *id)
2846 {
2847 	struct rdma_id_private *id_priv;
2848 	int ret;
2849 
2850 	id_priv = container_of(id, struct rdma_id_private, id);
2851 	if (!cma_has_cm_dev(id_priv))
2852 		return -EINVAL;
2853 
2854 	switch (rdma_node_get_transport(id->device->node_type)) {
2855 	case RDMA_TRANSPORT_IB:
2856 		ret = cma_modify_qp_err(id_priv);
2857 		if (ret)
2858 			goto out;
2859 		/* Initiate or respond to a disconnect. */
2860 		if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2861 			ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2862 		break;
2863 	case RDMA_TRANSPORT_IWARP:
2864 		ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2865 		break;
2866 	default:
2867 		ret = -EINVAL;
2868 		break;
2869 	}
2870 out:
2871 	return ret;
2872 }
2873 EXPORT_SYMBOL(rdma_disconnect);
2874 
2875 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2876 {
2877 	struct rdma_id_private *id_priv;
2878 	struct cma_multicast *mc = multicast->context;
2879 	struct rdma_cm_event event;
2880 	int ret;
2881 
2882 	id_priv = mc->id_priv;
2883 	if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
2884 	    cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
2885 		return 0;
2886 
2887 	mutex_lock(&id_priv->qp_mutex);
2888 	if (!status && id_priv->id.qp)
2889 		status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2890 					 multicast->rec.mlid);
2891 	mutex_unlock(&id_priv->qp_mutex);
2892 
2893 	memset(&event, 0, sizeof event);
2894 	event.status = status;
2895 	event.param.ud.private_data = mc->context;
2896 	if (!status) {
2897 		event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2898 		ib_init_ah_from_mcmember(id_priv->id.device,
2899 					 id_priv->id.port_num, &multicast->rec,
2900 					 &event.param.ud.ah_attr);
2901 		event.param.ud.qp_num = 0xFFFFFF;
2902 		event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2903 	} else
2904 		event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2905 
2906 	ret = id_priv->id.event_handler(&id_priv->id, &event);
2907 	if (ret) {
2908 		cma_exch(id_priv, RDMA_CM_DESTROYING);
2909 		mutex_unlock(&id_priv->handler_mutex);
2910 		rdma_destroy_id(&id_priv->id);
2911 		return 0;
2912 	}
2913 
2914 	mutex_unlock(&id_priv->handler_mutex);
2915 	return 0;
2916 }
2917 
2918 static void cma_set_mgid(struct rdma_id_private *id_priv,
2919 			 struct sockaddr *addr, union ib_gid *mgid)
2920 {
2921 	unsigned char mc_map[MAX_ADDR_LEN];
2922 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2923 	struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2924 	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2925 
2926 	if (cma_any_addr(addr)) {
2927 		memset(mgid, 0, sizeof *mgid);
2928 	} else if ((addr->sa_family == AF_INET6) &&
2929 		   ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2930 								 0xFF10A01B)) {
2931 		/* IPv6 address is an SA assigned MGID. */
2932 		memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2933 	} else if ((addr->sa_family == AF_INET6)) {
2934 		ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2935 		if (id_priv->id.ps == RDMA_PS_UDP)
2936 			mc_map[7] = 0x01;	/* Use RDMA CM signature */
2937 		*mgid = *(union ib_gid *) (mc_map + 4);
2938 	} else {
2939 		ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2940 		if (id_priv->id.ps == RDMA_PS_UDP)
2941 			mc_map[7] = 0x01;	/* Use RDMA CM signature */
2942 		*mgid = *(union ib_gid *) (mc_map + 4);
2943 	}
2944 }
2945 
2946 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2947 				 struct cma_multicast *mc)
2948 {
2949 	struct ib_sa_mcmember_rec rec;
2950 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2951 	ib_sa_comp_mask comp_mask;
2952 	int ret;
2953 
2954 	ib_addr_get_mgid(dev_addr, &rec.mgid);
2955 	ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2956 				     &rec.mgid, &rec);
2957 	if (ret)
2958 		return ret;
2959 
2960 	cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
2961 	if (id_priv->id.ps == RDMA_PS_UDP)
2962 		rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2963 	rdma_addr_get_sgid(dev_addr, &rec.port_gid);
2964 	rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2965 	rec.join_state = 1;
2966 
2967 	comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2968 		    IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2969 		    IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2970 		    IB_SA_MCMEMBER_REC_FLOW_LABEL |
2971 		    IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2972 
2973 	if (id_priv->id.ps == RDMA_PS_IPOIB)
2974 		comp_mask |= IB_SA_MCMEMBER_REC_RATE |
2975 			     IB_SA_MCMEMBER_REC_RATE_SELECTOR;
2976 
2977 	mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2978 						id_priv->id.port_num, &rec,
2979 						comp_mask, GFP_KERNEL,
2980 						cma_ib_mc_handler, mc);
2981 	if (IS_ERR(mc->multicast.ib))
2982 		return PTR_ERR(mc->multicast.ib);
2983 
2984 	return 0;
2985 }
2986 
2987 static void iboe_mcast_work_handler(struct work_struct *work)
2988 {
2989 	struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
2990 	struct cma_multicast *mc = mw->mc;
2991 	struct ib_sa_multicast *m = mc->multicast.ib;
2992 
2993 	mc->multicast.ib->context = mc;
2994 	cma_ib_mc_handler(0, m);
2995 	kref_put(&mc->mcref, release_mc);
2996 	kfree(mw);
2997 }
2998 
2999 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3000 {
3001 	struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3002 	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3003 
3004 	if (cma_any_addr(addr)) {
3005 		memset(mgid, 0, sizeof *mgid);
3006 	} else if (addr->sa_family == AF_INET6) {
3007 		memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3008 	} else {
3009 		mgid->raw[0] = 0xff;
3010 		mgid->raw[1] = 0x0e;
3011 		mgid->raw[2] = 0;
3012 		mgid->raw[3] = 0;
3013 		mgid->raw[4] = 0;
3014 		mgid->raw[5] = 0;
3015 		mgid->raw[6] = 0;
3016 		mgid->raw[7] = 0;
3017 		mgid->raw[8] = 0;
3018 		mgid->raw[9] = 0;
3019 		mgid->raw[10] = 0xff;
3020 		mgid->raw[11] = 0xff;
3021 		*(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3022 	}
3023 }
3024 
3025 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3026 				   struct cma_multicast *mc)
3027 {
3028 	struct iboe_mcast_work *work;
3029 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3030 	int err;
3031 	struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3032 	struct net_device *ndev = NULL;
3033 
3034 	if (cma_zero_addr((struct sockaddr *)&mc->addr))
3035 		return -EINVAL;
3036 
3037 	work = kzalloc(sizeof *work, GFP_KERNEL);
3038 	if (!work)
3039 		return -ENOMEM;
3040 
3041 	mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3042 	if (!mc->multicast.ib) {
3043 		err = -ENOMEM;
3044 		goto out1;
3045 	}
3046 
3047 	cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3048 
3049 	mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3050 	if (id_priv->id.ps == RDMA_PS_UDP)
3051 		mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3052 
3053 	if (dev_addr->bound_dev_if)
3054 		ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3055 	if (!ndev) {
3056 		err = -ENODEV;
3057 		goto out2;
3058 	}
3059 	mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3060 	mc->multicast.ib->rec.hop_limit = 1;
3061 	mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3062 	dev_put(ndev);
3063 	if (!mc->multicast.ib->rec.mtu) {
3064 		err = -EINVAL;
3065 		goto out2;
3066 	}
3067 	iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3068 	work->id = id_priv;
3069 	work->mc = mc;
3070 	INIT_WORK(&work->work, iboe_mcast_work_handler);
3071 	kref_get(&mc->mcref);
3072 	queue_work(cma_wq, &work->work);
3073 
3074 	return 0;
3075 
3076 out2:
3077 	kfree(mc->multicast.ib);
3078 out1:
3079 	kfree(work);
3080 	return err;
3081 }
3082 
3083 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3084 			void *context)
3085 {
3086 	struct rdma_id_private *id_priv;
3087 	struct cma_multicast *mc;
3088 	int ret;
3089 
3090 	id_priv = container_of(id, struct rdma_id_private, id);
3091 	if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3092 	    !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3093 		return -EINVAL;
3094 
3095 	mc = kmalloc(sizeof *mc, GFP_KERNEL);
3096 	if (!mc)
3097 		return -ENOMEM;
3098 
3099 	memcpy(&mc->addr, addr, ip_addr_size(addr));
3100 	mc->context = context;
3101 	mc->id_priv = id_priv;
3102 
3103 	spin_lock(&id_priv->lock);
3104 	list_add(&mc->list, &id_priv->mc_list);
3105 	spin_unlock(&id_priv->lock);
3106 
3107 	switch (rdma_node_get_transport(id->device->node_type)) {
3108 	case RDMA_TRANSPORT_IB:
3109 		switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3110 		case IB_LINK_LAYER_INFINIBAND:
3111 			ret = cma_join_ib_multicast(id_priv, mc);
3112 			break;
3113 		case IB_LINK_LAYER_ETHERNET:
3114 			kref_init(&mc->mcref);
3115 			ret = cma_iboe_join_multicast(id_priv, mc);
3116 			break;
3117 		default:
3118 			ret = -EINVAL;
3119 		}
3120 		break;
3121 	default:
3122 		ret = -ENOSYS;
3123 		break;
3124 	}
3125 
3126 	if (ret) {
3127 		spin_lock_irq(&id_priv->lock);
3128 		list_del(&mc->list);
3129 		spin_unlock_irq(&id_priv->lock);
3130 		kfree(mc);
3131 	}
3132 	return ret;
3133 }
3134 EXPORT_SYMBOL(rdma_join_multicast);
3135 
3136 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3137 {
3138 	struct rdma_id_private *id_priv;
3139 	struct cma_multicast *mc;
3140 
3141 	id_priv = container_of(id, struct rdma_id_private, id);
3142 	spin_lock_irq(&id_priv->lock);
3143 	list_for_each_entry(mc, &id_priv->mc_list, list) {
3144 		if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3145 			list_del(&mc->list);
3146 			spin_unlock_irq(&id_priv->lock);
3147 
3148 			if (id->qp)
3149 				ib_detach_mcast(id->qp,
3150 						&mc->multicast.ib->rec.mgid,
3151 						mc->multicast.ib->rec.mlid);
3152 			if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3153 				switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3154 				case IB_LINK_LAYER_INFINIBAND:
3155 					ib_sa_free_multicast(mc->multicast.ib);
3156 					kfree(mc);
3157 					break;
3158 				case IB_LINK_LAYER_ETHERNET:
3159 					kref_put(&mc->mcref, release_mc);
3160 					break;
3161 				default:
3162 					break;
3163 				}
3164 			}
3165 			return;
3166 		}
3167 	}
3168 	spin_unlock_irq(&id_priv->lock);
3169 }
3170 EXPORT_SYMBOL(rdma_leave_multicast);
3171 
3172 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3173 {
3174 	struct rdma_dev_addr *dev_addr;
3175 	struct cma_ndev_work *work;
3176 
3177 	dev_addr = &id_priv->id.route.addr.dev_addr;
3178 
3179 	if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3180 	    memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3181 		printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3182 		       ndev->name, &id_priv->id);
3183 		work = kzalloc(sizeof *work, GFP_KERNEL);
3184 		if (!work)
3185 			return -ENOMEM;
3186 
3187 		INIT_WORK(&work->work, cma_ndev_work_handler);
3188 		work->id = id_priv;
3189 		work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3190 		atomic_inc(&id_priv->refcount);
3191 		queue_work(cma_wq, &work->work);
3192 	}
3193 
3194 	return 0;
3195 }
3196 
3197 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3198 			       void *ctx)
3199 {
3200 	struct net_device *ndev = (struct net_device *)ctx;
3201 	struct cma_device *cma_dev;
3202 	struct rdma_id_private *id_priv;
3203 	int ret = NOTIFY_DONE;
3204 
3205 	if (dev_net(ndev) != &init_net)
3206 		return NOTIFY_DONE;
3207 
3208 	if (event != NETDEV_BONDING_FAILOVER)
3209 		return NOTIFY_DONE;
3210 
3211 	if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3212 		return NOTIFY_DONE;
3213 
3214 	mutex_lock(&lock);
3215 	list_for_each_entry(cma_dev, &dev_list, list)
3216 		list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3217 			ret = cma_netdev_change(ndev, id_priv);
3218 			if (ret)
3219 				goto out;
3220 		}
3221 
3222 out:
3223 	mutex_unlock(&lock);
3224 	return ret;
3225 }
3226 
3227 static struct notifier_block cma_nb = {
3228 	.notifier_call = cma_netdev_callback
3229 };
3230 
3231 static void cma_add_one(struct ib_device *device)
3232 {
3233 	struct cma_device *cma_dev;
3234 	struct rdma_id_private *id_priv;
3235 
3236 	cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3237 	if (!cma_dev)
3238 		return;
3239 
3240 	cma_dev->device = device;
3241 
3242 	init_completion(&cma_dev->comp);
3243 	atomic_set(&cma_dev->refcount, 1);
3244 	INIT_LIST_HEAD(&cma_dev->id_list);
3245 	ib_set_client_data(device, &cma_client, cma_dev);
3246 
3247 	mutex_lock(&lock);
3248 	list_add_tail(&cma_dev->list, &dev_list);
3249 	list_for_each_entry(id_priv, &listen_any_list, list)
3250 		cma_listen_on_dev(id_priv, cma_dev);
3251 	mutex_unlock(&lock);
3252 }
3253 
3254 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3255 {
3256 	struct rdma_cm_event event;
3257 	enum rdma_cm_state state;
3258 	int ret = 0;
3259 
3260 	/* Record that we want to remove the device */
3261 	state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3262 	if (state == RDMA_CM_DESTROYING)
3263 		return 0;
3264 
3265 	cma_cancel_operation(id_priv, state);
3266 	mutex_lock(&id_priv->handler_mutex);
3267 
3268 	/* Check for destruction from another callback. */
3269 	if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3270 		goto out;
3271 
3272 	memset(&event, 0, sizeof event);
3273 	event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3274 	ret = id_priv->id.event_handler(&id_priv->id, &event);
3275 out:
3276 	mutex_unlock(&id_priv->handler_mutex);
3277 	return ret;
3278 }
3279 
3280 static void cma_process_remove(struct cma_device *cma_dev)
3281 {
3282 	struct rdma_id_private *id_priv;
3283 	int ret;
3284 
3285 	mutex_lock(&lock);
3286 	while (!list_empty(&cma_dev->id_list)) {
3287 		id_priv = list_entry(cma_dev->id_list.next,
3288 				     struct rdma_id_private, list);
3289 
3290 		list_del(&id_priv->listen_list);
3291 		list_del_init(&id_priv->list);
3292 		atomic_inc(&id_priv->refcount);
3293 		mutex_unlock(&lock);
3294 
3295 		ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3296 		cma_deref_id(id_priv);
3297 		if (ret)
3298 			rdma_destroy_id(&id_priv->id);
3299 
3300 		mutex_lock(&lock);
3301 	}
3302 	mutex_unlock(&lock);
3303 
3304 	cma_deref_dev(cma_dev);
3305 	wait_for_completion(&cma_dev->comp);
3306 }
3307 
3308 static void cma_remove_one(struct ib_device *device)
3309 {
3310 	struct cma_device *cma_dev;
3311 
3312 	cma_dev = ib_get_client_data(device, &cma_client);
3313 	if (!cma_dev)
3314 		return;
3315 
3316 	mutex_lock(&lock);
3317 	list_del(&cma_dev->list);
3318 	mutex_unlock(&lock);
3319 
3320 	cma_process_remove(cma_dev);
3321 	kfree(cma_dev);
3322 }
3323 
3324 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3325 {
3326 	struct nlmsghdr *nlh;
3327 	struct rdma_cm_id_stats *id_stats;
3328 	struct rdma_id_private *id_priv;
3329 	struct rdma_cm_id *id = NULL;
3330 	struct cma_device *cma_dev;
3331 	int i_dev = 0, i_id = 0;
3332 
3333 	/*
3334 	 * We export all of the IDs as a sequence of messages.  Each
3335 	 * ID gets its own netlink message.
3336 	 */
3337 	mutex_lock(&lock);
3338 
3339 	list_for_each_entry(cma_dev, &dev_list, list) {
3340 		if (i_dev < cb->args[0]) {
3341 			i_dev++;
3342 			continue;
3343 		}
3344 
3345 		i_id = 0;
3346 		list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3347 			if (i_id < cb->args[1]) {
3348 				i_id++;
3349 				continue;
3350 			}
3351 
3352 			id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3353 						sizeof *id_stats, RDMA_NL_RDMA_CM,
3354 						RDMA_NL_RDMA_CM_ID_STATS);
3355 			if (!id_stats)
3356 				goto out;
3357 
3358 			memset(id_stats, 0, sizeof *id_stats);
3359 			id = &id_priv->id;
3360 			id_stats->node_type = id->route.addr.dev_addr.dev_type;
3361 			id_stats->port_num = id->port_num;
3362 			id_stats->bound_dev_if =
3363 				id->route.addr.dev_addr.bound_dev_if;
3364 
3365 			if (id->route.addr.src_addr.ss_family == AF_INET) {
3366 				if (ibnl_put_attr(skb, nlh,
3367 						  sizeof(struct sockaddr_in),
3368 						  &id->route.addr.src_addr,
3369 						  RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3370 					goto out;
3371 				}
3372 				if (ibnl_put_attr(skb, nlh,
3373 						  sizeof(struct sockaddr_in),
3374 						  &id->route.addr.dst_addr,
3375 						  RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3376 					goto out;
3377 				}
3378 			} else if (id->route.addr.src_addr.ss_family == AF_INET6) {
3379 				if (ibnl_put_attr(skb, nlh,
3380 						  sizeof(struct sockaddr_in6),
3381 						  &id->route.addr.src_addr,
3382 						  RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3383 					goto out;
3384 				}
3385 				if (ibnl_put_attr(skb, nlh,
3386 						  sizeof(struct sockaddr_in6),
3387 						  &id->route.addr.dst_addr,
3388 						  RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3389 					goto out;
3390 				}
3391 			}
3392 
3393 			id_stats->pid		= id_priv->owner;
3394 			id_stats->port_space	= id->ps;
3395 			id_stats->cm_state	= id_priv->state;
3396 			id_stats->qp_num	= id_priv->qp_num;
3397 			id_stats->qp_type	= id->qp_type;
3398 
3399 			i_id++;
3400 		}
3401 
3402 		cb->args[1] = 0;
3403 		i_dev++;
3404 	}
3405 
3406 out:
3407 	mutex_unlock(&lock);
3408 	cb->args[0] = i_dev;
3409 	cb->args[1] = i_id;
3410 
3411 	return skb->len;
3412 }
3413 
3414 static const struct ibnl_client_cbs cma_cb_table[] = {
3415 	[RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats },
3416 };
3417 
3418 static int __init cma_init(void)
3419 {
3420 	int ret;
3421 
3422 	cma_wq = create_singlethread_workqueue("rdma_cm");
3423 	if (!cma_wq)
3424 		return -ENOMEM;
3425 
3426 	ib_sa_register_client(&sa_client);
3427 	rdma_addr_register_client(&addr_client);
3428 	register_netdevice_notifier(&cma_nb);
3429 
3430 	ret = ib_register_client(&cma_client);
3431 	if (ret)
3432 		goto err;
3433 
3434 	if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3435 		printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3436 
3437 	return 0;
3438 
3439 err:
3440 	unregister_netdevice_notifier(&cma_nb);
3441 	rdma_addr_unregister_client(&addr_client);
3442 	ib_sa_unregister_client(&sa_client);
3443 	destroy_workqueue(cma_wq);
3444 	return ret;
3445 }
3446 
3447 static void __exit cma_cleanup(void)
3448 {
3449 	ibnl_remove_client(RDMA_NL_RDMA_CM);
3450 	ib_unregister_client(&cma_client);
3451 	unregister_netdevice_notifier(&cma_nb);
3452 	rdma_addr_unregister_client(&addr_client);
3453 	ib_sa_unregister_client(&sa_client);
3454 	destroy_workqueue(cma_wq);
3455 	idr_destroy(&sdp_ps);
3456 	idr_destroy(&tcp_ps);
3457 	idr_destroy(&udp_ps);
3458 	idr_destroy(&ipoib_ps);
3459 }
3460 
3461 module_init(cma_init);
3462 module_exit(cma_cleanup);
3463