xref: /linux/drivers/infiniband/core/cma.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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 #include <linux/module.h>
45 #include <net/route.h>
46 
47 #include <net/tcp.h>
48 #include <net/ipv6.h>
49 #include <net/ip_fib.h>
50 #include <net/ip6_route.h>
51 
52 #include <rdma/rdma_cm.h>
53 #include <rdma/rdma_cm_ib.h>
54 #include <rdma/rdma_netlink.h>
55 #include <rdma/ib.h>
56 #include <rdma/ib_cache.h>
57 #include <rdma/ib_cm.h>
58 #include <rdma/ib_sa.h>
59 #include <rdma/iw_cm.h>
60 
61 MODULE_AUTHOR("Sean Hefty");
62 MODULE_DESCRIPTION("Generic RDMA CM Agent");
63 MODULE_LICENSE("Dual BSD/GPL");
64 
65 #define CMA_CM_RESPONSE_TIMEOUT 20
66 #define CMA_MAX_CM_RETRIES 15
67 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
68 #define CMA_IBOE_PACKET_LIFETIME 18
69 
70 static const char * const cma_events[] = {
71 	[RDMA_CM_EVENT_ADDR_RESOLVED]	 = "address resolved",
72 	[RDMA_CM_EVENT_ADDR_ERROR]	 = "address error",
73 	[RDMA_CM_EVENT_ROUTE_RESOLVED]	 = "route resolved ",
74 	[RDMA_CM_EVENT_ROUTE_ERROR]	 = "route error",
75 	[RDMA_CM_EVENT_CONNECT_REQUEST]	 = "connect request",
76 	[RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
77 	[RDMA_CM_EVENT_CONNECT_ERROR]	 = "connect error",
78 	[RDMA_CM_EVENT_UNREACHABLE]	 = "unreachable",
79 	[RDMA_CM_EVENT_REJECTED]	 = "rejected",
80 	[RDMA_CM_EVENT_ESTABLISHED]	 = "established",
81 	[RDMA_CM_EVENT_DISCONNECTED]	 = "disconnected",
82 	[RDMA_CM_EVENT_DEVICE_REMOVAL]	 = "device removal",
83 	[RDMA_CM_EVENT_MULTICAST_JOIN]	 = "multicast join",
84 	[RDMA_CM_EVENT_MULTICAST_ERROR]	 = "multicast error",
85 	[RDMA_CM_EVENT_ADDR_CHANGE]	 = "address change",
86 	[RDMA_CM_EVENT_TIMEWAIT_EXIT]	 = "timewait exit",
87 };
88 
89 const char *rdma_event_msg(enum rdma_cm_event_type event)
90 {
91 	size_t index = event;
92 
93 	return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
94 			cma_events[index] : "unrecognized event";
95 }
96 EXPORT_SYMBOL(rdma_event_msg);
97 
98 static void cma_add_one(struct ib_device *device);
99 static void cma_remove_one(struct ib_device *device, void *client_data);
100 
101 static struct ib_client cma_client = {
102 	.name   = "cma",
103 	.add    = cma_add_one,
104 	.remove = cma_remove_one
105 };
106 
107 static struct ib_sa_client sa_client;
108 static struct rdma_addr_client addr_client;
109 static LIST_HEAD(dev_list);
110 static LIST_HEAD(listen_any_list);
111 static DEFINE_MUTEX(lock);
112 static struct workqueue_struct *cma_wq;
113 static DEFINE_IDR(tcp_ps);
114 static DEFINE_IDR(udp_ps);
115 static DEFINE_IDR(ipoib_ps);
116 static DEFINE_IDR(ib_ps);
117 
118 static struct idr *cma_idr(enum rdma_port_space ps)
119 {
120 	switch (ps) {
121 	case RDMA_PS_TCP:
122 		return &tcp_ps;
123 	case RDMA_PS_UDP:
124 		return &udp_ps;
125 	case RDMA_PS_IPOIB:
126 		return &ipoib_ps;
127 	case RDMA_PS_IB:
128 		return &ib_ps;
129 	default:
130 		return NULL;
131 	}
132 }
133 
134 struct cma_device {
135 	struct list_head	list;
136 	struct ib_device	*device;
137 	struct completion	comp;
138 	atomic_t		refcount;
139 	struct list_head	id_list;
140 };
141 
142 struct rdma_bind_list {
143 	enum rdma_port_space	ps;
144 	struct hlist_head	owners;
145 	unsigned short		port;
146 };
147 
148 static int cma_ps_alloc(enum rdma_port_space ps,
149 			struct rdma_bind_list *bind_list, int snum)
150 {
151 	struct idr *idr = cma_idr(ps);
152 
153 	return idr_alloc(idr, bind_list, snum, snum + 1, GFP_KERNEL);
154 }
155 
156 static struct rdma_bind_list *cma_ps_find(enum rdma_port_space ps, int snum)
157 {
158 	struct idr *idr = cma_idr(ps);
159 
160 	return idr_find(idr, snum);
161 }
162 
163 static void cma_ps_remove(enum rdma_port_space ps, int snum)
164 {
165 	struct idr *idr = cma_idr(ps);
166 
167 	idr_remove(idr, snum);
168 }
169 
170 enum {
171 	CMA_OPTION_AFONLY,
172 };
173 
174 /*
175  * Device removal can occur at anytime, so we need extra handling to
176  * serialize notifying the user of device removal with other callbacks.
177  * We do this by disabling removal notification while a callback is in process,
178  * and reporting it after the callback completes.
179  */
180 struct rdma_id_private {
181 	struct rdma_cm_id	id;
182 
183 	struct rdma_bind_list	*bind_list;
184 	struct hlist_node	node;
185 	struct list_head	list; /* listen_any_list or cma_device.list */
186 	struct list_head	listen_list; /* per device listens */
187 	struct cma_device	*cma_dev;
188 	struct list_head	mc_list;
189 
190 	int			internal_id;
191 	enum rdma_cm_state	state;
192 	spinlock_t		lock;
193 	struct mutex		qp_mutex;
194 
195 	struct completion	comp;
196 	atomic_t		refcount;
197 	struct mutex		handler_mutex;
198 
199 	int			backlog;
200 	int			timeout_ms;
201 	struct ib_sa_query	*query;
202 	int			query_id;
203 	union {
204 		struct ib_cm_id	*ib;
205 		struct iw_cm_id	*iw;
206 	} cm_id;
207 
208 	u32			seq_num;
209 	u32			qkey;
210 	u32			qp_num;
211 	pid_t			owner;
212 	u32			options;
213 	u8			srq;
214 	u8			tos;
215 	u8			reuseaddr;
216 	u8			afonly;
217 };
218 
219 struct cma_multicast {
220 	struct rdma_id_private *id_priv;
221 	union {
222 		struct ib_sa_multicast *ib;
223 	} multicast;
224 	struct list_head	list;
225 	void			*context;
226 	struct sockaddr_storage	addr;
227 	struct kref		mcref;
228 };
229 
230 struct cma_work {
231 	struct work_struct	work;
232 	struct rdma_id_private	*id;
233 	enum rdma_cm_state	old_state;
234 	enum rdma_cm_state	new_state;
235 	struct rdma_cm_event	event;
236 };
237 
238 struct cma_ndev_work {
239 	struct work_struct	work;
240 	struct rdma_id_private	*id;
241 	struct rdma_cm_event	event;
242 };
243 
244 struct iboe_mcast_work {
245 	struct work_struct	 work;
246 	struct rdma_id_private	*id;
247 	struct cma_multicast	*mc;
248 };
249 
250 union cma_ip_addr {
251 	struct in6_addr ip6;
252 	struct {
253 		__be32 pad[3];
254 		__be32 addr;
255 	} ip4;
256 };
257 
258 struct cma_hdr {
259 	u8 cma_version;
260 	u8 ip_version;	/* IP version: 7:4 */
261 	__be16 port;
262 	union cma_ip_addr src_addr;
263 	union cma_ip_addr dst_addr;
264 };
265 
266 #define CMA_VERSION 0x00
267 
268 struct cma_req_info {
269 	struct ib_device *device;
270 	int port;
271 	union ib_gid local_gid;
272 	__be64 service_id;
273 	u16 pkey;
274 	bool has_gid:1;
275 };
276 
277 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
278 {
279 	unsigned long flags;
280 	int ret;
281 
282 	spin_lock_irqsave(&id_priv->lock, flags);
283 	ret = (id_priv->state == comp);
284 	spin_unlock_irqrestore(&id_priv->lock, flags);
285 	return ret;
286 }
287 
288 static int cma_comp_exch(struct rdma_id_private *id_priv,
289 			 enum rdma_cm_state comp, enum rdma_cm_state exch)
290 {
291 	unsigned long flags;
292 	int ret;
293 
294 	spin_lock_irqsave(&id_priv->lock, flags);
295 	if ((ret = (id_priv->state == comp)))
296 		id_priv->state = exch;
297 	spin_unlock_irqrestore(&id_priv->lock, flags);
298 	return ret;
299 }
300 
301 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
302 				   enum rdma_cm_state exch)
303 {
304 	unsigned long flags;
305 	enum rdma_cm_state old;
306 
307 	spin_lock_irqsave(&id_priv->lock, flags);
308 	old = id_priv->state;
309 	id_priv->state = exch;
310 	spin_unlock_irqrestore(&id_priv->lock, flags);
311 	return old;
312 }
313 
314 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
315 {
316 	return hdr->ip_version >> 4;
317 }
318 
319 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
320 {
321 	hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
322 }
323 
324 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
325 			      struct cma_device *cma_dev)
326 {
327 	atomic_inc(&cma_dev->refcount);
328 	id_priv->cma_dev = cma_dev;
329 	id_priv->id.device = cma_dev->device;
330 	id_priv->id.route.addr.dev_addr.transport =
331 		rdma_node_get_transport(cma_dev->device->node_type);
332 	list_add_tail(&id_priv->list, &cma_dev->id_list);
333 }
334 
335 static inline void cma_deref_dev(struct cma_device *cma_dev)
336 {
337 	if (atomic_dec_and_test(&cma_dev->refcount))
338 		complete(&cma_dev->comp);
339 }
340 
341 static inline void release_mc(struct kref *kref)
342 {
343 	struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
344 
345 	kfree(mc->multicast.ib);
346 	kfree(mc);
347 }
348 
349 static void cma_release_dev(struct rdma_id_private *id_priv)
350 {
351 	mutex_lock(&lock);
352 	list_del(&id_priv->list);
353 	cma_deref_dev(id_priv->cma_dev);
354 	id_priv->cma_dev = NULL;
355 	mutex_unlock(&lock);
356 }
357 
358 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
359 {
360 	return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
361 }
362 
363 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
364 {
365 	return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
366 }
367 
368 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
369 {
370 	return id_priv->id.route.addr.src_addr.ss_family;
371 }
372 
373 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
374 {
375 	struct ib_sa_mcmember_rec rec;
376 	int ret = 0;
377 
378 	if (id_priv->qkey) {
379 		if (qkey && id_priv->qkey != qkey)
380 			return -EINVAL;
381 		return 0;
382 	}
383 
384 	if (qkey) {
385 		id_priv->qkey = qkey;
386 		return 0;
387 	}
388 
389 	switch (id_priv->id.ps) {
390 	case RDMA_PS_UDP:
391 	case RDMA_PS_IB:
392 		id_priv->qkey = RDMA_UDP_QKEY;
393 		break;
394 	case RDMA_PS_IPOIB:
395 		ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
396 		ret = ib_sa_get_mcmember_rec(id_priv->id.device,
397 					     id_priv->id.port_num, &rec.mgid,
398 					     &rec);
399 		if (!ret)
400 			id_priv->qkey = be32_to_cpu(rec.qkey);
401 		break;
402 	default:
403 		break;
404 	}
405 	return ret;
406 }
407 
408 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
409 {
410 	dev_addr->dev_type = ARPHRD_INFINIBAND;
411 	rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
412 	ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
413 }
414 
415 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
416 {
417 	int ret;
418 
419 	if (addr->sa_family != AF_IB) {
420 		ret = rdma_translate_ip(addr, dev_addr, NULL);
421 	} else {
422 		cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
423 		ret = 0;
424 	}
425 
426 	return ret;
427 }
428 
429 static inline int cma_validate_port(struct ib_device *device, u8 port,
430 				      union ib_gid *gid, int dev_type)
431 {
432 	u8 found_port;
433 	int ret = -ENODEV;
434 
435 	if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
436 		return ret;
437 
438 	if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
439 		return ret;
440 
441 	ret = ib_find_cached_gid(device, gid, &found_port, NULL);
442 	if (port != found_port)
443 		return -ENODEV;
444 
445 	return ret;
446 }
447 
448 static int cma_acquire_dev(struct rdma_id_private *id_priv,
449 			   struct rdma_id_private *listen_id_priv)
450 {
451 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
452 	struct cma_device *cma_dev;
453 	union ib_gid gid, iboe_gid, *gidp;
454 	int ret = -ENODEV;
455 	u8 port;
456 
457 	if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
458 	    id_priv->id.ps == RDMA_PS_IPOIB)
459 		return -EINVAL;
460 
461 	mutex_lock(&lock);
462 	rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
463 		    &iboe_gid);
464 
465 	memcpy(&gid, dev_addr->src_dev_addr +
466 	       rdma_addr_gid_offset(dev_addr), sizeof gid);
467 
468 	if (listen_id_priv) {
469 		cma_dev = listen_id_priv->cma_dev;
470 		port = listen_id_priv->id.port_num;
471 		gidp = rdma_protocol_roce(cma_dev->device, port) ?
472 		       &iboe_gid : &gid;
473 
474 		ret = cma_validate_port(cma_dev->device, port, gidp,
475 					dev_addr->dev_type);
476 		if (!ret) {
477 			id_priv->id.port_num = port;
478 			goto out;
479 		}
480 	}
481 
482 	list_for_each_entry(cma_dev, &dev_list, list) {
483 		for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
484 			if (listen_id_priv &&
485 			    listen_id_priv->cma_dev == cma_dev &&
486 			    listen_id_priv->id.port_num == port)
487 				continue;
488 
489 			gidp = rdma_protocol_roce(cma_dev->device, port) ?
490 			       &iboe_gid : &gid;
491 
492 			ret = cma_validate_port(cma_dev->device, port, gidp,
493 						dev_addr->dev_type);
494 			if (!ret) {
495 				id_priv->id.port_num = port;
496 				goto out;
497 			}
498 		}
499 	}
500 
501 out:
502 	if (!ret)
503 		cma_attach_to_dev(id_priv, cma_dev);
504 
505 	mutex_unlock(&lock);
506 	return ret;
507 }
508 
509 /*
510  * Select the source IB device and address to reach the destination IB address.
511  */
512 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
513 {
514 	struct cma_device *cma_dev, *cur_dev;
515 	struct sockaddr_ib *addr;
516 	union ib_gid gid, sgid, *dgid;
517 	u16 pkey, index;
518 	u8 p;
519 	int i;
520 
521 	cma_dev = NULL;
522 	addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
523 	dgid = (union ib_gid *) &addr->sib_addr;
524 	pkey = ntohs(addr->sib_pkey);
525 
526 	list_for_each_entry(cur_dev, &dev_list, list) {
527 		for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
528 			if (!rdma_cap_af_ib(cur_dev->device, p))
529 				continue;
530 
531 			if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
532 				continue;
533 
534 			for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i, &gid); i++) {
535 				if (!memcmp(&gid, dgid, sizeof(gid))) {
536 					cma_dev = cur_dev;
537 					sgid = gid;
538 					id_priv->id.port_num = p;
539 					goto found;
540 				}
541 
542 				if (!cma_dev && (gid.global.subnet_prefix ==
543 						 dgid->global.subnet_prefix)) {
544 					cma_dev = cur_dev;
545 					sgid = gid;
546 					id_priv->id.port_num = p;
547 				}
548 			}
549 		}
550 	}
551 
552 	if (!cma_dev)
553 		return -ENODEV;
554 
555 found:
556 	cma_attach_to_dev(id_priv, cma_dev);
557 	addr = (struct sockaddr_ib *) cma_src_addr(id_priv);
558 	memcpy(&addr->sib_addr, &sgid, sizeof sgid);
559 	cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
560 	return 0;
561 }
562 
563 static void cma_deref_id(struct rdma_id_private *id_priv)
564 {
565 	if (atomic_dec_and_test(&id_priv->refcount))
566 		complete(&id_priv->comp);
567 }
568 
569 static int cma_disable_callback(struct rdma_id_private *id_priv,
570 				enum rdma_cm_state state)
571 {
572 	mutex_lock(&id_priv->handler_mutex);
573 	if (id_priv->state != state) {
574 		mutex_unlock(&id_priv->handler_mutex);
575 		return -EINVAL;
576 	}
577 	return 0;
578 }
579 
580 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
581 				  void *context, enum rdma_port_space ps,
582 				  enum ib_qp_type qp_type)
583 {
584 	struct rdma_id_private *id_priv;
585 
586 	id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
587 	if (!id_priv)
588 		return ERR_PTR(-ENOMEM);
589 
590 	id_priv->owner = task_pid_nr(current);
591 	id_priv->state = RDMA_CM_IDLE;
592 	id_priv->id.context = context;
593 	id_priv->id.event_handler = event_handler;
594 	id_priv->id.ps = ps;
595 	id_priv->id.qp_type = qp_type;
596 	spin_lock_init(&id_priv->lock);
597 	mutex_init(&id_priv->qp_mutex);
598 	init_completion(&id_priv->comp);
599 	atomic_set(&id_priv->refcount, 1);
600 	mutex_init(&id_priv->handler_mutex);
601 	INIT_LIST_HEAD(&id_priv->listen_list);
602 	INIT_LIST_HEAD(&id_priv->mc_list);
603 	get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
604 
605 	return &id_priv->id;
606 }
607 EXPORT_SYMBOL(rdma_create_id);
608 
609 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
610 {
611 	struct ib_qp_attr qp_attr;
612 	int qp_attr_mask, ret;
613 
614 	qp_attr.qp_state = IB_QPS_INIT;
615 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
616 	if (ret)
617 		return ret;
618 
619 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
620 	if (ret)
621 		return ret;
622 
623 	qp_attr.qp_state = IB_QPS_RTR;
624 	ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
625 	if (ret)
626 		return ret;
627 
628 	qp_attr.qp_state = IB_QPS_RTS;
629 	qp_attr.sq_psn = 0;
630 	ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
631 
632 	return ret;
633 }
634 
635 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
636 {
637 	struct ib_qp_attr qp_attr;
638 	int qp_attr_mask, ret;
639 
640 	qp_attr.qp_state = IB_QPS_INIT;
641 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
642 	if (ret)
643 		return ret;
644 
645 	return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
646 }
647 
648 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
649 		   struct ib_qp_init_attr *qp_init_attr)
650 {
651 	struct rdma_id_private *id_priv;
652 	struct ib_qp *qp;
653 	int ret;
654 
655 	id_priv = container_of(id, struct rdma_id_private, id);
656 	if (id->device != pd->device)
657 		return -EINVAL;
658 
659 	qp = ib_create_qp(pd, qp_init_attr);
660 	if (IS_ERR(qp))
661 		return PTR_ERR(qp);
662 
663 	if (id->qp_type == IB_QPT_UD)
664 		ret = cma_init_ud_qp(id_priv, qp);
665 	else
666 		ret = cma_init_conn_qp(id_priv, qp);
667 	if (ret)
668 		goto err;
669 
670 	id->qp = qp;
671 	id_priv->qp_num = qp->qp_num;
672 	id_priv->srq = (qp->srq != NULL);
673 	return 0;
674 err:
675 	ib_destroy_qp(qp);
676 	return ret;
677 }
678 EXPORT_SYMBOL(rdma_create_qp);
679 
680 void rdma_destroy_qp(struct rdma_cm_id *id)
681 {
682 	struct rdma_id_private *id_priv;
683 
684 	id_priv = container_of(id, struct rdma_id_private, id);
685 	mutex_lock(&id_priv->qp_mutex);
686 	ib_destroy_qp(id_priv->id.qp);
687 	id_priv->id.qp = NULL;
688 	mutex_unlock(&id_priv->qp_mutex);
689 }
690 EXPORT_SYMBOL(rdma_destroy_qp);
691 
692 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
693 			     struct rdma_conn_param *conn_param)
694 {
695 	struct ib_qp_attr qp_attr;
696 	int qp_attr_mask, ret;
697 	union ib_gid sgid;
698 
699 	mutex_lock(&id_priv->qp_mutex);
700 	if (!id_priv->id.qp) {
701 		ret = 0;
702 		goto out;
703 	}
704 
705 	/* Need to update QP attributes from default values. */
706 	qp_attr.qp_state = IB_QPS_INIT;
707 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
708 	if (ret)
709 		goto out;
710 
711 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
712 	if (ret)
713 		goto out;
714 
715 	qp_attr.qp_state = IB_QPS_RTR;
716 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
717 	if (ret)
718 		goto out;
719 
720 	ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num,
721 			   qp_attr.ah_attr.grh.sgid_index, &sgid);
722 	if (ret)
723 		goto out;
724 
725 	BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
726 
727 	if (rdma_protocol_roce(id_priv->id.device, id_priv->id.port_num)) {
728 		ret = rdma_addr_find_smac_by_sgid(&sgid, qp_attr.smac, NULL);
729 
730 		if (ret)
731 			goto out;
732 	}
733 	if (conn_param)
734 		qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
735 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
736 out:
737 	mutex_unlock(&id_priv->qp_mutex);
738 	return ret;
739 }
740 
741 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
742 			     struct rdma_conn_param *conn_param)
743 {
744 	struct ib_qp_attr qp_attr;
745 	int qp_attr_mask, ret;
746 
747 	mutex_lock(&id_priv->qp_mutex);
748 	if (!id_priv->id.qp) {
749 		ret = 0;
750 		goto out;
751 	}
752 
753 	qp_attr.qp_state = IB_QPS_RTS;
754 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
755 	if (ret)
756 		goto out;
757 
758 	if (conn_param)
759 		qp_attr.max_rd_atomic = conn_param->initiator_depth;
760 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
761 out:
762 	mutex_unlock(&id_priv->qp_mutex);
763 	return ret;
764 }
765 
766 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
767 {
768 	struct ib_qp_attr qp_attr;
769 	int ret;
770 
771 	mutex_lock(&id_priv->qp_mutex);
772 	if (!id_priv->id.qp) {
773 		ret = 0;
774 		goto out;
775 	}
776 
777 	qp_attr.qp_state = IB_QPS_ERR;
778 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
779 out:
780 	mutex_unlock(&id_priv->qp_mutex);
781 	return ret;
782 }
783 
784 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
785 			       struct ib_qp_attr *qp_attr, int *qp_attr_mask)
786 {
787 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
788 	int ret;
789 	u16 pkey;
790 
791 	if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
792 		pkey = 0xffff;
793 	else
794 		pkey = ib_addr_get_pkey(dev_addr);
795 
796 	ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
797 				  pkey, &qp_attr->pkey_index);
798 	if (ret)
799 		return ret;
800 
801 	qp_attr->port_num = id_priv->id.port_num;
802 	*qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
803 
804 	if (id_priv->id.qp_type == IB_QPT_UD) {
805 		ret = cma_set_qkey(id_priv, 0);
806 		if (ret)
807 			return ret;
808 
809 		qp_attr->qkey = id_priv->qkey;
810 		*qp_attr_mask |= IB_QP_QKEY;
811 	} else {
812 		qp_attr->qp_access_flags = 0;
813 		*qp_attr_mask |= IB_QP_ACCESS_FLAGS;
814 	}
815 	return 0;
816 }
817 
818 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
819 		       int *qp_attr_mask)
820 {
821 	struct rdma_id_private *id_priv;
822 	int ret = 0;
823 
824 	id_priv = container_of(id, struct rdma_id_private, id);
825 	if (rdma_cap_ib_cm(id->device, id->port_num)) {
826 		if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
827 			ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
828 		else
829 			ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
830 						 qp_attr_mask);
831 
832 		if (qp_attr->qp_state == IB_QPS_RTR)
833 			qp_attr->rq_psn = id_priv->seq_num;
834 	} else if (rdma_cap_iw_cm(id->device, id->port_num)) {
835 		if (!id_priv->cm_id.iw) {
836 			qp_attr->qp_access_flags = 0;
837 			*qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
838 		} else
839 			ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
840 						 qp_attr_mask);
841 	} else
842 		ret = -ENOSYS;
843 
844 	return ret;
845 }
846 EXPORT_SYMBOL(rdma_init_qp_attr);
847 
848 static inline int cma_zero_addr(struct sockaddr *addr)
849 {
850 	switch (addr->sa_family) {
851 	case AF_INET:
852 		return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
853 	case AF_INET6:
854 		return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
855 	case AF_IB:
856 		return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
857 	default:
858 		return 0;
859 	}
860 }
861 
862 static inline int cma_loopback_addr(struct sockaddr *addr)
863 {
864 	switch (addr->sa_family) {
865 	case AF_INET:
866 		return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
867 	case AF_INET6:
868 		return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
869 	case AF_IB:
870 		return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
871 	default:
872 		return 0;
873 	}
874 }
875 
876 static inline int cma_any_addr(struct sockaddr *addr)
877 {
878 	return cma_zero_addr(addr) || cma_loopback_addr(addr);
879 }
880 
881 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
882 {
883 	if (src->sa_family != dst->sa_family)
884 		return -1;
885 
886 	switch (src->sa_family) {
887 	case AF_INET:
888 		return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
889 		       ((struct sockaddr_in *) dst)->sin_addr.s_addr;
890 	case AF_INET6:
891 		return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
892 				     &((struct sockaddr_in6 *) dst)->sin6_addr);
893 	default:
894 		return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
895 				   &((struct sockaddr_ib *) dst)->sib_addr);
896 	}
897 }
898 
899 static __be16 cma_port(struct sockaddr *addr)
900 {
901 	struct sockaddr_ib *sib;
902 
903 	switch (addr->sa_family) {
904 	case AF_INET:
905 		return ((struct sockaddr_in *) addr)->sin_port;
906 	case AF_INET6:
907 		return ((struct sockaddr_in6 *) addr)->sin6_port;
908 	case AF_IB:
909 		sib = (struct sockaddr_ib *) addr;
910 		return htons((u16) (be64_to_cpu(sib->sib_sid) &
911 				    be64_to_cpu(sib->sib_sid_mask)));
912 	default:
913 		return 0;
914 	}
915 }
916 
917 static inline int cma_any_port(struct sockaddr *addr)
918 {
919 	return !cma_port(addr);
920 }
921 
922 static void cma_save_ib_info(struct sockaddr *src_addr,
923 			     struct sockaddr *dst_addr,
924 			     struct rdma_cm_id *listen_id,
925 			     struct ib_sa_path_rec *path)
926 {
927 	struct sockaddr_ib *listen_ib, *ib;
928 
929 	listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
930 	if (src_addr) {
931 		ib = (struct sockaddr_ib *)src_addr;
932 		ib->sib_family = AF_IB;
933 		if (path) {
934 			ib->sib_pkey = path->pkey;
935 			ib->sib_flowinfo = path->flow_label;
936 			memcpy(&ib->sib_addr, &path->sgid, 16);
937 			ib->sib_sid = path->service_id;
938 			ib->sib_scope_id = 0;
939 		} else {
940 			ib->sib_pkey = listen_ib->sib_pkey;
941 			ib->sib_flowinfo = listen_ib->sib_flowinfo;
942 			ib->sib_addr = listen_ib->sib_addr;
943 			ib->sib_sid = listen_ib->sib_sid;
944 			ib->sib_scope_id = listen_ib->sib_scope_id;
945 		}
946 		ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
947 	}
948 	if (dst_addr) {
949 		ib = (struct sockaddr_ib *)dst_addr;
950 		ib->sib_family = AF_IB;
951 		if (path) {
952 			ib->sib_pkey = path->pkey;
953 			ib->sib_flowinfo = path->flow_label;
954 			memcpy(&ib->sib_addr, &path->dgid, 16);
955 		}
956 	}
957 }
958 
959 static void cma_save_ip4_info(struct sockaddr *src_addr,
960 			      struct sockaddr *dst_addr,
961 			      struct cma_hdr *hdr,
962 			      __be16 local_port)
963 {
964 	struct sockaddr_in *ip4;
965 
966 	if (src_addr) {
967 		ip4 = (struct sockaddr_in *)src_addr;
968 		ip4->sin_family = AF_INET;
969 		ip4->sin_addr.s_addr = hdr->dst_addr.ip4.addr;
970 		ip4->sin_port = local_port;
971 	}
972 
973 	if (dst_addr) {
974 		ip4 = (struct sockaddr_in *)dst_addr;
975 		ip4->sin_family = AF_INET;
976 		ip4->sin_addr.s_addr = hdr->src_addr.ip4.addr;
977 		ip4->sin_port = hdr->port;
978 	}
979 }
980 
981 static void cma_save_ip6_info(struct sockaddr *src_addr,
982 			      struct sockaddr *dst_addr,
983 			      struct cma_hdr *hdr,
984 			      __be16 local_port)
985 {
986 	struct sockaddr_in6 *ip6;
987 
988 	if (src_addr) {
989 		ip6 = (struct sockaddr_in6 *)src_addr;
990 		ip6->sin6_family = AF_INET6;
991 		ip6->sin6_addr = hdr->dst_addr.ip6;
992 		ip6->sin6_port = local_port;
993 	}
994 
995 	if (dst_addr) {
996 		ip6 = (struct sockaddr_in6 *)dst_addr;
997 		ip6->sin6_family = AF_INET6;
998 		ip6->sin6_addr = hdr->src_addr.ip6;
999 		ip6->sin6_port = hdr->port;
1000 	}
1001 }
1002 
1003 static u16 cma_port_from_service_id(__be64 service_id)
1004 {
1005 	return (u16)be64_to_cpu(service_id);
1006 }
1007 
1008 static int cma_save_ip_info(struct sockaddr *src_addr,
1009 			    struct sockaddr *dst_addr,
1010 			    struct ib_cm_event *ib_event,
1011 			    __be64 service_id)
1012 {
1013 	struct cma_hdr *hdr;
1014 	__be16 port;
1015 
1016 	hdr = ib_event->private_data;
1017 	if (hdr->cma_version != CMA_VERSION)
1018 		return -EINVAL;
1019 
1020 	port = htons(cma_port_from_service_id(service_id));
1021 
1022 	switch (cma_get_ip_ver(hdr)) {
1023 	case 4:
1024 		cma_save_ip4_info(src_addr, dst_addr, hdr, port);
1025 		break;
1026 	case 6:
1027 		cma_save_ip6_info(src_addr, dst_addr, hdr, port);
1028 		break;
1029 	default:
1030 		return -EAFNOSUPPORT;
1031 	}
1032 
1033 	return 0;
1034 }
1035 
1036 static int cma_save_net_info(struct sockaddr *src_addr,
1037 			     struct sockaddr *dst_addr,
1038 			     struct rdma_cm_id *listen_id,
1039 			     struct ib_cm_event *ib_event,
1040 			     sa_family_t sa_family, __be64 service_id)
1041 {
1042 	if (sa_family == AF_IB) {
1043 		if (ib_event->event == IB_CM_REQ_RECEIVED)
1044 			cma_save_ib_info(src_addr, dst_addr, listen_id,
1045 					 ib_event->param.req_rcvd.primary_path);
1046 		else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1047 			cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1048 		return 0;
1049 	}
1050 
1051 	return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1052 }
1053 
1054 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1055 			     struct cma_req_info *req)
1056 {
1057 	const struct ib_cm_req_event_param *req_param =
1058 		&ib_event->param.req_rcvd;
1059 	const struct ib_cm_sidr_req_event_param *sidr_param =
1060 		&ib_event->param.sidr_req_rcvd;
1061 
1062 	switch (ib_event->event) {
1063 	case IB_CM_REQ_RECEIVED:
1064 		req->device	= req_param->listen_id->device;
1065 		req->port	= req_param->port;
1066 		memcpy(&req->local_gid, &req_param->primary_path->sgid,
1067 		       sizeof(req->local_gid));
1068 		req->has_gid	= true;
1069 		req->service_id	= req_param->primary_path->service_id;
1070 		req->pkey	= be16_to_cpu(req_param->primary_path->pkey);
1071 		break;
1072 	case IB_CM_SIDR_REQ_RECEIVED:
1073 		req->device	= sidr_param->listen_id->device;
1074 		req->port	= sidr_param->port;
1075 		req->has_gid	= false;
1076 		req->service_id	= sidr_param->service_id;
1077 		req->pkey	= sidr_param->pkey;
1078 		break;
1079 	default:
1080 		return -EINVAL;
1081 	}
1082 
1083 	return 0;
1084 }
1085 
1086 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1087 				  const struct sockaddr_in *dst_addr,
1088 				  const struct sockaddr_in *src_addr)
1089 {
1090 	__be32 daddr = dst_addr->sin_addr.s_addr,
1091 	       saddr = src_addr->sin_addr.s_addr;
1092 	struct fib_result res;
1093 	struct flowi4 fl4;
1094 	int err;
1095 	bool ret;
1096 
1097 	if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1098 	    ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1099 	    ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1100 	    ipv4_is_loopback(saddr))
1101 		return false;
1102 
1103 	memset(&fl4, 0, sizeof(fl4));
1104 	fl4.flowi4_iif = net_dev->ifindex;
1105 	fl4.daddr = daddr;
1106 	fl4.saddr = saddr;
1107 
1108 	rcu_read_lock();
1109 	err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1110 	if (err)
1111 		return false;
1112 
1113 	ret = FIB_RES_DEV(res) == net_dev;
1114 	rcu_read_unlock();
1115 
1116 	return ret;
1117 }
1118 
1119 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1120 				  const struct sockaddr_in6 *dst_addr,
1121 				  const struct sockaddr_in6 *src_addr)
1122 {
1123 #if IS_ENABLED(CONFIG_IPV6)
1124 	const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1125 			   IPV6_ADDR_LINKLOCAL;
1126 	struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1127 					 &src_addr->sin6_addr, net_dev->ifindex,
1128 					 strict);
1129 	bool ret;
1130 
1131 	if (!rt)
1132 		return false;
1133 
1134 	ret = rt->rt6i_idev->dev == net_dev;
1135 	ip6_rt_put(rt);
1136 
1137 	return ret;
1138 #else
1139 	return false;
1140 #endif
1141 }
1142 
1143 static bool validate_net_dev(struct net_device *net_dev,
1144 			     const struct sockaddr *daddr,
1145 			     const struct sockaddr *saddr)
1146 {
1147 	const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1148 	const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1149 	const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1150 	const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1151 
1152 	switch (daddr->sa_family) {
1153 	case AF_INET:
1154 		return saddr->sa_family == AF_INET &&
1155 		       validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1156 
1157 	case AF_INET6:
1158 		return saddr->sa_family == AF_INET6 &&
1159 		       validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1160 
1161 	default:
1162 		return false;
1163 	}
1164 }
1165 
1166 static struct net_device *cma_get_net_dev(struct ib_cm_event *ib_event,
1167 					  const struct cma_req_info *req)
1168 {
1169 	struct sockaddr_storage listen_addr_storage, src_addr_storage;
1170 	struct sockaddr *listen_addr = (struct sockaddr *)&listen_addr_storage,
1171 			*src_addr = (struct sockaddr *)&src_addr_storage;
1172 	struct net_device *net_dev;
1173 	const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1174 	int err;
1175 
1176 	err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1177 			       req->service_id);
1178 	if (err)
1179 		return ERR_PTR(err);
1180 
1181 	net_dev = ib_get_net_dev_by_params(req->device, req->port, req->pkey,
1182 					   gid, listen_addr);
1183 	if (!net_dev)
1184 		return ERR_PTR(-ENODEV);
1185 
1186 	if (!validate_net_dev(net_dev, listen_addr, src_addr)) {
1187 		dev_put(net_dev);
1188 		return ERR_PTR(-EHOSTUNREACH);
1189 	}
1190 
1191 	return net_dev;
1192 }
1193 
1194 static enum rdma_port_space rdma_ps_from_service_id(__be64 service_id)
1195 {
1196 	return (be64_to_cpu(service_id) >> 16) & 0xffff;
1197 }
1198 
1199 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1200 				   const struct cma_hdr *hdr)
1201 {
1202 	struct sockaddr *addr = cma_src_addr(id_priv);
1203 	__be32 ip4_addr;
1204 	struct in6_addr ip6_addr;
1205 
1206 	if (cma_any_addr(addr) && !id_priv->afonly)
1207 		return true;
1208 
1209 	switch (addr->sa_family) {
1210 	case AF_INET:
1211 		ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1212 		if (cma_get_ip_ver(hdr) != 4)
1213 			return false;
1214 		if (!cma_any_addr(addr) &&
1215 		    hdr->dst_addr.ip4.addr != ip4_addr)
1216 			return false;
1217 		break;
1218 	case AF_INET6:
1219 		ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1220 		if (cma_get_ip_ver(hdr) != 6)
1221 			return false;
1222 		if (!cma_any_addr(addr) &&
1223 		    memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1224 			return false;
1225 		break;
1226 	case AF_IB:
1227 		return true;
1228 	default:
1229 		return false;
1230 	}
1231 
1232 	return true;
1233 }
1234 
1235 static bool cma_protocol_roce_dev_port(struct ib_device *device, int port_num)
1236 {
1237 	enum rdma_link_layer ll = rdma_port_get_link_layer(device, port_num);
1238 	enum rdma_transport_type transport =
1239 		rdma_node_get_transport(device->node_type);
1240 
1241 	return ll == IB_LINK_LAYER_ETHERNET && transport == RDMA_TRANSPORT_IB;
1242 }
1243 
1244 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1245 {
1246 	struct ib_device *device = id->device;
1247 	const int port_num = id->port_num ?: rdma_start_port(device);
1248 
1249 	return cma_protocol_roce_dev_port(device, port_num);
1250 }
1251 
1252 static bool cma_match_net_dev(const struct rdma_id_private *id_priv,
1253 			      const struct net_device *net_dev)
1254 {
1255 	const struct rdma_addr *addr = &id_priv->id.route.addr;
1256 
1257 	if (!net_dev)
1258 		/* This request is an AF_IB request or a RoCE request */
1259 		return addr->src_addr.ss_family == AF_IB ||
1260 		       cma_protocol_roce(&id_priv->id);
1261 
1262 	return !addr->dev_addr.bound_dev_if ||
1263 	       (net_eq(dev_net(net_dev), &init_net) &&
1264 		addr->dev_addr.bound_dev_if == net_dev->ifindex);
1265 }
1266 
1267 static struct rdma_id_private *cma_find_listener(
1268 		const struct rdma_bind_list *bind_list,
1269 		const struct ib_cm_id *cm_id,
1270 		const struct ib_cm_event *ib_event,
1271 		const struct cma_req_info *req,
1272 		const struct net_device *net_dev)
1273 {
1274 	struct rdma_id_private *id_priv, *id_priv_dev;
1275 
1276 	if (!bind_list)
1277 		return ERR_PTR(-EINVAL);
1278 
1279 	hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1280 		if (cma_match_private_data(id_priv, ib_event->private_data)) {
1281 			if (id_priv->id.device == cm_id->device &&
1282 			    cma_match_net_dev(id_priv, net_dev))
1283 				return id_priv;
1284 			list_for_each_entry(id_priv_dev,
1285 					    &id_priv->listen_list,
1286 					    listen_list) {
1287 				if (id_priv_dev->id.device == cm_id->device &&
1288 				    cma_match_net_dev(id_priv_dev, net_dev))
1289 					return id_priv_dev;
1290 			}
1291 		}
1292 	}
1293 
1294 	return ERR_PTR(-EINVAL);
1295 }
1296 
1297 static struct rdma_id_private *cma_id_from_event(struct ib_cm_id *cm_id,
1298 						 struct ib_cm_event *ib_event,
1299 						 struct net_device **net_dev)
1300 {
1301 	struct cma_req_info req;
1302 	struct rdma_bind_list *bind_list;
1303 	struct rdma_id_private *id_priv;
1304 	int err;
1305 
1306 	err = cma_save_req_info(ib_event, &req);
1307 	if (err)
1308 		return ERR_PTR(err);
1309 
1310 	*net_dev = cma_get_net_dev(ib_event, &req);
1311 	if (IS_ERR(*net_dev)) {
1312 		if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1313 			/* Assuming the protocol is AF_IB */
1314 			*net_dev = NULL;
1315 		} else if (cma_protocol_roce_dev_port(req.device, req.port)) {
1316 			/* TODO find the net dev matching the request parameters
1317 			 * through the RoCE GID table */
1318 			*net_dev = NULL;
1319 		} else {
1320 			return ERR_CAST(*net_dev);
1321 		}
1322 	}
1323 
1324 	bind_list = cma_ps_find(rdma_ps_from_service_id(req.service_id),
1325 				cma_port_from_service_id(req.service_id));
1326 	id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev);
1327 	if (IS_ERR(id_priv) && *net_dev) {
1328 		dev_put(*net_dev);
1329 		*net_dev = NULL;
1330 	}
1331 
1332 	return id_priv;
1333 }
1334 
1335 static inline int cma_user_data_offset(struct rdma_id_private *id_priv)
1336 {
1337 	return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1338 }
1339 
1340 static void cma_cancel_route(struct rdma_id_private *id_priv)
1341 {
1342 	if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1343 		if (id_priv->query)
1344 			ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1345 	}
1346 }
1347 
1348 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1349 {
1350 	struct rdma_id_private *dev_id_priv;
1351 
1352 	/*
1353 	 * Remove from listen_any_list to prevent added devices from spawning
1354 	 * additional listen requests.
1355 	 */
1356 	mutex_lock(&lock);
1357 	list_del(&id_priv->list);
1358 
1359 	while (!list_empty(&id_priv->listen_list)) {
1360 		dev_id_priv = list_entry(id_priv->listen_list.next,
1361 					 struct rdma_id_private, listen_list);
1362 		/* sync with device removal to avoid duplicate destruction */
1363 		list_del_init(&dev_id_priv->list);
1364 		list_del(&dev_id_priv->listen_list);
1365 		mutex_unlock(&lock);
1366 
1367 		rdma_destroy_id(&dev_id_priv->id);
1368 		mutex_lock(&lock);
1369 	}
1370 	mutex_unlock(&lock);
1371 }
1372 
1373 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1374 				 enum rdma_cm_state state)
1375 {
1376 	switch (state) {
1377 	case RDMA_CM_ADDR_QUERY:
1378 		rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1379 		break;
1380 	case RDMA_CM_ROUTE_QUERY:
1381 		cma_cancel_route(id_priv);
1382 		break;
1383 	case RDMA_CM_LISTEN:
1384 		if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1385 			cma_cancel_listens(id_priv);
1386 		break;
1387 	default:
1388 		break;
1389 	}
1390 }
1391 
1392 static void cma_release_port(struct rdma_id_private *id_priv)
1393 {
1394 	struct rdma_bind_list *bind_list = id_priv->bind_list;
1395 
1396 	if (!bind_list)
1397 		return;
1398 
1399 	mutex_lock(&lock);
1400 	hlist_del(&id_priv->node);
1401 	if (hlist_empty(&bind_list->owners)) {
1402 		cma_ps_remove(bind_list->ps, bind_list->port);
1403 		kfree(bind_list);
1404 	}
1405 	mutex_unlock(&lock);
1406 }
1407 
1408 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1409 {
1410 	struct cma_multicast *mc;
1411 
1412 	while (!list_empty(&id_priv->mc_list)) {
1413 		mc = container_of(id_priv->mc_list.next,
1414 				  struct cma_multicast, list);
1415 		list_del(&mc->list);
1416 		if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1417 				      id_priv->id.port_num)) {
1418 			ib_sa_free_multicast(mc->multicast.ib);
1419 			kfree(mc);
1420 		} else
1421 			kref_put(&mc->mcref, release_mc);
1422 	}
1423 }
1424 
1425 void rdma_destroy_id(struct rdma_cm_id *id)
1426 {
1427 	struct rdma_id_private *id_priv;
1428 	enum rdma_cm_state state;
1429 
1430 	id_priv = container_of(id, struct rdma_id_private, id);
1431 	state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1432 	cma_cancel_operation(id_priv, state);
1433 
1434 	/*
1435 	 * Wait for any active callback to finish.  New callbacks will find
1436 	 * the id_priv state set to destroying and abort.
1437 	 */
1438 	mutex_lock(&id_priv->handler_mutex);
1439 	mutex_unlock(&id_priv->handler_mutex);
1440 
1441 	if (id_priv->cma_dev) {
1442 		if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1443 			if (id_priv->cm_id.ib)
1444 				ib_destroy_cm_id(id_priv->cm_id.ib);
1445 		} else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1446 			if (id_priv->cm_id.iw)
1447 				iw_destroy_cm_id(id_priv->cm_id.iw);
1448 		}
1449 		cma_leave_mc_groups(id_priv);
1450 		cma_release_dev(id_priv);
1451 	}
1452 
1453 	cma_release_port(id_priv);
1454 	cma_deref_id(id_priv);
1455 	wait_for_completion(&id_priv->comp);
1456 
1457 	if (id_priv->internal_id)
1458 		cma_deref_id(id_priv->id.context);
1459 
1460 	kfree(id_priv->id.route.path_rec);
1461 	kfree(id_priv);
1462 }
1463 EXPORT_SYMBOL(rdma_destroy_id);
1464 
1465 static int cma_rep_recv(struct rdma_id_private *id_priv)
1466 {
1467 	int ret;
1468 
1469 	ret = cma_modify_qp_rtr(id_priv, NULL);
1470 	if (ret)
1471 		goto reject;
1472 
1473 	ret = cma_modify_qp_rts(id_priv, NULL);
1474 	if (ret)
1475 		goto reject;
1476 
1477 	ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1478 	if (ret)
1479 		goto reject;
1480 
1481 	return 0;
1482 reject:
1483 	cma_modify_qp_err(id_priv);
1484 	ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1485 		       NULL, 0, NULL, 0);
1486 	return ret;
1487 }
1488 
1489 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1490 				   struct ib_cm_rep_event_param *rep_data,
1491 				   void *private_data)
1492 {
1493 	event->param.conn.private_data = private_data;
1494 	event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1495 	event->param.conn.responder_resources = rep_data->responder_resources;
1496 	event->param.conn.initiator_depth = rep_data->initiator_depth;
1497 	event->param.conn.flow_control = rep_data->flow_control;
1498 	event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1499 	event->param.conn.srq = rep_data->srq;
1500 	event->param.conn.qp_num = rep_data->remote_qpn;
1501 }
1502 
1503 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1504 {
1505 	struct rdma_id_private *id_priv = cm_id->context;
1506 	struct rdma_cm_event event;
1507 	int ret = 0;
1508 
1509 	if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1510 		cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1511 	    (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1512 		cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1513 		return 0;
1514 
1515 	memset(&event, 0, sizeof event);
1516 	switch (ib_event->event) {
1517 	case IB_CM_REQ_ERROR:
1518 	case IB_CM_REP_ERROR:
1519 		event.event = RDMA_CM_EVENT_UNREACHABLE;
1520 		event.status = -ETIMEDOUT;
1521 		break;
1522 	case IB_CM_REP_RECEIVED:
1523 		if (id_priv->id.qp) {
1524 			event.status = cma_rep_recv(id_priv);
1525 			event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1526 						     RDMA_CM_EVENT_ESTABLISHED;
1527 		} else {
1528 			event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1529 		}
1530 		cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1531 				       ib_event->private_data);
1532 		break;
1533 	case IB_CM_RTU_RECEIVED:
1534 	case IB_CM_USER_ESTABLISHED:
1535 		event.event = RDMA_CM_EVENT_ESTABLISHED;
1536 		break;
1537 	case IB_CM_DREQ_ERROR:
1538 		event.status = -ETIMEDOUT; /* fall through */
1539 	case IB_CM_DREQ_RECEIVED:
1540 	case IB_CM_DREP_RECEIVED:
1541 		if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1542 				   RDMA_CM_DISCONNECT))
1543 			goto out;
1544 		event.event = RDMA_CM_EVENT_DISCONNECTED;
1545 		break;
1546 	case IB_CM_TIMEWAIT_EXIT:
1547 		event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1548 		break;
1549 	case IB_CM_MRA_RECEIVED:
1550 		/* ignore event */
1551 		goto out;
1552 	case IB_CM_REJ_RECEIVED:
1553 		cma_modify_qp_err(id_priv);
1554 		event.status = ib_event->param.rej_rcvd.reason;
1555 		event.event = RDMA_CM_EVENT_REJECTED;
1556 		event.param.conn.private_data = ib_event->private_data;
1557 		event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1558 		break;
1559 	default:
1560 		printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1561 		       ib_event->event);
1562 		goto out;
1563 	}
1564 
1565 	ret = id_priv->id.event_handler(&id_priv->id, &event);
1566 	if (ret) {
1567 		/* Destroy the CM ID by returning a non-zero value. */
1568 		id_priv->cm_id.ib = NULL;
1569 		cma_exch(id_priv, RDMA_CM_DESTROYING);
1570 		mutex_unlock(&id_priv->handler_mutex);
1571 		rdma_destroy_id(&id_priv->id);
1572 		return ret;
1573 	}
1574 out:
1575 	mutex_unlock(&id_priv->handler_mutex);
1576 	return ret;
1577 }
1578 
1579 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1580 					       struct ib_cm_event *ib_event,
1581 					       struct net_device *net_dev)
1582 {
1583 	struct rdma_id_private *id_priv;
1584 	struct rdma_cm_id *id;
1585 	struct rdma_route *rt;
1586 	const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1587 	const __be64 service_id =
1588 		      ib_event->param.req_rcvd.primary_path->service_id;
1589 	int ret;
1590 
1591 	id = rdma_create_id(listen_id->event_handler, listen_id->context,
1592 			    listen_id->ps, ib_event->param.req_rcvd.qp_type);
1593 	if (IS_ERR(id))
1594 		return NULL;
1595 
1596 	id_priv = container_of(id, struct rdma_id_private, id);
1597 	if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1598 			      (struct sockaddr *)&id->route.addr.dst_addr,
1599 			      listen_id, ib_event, ss_family, service_id))
1600 		goto err;
1601 
1602 	rt = &id->route;
1603 	rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1604 	rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1605 			       GFP_KERNEL);
1606 	if (!rt->path_rec)
1607 		goto err;
1608 
1609 	rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1610 	if (rt->num_paths == 2)
1611 		rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1612 
1613 	if (net_dev) {
1614 		ret = rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL);
1615 		if (ret)
1616 			goto err;
1617 	} else {
1618 		if (!cma_protocol_roce(listen_id) &&
1619 		    cma_any_addr(cma_src_addr(id_priv))) {
1620 			rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1621 			rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1622 			ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1623 		} else if (!cma_any_addr(cma_src_addr(id_priv))) {
1624 			ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1625 			if (ret)
1626 				goto err;
1627 		}
1628 	}
1629 	rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1630 
1631 	id_priv->state = RDMA_CM_CONNECT;
1632 	return id_priv;
1633 
1634 err:
1635 	rdma_destroy_id(id);
1636 	return NULL;
1637 }
1638 
1639 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1640 					      struct ib_cm_event *ib_event,
1641 					      struct net_device *net_dev)
1642 {
1643 	struct rdma_id_private *id_priv;
1644 	struct rdma_cm_id *id;
1645 	const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1646 	int ret;
1647 
1648 	id = rdma_create_id(listen_id->event_handler, listen_id->context,
1649 			    listen_id->ps, IB_QPT_UD);
1650 	if (IS_ERR(id))
1651 		return NULL;
1652 
1653 	id_priv = container_of(id, struct rdma_id_private, id);
1654 	if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1655 			      (struct sockaddr *)&id->route.addr.dst_addr,
1656 			      listen_id, ib_event, ss_family,
1657 			      ib_event->param.sidr_req_rcvd.service_id))
1658 		goto err;
1659 
1660 	if (net_dev) {
1661 		ret = rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL);
1662 		if (ret)
1663 			goto err;
1664 	} else {
1665 		if (!cma_any_addr(cma_src_addr(id_priv))) {
1666 			ret = cma_translate_addr(cma_src_addr(id_priv),
1667 						 &id->route.addr.dev_addr);
1668 			if (ret)
1669 				goto err;
1670 		}
1671 	}
1672 
1673 	id_priv->state = RDMA_CM_CONNECT;
1674 	return id_priv;
1675 err:
1676 	rdma_destroy_id(id);
1677 	return NULL;
1678 }
1679 
1680 static void cma_set_req_event_data(struct rdma_cm_event *event,
1681 				   struct ib_cm_req_event_param *req_data,
1682 				   void *private_data, int offset)
1683 {
1684 	event->param.conn.private_data = private_data + offset;
1685 	event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1686 	event->param.conn.responder_resources = req_data->responder_resources;
1687 	event->param.conn.initiator_depth = req_data->initiator_depth;
1688 	event->param.conn.flow_control = req_data->flow_control;
1689 	event->param.conn.retry_count = req_data->retry_count;
1690 	event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1691 	event->param.conn.srq = req_data->srq;
1692 	event->param.conn.qp_num = req_data->remote_qpn;
1693 }
1694 
1695 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1696 {
1697 	return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1698 		 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1699 		((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1700 		 (id->qp_type == IB_QPT_UD)) ||
1701 		(!id->qp_type));
1702 }
1703 
1704 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1705 {
1706 	struct rdma_id_private *listen_id, *conn_id;
1707 	struct rdma_cm_event event;
1708 	struct net_device *net_dev;
1709 	int offset, ret;
1710 
1711 	listen_id = cma_id_from_event(cm_id, ib_event, &net_dev);
1712 	if (IS_ERR(listen_id))
1713 		return PTR_ERR(listen_id);
1714 
1715 	if (!cma_check_req_qp_type(&listen_id->id, ib_event)) {
1716 		ret = -EINVAL;
1717 		goto net_dev_put;
1718 	}
1719 
1720 	if (cma_disable_callback(listen_id, RDMA_CM_LISTEN)) {
1721 		ret = -ECONNABORTED;
1722 		goto net_dev_put;
1723 	}
1724 
1725 	memset(&event, 0, sizeof event);
1726 	offset = cma_user_data_offset(listen_id);
1727 	event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1728 	if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1729 		conn_id = cma_new_udp_id(&listen_id->id, ib_event, net_dev);
1730 		event.param.ud.private_data = ib_event->private_data + offset;
1731 		event.param.ud.private_data_len =
1732 				IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1733 	} else {
1734 		conn_id = cma_new_conn_id(&listen_id->id, ib_event, net_dev);
1735 		cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1736 				       ib_event->private_data, offset);
1737 	}
1738 	if (!conn_id) {
1739 		ret = -ENOMEM;
1740 		goto err1;
1741 	}
1742 
1743 	mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1744 	ret = cma_acquire_dev(conn_id, listen_id);
1745 	if (ret)
1746 		goto err2;
1747 
1748 	conn_id->cm_id.ib = cm_id;
1749 	cm_id->context = conn_id;
1750 	cm_id->cm_handler = cma_ib_handler;
1751 
1752 	/*
1753 	 * Protect against the user destroying conn_id from another thread
1754 	 * until we're done accessing it.
1755 	 */
1756 	atomic_inc(&conn_id->refcount);
1757 	ret = conn_id->id.event_handler(&conn_id->id, &event);
1758 	if (ret)
1759 		goto err3;
1760 	/*
1761 	 * Acquire mutex to prevent user executing rdma_destroy_id()
1762 	 * while we're accessing the cm_id.
1763 	 */
1764 	mutex_lock(&lock);
1765 	if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
1766 	    (conn_id->id.qp_type != IB_QPT_UD))
1767 		ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1768 	mutex_unlock(&lock);
1769 	mutex_unlock(&conn_id->handler_mutex);
1770 	mutex_unlock(&listen_id->handler_mutex);
1771 	cma_deref_id(conn_id);
1772 	if (net_dev)
1773 		dev_put(net_dev);
1774 	return 0;
1775 
1776 err3:
1777 	cma_deref_id(conn_id);
1778 	/* Destroy the CM ID by returning a non-zero value. */
1779 	conn_id->cm_id.ib = NULL;
1780 err2:
1781 	cma_exch(conn_id, RDMA_CM_DESTROYING);
1782 	mutex_unlock(&conn_id->handler_mutex);
1783 err1:
1784 	mutex_unlock(&listen_id->handler_mutex);
1785 	if (conn_id)
1786 		rdma_destroy_id(&conn_id->id);
1787 
1788 net_dev_put:
1789 	if (net_dev)
1790 		dev_put(net_dev);
1791 
1792 	return ret;
1793 }
1794 
1795 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
1796 {
1797 	if (addr->sa_family == AF_IB)
1798 		return ((struct sockaddr_ib *) addr)->sib_sid;
1799 
1800 	return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
1801 }
1802 EXPORT_SYMBOL(rdma_get_service_id);
1803 
1804 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1805 {
1806 	struct rdma_id_private *id_priv = iw_id->context;
1807 	struct rdma_cm_event event;
1808 	int ret = 0;
1809 	struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1810 	struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1811 
1812 	if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1813 		return 0;
1814 
1815 	memset(&event, 0, sizeof event);
1816 	switch (iw_event->event) {
1817 	case IW_CM_EVENT_CLOSE:
1818 		event.event = RDMA_CM_EVENT_DISCONNECTED;
1819 		break;
1820 	case IW_CM_EVENT_CONNECT_REPLY:
1821 		memcpy(cma_src_addr(id_priv), laddr,
1822 		       rdma_addr_size(laddr));
1823 		memcpy(cma_dst_addr(id_priv), raddr,
1824 		       rdma_addr_size(raddr));
1825 		switch (iw_event->status) {
1826 		case 0:
1827 			event.event = RDMA_CM_EVENT_ESTABLISHED;
1828 			event.param.conn.initiator_depth = iw_event->ird;
1829 			event.param.conn.responder_resources = iw_event->ord;
1830 			break;
1831 		case -ECONNRESET:
1832 		case -ECONNREFUSED:
1833 			event.event = RDMA_CM_EVENT_REJECTED;
1834 			break;
1835 		case -ETIMEDOUT:
1836 			event.event = RDMA_CM_EVENT_UNREACHABLE;
1837 			break;
1838 		default:
1839 			event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1840 			break;
1841 		}
1842 		break;
1843 	case IW_CM_EVENT_ESTABLISHED:
1844 		event.event = RDMA_CM_EVENT_ESTABLISHED;
1845 		event.param.conn.initiator_depth = iw_event->ird;
1846 		event.param.conn.responder_resources = iw_event->ord;
1847 		break;
1848 	default:
1849 		BUG_ON(1);
1850 	}
1851 
1852 	event.status = iw_event->status;
1853 	event.param.conn.private_data = iw_event->private_data;
1854 	event.param.conn.private_data_len = iw_event->private_data_len;
1855 	ret = id_priv->id.event_handler(&id_priv->id, &event);
1856 	if (ret) {
1857 		/* Destroy the CM ID by returning a non-zero value. */
1858 		id_priv->cm_id.iw = NULL;
1859 		cma_exch(id_priv, RDMA_CM_DESTROYING);
1860 		mutex_unlock(&id_priv->handler_mutex);
1861 		rdma_destroy_id(&id_priv->id);
1862 		return ret;
1863 	}
1864 
1865 	mutex_unlock(&id_priv->handler_mutex);
1866 	return ret;
1867 }
1868 
1869 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1870 			       struct iw_cm_event *iw_event)
1871 {
1872 	struct rdma_cm_id *new_cm_id;
1873 	struct rdma_id_private *listen_id, *conn_id;
1874 	struct rdma_cm_event event;
1875 	int ret;
1876 	struct ib_device_attr attr;
1877 	struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1878 	struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1879 
1880 	listen_id = cm_id->context;
1881 	if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1882 		return -ECONNABORTED;
1883 
1884 	/* Create a new RDMA id for the new IW CM ID */
1885 	new_cm_id = rdma_create_id(listen_id->id.event_handler,
1886 				   listen_id->id.context,
1887 				   RDMA_PS_TCP, IB_QPT_RC);
1888 	if (IS_ERR(new_cm_id)) {
1889 		ret = -ENOMEM;
1890 		goto out;
1891 	}
1892 	conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1893 	mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1894 	conn_id->state = RDMA_CM_CONNECT;
1895 
1896 	ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL);
1897 	if (ret) {
1898 		mutex_unlock(&conn_id->handler_mutex);
1899 		rdma_destroy_id(new_cm_id);
1900 		goto out;
1901 	}
1902 
1903 	ret = cma_acquire_dev(conn_id, listen_id);
1904 	if (ret) {
1905 		mutex_unlock(&conn_id->handler_mutex);
1906 		rdma_destroy_id(new_cm_id);
1907 		goto out;
1908 	}
1909 
1910 	conn_id->cm_id.iw = cm_id;
1911 	cm_id->context = conn_id;
1912 	cm_id->cm_handler = cma_iw_handler;
1913 
1914 	memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
1915 	memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
1916 
1917 	ret = ib_query_device(conn_id->id.device, &attr);
1918 	if (ret) {
1919 		mutex_unlock(&conn_id->handler_mutex);
1920 		rdma_destroy_id(new_cm_id);
1921 		goto out;
1922 	}
1923 
1924 	memset(&event, 0, sizeof event);
1925 	event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1926 	event.param.conn.private_data = iw_event->private_data;
1927 	event.param.conn.private_data_len = iw_event->private_data_len;
1928 	event.param.conn.initiator_depth = iw_event->ird;
1929 	event.param.conn.responder_resources = iw_event->ord;
1930 
1931 	/*
1932 	 * Protect against the user destroying conn_id from another thread
1933 	 * until we're done accessing it.
1934 	 */
1935 	atomic_inc(&conn_id->refcount);
1936 	ret = conn_id->id.event_handler(&conn_id->id, &event);
1937 	if (ret) {
1938 		/* User wants to destroy the CM ID */
1939 		conn_id->cm_id.iw = NULL;
1940 		cma_exch(conn_id, RDMA_CM_DESTROYING);
1941 		mutex_unlock(&conn_id->handler_mutex);
1942 		cma_deref_id(conn_id);
1943 		rdma_destroy_id(&conn_id->id);
1944 		goto out;
1945 	}
1946 
1947 	mutex_unlock(&conn_id->handler_mutex);
1948 	cma_deref_id(conn_id);
1949 
1950 out:
1951 	mutex_unlock(&listen_id->handler_mutex);
1952 	return ret;
1953 }
1954 
1955 static int cma_ib_listen(struct rdma_id_private *id_priv)
1956 {
1957 	struct sockaddr *addr;
1958 	struct ib_cm_id	*id;
1959 	__be64 svc_id;
1960 
1961 	addr = cma_src_addr(id_priv);
1962 	svc_id = rdma_get_service_id(&id_priv->id, addr);
1963 	id = ib_cm_insert_listen(id_priv->id.device, cma_req_handler, svc_id);
1964 	if (IS_ERR(id))
1965 		return PTR_ERR(id);
1966 	id_priv->cm_id.ib = id;
1967 
1968 	return 0;
1969 }
1970 
1971 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1972 {
1973 	int ret;
1974 	struct iw_cm_id	*id;
1975 
1976 	id = iw_create_cm_id(id_priv->id.device,
1977 			     iw_conn_req_handler,
1978 			     id_priv);
1979 	if (IS_ERR(id))
1980 		return PTR_ERR(id);
1981 
1982 	id->tos = id_priv->tos;
1983 	id_priv->cm_id.iw = id;
1984 
1985 	memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
1986 	       rdma_addr_size(cma_src_addr(id_priv)));
1987 
1988 	ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1989 
1990 	if (ret) {
1991 		iw_destroy_cm_id(id_priv->cm_id.iw);
1992 		id_priv->cm_id.iw = NULL;
1993 	}
1994 
1995 	return ret;
1996 }
1997 
1998 static int cma_listen_handler(struct rdma_cm_id *id,
1999 			      struct rdma_cm_event *event)
2000 {
2001 	struct rdma_id_private *id_priv = id->context;
2002 
2003 	id->context = id_priv->id.context;
2004 	id->event_handler = id_priv->id.event_handler;
2005 	return id_priv->id.event_handler(id, event);
2006 }
2007 
2008 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2009 			      struct cma_device *cma_dev)
2010 {
2011 	struct rdma_id_private *dev_id_priv;
2012 	struct rdma_cm_id *id;
2013 	int ret;
2014 
2015 	if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2016 		return;
2017 
2018 	id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
2019 			    id_priv->id.qp_type);
2020 	if (IS_ERR(id))
2021 		return;
2022 
2023 	dev_id_priv = container_of(id, struct rdma_id_private, id);
2024 
2025 	dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2026 	memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2027 	       rdma_addr_size(cma_src_addr(id_priv)));
2028 
2029 	cma_attach_to_dev(dev_id_priv, cma_dev);
2030 	list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2031 	atomic_inc(&id_priv->refcount);
2032 	dev_id_priv->internal_id = 1;
2033 	dev_id_priv->afonly = id_priv->afonly;
2034 
2035 	ret = rdma_listen(id, id_priv->backlog);
2036 	if (ret)
2037 		printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
2038 		       "listening on device %s\n", ret, cma_dev->device->name);
2039 }
2040 
2041 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2042 {
2043 	struct cma_device *cma_dev;
2044 
2045 	mutex_lock(&lock);
2046 	list_add_tail(&id_priv->list, &listen_any_list);
2047 	list_for_each_entry(cma_dev, &dev_list, list)
2048 		cma_listen_on_dev(id_priv, cma_dev);
2049 	mutex_unlock(&lock);
2050 }
2051 
2052 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2053 {
2054 	struct rdma_id_private *id_priv;
2055 
2056 	id_priv = container_of(id, struct rdma_id_private, id);
2057 	id_priv->tos = (u8) tos;
2058 }
2059 EXPORT_SYMBOL(rdma_set_service_type);
2060 
2061 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
2062 			      void *context)
2063 {
2064 	struct cma_work *work = context;
2065 	struct rdma_route *route;
2066 
2067 	route = &work->id->id.route;
2068 
2069 	if (!status) {
2070 		route->num_paths = 1;
2071 		*route->path_rec = *path_rec;
2072 	} else {
2073 		work->old_state = RDMA_CM_ROUTE_QUERY;
2074 		work->new_state = RDMA_CM_ADDR_RESOLVED;
2075 		work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2076 		work->event.status = status;
2077 	}
2078 
2079 	queue_work(cma_wq, &work->work);
2080 }
2081 
2082 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
2083 			      struct cma_work *work)
2084 {
2085 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2086 	struct ib_sa_path_rec path_rec;
2087 	ib_sa_comp_mask comp_mask;
2088 	struct sockaddr_in6 *sin6;
2089 	struct sockaddr_ib *sib;
2090 
2091 	memset(&path_rec, 0, sizeof path_rec);
2092 	rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2093 	rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2094 	path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2095 	path_rec.numb_path = 1;
2096 	path_rec.reversible = 1;
2097 	path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2098 
2099 	comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2100 		    IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2101 		    IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2102 
2103 	switch (cma_family(id_priv)) {
2104 	case AF_INET:
2105 		path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2106 		comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2107 		break;
2108 	case AF_INET6:
2109 		sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2110 		path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2111 		comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2112 		break;
2113 	case AF_IB:
2114 		sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2115 		path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2116 		comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2117 		break;
2118 	}
2119 
2120 	id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2121 					       id_priv->id.port_num, &path_rec,
2122 					       comp_mask, timeout_ms,
2123 					       GFP_KERNEL, cma_query_handler,
2124 					       work, &id_priv->query);
2125 
2126 	return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2127 }
2128 
2129 static void cma_work_handler(struct work_struct *_work)
2130 {
2131 	struct cma_work *work = container_of(_work, struct cma_work, work);
2132 	struct rdma_id_private *id_priv = work->id;
2133 	int destroy = 0;
2134 
2135 	mutex_lock(&id_priv->handler_mutex);
2136 	if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2137 		goto out;
2138 
2139 	if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2140 		cma_exch(id_priv, RDMA_CM_DESTROYING);
2141 		destroy = 1;
2142 	}
2143 out:
2144 	mutex_unlock(&id_priv->handler_mutex);
2145 	cma_deref_id(id_priv);
2146 	if (destroy)
2147 		rdma_destroy_id(&id_priv->id);
2148 	kfree(work);
2149 }
2150 
2151 static void cma_ndev_work_handler(struct work_struct *_work)
2152 {
2153 	struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2154 	struct rdma_id_private *id_priv = work->id;
2155 	int destroy = 0;
2156 
2157 	mutex_lock(&id_priv->handler_mutex);
2158 	if (id_priv->state == RDMA_CM_DESTROYING ||
2159 	    id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2160 		goto out;
2161 
2162 	if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2163 		cma_exch(id_priv, RDMA_CM_DESTROYING);
2164 		destroy = 1;
2165 	}
2166 
2167 out:
2168 	mutex_unlock(&id_priv->handler_mutex);
2169 	cma_deref_id(id_priv);
2170 	if (destroy)
2171 		rdma_destroy_id(&id_priv->id);
2172 	kfree(work);
2173 }
2174 
2175 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
2176 {
2177 	struct rdma_route *route = &id_priv->id.route;
2178 	struct cma_work *work;
2179 	int ret;
2180 
2181 	work = kzalloc(sizeof *work, GFP_KERNEL);
2182 	if (!work)
2183 		return -ENOMEM;
2184 
2185 	work->id = id_priv;
2186 	INIT_WORK(&work->work, cma_work_handler);
2187 	work->old_state = RDMA_CM_ROUTE_QUERY;
2188 	work->new_state = RDMA_CM_ROUTE_RESOLVED;
2189 	work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2190 
2191 	route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2192 	if (!route->path_rec) {
2193 		ret = -ENOMEM;
2194 		goto err1;
2195 	}
2196 
2197 	ret = cma_query_ib_route(id_priv, timeout_ms, work);
2198 	if (ret)
2199 		goto err2;
2200 
2201 	return 0;
2202 err2:
2203 	kfree(route->path_rec);
2204 	route->path_rec = NULL;
2205 err1:
2206 	kfree(work);
2207 	return ret;
2208 }
2209 
2210 int rdma_set_ib_paths(struct rdma_cm_id *id,
2211 		      struct ib_sa_path_rec *path_rec, int num_paths)
2212 {
2213 	struct rdma_id_private *id_priv;
2214 	int ret;
2215 
2216 	id_priv = container_of(id, struct rdma_id_private, id);
2217 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2218 			   RDMA_CM_ROUTE_RESOLVED))
2219 		return -EINVAL;
2220 
2221 	id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
2222 				     GFP_KERNEL);
2223 	if (!id->route.path_rec) {
2224 		ret = -ENOMEM;
2225 		goto err;
2226 	}
2227 
2228 	id->route.num_paths = num_paths;
2229 	return 0;
2230 err:
2231 	cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2232 	return ret;
2233 }
2234 EXPORT_SYMBOL(rdma_set_ib_paths);
2235 
2236 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2237 {
2238 	struct cma_work *work;
2239 
2240 	work = kzalloc(sizeof *work, GFP_KERNEL);
2241 	if (!work)
2242 		return -ENOMEM;
2243 
2244 	work->id = id_priv;
2245 	INIT_WORK(&work->work, cma_work_handler);
2246 	work->old_state = RDMA_CM_ROUTE_QUERY;
2247 	work->new_state = RDMA_CM_ROUTE_RESOLVED;
2248 	work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2249 	queue_work(cma_wq, &work->work);
2250 	return 0;
2251 }
2252 
2253 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2254 {
2255 	int prio;
2256 	struct net_device *dev;
2257 
2258 	prio = rt_tos2priority(tos);
2259 	dev = ndev->priv_flags & IFF_802_1Q_VLAN ?
2260 		vlan_dev_real_dev(ndev) : ndev;
2261 
2262 	if (dev->num_tc)
2263 		return netdev_get_prio_tc_map(dev, prio);
2264 
2265 #if IS_ENABLED(CONFIG_VLAN_8021Q)
2266 	if (ndev->priv_flags & IFF_802_1Q_VLAN)
2267 		return (vlan_dev_get_egress_qos_mask(ndev, prio) &
2268 			VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2269 #endif
2270 	return 0;
2271 }
2272 
2273 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2274 {
2275 	struct rdma_route *route = &id_priv->id.route;
2276 	struct rdma_addr *addr = &route->addr;
2277 	struct cma_work *work;
2278 	int ret;
2279 	struct net_device *ndev = NULL;
2280 
2281 
2282 	work = kzalloc(sizeof *work, GFP_KERNEL);
2283 	if (!work)
2284 		return -ENOMEM;
2285 
2286 	work->id = id_priv;
2287 	INIT_WORK(&work->work, cma_work_handler);
2288 
2289 	route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2290 	if (!route->path_rec) {
2291 		ret = -ENOMEM;
2292 		goto err1;
2293 	}
2294 
2295 	route->num_paths = 1;
2296 
2297 	if (addr->dev_addr.bound_dev_if)
2298 		ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
2299 	if (!ndev) {
2300 		ret = -ENODEV;
2301 		goto err2;
2302 	}
2303 
2304 	route->path_rec->vlan_id = rdma_vlan_dev_vlan_id(ndev);
2305 	memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN);
2306 	memcpy(route->path_rec->smac, ndev->dev_addr, ndev->addr_len);
2307 
2308 	rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2309 		    &route->path_rec->sgid);
2310 	rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2311 		    &route->path_rec->dgid);
2312 
2313 	route->path_rec->hop_limit = 1;
2314 	route->path_rec->reversible = 1;
2315 	route->path_rec->pkey = cpu_to_be16(0xffff);
2316 	route->path_rec->mtu_selector = IB_SA_EQ;
2317 	route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos);
2318 	route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2319 	route->path_rec->rate_selector = IB_SA_EQ;
2320 	route->path_rec->rate = iboe_get_rate(ndev);
2321 	dev_put(ndev);
2322 	route->path_rec->packet_life_time_selector = IB_SA_EQ;
2323 	route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2324 	if (!route->path_rec->mtu) {
2325 		ret = -EINVAL;
2326 		goto err2;
2327 	}
2328 
2329 	work->old_state = RDMA_CM_ROUTE_QUERY;
2330 	work->new_state = RDMA_CM_ROUTE_RESOLVED;
2331 	work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2332 	work->event.status = 0;
2333 
2334 	queue_work(cma_wq, &work->work);
2335 
2336 	return 0;
2337 
2338 err2:
2339 	kfree(route->path_rec);
2340 	route->path_rec = NULL;
2341 err1:
2342 	kfree(work);
2343 	return ret;
2344 }
2345 
2346 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2347 {
2348 	struct rdma_id_private *id_priv;
2349 	int ret;
2350 
2351 	id_priv = container_of(id, struct rdma_id_private, id);
2352 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2353 		return -EINVAL;
2354 
2355 	atomic_inc(&id_priv->refcount);
2356 	if (rdma_cap_ib_sa(id->device, id->port_num))
2357 		ret = cma_resolve_ib_route(id_priv, timeout_ms);
2358 	else if (rdma_protocol_roce(id->device, id->port_num))
2359 		ret = cma_resolve_iboe_route(id_priv);
2360 	else if (rdma_protocol_iwarp(id->device, id->port_num))
2361 		ret = cma_resolve_iw_route(id_priv, timeout_ms);
2362 	else
2363 		ret = -ENOSYS;
2364 
2365 	if (ret)
2366 		goto err;
2367 
2368 	return 0;
2369 err:
2370 	cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2371 	cma_deref_id(id_priv);
2372 	return ret;
2373 }
2374 EXPORT_SYMBOL(rdma_resolve_route);
2375 
2376 static void cma_set_loopback(struct sockaddr *addr)
2377 {
2378 	switch (addr->sa_family) {
2379 	case AF_INET:
2380 		((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2381 		break;
2382 	case AF_INET6:
2383 		ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2384 			      0, 0, 0, htonl(1));
2385 		break;
2386 	default:
2387 		ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2388 			    0, 0, 0, htonl(1));
2389 		break;
2390 	}
2391 }
2392 
2393 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2394 {
2395 	struct cma_device *cma_dev, *cur_dev;
2396 	struct ib_port_attr port_attr;
2397 	union ib_gid gid;
2398 	u16 pkey;
2399 	int ret;
2400 	u8 p;
2401 
2402 	cma_dev = NULL;
2403 	mutex_lock(&lock);
2404 	list_for_each_entry(cur_dev, &dev_list, list) {
2405 		if (cma_family(id_priv) == AF_IB &&
2406 		    !rdma_cap_ib_cm(cur_dev->device, 1))
2407 			continue;
2408 
2409 		if (!cma_dev)
2410 			cma_dev = cur_dev;
2411 
2412 		for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2413 			if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2414 			    port_attr.state == IB_PORT_ACTIVE) {
2415 				cma_dev = cur_dev;
2416 				goto port_found;
2417 			}
2418 		}
2419 	}
2420 
2421 	if (!cma_dev) {
2422 		ret = -ENODEV;
2423 		goto out;
2424 	}
2425 
2426 	p = 1;
2427 
2428 port_found:
2429 	ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
2430 	if (ret)
2431 		goto out;
2432 
2433 	ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2434 	if (ret)
2435 		goto out;
2436 
2437 	id_priv->id.route.addr.dev_addr.dev_type =
2438 		(rdma_protocol_ib(cma_dev->device, p)) ?
2439 		ARPHRD_INFINIBAND : ARPHRD_ETHER;
2440 
2441 	rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2442 	ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2443 	id_priv->id.port_num = p;
2444 	cma_attach_to_dev(id_priv, cma_dev);
2445 	cma_set_loopback(cma_src_addr(id_priv));
2446 out:
2447 	mutex_unlock(&lock);
2448 	return ret;
2449 }
2450 
2451 static void addr_handler(int status, struct sockaddr *src_addr,
2452 			 struct rdma_dev_addr *dev_addr, void *context)
2453 {
2454 	struct rdma_id_private *id_priv = context;
2455 	struct rdma_cm_event event;
2456 
2457 	memset(&event, 0, sizeof event);
2458 	mutex_lock(&id_priv->handler_mutex);
2459 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2460 			   RDMA_CM_ADDR_RESOLVED))
2461 		goto out;
2462 
2463 	memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2464 	if (!status && !id_priv->cma_dev)
2465 		status = cma_acquire_dev(id_priv, NULL);
2466 
2467 	if (status) {
2468 		if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2469 				   RDMA_CM_ADDR_BOUND))
2470 			goto out;
2471 		event.event = RDMA_CM_EVENT_ADDR_ERROR;
2472 		event.status = status;
2473 	} else
2474 		event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2475 
2476 	if (id_priv->id.event_handler(&id_priv->id, &event)) {
2477 		cma_exch(id_priv, RDMA_CM_DESTROYING);
2478 		mutex_unlock(&id_priv->handler_mutex);
2479 		cma_deref_id(id_priv);
2480 		rdma_destroy_id(&id_priv->id);
2481 		return;
2482 	}
2483 out:
2484 	mutex_unlock(&id_priv->handler_mutex);
2485 	cma_deref_id(id_priv);
2486 }
2487 
2488 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2489 {
2490 	struct cma_work *work;
2491 	union ib_gid gid;
2492 	int ret;
2493 
2494 	work = kzalloc(sizeof *work, GFP_KERNEL);
2495 	if (!work)
2496 		return -ENOMEM;
2497 
2498 	if (!id_priv->cma_dev) {
2499 		ret = cma_bind_loopback(id_priv);
2500 		if (ret)
2501 			goto err;
2502 	}
2503 
2504 	rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2505 	rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2506 
2507 	work->id = id_priv;
2508 	INIT_WORK(&work->work, cma_work_handler);
2509 	work->old_state = RDMA_CM_ADDR_QUERY;
2510 	work->new_state = RDMA_CM_ADDR_RESOLVED;
2511 	work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2512 	queue_work(cma_wq, &work->work);
2513 	return 0;
2514 err:
2515 	kfree(work);
2516 	return ret;
2517 }
2518 
2519 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2520 {
2521 	struct cma_work *work;
2522 	int ret;
2523 
2524 	work = kzalloc(sizeof *work, GFP_KERNEL);
2525 	if (!work)
2526 		return -ENOMEM;
2527 
2528 	if (!id_priv->cma_dev) {
2529 		ret = cma_resolve_ib_dev(id_priv);
2530 		if (ret)
2531 			goto err;
2532 	}
2533 
2534 	rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2535 		&(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2536 
2537 	work->id = id_priv;
2538 	INIT_WORK(&work->work, cma_work_handler);
2539 	work->old_state = RDMA_CM_ADDR_QUERY;
2540 	work->new_state = RDMA_CM_ADDR_RESOLVED;
2541 	work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2542 	queue_work(cma_wq, &work->work);
2543 	return 0;
2544 err:
2545 	kfree(work);
2546 	return ret;
2547 }
2548 
2549 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2550 			 struct sockaddr *dst_addr)
2551 {
2552 	if (!src_addr || !src_addr->sa_family) {
2553 		src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2554 		src_addr->sa_family = dst_addr->sa_family;
2555 		if (dst_addr->sa_family == AF_INET6) {
2556 			struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
2557 			struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
2558 			src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
2559 			if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
2560 				id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
2561 		} else if (dst_addr->sa_family == AF_IB) {
2562 			((struct sockaddr_ib *) src_addr)->sib_pkey =
2563 				((struct sockaddr_ib *) dst_addr)->sib_pkey;
2564 		}
2565 	}
2566 	return rdma_bind_addr(id, src_addr);
2567 }
2568 
2569 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2570 		      struct sockaddr *dst_addr, int timeout_ms)
2571 {
2572 	struct rdma_id_private *id_priv;
2573 	int ret;
2574 
2575 	id_priv = container_of(id, struct rdma_id_private, id);
2576 	if (id_priv->state == RDMA_CM_IDLE) {
2577 		ret = cma_bind_addr(id, src_addr, dst_addr);
2578 		if (ret)
2579 			return ret;
2580 	}
2581 
2582 	if (cma_family(id_priv) != dst_addr->sa_family)
2583 		return -EINVAL;
2584 
2585 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2586 		return -EINVAL;
2587 
2588 	atomic_inc(&id_priv->refcount);
2589 	memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2590 	if (cma_any_addr(dst_addr)) {
2591 		ret = cma_resolve_loopback(id_priv);
2592 	} else {
2593 		if (dst_addr->sa_family == AF_IB) {
2594 			ret = cma_resolve_ib_addr(id_priv);
2595 		} else {
2596 			ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2597 					      dst_addr, &id->route.addr.dev_addr,
2598 					      timeout_ms, addr_handler, id_priv);
2599 		}
2600 	}
2601 	if (ret)
2602 		goto err;
2603 
2604 	return 0;
2605 err:
2606 	cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2607 	cma_deref_id(id_priv);
2608 	return ret;
2609 }
2610 EXPORT_SYMBOL(rdma_resolve_addr);
2611 
2612 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2613 {
2614 	struct rdma_id_private *id_priv;
2615 	unsigned long flags;
2616 	int ret;
2617 
2618 	id_priv = container_of(id, struct rdma_id_private, id);
2619 	spin_lock_irqsave(&id_priv->lock, flags);
2620 	if (reuse || id_priv->state == RDMA_CM_IDLE) {
2621 		id_priv->reuseaddr = reuse;
2622 		ret = 0;
2623 	} else {
2624 		ret = -EINVAL;
2625 	}
2626 	spin_unlock_irqrestore(&id_priv->lock, flags);
2627 	return ret;
2628 }
2629 EXPORT_SYMBOL(rdma_set_reuseaddr);
2630 
2631 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2632 {
2633 	struct rdma_id_private *id_priv;
2634 	unsigned long flags;
2635 	int ret;
2636 
2637 	id_priv = container_of(id, struct rdma_id_private, id);
2638 	spin_lock_irqsave(&id_priv->lock, flags);
2639 	if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2640 		id_priv->options |= (1 << CMA_OPTION_AFONLY);
2641 		id_priv->afonly = afonly;
2642 		ret = 0;
2643 	} else {
2644 		ret = -EINVAL;
2645 	}
2646 	spin_unlock_irqrestore(&id_priv->lock, flags);
2647 	return ret;
2648 }
2649 EXPORT_SYMBOL(rdma_set_afonly);
2650 
2651 static void cma_bind_port(struct rdma_bind_list *bind_list,
2652 			  struct rdma_id_private *id_priv)
2653 {
2654 	struct sockaddr *addr;
2655 	struct sockaddr_ib *sib;
2656 	u64 sid, mask;
2657 	__be16 port;
2658 
2659 	addr = cma_src_addr(id_priv);
2660 	port = htons(bind_list->port);
2661 
2662 	switch (addr->sa_family) {
2663 	case AF_INET:
2664 		((struct sockaddr_in *) addr)->sin_port = port;
2665 		break;
2666 	case AF_INET6:
2667 		((struct sockaddr_in6 *) addr)->sin6_port = port;
2668 		break;
2669 	case AF_IB:
2670 		sib = (struct sockaddr_ib *) addr;
2671 		sid = be64_to_cpu(sib->sib_sid);
2672 		mask = be64_to_cpu(sib->sib_sid_mask);
2673 		sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2674 		sib->sib_sid_mask = cpu_to_be64(~0ULL);
2675 		break;
2676 	}
2677 	id_priv->bind_list = bind_list;
2678 	hlist_add_head(&id_priv->node, &bind_list->owners);
2679 }
2680 
2681 static int cma_alloc_port(enum rdma_port_space ps,
2682 			  struct rdma_id_private *id_priv, unsigned short snum)
2683 {
2684 	struct rdma_bind_list *bind_list;
2685 	int ret;
2686 
2687 	bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2688 	if (!bind_list)
2689 		return -ENOMEM;
2690 
2691 	ret = cma_ps_alloc(ps, bind_list, snum);
2692 	if (ret < 0)
2693 		goto err;
2694 
2695 	bind_list->ps = ps;
2696 	bind_list->port = (unsigned short)ret;
2697 	cma_bind_port(bind_list, id_priv);
2698 	return 0;
2699 err:
2700 	kfree(bind_list);
2701 	return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2702 }
2703 
2704 static int cma_alloc_any_port(enum rdma_port_space ps,
2705 			      struct rdma_id_private *id_priv)
2706 {
2707 	static unsigned int last_used_port;
2708 	int low, high, remaining;
2709 	unsigned int rover;
2710 
2711 	inet_get_local_port_range(&init_net, &low, &high);
2712 	remaining = (high - low) + 1;
2713 	rover = prandom_u32() % remaining + low;
2714 retry:
2715 	if (last_used_port != rover &&
2716 	    !cma_ps_find(ps, (unsigned short)rover)) {
2717 		int ret = cma_alloc_port(ps, id_priv, rover);
2718 		/*
2719 		 * Remember previously used port number in order to avoid
2720 		 * re-using same port immediately after it is closed.
2721 		 */
2722 		if (!ret)
2723 			last_used_port = rover;
2724 		if (ret != -EADDRNOTAVAIL)
2725 			return ret;
2726 	}
2727 	if (--remaining) {
2728 		rover++;
2729 		if ((rover < low) || (rover > high))
2730 			rover = low;
2731 		goto retry;
2732 	}
2733 	return -EADDRNOTAVAIL;
2734 }
2735 
2736 /*
2737  * Check that the requested port is available.  This is called when trying to
2738  * bind to a specific port, or when trying to listen on a bound port.  In
2739  * the latter case, the provided id_priv may already be on the bind_list, but
2740  * we still need to check that it's okay to start listening.
2741  */
2742 static int cma_check_port(struct rdma_bind_list *bind_list,
2743 			  struct rdma_id_private *id_priv, uint8_t reuseaddr)
2744 {
2745 	struct rdma_id_private *cur_id;
2746 	struct sockaddr *addr, *cur_addr;
2747 
2748 	addr = cma_src_addr(id_priv);
2749 	hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2750 		if (id_priv == cur_id)
2751 			continue;
2752 
2753 		if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2754 		    cur_id->reuseaddr)
2755 			continue;
2756 
2757 		cur_addr = cma_src_addr(cur_id);
2758 		if (id_priv->afonly && cur_id->afonly &&
2759 		    (addr->sa_family != cur_addr->sa_family))
2760 			continue;
2761 
2762 		if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2763 			return -EADDRNOTAVAIL;
2764 
2765 		if (!cma_addr_cmp(addr, cur_addr))
2766 			return -EADDRINUSE;
2767 	}
2768 	return 0;
2769 }
2770 
2771 static int cma_use_port(enum rdma_port_space ps,
2772 			struct rdma_id_private *id_priv)
2773 {
2774 	struct rdma_bind_list *bind_list;
2775 	unsigned short snum;
2776 	int ret;
2777 
2778 	snum = ntohs(cma_port(cma_src_addr(id_priv)));
2779 	if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2780 		return -EACCES;
2781 
2782 	bind_list = cma_ps_find(ps, snum);
2783 	if (!bind_list) {
2784 		ret = cma_alloc_port(ps, id_priv, snum);
2785 	} else {
2786 		ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2787 		if (!ret)
2788 			cma_bind_port(bind_list, id_priv);
2789 	}
2790 	return ret;
2791 }
2792 
2793 static int cma_bind_listen(struct rdma_id_private *id_priv)
2794 {
2795 	struct rdma_bind_list *bind_list = id_priv->bind_list;
2796 	int ret = 0;
2797 
2798 	mutex_lock(&lock);
2799 	if (bind_list->owners.first->next)
2800 		ret = cma_check_port(bind_list, id_priv, 0);
2801 	mutex_unlock(&lock);
2802 	return ret;
2803 }
2804 
2805 static enum rdma_port_space cma_select_inet_ps(
2806 		struct rdma_id_private *id_priv)
2807 {
2808 	switch (id_priv->id.ps) {
2809 	case RDMA_PS_TCP:
2810 	case RDMA_PS_UDP:
2811 	case RDMA_PS_IPOIB:
2812 	case RDMA_PS_IB:
2813 		return id_priv->id.ps;
2814 	default:
2815 
2816 		return 0;
2817 	}
2818 }
2819 
2820 static enum rdma_port_space cma_select_ib_ps(struct rdma_id_private *id_priv)
2821 {
2822 	enum rdma_port_space ps = 0;
2823 	struct sockaddr_ib *sib;
2824 	u64 sid_ps, mask, sid;
2825 
2826 	sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2827 	mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
2828 	sid = be64_to_cpu(sib->sib_sid) & mask;
2829 
2830 	if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
2831 		sid_ps = RDMA_IB_IP_PS_IB;
2832 		ps = RDMA_PS_IB;
2833 	} else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
2834 		   (sid == (RDMA_IB_IP_PS_TCP & mask))) {
2835 		sid_ps = RDMA_IB_IP_PS_TCP;
2836 		ps = RDMA_PS_TCP;
2837 	} else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
2838 		   (sid == (RDMA_IB_IP_PS_UDP & mask))) {
2839 		sid_ps = RDMA_IB_IP_PS_UDP;
2840 		ps = RDMA_PS_UDP;
2841 	}
2842 
2843 	if (ps) {
2844 		sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
2845 		sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
2846 						be64_to_cpu(sib->sib_sid_mask));
2847 	}
2848 	return ps;
2849 }
2850 
2851 static int cma_get_port(struct rdma_id_private *id_priv)
2852 {
2853 	enum rdma_port_space ps;
2854 	int ret;
2855 
2856 	if (cma_family(id_priv) != AF_IB)
2857 		ps = cma_select_inet_ps(id_priv);
2858 	else
2859 		ps = cma_select_ib_ps(id_priv);
2860 	if (!ps)
2861 		return -EPROTONOSUPPORT;
2862 
2863 	mutex_lock(&lock);
2864 	if (cma_any_port(cma_src_addr(id_priv)))
2865 		ret = cma_alloc_any_port(ps, id_priv);
2866 	else
2867 		ret = cma_use_port(ps, id_priv);
2868 	mutex_unlock(&lock);
2869 
2870 	return ret;
2871 }
2872 
2873 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2874 			       struct sockaddr *addr)
2875 {
2876 #if IS_ENABLED(CONFIG_IPV6)
2877 	struct sockaddr_in6 *sin6;
2878 
2879 	if (addr->sa_family != AF_INET6)
2880 		return 0;
2881 
2882 	sin6 = (struct sockaddr_in6 *) addr;
2883 
2884 	if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
2885 		return 0;
2886 
2887 	if (!sin6->sin6_scope_id)
2888 			return -EINVAL;
2889 
2890 	dev_addr->bound_dev_if = sin6->sin6_scope_id;
2891 #endif
2892 	return 0;
2893 }
2894 
2895 int rdma_listen(struct rdma_cm_id *id, int backlog)
2896 {
2897 	struct rdma_id_private *id_priv;
2898 	int ret;
2899 
2900 	id_priv = container_of(id, struct rdma_id_private, id);
2901 	if (id_priv->state == RDMA_CM_IDLE) {
2902 		id->route.addr.src_addr.ss_family = AF_INET;
2903 		ret = rdma_bind_addr(id, cma_src_addr(id_priv));
2904 		if (ret)
2905 			return ret;
2906 	}
2907 
2908 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2909 		return -EINVAL;
2910 
2911 	if (id_priv->reuseaddr) {
2912 		ret = cma_bind_listen(id_priv);
2913 		if (ret)
2914 			goto err;
2915 	}
2916 
2917 	id_priv->backlog = backlog;
2918 	if (id->device) {
2919 		if (rdma_cap_ib_cm(id->device, 1)) {
2920 			ret = cma_ib_listen(id_priv);
2921 			if (ret)
2922 				goto err;
2923 		} else if (rdma_cap_iw_cm(id->device, 1)) {
2924 			ret = cma_iw_listen(id_priv, backlog);
2925 			if (ret)
2926 				goto err;
2927 		} else {
2928 			ret = -ENOSYS;
2929 			goto err;
2930 		}
2931 	} else
2932 		cma_listen_on_all(id_priv);
2933 
2934 	return 0;
2935 err:
2936 	id_priv->backlog = 0;
2937 	cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2938 	return ret;
2939 }
2940 EXPORT_SYMBOL(rdma_listen);
2941 
2942 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2943 {
2944 	struct rdma_id_private *id_priv;
2945 	int ret;
2946 
2947 	if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
2948 	    addr->sa_family != AF_IB)
2949 		return -EAFNOSUPPORT;
2950 
2951 	id_priv = container_of(id, struct rdma_id_private, id);
2952 	if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2953 		return -EINVAL;
2954 
2955 	ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2956 	if (ret)
2957 		goto err1;
2958 
2959 	memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
2960 	if (!cma_any_addr(addr)) {
2961 		ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
2962 		if (ret)
2963 			goto err1;
2964 
2965 		ret = cma_acquire_dev(id_priv, NULL);
2966 		if (ret)
2967 			goto err1;
2968 	}
2969 
2970 	if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
2971 		if (addr->sa_family == AF_INET)
2972 			id_priv->afonly = 1;
2973 #if IS_ENABLED(CONFIG_IPV6)
2974 		else if (addr->sa_family == AF_INET6)
2975 			id_priv->afonly = init_net.ipv6.sysctl.bindv6only;
2976 #endif
2977 	}
2978 	ret = cma_get_port(id_priv);
2979 	if (ret)
2980 		goto err2;
2981 
2982 	return 0;
2983 err2:
2984 	if (id_priv->cma_dev)
2985 		cma_release_dev(id_priv);
2986 err1:
2987 	cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2988 	return ret;
2989 }
2990 EXPORT_SYMBOL(rdma_bind_addr);
2991 
2992 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
2993 {
2994 	struct cma_hdr *cma_hdr;
2995 
2996 	cma_hdr = hdr;
2997 	cma_hdr->cma_version = CMA_VERSION;
2998 	if (cma_family(id_priv) == AF_INET) {
2999 		struct sockaddr_in *src4, *dst4;
3000 
3001 		src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3002 		dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3003 
3004 		cma_set_ip_ver(cma_hdr, 4);
3005 		cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3006 		cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3007 		cma_hdr->port = src4->sin_port;
3008 	} else if (cma_family(id_priv) == AF_INET6) {
3009 		struct sockaddr_in6 *src6, *dst6;
3010 
3011 		src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3012 		dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3013 
3014 		cma_set_ip_ver(cma_hdr, 6);
3015 		cma_hdr->src_addr.ip6 = src6->sin6_addr;
3016 		cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3017 		cma_hdr->port = src6->sin6_port;
3018 	}
3019 	return 0;
3020 }
3021 
3022 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3023 				struct ib_cm_event *ib_event)
3024 {
3025 	struct rdma_id_private *id_priv = cm_id->context;
3026 	struct rdma_cm_event event;
3027 	struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
3028 	int ret = 0;
3029 
3030 	if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
3031 		return 0;
3032 
3033 	memset(&event, 0, sizeof event);
3034 	switch (ib_event->event) {
3035 	case IB_CM_SIDR_REQ_ERROR:
3036 		event.event = RDMA_CM_EVENT_UNREACHABLE;
3037 		event.status = -ETIMEDOUT;
3038 		break;
3039 	case IB_CM_SIDR_REP_RECEIVED:
3040 		event.param.ud.private_data = ib_event->private_data;
3041 		event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3042 		if (rep->status != IB_SIDR_SUCCESS) {
3043 			event.event = RDMA_CM_EVENT_UNREACHABLE;
3044 			event.status = ib_event->param.sidr_rep_rcvd.status;
3045 			break;
3046 		}
3047 		ret = cma_set_qkey(id_priv, rep->qkey);
3048 		if (ret) {
3049 			event.event = RDMA_CM_EVENT_ADDR_ERROR;
3050 			event.status = ret;
3051 			break;
3052 		}
3053 		ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
3054 				     id_priv->id.route.path_rec,
3055 				     &event.param.ud.ah_attr);
3056 		event.param.ud.qp_num = rep->qpn;
3057 		event.param.ud.qkey = rep->qkey;
3058 		event.event = RDMA_CM_EVENT_ESTABLISHED;
3059 		event.status = 0;
3060 		break;
3061 	default:
3062 		printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
3063 		       ib_event->event);
3064 		goto out;
3065 	}
3066 
3067 	ret = id_priv->id.event_handler(&id_priv->id, &event);
3068 	if (ret) {
3069 		/* Destroy the CM ID by returning a non-zero value. */
3070 		id_priv->cm_id.ib = NULL;
3071 		cma_exch(id_priv, RDMA_CM_DESTROYING);
3072 		mutex_unlock(&id_priv->handler_mutex);
3073 		rdma_destroy_id(&id_priv->id);
3074 		return ret;
3075 	}
3076 out:
3077 	mutex_unlock(&id_priv->handler_mutex);
3078 	return ret;
3079 }
3080 
3081 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3082 			      struct rdma_conn_param *conn_param)
3083 {
3084 	struct ib_cm_sidr_req_param req;
3085 	struct ib_cm_id	*id;
3086 	void *private_data;
3087 	int offset, ret;
3088 
3089 	memset(&req, 0, sizeof req);
3090 	offset = cma_user_data_offset(id_priv);
3091 	req.private_data_len = offset + conn_param->private_data_len;
3092 	if (req.private_data_len < conn_param->private_data_len)
3093 		return -EINVAL;
3094 
3095 	if (req.private_data_len) {
3096 		private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3097 		if (!private_data)
3098 			return -ENOMEM;
3099 	} else {
3100 		private_data = NULL;
3101 	}
3102 
3103 	if (conn_param->private_data && conn_param->private_data_len)
3104 		memcpy(private_data + offset, conn_param->private_data,
3105 		       conn_param->private_data_len);
3106 
3107 	if (private_data) {
3108 		ret = cma_format_hdr(private_data, id_priv);
3109 		if (ret)
3110 			goto out;
3111 		req.private_data = private_data;
3112 	}
3113 
3114 	id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3115 			     id_priv);
3116 	if (IS_ERR(id)) {
3117 		ret = PTR_ERR(id);
3118 		goto out;
3119 	}
3120 	id_priv->cm_id.ib = id;
3121 
3122 	req.path = id_priv->id.route.path_rec;
3123 	req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3124 	req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3125 	req.max_cm_retries = CMA_MAX_CM_RETRIES;
3126 
3127 	ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3128 	if (ret) {
3129 		ib_destroy_cm_id(id_priv->cm_id.ib);
3130 		id_priv->cm_id.ib = NULL;
3131 	}
3132 out:
3133 	kfree(private_data);
3134 	return ret;
3135 }
3136 
3137 static int cma_connect_ib(struct rdma_id_private *id_priv,
3138 			  struct rdma_conn_param *conn_param)
3139 {
3140 	struct ib_cm_req_param req;
3141 	struct rdma_route *route;
3142 	void *private_data;
3143 	struct ib_cm_id	*id;
3144 	int offset, ret;
3145 
3146 	memset(&req, 0, sizeof req);
3147 	offset = cma_user_data_offset(id_priv);
3148 	req.private_data_len = offset + conn_param->private_data_len;
3149 	if (req.private_data_len < conn_param->private_data_len)
3150 		return -EINVAL;
3151 
3152 	if (req.private_data_len) {
3153 		private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3154 		if (!private_data)
3155 			return -ENOMEM;
3156 	} else {
3157 		private_data = NULL;
3158 	}
3159 
3160 	if (conn_param->private_data && conn_param->private_data_len)
3161 		memcpy(private_data + offset, conn_param->private_data,
3162 		       conn_param->private_data_len);
3163 
3164 	id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3165 	if (IS_ERR(id)) {
3166 		ret = PTR_ERR(id);
3167 		goto out;
3168 	}
3169 	id_priv->cm_id.ib = id;
3170 
3171 	route = &id_priv->id.route;
3172 	if (private_data) {
3173 		ret = cma_format_hdr(private_data, id_priv);
3174 		if (ret)
3175 			goto out;
3176 		req.private_data = private_data;
3177 	}
3178 
3179 	req.primary_path = &route->path_rec[0];
3180 	if (route->num_paths == 2)
3181 		req.alternate_path = &route->path_rec[1];
3182 
3183 	req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3184 	req.qp_num = id_priv->qp_num;
3185 	req.qp_type = id_priv->id.qp_type;
3186 	req.starting_psn = id_priv->seq_num;
3187 	req.responder_resources = conn_param->responder_resources;
3188 	req.initiator_depth = conn_param->initiator_depth;
3189 	req.flow_control = conn_param->flow_control;
3190 	req.retry_count = min_t(u8, 7, conn_param->retry_count);
3191 	req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3192 	req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3193 	req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3194 	req.max_cm_retries = CMA_MAX_CM_RETRIES;
3195 	req.srq = id_priv->srq ? 1 : 0;
3196 
3197 	ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3198 out:
3199 	if (ret && !IS_ERR(id)) {
3200 		ib_destroy_cm_id(id);
3201 		id_priv->cm_id.ib = NULL;
3202 	}
3203 
3204 	kfree(private_data);
3205 	return ret;
3206 }
3207 
3208 static int cma_connect_iw(struct rdma_id_private *id_priv,
3209 			  struct rdma_conn_param *conn_param)
3210 {
3211 	struct iw_cm_id *cm_id;
3212 	int ret;
3213 	struct iw_cm_conn_param iw_param;
3214 
3215 	cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3216 	if (IS_ERR(cm_id))
3217 		return PTR_ERR(cm_id);
3218 
3219 	cm_id->tos = id_priv->tos;
3220 	id_priv->cm_id.iw = cm_id;
3221 
3222 	memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3223 	       rdma_addr_size(cma_src_addr(id_priv)));
3224 	memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3225 	       rdma_addr_size(cma_dst_addr(id_priv)));
3226 
3227 	ret = cma_modify_qp_rtr(id_priv, conn_param);
3228 	if (ret)
3229 		goto out;
3230 
3231 	if (conn_param) {
3232 		iw_param.ord = conn_param->initiator_depth;
3233 		iw_param.ird = conn_param->responder_resources;
3234 		iw_param.private_data = conn_param->private_data;
3235 		iw_param.private_data_len = conn_param->private_data_len;
3236 		iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3237 	} else {
3238 		memset(&iw_param, 0, sizeof iw_param);
3239 		iw_param.qpn = id_priv->qp_num;
3240 	}
3241 	ret = iw_cm_connect(cm_id, &iw_param);
3242 out:
3243 	if (ret) {
3244 		iw_destroy_cm_id(cm_id);
3245 		id_priv->cm_id.iw = NULL;
3246 	}
3247 	return ret;
3248 }
3249 
3250 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3251 {
3252 	struct rdma_id_private *id_priv;
3253 	int ret;
3254 
3255 	id_priv = container_of(id, struct rdma_id_private, id);
3256 	if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3257 		return -EINVAL;
3258 
3259 	if (!id->qp) {
3260 		id_priv->qp_num = conn_param->qp_num;
3261 		id_priv->srq = conn_param->srq;
3262 	}
3263 
3264 	if (rdma_cap_ib_cm(id->device, id->port_num)) {
3265 		if (id->qp_type == IB_QPT_UD)
3266 			ret = cma_resolve_ib_udp(id_priv, conn_param);
3267 		else
3268 			ret = cma_connect_ib(id_priv, conn_param);
3269 	} else if (rdma_cap_iw_cm(id->device, id->port_num))
3270 		ret = cma_connect_iw(id_priv, conn_param);
3271 	else
3272 		ret = -ENOSYS;
3273 	if (ret)
3274 		goto err;
3275 
3276 	return 0;
3277 err:
3278 	cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3279 	return ret;
3280 }
3281 EXPORT_SYMBOL(rdma_connect);
3282 
3283 static int cma_accept_ib(struct rdma_id_private *id_priv,
3284 			 struct rdma_conn_param *conn_param)
3285 {
3286 	struct ib_cm_rep_param rep;
3287 	int ret;
3288 
3289 	ret = cma_modify_qp_rtr(id_priv, conn_param);
3290 	if (ret)
3291 		goto out;
3292 
3293 	ret = cma_modify_qp_rts(id_priv, conn_param);
3294 	if (ret)
3295 		goto out;
3296 
3297 	memset(&rep, 0, sizeof rep);
3298 	rep.qp_num = id_priv->qp_num;
3299 	rep.starting_psn = id_priv->seq_num;
3300 	rep.private_data = conn_param->private_data;
3301 	rep.private_data_len = conn_param->private_data_len;
3302 	rep.responder_resources = conn_param->responder_resources;
3303 	rep.initiator_depth = conn_param->initiator_depth;
3304 	rep.failover_accepted = 0;
3305 	rep.flow_control = conn_param->flow_control;
3306 	rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3307 	rep.srq = id_priv->srq ? 1 : 0;
3308 
3309 	ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3310 out:
3311 	return ret;
3312 }
3313 
3314 static int cma_accept_iw(struct rdma_id_private *id_priv,
3315 		  struct rdma_conn_param *conn_param)
3316 {
3317 	struct iw_cm_conn_param iw_param;
3318 	int ret;
3319 
3320 	ret = cma_modify_qp_rtr(id_priv, conn_param);
3321 	if (ret)
3322 		return ret;
3323 
3324 	iw_param.ord = conn_param->initiator_depth;
3325 	iw_param.ird = conn_param->responder_resources;
3326 	iw_param.private_data = conn_param->private_data;
3327 	iw_param.private_data_len = conn_param->private_data_len;
3328 	if (id_priv->id.qp) {
3329 		iw_param.qpn = id_priv->qp_num;
3330 	} else
3331 		iw_param.qpn = conn_param->qp_num;
3332 
3333 	return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3334 }
3335 
3336 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3337 			     enum ib_cm_sidr_status status, u32 qkey,
3338 			     const void *private_data, int private_data_len)
3339 {
3340 	struct ib_cm_sidr_rep_param rep;
3341 	int ret;
3342 
3343 	memset(&rep, 0, sizeof rep);
3344 	rep.status = status;
3345 	if (status == IB_SIDR_SUCCESS) {
3346 		ret = cma_set_qkey(id_priv, qkey);
3347 		if (ret)
3348 			return ret;
3349 		rep.qp_num = id_priv->qp_num;
3350 		rep.qkey = id_priv->qkey;
3351 	}
3352 	rep.private_data = private_data;
3353 	rep.private_data_len = private_data_len;
3354 
3355 	return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3356 }
3357 
3358 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3359 {
3360 	struct rdma_id_private *id_priv;
3361 	int ret;
3362 
3363 	id_priv = container_of(id, struct rdma_id_private, id);
3364 
3365 	id_priv->owner = task_pid_nr(current);
3366 
3367 	if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3368 		return -EINVAL;
3369 
3370 	if (!id->qp && conn_param) {
3371 		id_priv->qp_num = conn_param->qp_num;
3372 		id_priv->srq = conn_param->srq;
3373 	}
3374 
3375 	if (rdma_cap_ib_cm(id->device, id->port_num)) {
3376 		if (id->qp_type == IB_QPT_UD) {
3377 			if (conn_param)
3378 				ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3379 							conn_param->qkey,
3380 							conn_param->private_data,
3381 							conn_param->private_data_len);
3382 			else
3383 				ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3384 							0, NULL, 0);
3385 		} else {
3386 			if (conn_param)
3387 				ret = cma_accept_ib(id_priv, conn_param);
3388 			else
3389 				ret = cma_rep_recv(id_priv);
3390 		}
3391 	} else if (rdma_cap_iw_cm(id->device, id->port_num))
3392 		ret = cma_accept_iw(id_priv, conn_param);
3393 	else
3394 		ret = -ENOSYS;
3395 
3396 	if (ret)
3397 		goto reject;
3398 
3399 	return 0;
3400 reject:
3401 	cma_modify_qp_err(id_priv);
3402 	rdma_reject(id, NULL, 0);
3403 	return ret;
3404 }
3405 EXPORT_SYMBOL(rdma_accept);
3406 
3407 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3408 {
3409 	struct rdma_id_private *id_priv;
3410 	int ret;
3411 
3412 	id_priv = container_of(id, struct rdma_id_private, id);
3413 	if (!id_priv->cm_id.ib)
3414 		return -EINVAL;
3415 
3416 	switch (id->device->node_type) {
3417 	case RDMA_NODE_IB_CA:
3418 		ret = ib_cm_notify(id_priv->cm_id.ib, event);
3419 		break;
3420 	default:
3421 		ret = 0;
3422 		break;
3423 	}
3424 	return ret;
3425 }
3426 EXPORT_SYMBOL(rdma_notify);
3427 
3428 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3429 		u8 private_data_len)
3430 {
3431 	struct rdma_id_private *id_priv;
3432 	int ret;
3433 
3434 	id_priv = container_of(id, struct rdma_id_private, id);
3435 	if (!id_priv->cm_id.ib)
3436 		return -EINVAL;
3437 
3438 	if (rdma_cap_ib_cm(id->device, id->port_num)) {
3439 		if (id->qp_type == IB_QPT_UD)
3440 			ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3441 						private_data, private_data_len);
3442 		else
3443 			ret = ib_send_cm_rej(id_priv->cm_id.ib,
3444 					     IB_CM_REJ_CONSUMER_DEFINED, NULL,
3445 					     0, private_data, private_data_len);
3446 	} else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3447 		ret = iw_cm_reject(id_priv->cm_id.iw,
3448 				   private_data, private_data_len);
3449 	} else
3450 		ret = -ENOSYS;
3451 
3452 	return ret;
3453 }
3454 EXPORT_SYMBOL(rdma_reject);
3455 
3456 int rdma_disconnect(struct rdma_cm_id *id)
3457 {
3458 	struct rdma_id_private *id_priv;
3459 	int ret;
3460 
3461 	id_priv = container_of(id, struct rdma_id_private, id);
3462 	if (!id_priv->cm_id.ib)
3463 		return -EINVAL;
3464 
3465 	if (rdma_cap_ib_cm(id->device, id->port_num)) {
3466 		ret = cma_modify_qp_err(id_priv);
3467 		if (ret)
3468 			goto out;
3469 		/* Initiate or respond to a disconnect. */
3470 		if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3471 			ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3472 	} else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3473 		ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3474 	} else
3475 		ret = -EINVAL;
3476 
3477 out:
3478 	return ret;
3479 }
3480 EXPORT_SYMBOL(rdma_disconnect);
3481 
3482 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3483 {
3484 	struct rdma_id_private *id_priv;
3485 	struct cma_multicast *mc = multicast->context;
3486 	struct rdma_cm_event event;
3487 	int ret;
3488 
3489 	id_priv = mc->id_priv;
3490 	if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
3491 	    cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
3492 		return 0;
3493 
3494 	if (!status)
3495 		status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3496 	mutex_lock(&id_priv->qp_mutex);
3497 	if (!status && id_priv->id.qp)
3498 		status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3499 					 be16_to_cpu(multicast->rec.mlid));
3500 	mutex_unlock(&id_priv->qp_mutex);
3501 
3502 	memset(&event, 0, sizeof event);
3503 	event.status = status;
3504 	event.param.ud.private_data = mc->context;
3505 	if (!status) {
3506 		event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3507 		ib_init_ah_from_mcmember(id_priv->id.device,
3508 					 id_priv->id.port_num, &multicast->rec,
3509 					 &event.param.ud.ah_attr);
3510 		event.param.ud.qp_num = 0xFFFFFF;
3511 		event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3512 	} else
3513 		event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3514 
3515 	ret = id_priv->id.event_handler(&id_priv->id, &event);
3516 	if (ret) {
3517 		cma_exch(id_priv, RDMA_CM_DESTROYING);
3518 		mutex_unlock(&id_priv->handler_mutex);
3519 		rdma_destroy_id(&id_priv->id);
3520 		return 0;
3521 	}
3522 
3523 	mutex_unlock(&id_priv->handler_mutex);
3524 	return 0;
3525 }
3526 
3527 static void cma_set_mgid(struct rdma_id_private *id_priv,
3528 			 struct sockaddr *addr, union ib_gid *mgid)
3529 {
3530 	unsigned char mc_map[MAX_ADDR_LEN];
3531 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3532 	struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3533 	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3534 
3535 	if (cma_any_addr(addr)) {
3536 		memset(mgid, 0, sizeof *mgid);
3537 	} else if ((addr->sa_family == AF_INET6) &&
3538 		   ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3539 								 0xFF10A01B)) {
3540 		/* IPv6 address is an SA assigned MGID. */
3541 		memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3542 	} else if (addr->sa_family == AF_IB) {
3543 		memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3544 	} else if ((addr->sa_family == AF_INET6)) {
3545 		ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3546 		if (id_priv->id.ps == RDMA_PS_UDP)
3547 			mc_map[7] = 0x01;	/* Use RDMA CM signature */
3548 		*mgid = *(union ib_gid *) (mc_map + 4);
3549 	} else {
3550 		ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3551 		if (id_priv->id.ps == RDMA_PS_UDP)
3552 			mc_map[7] = 0x01;	/* Use RDMA CM signature */
3553 		*mgid = *(union ib_gid *) (mc_map + 4);
3554 	}
3555 }
3556 
3557 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3558 				 struct cma_multicast *mc)
3559 {
3560 	struct ib_sa_mcmember_rec rec;
3561 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3562 	ib_sa_comp_mask comp_mask;
3563 	int ret;
3564 
3565 	ib_addr_get_mgid(dev_addr, &rec.mgid);
3566 	ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3567 				     &rec.mgid, &rec);
3568 	if (ret)
3569 		return ret;
3570 
3571 	ret = cma_set_qkey(id_priv, 0);
3572 	if (ret)
3573 		return ret;
3574 
3575 	cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3576 	rec.qkey = cpu_to_be32(id_priv->qkey);
3577 	rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3578 	rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3579 	rec.join_state = 1;
3580 
3581 	comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3582 		    IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3583 		    IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3584 		    IB_SA_MCMEMBER_REC_FLOW_LABEL |
3585 		    IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3586 
3587 	if (id_priv->id.ps == RDMA_PS_IPOIB)
3588 		comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3589 			     IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3590 			     IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3591 			     IB_SA_MCMEMBER_REC_MTU |
3592 			     IB_SA_MCMEMBER_REC_HOP_LIMIT;
3593 
3594 	mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3595 						id_priv->id.port_num, &rec,
3596 						comp_mask, GFP_KERNEL,
3597 						cma_ib_mc_handler, mc);
3598 	return PTR_ERR_OR_ZERO(mc->multicast.ib);
3599 }
3600 
3601 static void iboe_mcast_work_handler(struct work_struct *work)
3602 {
3603 	struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3604 	struct cma_multicast *mc = mw->mc;
3605 	struct ib_sa_multicast *m = mc->multicast.ib;
3606 
3607 	mc->multicast.ib->context = mc;
3608 	cma_ib_mc_handler(0, m);
3609 	kref_put(&mc->mcref, release_mc);
3610 	kfree(mw);
3611 }
3612 
3613 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3614 {
3615 	struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3616 	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3617 
3618 	if (cma_any_addr(addr)) {
3619 		memset(mgid, 0, sizeof *mgid);
3620 	} else if (addr->sa_family == AF_INET6) {
3621 		memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3622 	} else {
3623 		mgid->raw[0] = 0xff;
3624 		mgid->raw[1] = 0x0e;
3625 		mgid->raw[2] = 0;
3626 		mgid->raw[3] = 0;
3627 		mgid->raw[4] = 0;
3628 		mgid->raw[5] = 0;
3629 		mgid->raw[6] = 0;
3630 		mgid->raw[7] = 0;
3631 		mgid->raw[8] = 0;
3632 		mgid->raw[9] = 0;
3633 		mgid->raw[10] = 0xff;
3634 		mgid->raw[11] = 0xff;
3635 		*(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3636 	}
3637 }
3638 
3639 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3640 				   struct cma_multicast *mc)
3641 {
3642 	struct iboe_mcast_work *work;
3643 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3644 	int err;
3645 	struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3646 	struct net_device *ndev = NULL;
3647 
3648 	if (cma_zero_addr((struct sockaddr *)&mc->addr))
3649 		return -EINVAL;
3650 
3651 	work = kzalloc(sizeof *work, GFP_KERNEL);
3652 	if (!work)
3653 		return -ENOMEM;
3654 
3655 	mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3656 	if (!mc->multicast.ib) {
3657 		err = -ENOMEM;
3658 		goto out1;
3659 	}
3660 
3661 	cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3662 
3663 	mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3664 	if (id_priv->id.ps == RDMA_PS_UDP)
3665 		mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3666 
3667 	if (dev_addr->bound_dev_if)
3668 		ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3669 	if (!ndev) {
3670 		err = -ENODEV;
3671 		goto out2;
3672 	}
3673 	mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3674 	mc->multicast.ib->rec.hop_limit = 1;
3675 	mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3676 	dev_put(ndev);
3677 	if (!mc->multicast.ib->rec.mtu) {
3678 		err = -EINVAL;
3679 		goto out2;
3680 	}
3681 	rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3682 		    &mc->multicast.ib->rec.port_gid);
3683 	work->id = id_priv;
3684 	work->mc = mc;
3685 	INIT_WORK(&work->work, iboe_mcast_work_handler);
3686 	kref_get(&mc->mcref);
3687 	queue_work(cma_wq, &work->work);
3688 
3689 	return 0;
3690 
3691 out2:
3692 	kfree(mc->multicast.ib);
3693 out1:
3694 	kfree(work);
3695 	return err;
3696 }
3697 
3698 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3699 			void *context)
3700 {
3701 	struct rdma_id_private *id_priv;
3702 	struct cma_multicast *mc;
3703 	int ret;
3704 
3705 	id_priv = container_of(id, struct rdma_id_private, id);
3706 	if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3707 	    !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3708 		return -EINVAL;
3709 
3710 	mc = kmalloc(sizeof *mc, GFP_KERNEL);
3711 	if (!mc)
3712 		return -ENOMEM;
3713 
3714 	memcpy(&mc->addr, addr, rdma_addr_size(addr));
3715 	mc->context = context;
3716 	mc->id_priv = id_priv;
3717 
3718 	spin_lock(&id_priv->lock);
3719 	list_add(&mc->list, &id_priv->mc_list);
3720 	spin_unlock(&id_priv->lock);
3721 
3722 	if (rdma_protocol_roce(id->device, id->port_num)) {
3723 		kref_init(&mc->mcref);
3724 		ret = cma_iboe_join_multicast(id_priv, mc);
3725 	} else if (rdma_cap_ib_mcast(id->device, id->port_num))
3726 		ret = cma_join_ib_multicast(id_priv, mc);
3727 	else
3728 		ret = -ENOSYS;
3729 
3730 	if (ret) {
3731 		spin_lock_irq(&id_priv->lock);
3732 		list_del(&mc->list);
3733 		spin_unlock_irq(&id_priv->lock);
3734 		kfree(mc);
3735 	}
3736 	return ret;
3737 }
3738 EXPORT_SYMBOL(rdma_join_multicast);
3739 
3740 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3741 {
3742 	struct rdma_id_private *id_priv;
3743 	struct cma_multicast *mc;
3744 
3745 	id_priv = container_of(id, struct rdma_id_private, id);
3746 	spin_lock_irq(&id_priv->lock);
3747 	list_for_each_entry(mc, &id_priv->mc_list, list) {
3748 		if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
3749 			list_del(&mc->list);
3750 			spin_unlock_irq(&id_priv->lock);
3751 
3752 			if (id->qp)
3753 				ib_detach_mcast(id->qp,
3754 						&mc->multicast.ib->rec.mgid,
3755 						be16_to_cpu(mc->multicast.ib->rec.mlid));
3756 
3757 			BUG_ON(id_priv->cma_dev->device != id->device);
3758 
3759 			if (rdma_cap_ib_mcast(id->device, id->port_num)) {
3760 				ib_sa_free_multicast(mc->multicast.ib);
3761 				kfree(mc);
3762 			} else if (rdma_protocol_roce(id->device, id->port_num))
3763 				kref_put(&mc->mcref, release_mc);
3764 
3765 			return;
3766 		}
3767 	}
3768 	spin_unlock_irq(&id_priv->lock);
3769 }
3770 EXPORT_SYMBOL(rdma_leave_multicast);
3771 
3772 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3773 {
3774 	struct rdma_dev_addr *dev_addr;
3775 	struct cma_ndev_work *work;
3776 
3777 	dev_addr = &id_priv->id.route.addr.dev_addr;
3778 
3779 	if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3780 	    memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3781 		printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3782 		       ndev->name, &id_priv->id);
3783 		work = kzalloc(sizeof *work, GFP_KERNEL);
3784 		if (!work)
3785 			return -ENOMEM;
3786 
3787 		INIT_WORK(&work->work, cma_ndev_work_handler);
3788 		work->id = id_priv;
3789 		work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3790 		atomic_inc(&id_priv->refcount);
3791 		queue_work(cma_wq, &work->work);
3792 	}
3793 
3794 	return 0;
3795 }
3796 
3797 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3798 			       void *ptr)
3799 {
3800 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
3801 	struct cma_device *cma_dev;
3802 	struct rdma_id_private *id_priv;
3803 	int ret = NOTIFY_DONE;
3804 
3805 	if (dev_net(ndev) != &init_net)
3806 		return NOTIFY_DONE;
3807 
3808 	if (event != NETDEV_BONDING_FAILOVER)
3809 		return NOTIFY_DONE;
3810 
3811 	if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3812 		return NOTIFY_DONE;
3813 
3814 	mutex_lock(&lock);
3815 	list_for_each_entry(cma_dev, &dev_list, list)
3816 		list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3817 			ret = cma_netdev_change(ndev, id_priv);
3818 			if (ret)
3819 				goto out;
3820 		}
3821 
3822 out:
3823 	mutex_unlock(&lock);
3824 	return ret;
3825 }
3826 
3827 static struct notifier_block cma_nb = {
3828 	.notifier_call = cma_netdev_callback
3829 };
3830 
3831 static void cma_add_one(struct ib_device *device)
3832 {
3833 	struct cma_device *cma_dev;
3834 	struct rdma_id_private *id_priv;
3835 
3836 	cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3837 	if (!cma_dev)
3838 		return;
3839 
3840 	cma_dev->device = device;
3841 
3842 	init_completion(&cma_dev->comp);
3843 	atomic_set(&cma_dev->refcount, 1);
3844 	INIT_LIST_HEAD(&cma_dev->id_list);
3845 	ib_set_client_data(device, &cma_client, cma_dev);
3846 
3847 	mutex_lock(&lock);
3848 	list_add_tail(&cma_dev->list, &dev_list);
3849 	list_for_each_entry(id_priv, &listen_any_list, list)
3850 		cma_listen_on_dev(id_priv, cma_dev);
3851 	mutex_unlock(&lock);
3852 }
3853 
3854 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3855 {
3856 	struct rdma_cm_event event;
3857 	enum rdma_cm_state state;
3858 	int ret = 0;
3859 
3860 	/* Record that we want to remove the device */
3861 	state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3862 	if (state == RDMA_CM_DESTROYING)
3863 		return 0;
3864 
3865 	cma_cancel_operation(id_priv, state);
3866 	mutex_lock(&id_priv->handler_mutex);
3867 
3868 	/* Check for destruction from another callback. */
3869 	if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3870 		goto out;
3871 
3872 	memset(&event, 0, sizeof event);
3873 	event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3874 	ret = id_priv->id.event_handler(&id_priv->id, &event);
3875 out:
3876 	mutex_unlock(&id_priv->handler_mutex);
3877 	return ret;
3878 }
3879 
3880 static void cma_process_remove(struct cma_device *cma_dev)
3881 {
3882 	struct rdma_id_private *id_priv;
3883 	int ret;
3884 
3885 	mutex_lock(&lock);
3886 	while (!list_empty(&cma_dev->id_list)) {
3887 		id_priv = list_entry(cma_dev->id_list.next,
3888 				     struct rdma_id_private, list);
3889 
3890 		list_del(&id_priv->listen_list);
3891 		list_del_init(&id_priv->list);
3892 		atomic_inc(&id_priv->refcount);
3893 		mutex_unlock(&lock);
3894 
3895 		ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3896 		cma_deref_id(id_priv);
3897 		if (ret)
3898 			rdma_destroy_id(&id_priv->id);
3899 
3900 		mutex_lock(&lock);
3901 	}
3902 	mutex_unlock(&lock);
3903 
3904 	cma_deref_dev(cma_dev);
3905 	wait_for_completion(&cma_dev->comp);
3906 }
3907 
3908 static void cma_remove_one(struct ib_device *device, void *client_data)
3909 {
3910 	struct cma_device *cma_dev = client_data;
3911 
3912 	if (!cma_dev)
3913 		return;
3914 
3915 	mutex_lock(&lock);
3916 	list_del(&cma_dev->list);
3917 	mutex_unlock(&lock);
3918 
3919 	cma_process_remove(cma_dev);
3920 	kfree(cma_dev);
3921 }
3922 
3923 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3924 {
3925 	struct nlmsghdr *nlh;
3926 	struct rdma_cm_id_stats *id_stats;
3927 	struct rdma_id_private *id_priv;
3928 	struct rdma_cm_id *id = NULL;
3929 	struct cma_device *cma_dev;
3930 	int i_dev = 0, i_id = 0;
3931 
3932 	/*
3933 	 * We export all of the IDs as a sequence of messages.  Each
3934 	 * ID gets its own netlink message.
3935 	 */
3936 	mutex_lock(&lock);
3937 
3938 	list_for_each_entry(cma_dev, &dev_list, list) {
3939 		if (i_dev < cb->args[0]) {
3940 			i_dev++;
3941 			continue;
3942 		}
3943 
3944 		i_id = 0;
3945 		list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3946 			if (i_id < cb->args[1]) {
3947 				i_id++;
3948 				continue;
3949 			}
3950 
3951 			id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3952 						sizeof *id_stats, RDMA_NL_RDMA_CM,
3953 						RDMA_NL_RDMA_CM_ID_STATS,
3954 						NLM_F_MULTI);
3955 			if (!id_stats)
3956 				goto out;
3957 
3958 			memset(id_stats, 0, sizeof *id_stats);
3959 			id = &id_priv->id;
3960 			id_stats->node_type = id->route.addr.dev_addr.dev_type;
3961 			id_stats->port_num = id->port_num;
3962 			id_stats->bound_dev_if =
3963 				id->route.addr.dev_addr.bound_dev_if;
3964 
3965 			if (ibnl_put_attr(skb, nlh,
3966 					  rdma_addr_size(cma_src_addr(id_priv)),
3967 					  cma_src_addr(id_priv),
3968 					  RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
3969 				goto out;
3970 			if (ibnl_put_attr(skb, nlh,
3971 					  rdma_addr_size(cma_src_addr(id_priv)),
3972 					  cma_dst_addr(id_priv),
3973 					  RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
3974 				goto out;
3975 
3976 			id_stats->pid		= id_priv->owner;
3977 			id_stats->port_space	= id->ps;
3978 			id_stats->cm_state	= id_priv->state;
3979 			id_stats->qp_num	= id_priv->qp_num;
3980 			id_stats->qp_type	= id->qp_type;
3981 
3982 			i_id++;
3983 		}
3984 
3985 		cb->args[1] = 0;
3986 		i_dev++;
3987 	}
3988 
3989 out:
3990 	mutex_unlock(&lock);
3991 	cb->args[0] = i_dev;
3992 	cb->args[1] = i_id;
3993 
3994 	return skb->len;
3995 }
3996 
3997 static const struct ibnl_client_cbs cma_cb_table[] = {
3998 	[RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
3999 				       .module = THIS_MODULE },
4000 };
4001 
4002 static int __init cma_init(void)
4003 {
4004 	int ret;
4005 
4006 	cma_wq = create_singlethread_workqueue("rdma_cm");
4007 	if (!cma_wq)
4008 		return -ENOMEM;
4009 
4010 	ib_sa_register_client(&sa_client);
4011 	rdma_addr_register_client(&addr_client);
4012 	register_netdevice_notifier(&cma_nb);
4013 
4014 	ret = ib_register_client(&cma_client);
4015 	if (ret)
4016 		goto err;
4017 
4018 	if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
4019 		printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
4020 
4021 	return 0;
4022 
4023 err:
4024 	unregister_netdevice_notifier(&cma_nb);
4025 	rdma_addr_unregister_client(&addr_client);
4026 	ib_sa_unregister_client(&sa_client);
4027 	destroy_workqueue(cma_wq);
4028 	return ret;
4029 }
4030 
4031 static void __exit cma_cleanup(void)
4032 {
4033 	ibnl_remove_client(RDMA_NL_RDMA_CM);
4034 	ib_unregister_client(&cma_client);
4035 	unregister_netdevice_notifier(&cma_nb);
4036 	rdma_addr_unregister_client(&addr_client);
4037 	ib_sa_unregister_client(&sa_client);
4038 	destroy_workqueue(cma_wq);
4039 	idr_destroy(&tcp_ps);
4040 	idr_destroy(&udp_ps);
4041 	idr_destroy(&ipoib_ps);
4042 	idr_destroy(&ib_ps);
4043 }
4044 
4045 module_init(cma_init);
4046 module_exit(cma_cleanup);
4047