xref: /openbmc/linux/drivers/infiniband/core/cma.c (revision 8365a898)
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
4  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
5  * Copyright (c) 1999-2019, Mellanox Technologies, Inc. All rights reserved.
6  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
7  */
8 
9 #include <linux/completion.h>
10 #include <linux/in.h>
11 #include <linux/in6.h>
12 #include <linux/mutex.h>
13 #include <linux/random.h>
14 #include <linux/igmp.h>
15 #include <linux/xarray.h>
16 #include <linux/inetdevice.h>
17 #include <linux/slab.h>
18 #include <linux/module.h>
19 #include <net/route.h>
20 
21 #include <net/net_namespace.h>
22 #include <net/netns/generic.h>
23 #include <net/tcp.h>
24 #include <net/ipv6.h>
25 #include <net/ip_fib.h>
26 #include <net/ip6_route.h>
27 
28 #include <rdma/rdma_cm.h>
29 #include <rdma/rdma_cm_ib.h>
30 #include <rdma/rdma_netlink.h>
31 #include <rdma/ib.h>
32 #include <rdma/ib_cache.h>
33 #include <rdma/ib_cm.h>
34 #include <rdma/ib_sa.h>
35 #include <rdma/iw_cm.h>
36 
37 #include "core_priv.h"
38 #include "cma_priv.h"
39 #include "cma_trace.h"
40 
41 MODULE_AUTHOR("Sean Hefty");
42 MODULE_DESCRIPTION("Generic RDMA CM Agent");
43 MODULE_LICENSE("Dual BSD/GPL");
44 
45 #define CMA_CM_RESPONSE_TIMEOUT 20
46 #define CMA_QUERY_CLASSPORT_INFO_TIMEOUT 3000
47 #define CMA_MAX_CM_RETRIES 15
48 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
49 #define CMA_IBOE_PACKET_LIFETIME 18
50 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
51 
52 static const char * const cma_events[] = {
53 	[RDMA_CM_EVENT_ADDR_RESOLVED]	 = "address resolved",
54 	[RDMA_CM_EVENT_ADDR_ERROR]	 = "address error",
55 	[RDMA_CM_EVENT_ROUTE_RESOLVED]	 = "route resolved ",
56 	[RDMA_CM_EVENT_ROUTE_ERROR]	 = "route error",
57 	[RDMA_CM_EVENT_CONNECT_REQUEST]	 = "connect request",
58 	[RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
59 	[RDMA_CM_EVENT_CONNECT_ERROR]	 = "connect error",
60 	[RDMA_CM_EVENT_UNREACHABLE]	 = "unreachable",
61 	[RDMA_CM_EVENT_REJECTED]	 = "rejected",
62 	[RDMA_CM_EVENT_ESTABLISHED]	 = "established",
63 	[RDMA_CM_EVENT_DISCONNECTED]	 = "disconnected",
64 	[RDMA_CM_EVENT_DEVICE_REMOVAL]	 = "device removal",
65 	[RDMA_CM_EVENT_MULTICAST_JOIN]	 = "multicast join",
66 	[RDMA_CM_EVENT_MULTICAST_ERROR]	 = "multicast error",
67 	[RDMA_CM_EVENT_ADDR_CHANGE]	 = "address change",
68 	[RDMA_CM_EVENT_TIMEWAIT_EXIT]	 = "timewait exit",
69 };
70 
71 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
72 {
73 	size_t index = event;
74 
75 	return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
76 			cma_events[index] : "unrecognized event";
77 }
78 EXPORT_SYMBOL(rdma_event_msg);
79 
80 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
81 						int reason)
82 {
83 	if (rdma_ib_or_roce(id->device, id->port_num))
84 		return ibcm_reject_msg(reason);
85 
86 	if (rdma_protocol_iwarp(id->device, id->port_num))
87 		return iwcm_reject_msg(reason);
88 
89 	WARN_ON_ONCE(1);
90 	return "unrecognized transport";
91 }
92 EXPORT_SYMBOL(rdma_reject_msg);
93 
94 /**
95  * rdma_is_consumer_reject - return true if the consumer rejected the connect
96  *                           request.
97  * @id: Communication identifier that received the REJECT event.
98  * @reason: Value returned in the REJECT event status field.
99  */
100 static bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
101 {
102 	if (rdma_ib_or_roce(id->device, id->port_num))
103 		return reason == IB_CM_REJ_CONSUMER_DEFINED;
104 
105 	if (rdma_protocol_iwarp(id->device, id->port_num))
106 		return reason == -ECONNREFUSED;
107 
108 	WARN_ON_ONCE(1);
109 	return false;
110 }
111 
112 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
113 				      struct rdma_cm_event *ev, u8 *data_len)
114 {
115 	const void *p;
116 
117 	if (rdma_is_consumer_reject(id, ev->status)) {
118 		*data_len = ev->param.conn.private_data_len;
119 		p = ev->param.conn.private_data;
120 	} else {
121 		*data_len = 0;
122 		p = NULL;
123 	}
124 	return p;
125 }
126 EXPORT_SYMBOL(rdma_consumer_reject_data);
127 
128 /**
129  * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
130  * @id: Communication Identifier
131  */
132 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
133 {
134 	struct rdma_id_private *id_priv;
135 
136 	id_priv = container_of(id, struct rdma_id_private, id);
137 	if (id->device->node_type == RDMA_NODE_RNIC)
138 		return id_priv->cm_id.iw;
139 	return NULL;
140 }
141 EXPORT_SYMBOL(rdma_iw_cm_id);
142 
143 /**
144  * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
145  * @res: rdma resource tracking entry pointer
146  */
147 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
148 {
149 	struct rdma_id_private *id_priv =
150 		container_of(res, struct rdma_id_private, res);
151 
152 	return &id_priv->id;
153 }
154 EXPORT_SYMBOL(rdma_res_to_id);
155 
156 static int cma_add_one(struct ib_device *device);
157 static void cma_remove_one(struct ib_device *device, void *client_data);
158 
159 static struct ib_client cma_client = {
160 	.name   = "cma",
161 	.add    = cma_add_one,
162 	.remove = cma_remove_one
163 };
164 
165 static struct ib_sa_client sa_client;
166 static LIST_HEAD(dev_list);
167 static LIST_HEAD(listen_any_list);
168 static DEFINE_MUTEX(lock);
169 static struct workqueue_struct *cma_wq;
170 static unsigned int cma_pernet_id;
171 
172 struct cma_pernet {
173 	struct xarray tcp_ps;
174 	struct xarray udp_ps;
175 	struct xarray ipoib_ps;
176 	struct xarray ib_ps;
177 };
178 
179 static struct cma_pernet *cma_pernet(struct net *net)
180 {
181 	return net_generic(net, cma_pernet_id);
182 }
183 
184 static
185 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
186 {
187 	struct cma_pernet *pernet = cma_pernet(net);
188 
189 	switch (ps) {
190 	case RDMA_PS_TCP:
191 		return &pernet->tcp_ps;
192 	case RDMA_PS_UDP:
193 		return &pernet->udp_ps;
194 	case RDMA_PS_IPOIB:
195 		return &pernet->ipoib_ps;
196 	case RDMA_PS_IB:
197 		return &pernet->ib_ps;
198 	default:
199 		return NULL;
200 	}
201 }
202 
203 struct cma_device {
204 	struct list_head	list;
205 	struct ib_device	*device;
206 	struct completion	comp;
207 	refcount_t refcount;
208 	struct list_head	id_list;
209 	enum ib_gid_type	*default_gid_type;
210 	u8			*default_roce_tos;
211 };
212 
213 struct rdma_bind_list {
214 	enum rdma_ucm_port_space ps;
215 	struct hlist_head	owners;
216 	unsigned short		port;
217 };
218 
219 struct class_port_info_context {
220 	struct ib_class_port_info	*class_port_info;
221 	struct ib_device		*device;
222 	struct completion		done;
223 	struct ib_sa_query		*sa_query;
224 	u8				port_num;
225 };
226 
227 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
228 			struct rdma_bind_list *bind_list, int snum)
229 {
230 	struct xarray *xa = cma_pernet_xa(net, ps);
231 
232 	return xa_insert(xa, snum, bind_list, GFP_KERNEL);
233 }
234 
235 static struct rdma_bind_list *cma_ps_find(struct net *net,
236 					  enum rdma_ucm_port_space ps, int snum)
237 {
238 	struct xarray *xa = cma_pernet_xa(net, ps);
239 
240 	return xa_load(xa, snum);
241 }
242 
243 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
244 			  int snum)
245 {
246 	struct xarray *xa = cma_pernet_xa(net, ps);
247 
248 	xa_erase(xa, snum);
249 }
250 
251 enum {
252 	CMA_OPTION_AFONLY,
253 };
254 
255 void cma_dev_get(struct cma_device *cma_dev)
256 {
257 	refcount_inc(&cma_dev->refcount);
258 }
259 
260 void cma_dev_put(struct cma_device *cma_dev)
261 {
262 	if (refcount_dec_and_test(&cma_dev->refcount))
263 		complete(&cma_dev->comp);
264 }
265 
266 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter	filter,
267 					     void		*cookie)
268 {
269 	struct cma_device *cma_dev;
270 	struct cma_device *found_cma_dev = NULL;
271 
272 	mutex_lock(&lock);
273 
274 	list_for_each_entry(cma_dev, &dev_list, list)
275 		if (filter(cma_dev->device, cookie)) {
276 			found_cma_dev = cma_dev;
277 			break;
278 		}
279 
280 	if (found_cma_dev)
281 		cma_dev_get(found_cma_dev);
282 	mutex_unlock(&lock);
283 	return found_cma_dev;
284 }
285 
286 int cma_get_default_gid_type(struct cma_device *cma_dev,
287 			     unsigned int port)
288 {
289 	if (!rdma_is_port_valid(cma_dev->device, port))
290 		return -EINVAL;
291 
292 	return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
293 }
294 
295 int cma_set_default_gid_type(struct cma_device *cma_dev,
296 			     unsigned int port,
297 			     enum ib_gid_type default_gid_type)
298 {
299 	unsigned long supported_gids;
300 
301 	if (!rdma_is_port_valid(cma_dev->device, port))
302 		return -EINVAL;
303 
304 	supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
305 
306 	if (!(supported_gids & 1 << default_gid_type))
307 		return -EINVAL;
308 
309 	cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
310 		default_gid_type;
311 
312 	return 0;
313 }
314 
315 int cma_get_default_roce_tos(struct cma_device *cma_dev, unsigned int port)
316 {
317 	if (!rdma_is_port_valid(cma_dev->device, port))
318 		return -EINVAL;
319 
320 	return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
321 }
322 
323 int cma_set_default_roce_tos(struct cma_device *cma_dev, unsigned int port,
324 			     u8 default_roce_tos)
325 {
326 	if (!rdma_is_port_valid(cma_dev->device, port))
327 		return -EINVAL;
328 
329 	cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
330 		 default_roce_tos;
331 
332 	return 0;
333 }
334 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
335 {
336 	return cma_dev->device;
337 }
338 
339 /*
340  * Device removal can occur at anytime, so we need extra handling to
341  * serialize notifying the user of device removal with other callbacks.
342  * We do this by disabling removal notification while a callback is in process,
343  * and reporting it after the callback completes.
344  */
345 
346 struct cma_multicast {
347 	struct rdma_id_private *id_priv;
348 	union {
349 		struct ib_sa_multicast *ib;
350 	} multicast;
351 	struct list_head	list;
352 	void			*context;
353 	struct sockaddr_storage	addr;
354 	struct kref		mcref;
355 	u8			join_state;
356 };
357 
358 struct cma_work {
359 	struct work_struct	work;
360 	struct rdma_id_private	*id;
361 	enum rdma_cm_state	old_state;
362 	enum rdma_cm_state	new_state;
363 	struct rdma_cm_event	event;
364 };
365 
366 struct cma_ndev_work {
367 	struct work_struct	work;
368 	struct rdma_id_private	*id;
369 	struct rdma_cm_event	event;
370 };
371 
372 struct iboe_mcast_work {
373 	struct work_struct	 work;
374 	struct rdma_id_private	*id;
375 	struct cma_multicast	*mc;
376 };
377 
378 union cma_ip_addr {
379 	struct in6_addr ip6;
380 	struct {
381 		__be32 pad[3];
382 		__be32 addr;
383 	} ip4;
384 };
385 
386 struct cma_hdr {
387 	u8 cma_version;
388 	u8 ip_version;	/* IP version: 7:4 */
389 	__be16 port;
390 	union cma_ip_addr src_addr;
391 	union cma_ip_addr dst_addr;
392 };
393 
394 #define CMA_VERSION 0x00
395 
396 struct cma_req_info {
397 	struct sockaddr_storage listen_addr_storage;
398 	struct sockaddr_storage src_addr_storage;
399 	struct ib_device *device;
400 	union ib_gid local_gid;
401 	__be64 service_id;
402 	int port;
403 	bool has_gid;
404 	u16 pkey;
405 };
406 
407 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
408 {
409 	unsigned long flags;
410 	int ret;
411 
412 	spin_lock_irqsave(&id_priv->lock, flags);
413 	ret = (id_priv->state == comp);
414 	spin_unlock_irqrestore(&id_priv->lock, flags);
415 	return ret;
416 }
417 
418 static int cma_comp_exch(struct rdma_id_private *id_priv,
419 			 enum rdma_cm_state comp, enum rdma_cm_state exch)
420 {
421 	unsigned long flags;
422 	int ret;
423 
424 	spin_lock_irqsave(&id_priv->lock, flags);
425 	if ((ret = (id_priv->state == comp)))
426 		id_priv->state = exch;
427 	spin_unlock_irqrestore(&id_priv->lock, flags);
428 	return ret;
429 }
430 
431 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
432 				   enum rdma_cm_state exch)
433 {
434 	unsigned long flags;
435 	enum rdma_cm_state old;
436 
437 	spin_lock_irqsave(&id_priv->lock, flags);
438 	old = id_priv->state;
439 	id_priv->state = exch;
440 	spin_unlock_irqrestore(&id_priv->lock, flags);
441 	return old;
442 }
443 
444 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
445 {
446 	return hdr->ip_version >> 4;
447 }
448 
449 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
450 {
451 	hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
452 }
453 
454 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
455 {
456 	struct in_device *in_dev = NULL;
457 
458 	if (ndev) {
459 		rtnl_lock();
460 		in_dev = __in_dev_get_rtnl(ndev);
461 		if (in_dev) {
462 			if (join)
463 				ip_mc_inc_group(in_dev,
464 						*(__be32 *)(mgid->raw + 12));
465 			else
466 				ip_mc_dec_group(in_dev,
467 						*(__be32 *)(mgid->raw + 12));
468 		}
469 		rtnl_unlock();
470 	}
471 	return (in_dev) ? 0 : -ENODEV;
472 }
473 
474 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
475 			       struct cma_device *cma_dev)
476 {
477 	cma_dev_get(cma_dev);
478 	id_priv->cma_dev = cma_dev;
479 	id_priv->id.device = cma_dev->device;
480 	id_priv->id.route.addr.dev_addr.transport =
481 		rdma_node_get_transport(cma_dev->device->node_type);
482 	list_add_tail(&id_priv->list, &cma_dev->id_list);
483 	if (id_priv->res.kern_name)
484 		rdma_restrack_kadd(&id_priv->res);
485 	else
486 		rdma_restrack_uadd(&id_priv->res);
487 	trace_cm_id_attach(id_priv, cma_dev->device);
488 }
489 
490 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
491 			      struct cma_device *cma_dev)
492 {
493 	_cma_attach_to_dev(id_priv, cma_dev);
494 	id_priv->gid_type =
495 		cma_dev->default_gid_type[id_priv->id.port_num -
496 					  rdma_start_port(cma_dev->device)];
497 }
498 
499 static inline void release_mc(struct kref *kref)
500 {
501 	struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
502 
503 	kfree(mc->multicast.ib);
504 	kfree(mc);
505 }
506 
507 static void cma_release_dev(struct rdma_id_private *id_priv)
508 {
509 	mutex_lock(&lock);
510 	list_del(&id_priv->list);
511 	cma_dev_put(id_priv->cma_dev);
512 	id_priv->cma_dev = NULL;
513 	mutex_unlock(&lock);
514 }
515 
516 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
517 {
518 	return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
519 }
520 
521 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
522 {
523 	return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
524 }
525 
526 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
527 {
528 	return id_priv->id.route.addr.src_addr.ss_family;
529 }
530 
531 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
532 {
533 	struct ib_sa_mcmember_rec rec;
534 	int ret = 0;
535 
536 	if (id_priv->qkey) {
537 		if (qkey && id_priv->qkey != qkey)
538 			return -EINVAL;
539 		return 0;
540 	}
541 
542 	if (qkey) {
543 		id_priv->qkey = qkey;
544 		return 0;
545 	}
546 
547 	switch (id_priv->id.ps) {
548 	case RDMA_PS_UDP:
549 	case RDMA_PS_IB:
550 		id_priv->qkey = RDMA_UDP_QKEY;
551 		break;
552 	case RDMA_PS_IPOIB:
553 		ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
554 		ret = ib_sa_get_mcmember_rec(id_priv->id.device,
555 					     id_priv->id.port_num, &rec.mgid,
556 					     &rec);
557 		if (!ret)
558 			id_priv->qkey = be32_to_cpu(rec.qkey);
559 		break;
560 	default:
561 		break;
562 	}
563 	return ret;
564 }
565 
566 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
567 {
568 	dev_addr->dev_type = ARPHRD_INFINIBAND;
569 	rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
570 	ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
571 }
572 
573 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
574 {
575 	int ret;
576 
577 	if (addr->sa_family != AF_IB) {
578 		ret = rdma_translate_ip(addr, dev_addr);
579 	} else {
580 		cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
581 		ret = 0;
582 	}
583 
584 	return ret;
585 }
586 
587 static const struct ib_gid_attr *
588 cma_validate_port(struct ib_device *device, u8 port,
589 		  enum ib_gid_type gid_type,
590 		  union ib_gid *gid,
591 		  struct rdma_id_private *id_priv)
592 {
593 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
594 	int bound_if_index = dev_addr->bound_dev_if;
595 	const struct ib_gid_attr *sgid_attr;
596 	int dev_type = dev_addr->dev_type;
597 	struct net_device *ndev = NULL;
598 
599 	if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
600 		return ERR_PTR(-ENODEV);
601 
602 	if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
603 		return ERR_PTR(-ENODEV);
604 
605 	if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
606 		return ERR_PTR(-ENODEV);
607 
608 	if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
609 		ndev = dev_get_by_index(dev_addr->net, bound_if_index);
610 		if (!ndev)
611 			return ERR_PTR(-ENODEV);
612 	} else {
613 		gid_type = IB_GID_TYPE_IB;
614 	}
615 
616 	sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
617 	if (ndev)
618 		dev_put(ndev);
619 	return sgid_attr;
620 }
621 
622 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
623 			       const struct ib_gid_attr *sgid_attr)
624 {
625 	WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
626 	id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
627 }
628 
629 /**
630  * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
631  * based on source ip address.
632  * @id_priv:	cm_id which should be bound to cma device
633  *
634  * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
635  * based on source IP address. It returns 0 on success or error code otherwise.
636  * It is applicable to active and passive side cm_id.
637  */
638 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
639 {
640 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
641 	const struct ib_gid_attr *sgid_attr;
642 	union ib_gid gid, iboe_gid, *gidp;
643 	struct cma_device *cma_dev;
644 	enum ib_gid_type gid_type;
645 	int ret = -ENODEV;
646 	unsigned int port;
647 
648 	if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
649 	    id_priv->id.ps == RDMA_PS_IPOIB)
650 		return -EINVAL;
651 
652 	rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
653 		    &iboe_gid);
654 
655 	memcpy(&gid, dev_addr->src_dev_addr +
656 	       rdma_addr_gid_offset(dev_addr), sizeof(gid));
657 
658 	mutex_lock(&lock);
659 	list_for_each_entry(cma_dev, &dev_list, list) {
660 		rdma_for_each_port (cma_dev->device, port) {
661 			gidp = rdma_protocol_roce(cma_dev->device, port) ?
662 			       &iboe_gid : &gid;
663 			gid_type = cma_dev->default_gid_type[port - 1];
664 			sgid_attr = cma_validate_port(cma_dev->device, port,
665 						      gid_type, gidp, id_priv);
666 			if (!IS_ERR(sgid_attr)) {
667 				id_priv->id.port_num = port;
668 				cma_bind_sgid_attr(id_priv, sgid_attr);
669 				cma_attach_to_dev(id_priv, cma_dev);
670 				ret = 0;
671 				goto out;
672 			}
673 		}
674 	}
675 out:
676 	mutex_unlock(&lock);
677 	return ret;
678 }
679 
680 /**
681  * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
682  * @id_priv:		cm id to bind to cma device
683  * @listen_id_priv:	listener cm id to match against
684  * @req:		Pointer to req structure containaining incoming
685  *			request information
686  * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
687  * rdma device matches for listen_id and incoming request. It also verifies
688  * that a GID table entry is present for the source address.
689  * Returns 0 on success, or returns error code otherwise.
690  */
691 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
692 			      const struct rdma_id_private *listen_id_priv,
693 			      struct cma_req_info *req)
694 {
695 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
696 	const struct ib_gid_attr *sgid_attr;
697 	enum ib_gid_type gid_type;
698 	union ib_gid gid;
699 
700 	if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
701 	    id_priv->id.ps == RDMA_PS_IPOIB)
702 		return -EINVAL;
703 
704 	if (rdma_protocol_roce(req->device, req->port))
705 		rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
706 			    &gid);
707 	else
708 		memcpy(&gid, dev_addr->src_dev_addr +
709 		       rdma_addr_gid_offset(dev_addr), sizeof(gid));
710 
711 	gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
712 	sgid_attr = cma_validate_port(req->device, req->port,
713 				      gid_type, &gid, id_priv);
714 	if (IS_ERR(sgid_attr))
715 		return PTR_ERR(sgid_attr);
716 
717 	id_priv->id.port_num = req->port;
718 	cma_bind_sgid_attr(id_priv, sgid_attr);
719 	/* Need to acquire lock to protect against reader
720 	 * of cma_dev->id_list such as cma_netdev_callback() and
721 	 * cma_process_remove().
722 	 */
723 	mutex_lock(&lock);
724 	cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
725 	mutex_unlock(&lock);
726 	return 0;
727 }
728 
729 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
730 			      const struct rdma_id_private *listen_id_priv)
731 {
732 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
733 	const struct ib_gid_attr *sgid_attr;
734 	struct cma_device *cma_dev;
735 	enum ib_gid_type gid_type;
736 	int ret = -ENODEV;
737 	unsigned int port;
738 	union ib_gid gid;
739 
740 	if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
741 	    id_priv->id.ps == RDMA_PS_IPOIB)
742 		return -EINVAL;
743 
744 	memcpy(&gid, dev_addr->src_dev_addr +
745 	       rdma_addr_gid_offset(dev_addr), sizeof(gid));
746 
747 	mutex_lock(&lock);
748 
749 	cma_dev = listen_id_priv->cma_dev;
750 	port = listen_id_priv->id.port_num;
751 	gid_type = listen_id_priv->gid_type;
752 	sgid_attr = cma_validate_port(cma_dev->device, port,
753 				      gid_type, &gid, id_priv);
754 	if (!IS_ERR(sgid_attr)) {
755 		id_priv->id.port_num = port;
756 		cma_bind_sgid_attr(id_priv, sgid_attr);
757 		ret = 0;
758 		goto out;
759 	}
760 
761 	list_for_each_entry(cma_dev, &dev_list, list) {
762 		rdma_for_each_port (cma_dev->device, port) {
763 			if (listen_id_priv->cma_dev == cma_dev &&
764 			    listen_id_priv->id.port_num == port)
765 				continue;
766 
767 			gid_type = cma_dev->default_gid_type[port - 1];
768 			sgid_attr = cma_validate_port(cma_dev->device, port,
769 						      gid_type, &gid, id_priv);
770 			if (!IS_ERR(sgid_attr)) {
771 				id_priv->id.port_num = port;
772 				cma_bind_sgid_attr(id_priv, sgid_attr);
773 				ret = 0;
774 				goto out;
775 			}
776 		}
777 	}
778 
779 out:
780 	if (!ret)
781 		cma_attach_to_dev(id_priv, cma_dev);
782 
783 	mutex_unlock(&lock);
784 	return ret;
785 }
786 
787 /*
788  * Select the source IB device and address to reach the destination IB address.
789  */
790 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
791 {
792 	struct cma_device *cma_dev, *cur_dev;
793 	struct sockaddr_ib *addr;
794 	union ib_gid gid, sgid, *dgid;
795 	unsigned int p;
796 	u16 pkey, index;
797 	enum ib_port_state port_state;
798 	int i;
799 
800 	cma_dev = NULL;
801 	addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
802 	dgid = (union ib_gid *) &addr->sib_addr;
803 	pkey = ntohs(addr->sib_pkey);
804 
805 	mutex_lock(&lock);
806 	list_for_each_entry(cur_dev, &dev_list, list) {
807 		rdma_for_each_port (cur_dev->device, p) {
808 			if (!rdma_cap_af_ib(cur_dev->device, p))
809 				continue;
810 
811 			if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
812 				continue;
813 
814 			if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
815 				continue;
816 			for (i = 0; !rdma_query_gid(cur_dev->device,
817 						    p, i, &gid);
818 			     i++) {
819 				if (!memcmp(&gid, dgid, sizeof(gid))) {
820 					cma_dev = cur_dev;
821 					sgid = gid;
822 					id_priv->id.port_num = p;
823 					goto found;
824 				}
825 
826 				if (!cma_dev && (gid.global.subnet_prefix ==
827 				    dgid->global.subnet_prefix) &&
828 				    port_state == IB_PORT_ACTIVE) {
829 					cma_dev = cur_dev;
830 					sgid = gid;
831 					id_priv->id.port_num = p;
832 					goto found;
833 				}
834 			}
835 		}
836 	}
837 	mutex_unlock(&lock);
838 	return -ENODEV;
839 
840 found:
841 	cma_attach_to_dev(id_priv, cma_dev);
842 	mutex_unlock(&lock);
843 	addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
844 	memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
845 	cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
846 	return 0;
847 }
848 
849 static void cma_id_get(struct rdma_id_private *id_priv)
850 {
851 	refcount_inc(&id_priv->refcount);
852 }
853 
854 static void cma_id_put(struct rdma_id_private *id_priv)
855 {
856 	if (refcount_dec_and_test(&id_priv->refcount))
857 		complete(&id_priv->comp);
858 }
859 
860 struct rdma_cm_id *__rdma_create_id(struct net *net,
861 				    rdma_cm_event_handler event_handler,
862 				    void *context, enum rdma_ucm_port_space ps,
863 				    enum ib_qp_type qp_type, const char *caller)
864 {
865 	struct rdma_id_private *id_priv;
866 
867 	id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
868 	if (!id_priv)
869 		return ERR_PTR(-ENOMEM);
870 
871 	rdma_restrack_set_task(&id_priv->res, caller);
872 	id_priv->res.type = RDMA_RESTRACK_CM_ID;
873 	id_priv->state = RDMA_CM_IDLE;
874 	id_priv->id.context = context;
875 	id_priv->id.event_handler = event_handler;
876 	id_priv->id.ps = ps;
877 	id_priv->id.qp_type = qp_type;
878 	id_priv->tos_set = false;
879 	id_priv->timeout_set = false;
880 	id_priv->gid_type = IB_GID_TYPE_IB;
881 	spin_lock_init(&id_priv->lock);
882 	mutex_init(&id_priv->qp_mutex);
883 	init_completion(&id_priv->comp);
884 	refcount_set(&id_priv->refcount, 1);
885 	mutex_init(&id_priv->handler_mutex);
886 	INIT_LIST_HEAD(&id_priv->listen_list);
887 	INIT_LIST_HEAD(&id_priv->mc_list);
888 	get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
889 	id_priv->id.route.addr.dev_addr.net = get_net(net);
890 	id_priv->seq_num &= 0x00ffffff;
891 
892 	return &id_priv->id;
893 }
894 EXPORT_SYMBOL(__rdma_create_id);
895 
896 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
897 {
898 	struct ib_qp_attr qp_attr;
899 	int qp_attr_mask, ret;
900 
901 	qp_attr.qp_state = IB_QPS_INIT;
902 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
903 	if (ret)
904 		return ret;
905 
906 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
907 	if (ret)
908 		return ret;
909 
910 	qp_attr.qp_state = IB_QPS_RTR;
911 	ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
912 	if (ret)
913 		return ret;
914 
915 	qp_attr.qp_state = IB_QPS_RTS;
916 	qp_attr.sq_psn = 0;
917 	ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
918 
919 	return ret;
920 }
921 
922 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
923 {
924 	struct ib_qp_attr qp_attr;
925 	int qp_attr_mask, ret;
926 
927 	qp_attr.qp_state = IB_QPS_INIT;
928 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
929 	if (ret)
930 		return ret;
931 
932 	return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
933 }
934 
935 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
936 		   struct ib_qp_init_attr *qp_init_attr)
937 {
938 	struct rdma_id_private *id_priv;
939 	struct ib_qp *qp;
940 	int ret;
941 
942 	id_priv = container_of(id, struct rdma_id_private, id);
943 	if (id->device != pd->device) {
944 		ret = -EINVAL;
945 		goto out_err;
946 	}
947 
948 	qp_init_attr->port_num = id->port_num;
949 	qp = ib_create_qp(pd, qp_init_attr);
950 	if (IS_ERR(qp)) {
951 		ret = PTR_ERR(qp);
952 		goto out_err;
953 	}
954 
955 	if (id->qp_type == IB_QPT_UD)
956 		ret = cma_init_ud_qp(id_priv, qp);
957 	else
958 		ret = cma_init_conn_qp(id_priv, qp);
959 	if (ret)
960 		goto out_destroy;
961 
962 	id->qp = qp;
963 	id_priv->qp_num = qp->qp_num;
964 	id_priv->srq = (qp->srq != NULL);
965 	trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
966 	return 0;
967 out_destroy:
968 	ib_destroy_qp(qp);
969 out_err:
970 	trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
971 	return ret;
972 }
973 EXPORT_SYMBOL(rdma_create_qp);
974 
975 void rdma_destroy_qp(struct rdma_cm_id *id)
976 {
977 	struct rdma_id_private *id_priv;
978 
979 	id_priv = container_of(id, struct rdma_id_private, id);
980 	trace_cm_qp_destroy(id_priv);
981 	mutex_lock(&id_priv->qp_mutex);
982 	ib_destroy_qp(id_priv->id.qp);
983 	id_priv->id.qp = NULL;
984 	mutex_unlock(&id_priv->qp_mutex);
985 }
986 EXPORT_SYMBOL(rdma_destroy_qp);
987 
988 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
989 			     struct rdma_conn_param *conn_param)
990 {
991 	struct ib_qp_attr qp_attr;
992 	int qp_attr_mask, ret;
993 
994 	mutex_lock(&id_priv->qp_mutex);
995 	if (!id_priv->id.qp) {
996 		ret = 0;
997 		goto out;
998 	}
999 
1000 	/* Need to update QP attributes from default values. */
1001 	qp_attr.qp_state = IB_QPS_INIT;
1002 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1003 	if (ret)
1004 		goto out;
1005 
1006 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1007 	if (ret)
1008 		goto out;
1009 
1010 	qp_attr.qp_state = IB_QPS_RTR;
1011 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1012 	if (ret)
1013 		goto out;
1014 
1015 	BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1016 
1017 	if (conn_param)
1018 		qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1019 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1020 out:
1021 	mutex_unlock(&id_priv->qp_mutex);
1022 	return ret;
1023 }
1024 
1025 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1026 			     struct rdma_conn_param *conn_param)
1027 {
1028 	struct ib_qp_attr qp_attr;
1029 	int qp_attr_mask, ret;
1030 
1031 	mutex_lock(&id_priv->qp_mutex);
1032 	if (!id_priv->id.qp) {
1033 		ret = 0;
1034 		goto out;
1035 	}
1036 
1037 	qp_attr.qp_state = IB_QPS_RTS;
1038 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1039 	if (ret)
1040 		goto out;
1041 
1042 	if (conn_param)
1043 		qp_attr.max_rd_atomic = conn_param->initiator_depth;
1044 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1045 out:
1046 	mutex_unlock(&id_priv->qp_mutex);
1047 	return ret;
1048 }
1049 
1050 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1051 {
1052 	struct ib_qp_attr qp_attr;
1053 	int ret;
1054 
1055 	mutex_lock(&id_priv->qp_mutex);
1056 	if (!id_priv->id.qp) {
1057 		ret = 0;
1058 		goto out;
1059 	}
1060 
1061 	qp_attr.qp_state = IB_QPS_ERR;
1062 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1063 out:
1064 	mutex_unlock(&id_priv->qp_mutex);
1065 	return ret;
1066 }
1067 
1068 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1069 			       struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1070 {
1071 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1072 	int ret;
1073 	u16 pkey;
1074 
1075 	if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1076 		pkey = 0xffff;
1077 	else
1078 		pkey = ib_addr_get_pkey(dev_addr);
1079 
1080 	ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1081 				  pkey, &qp_attr->pkey_index);
1082 	if (ret)
1083 		return ret;
1084 
1085 	qp_attr->port_num = id_priv->id.port_num;
1086 	*qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1087 
1088 	if (id_priv->id.qp_type == IB_QPT_UD) {
1089 		ret = cma_set_qkey(id_priv, 0);
1090 		if (ret)
1091 			return ret;
1092 
1093 		qp_attr->qkey = id_priv->qkey;
1094 		*qp_attr_mask |= IB_QP_QKEY;
1095 	} else {
1096 		qp_attr->qp_access_flags = 0;
1097 		*qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1098 	}
1099 	return 0;
1100 }
1101 
1102 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1103 		       int *qp_attr_mask)
1104 {
1105 	struct rdma_id_private *id_priv;
1106 	int ret = 0;
1107 
1108 	id_priv = container_of(id, struct rdma_id_private, id);
1109 	if (rdma_cap_ib_cm(id->device, id->port_num)) {
1110 		if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1111 			ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1112 		else
1113 			ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1114 						 qp_attr_mask);
1115 
1116 		if (qp_attr->qp_state == IB_QPS_RTR)
1117 			qp_attr->rq_psn = id_priv->seq_num;
1118 	} else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1119 		if (!id_priv->cm_id.iw) {
1120 			qp_attr->qp_access_flags = 0;
1121 			*qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1122 		} else
1123 			ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1124 						 qp_attr_mask);
1125 		qp_attr->port_num = id_priv->id.port_num;
1126 		*qp_attr_mask |= IB_QP_PORT;
1127 	} else
1128 		ret = -ENOSYS;
1129 
1130 	if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1131 		qp_attr->timeout = id_priv->timeout;
1132 
1133 	return ret;
1134 }
1135 EXPORT_SYMBOL(rdma_init_qp_attr);
1136 
1137 static inline bool cma_zero_addr(const struct sockaddr *addr)
1138 {
1139 	switch (addr->sa_family) {
1140 	case AF_INET:
1141 		return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1142 	case AF_INET6:
1143 		return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1144 	case AF_IB:
1145 		return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1146 	default:
1147 		return false;
1148 	}
1149 }
1150 
1151 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1152 {
1153 	switch (addr->sa_family) {
1154 	case AF_INET:
1155 		return ipv4_is_loopback(
1156 			((struct sockaddr_in *)addr)->sin_addr.s_addr);
1157 	case AF_INET6:
1158 		return ipv6_addr_loopback(
1159 			&((struct sockaddr_in6 *)addr)->sin6_addr);
1160 	case AF_IB:
1161 		return ib_addr_loopback(
1162 			&((struct sockaddr_ib *)addr)->sib_addr);
1163 	default:
1164 		return false;
1165 	}
1166 }
1167 
1168 static inline bool cma_any_addr(const struct sockaddr *addr)
1169 {
1170 	return cma_zero_addr(addr) || cma_loopback_addr(addr);
1171 }
1172 
1173 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1174 {
1175 	if (src->sa_family != dst->sa_family)
1176 		return -1;
1177 
1178 	switch (src->sa_family) {
1179 	case AF_INET:
1180 		return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1181 		       ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1182 	case AF_INET6: {
1183 		struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1184 		struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1185 		bool link_local;
1186 
1187 		if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1188 					  &dst_addr6->sin6_addr))
1189 			return 1;
1190 		link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1191 			     IPV6_ADDR_LINKLOCAL;
1192 		/* Link local must match their scope_ids */
1193 		return link_local ? (src_addr6->sin6_scope_id !=
1194 				     dst_addr6->sin6_scope_id) :
1195 				    0;
1196 	}
1197 
1198 	default:
1199 		return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1200 				   &((struct sockaddr_ib *) dst)->sib_addr);
1201 	}
1202 }
1203 
1204 static __be16 cma_port(const struct sockaddr *addr)
1205 {
1206 	struct sockaddr_ib *sib;
1207 
1208 	switch (addr->sa_family) {
1209 	case AF_INET:
1210 		return ((struct sockaddr_in *) addr)->sin_port;
1211 	case AF_INET6:
1212 		return ((struct sockaddr_in6 *) addr)->sin6_port;
1213 	case AF_IB:
1214 		sib = (struct sockaddr_ib *) addr;
1215 		return htons((u16) (be64_to_cpu(sib->sib_sid) &
1216 				    be64_to_cpu(sib->sib_sid_mask)));
1217 	default:
1218 		return 0;
1219 	}
1220 }
1221 
1222 static inline int cma_any_port(const struct sockaddr *addr)
1223 {
1224 	return !cma_port(addr);
1225 }
1226 
1227 static void cma_save_ib_info(struct sockaddr *src_addr,
1228 			     struct sockaddr *dst_addr,
1229 			     const struct rdma_cm_id *listen_id,
1230 			     const struct sa_path_rec *path)
1231 {
1232 	struct sockaddr_ib *listen_ib, *ib;
1233 
1234 	listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1235 	if (src_addr) {
1236 		ib = (struct sockaddr_ib *)src_addr;
1237 		ib->sib_family = AF_IB;
1238 		if (path) {
1239 			ib->sib_pkey = path->pkey;
1240 			ib->sib_flowinfo = path->flow_label;
1241 			memcpy(&ib->sib_addr, &path->sgid, 16);
1242 			ib->sib_sid = path->service_id;
1243 			ib->sib_scope_id = 0;
1244 		} else {
1245 			ib->sib_pkey = listen_ib->sib_pkey;
1246 			ib->sib_flowinfo = listen_ib->sib_flowinfo;
1247 			ib->sib_addr = listen_ib->sib_addr;
1248 			ib->sib_sid = listen_ib->sib_sid;
1249 			ib->sib_scope_id = listen_ib->sib_scope_id;
1250 		}
1251 		ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1252 	}
1253 	if (dst_addr) {
1254 		ib = (struct sockaddr_ib *)dst_addr;
1255 		ib->sib_family = AF_IB;
1256 		if (path) {
1257 			ib->sib_pkey = path->pkey;
1258 			ib->sib_flowinfo = path->flow_label;
1259 			memcpy(&ib->sib_addr, &path->dgid, 16);
1260 		}
1261 	}
1262 }
1263 
1264 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1265 			      struct sockaddr_in *dst_addr,
1266 			      struct cma_hdr *hdr,
1267 			      __be16 local_port)
1268 {
1269 	if (src_addr) {
1270 		*src_addr = (struct sockaddr_in) {
1271 			.sin_family = AF_INET,
1272 			.sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1273 			.sin_port = local_port,
1274 		};
1275 	}
1276 
1277 	if (dst_addr) {
1278 		*dst_addr = (struct sockaddr_in) {
1279 			.sin_family = AF_INET,
1280 			.sin_addr.s_addr = hdr->src_addr.ip4.addr,
1281 			.sin_port = hdr->port,
1282 		};
1283 	}
1284 }
1285 
1286 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1287 			      struct sockaddr_in6 *dst_addr,
1288 			      struct cma_hdr *hdr,
1289 			      __be16 local_port)
1290 {
1291 	if (src_addr) {
1292 		*src_addr = (struct sockaddr_in6) {
1293 			.sin6_family = AF_INET6,
1294 			.sin6_addr = hdr->dst_addr.ip6,
1295 			.sin6_port = local_port,
1296 		};
1297 	}
1298 
1299 	if (dst_addr) {
1300 		*dst_addr = (struct sockaddr_in6) {
1301 			.sin6_family = AF_INET6,
1302 			.sin6_addr = hdr->src_addr.ip6,
1303 			.sin6_port = hdr->port,
1304 		};
1305 	}
1306 }
1307 
1308 static u16 cma_port_from_service_id(__be64 service_id)
1309 {
1310 	return (u16)be64_to_cpu(service_id);
1311 }
1312 
1313 static int cma_save_ip_info(struct sockaddr *src_addr,
1314 			    struct sockaddr *dst_addr,
1315 			    const struct ib_cm_event *ib_event,
1316 			    __be64 service_id)
1317 {
1318 	struct cma_hdr *hdr;
1319 	__be16 port;
1320 
1321 	hdr = ib_event->private_data;
1322 	if (hdr->cma_version != CMA_VERSION)
1323 		return -EINVAL;
1324 
1325 	port = htons(cma_port_from_service_id(service_id));
1326 
1327 	switch (cma_get_ip_ver(hdr)) {
1328 	case 4:
1329 		cma_save_ip4_info((struct sockaddr_in *)src_addr,
1330 				  (struct sockaddr_in *)dst_addr, hdr, port);
1331 		break;
1332 	case 6:
1333 		cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1334 				  (struct sockaddr_in6 *)dst_addr, hdr, port);
1335 		break;
1336 	default:
1337 		return -EAFNOSUPPORT;
1338 	}
1339 
1340 	return 0;
1341 }
1342 
1343 static int cma_save_net_info(struct sockaddr *src_addr,
1344 			     struct sockaddr *dst_addr,
1345 			     const struct rdma_cm_id *listen_id,
1346 			     const struct ib_cm_event *ib_event,
1347 			     sa_family_t sa_family, __be64 service_id)
1348 {
1349 	if (sa_family == AF_IB) {
1350 		if (ib_event->event == IB_CM_REQ_RECEIVED)
1351 			cma_save_ib_info(src_addr, dst_addr, listen_id,
1352 					 ib_event->param.req_rcvd.primary_path);
1353 		else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1354 			cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1355 		return 0;
1356 	}
1357 
1358 	return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1359 }
1360 
1361 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1362 			     struct cma_req_info *req)
1363 {
1364 	const struct ib_cm_req_event_param *req_param =
1365 		&ib_event->param.req_rcvd;
1366 	const struct ib_cm_sidr_req_event_param *sidr_param =
1367 		&ib_event->param.sidr_req_rcvd;
1368 
1369 	switch (ib_event->event) {
1370 	case IB_CM_REQ_RECEIVED:
1371 		req->device	= req_param->listen_id->device;
1372 		req->port	= req_param->port;
1373 		memcpy(&req->local_gid, &req_param->primary_path->sgid,
1374 		       sizeof(req->local_gid));
1375 		req->has_gid	= true;
1376 		req->service_id = req_param->primary_path->service_id;
1377 		req->pkey	= be16_to_cpu(req_param->primary_path->pkey);
1378 		if (req->pkey != req_param->bth_pkey)
1379 			pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1380 					    "RDMA CMA: in the future this may cause the request to be dropped\n",
1381 					    req_param->bth_pkey, req->pkey);
1382 		break;
1383 	case IB_CM_SIDR_REQ_RECEIVED:
1384 		req->device	= sidr_param->listen_id->device;
1385 		req->port	= sidr_param->port;
1386 		req->has_gid	= false;
1387 		req->service_id	= sidr_param->service_id;
1388 		req->pkey	= sidr_param->pkey;
1389 		if (req->pkey != sidr_param->bth_pkey)
1390 			pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1391 					    "RDMA CMA: in the future this may cause the request to be dropped\n",
1392 					    sidr_param->bth_pkey, req->pkey);
1393 		break;
1394 	default:
1395 		return -EINVAL;
1396 	}
1397 
1398 	return 0;
1399 }
1400 
1401 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1402 				  const struct sockaddr_in *dst_addr,
1403 				  const struct sockaddr_in *src_addr)
1404 {
1405 	__be32 daddr = dst_addr->sin_addr.s_addr,
1406 	       saddr = src_addr->sin_addr.s_addr;
1407 	struct fib_result res;
1408 	struct flowi4 fl4;
1409 	int err;
1410 	bool ret;
1411 
1412 	if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1413 	    ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1414 	    ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1415 	    ipv4_is_loopback(saddr))
1416 		return false;
1417 
1418 	memset(&fl4, 0, sizeof(fl4));
1419 	fl4.flowi4_iif = net_dev->ifindex;
1420 	fl4.daddr = daddr;
1421 	fl4.saddr = saddr;
1422 
1423 	rcu_read_lock();
1424 	err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1425 	ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1426 	rcu_read_unlock();
1427 
1428 	return ret;
1429 }
1430 
1431 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1432 				  const struct sockaddr_in6 *dst_addr,
1433 				  const struct sockaddr_in6 *src_addr)
1434 {
1435 #if IS_ENABLED(CONFIG_IPV6)
1436 	const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1437 			   IPV6_ADDR_LINKLOCAL;
1438 	struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1439 					 &src_addr->sin6_addr, net_dev->ifindex,
1440 					 NULL, strict);
1441 	bool ret;
1442 
1443 	if (!rt)
1444 		return false;
1445 
1446 	ret = rt->rt6i_idev->dev == net_dev;
1447 	ip6_rt_put(rt);
1448 
1449 	return ret;
1450 #else
1451 	return false;
1452 #endif
1453 }
1454 
1455 static bool validate_net_dev(struct net_device *net_dev,
1456 			     const struct sockaddr *daddr,
1457 			     const struct sockaddr *saddr)
1458 {
1459 	const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1460 	const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1461 	const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1462 	const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1463 
1464 	switch (daddr->sa_family) {
1465 	case AF_INET:
1466 		return saddr->sa_family == AF_INET &&
1467 		       validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1468 
1469 	case AF_INET6:
1470 		return saddr->sa_family == AF_INET6 &&
1471 		       validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1472 
1473 	default:
1474 		return false;
1475 	}
1476 }
1477 
1478 static struct net_device *
1479 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1480 {
1481 	const struct ib_gid_attr *sgid_attr = NULL;
1482 	struct net_device *ndev;
1483 
1484 	if (ib_event->event == IB_CM_REQ_RECEIVED)
1485 		sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1486 	else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1487 		sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1488 
1489 	if (!sgid_attr)
1490 		return NULL;
1491 
1492 	rcu_read_lock();
1493 	ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1494 	if (IS_ERR(ndev))
1495 		ndev = NULL;
1496 	else
1497 		dev_hold(ndev);
1498 	rcu_read_unlock();
1499 	return ndev;
1500 }
1501 
1502 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1503 					  struct cma_req_info *req)
1504 {
1505 	struct sockaddr *listen_addr =
1506 			(struct sockaddr *)&req->listen_addr_storage;
1507 	struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1508 	struct net_device *net_dev;
1509 	const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1510 	int err;
1511 
1512 	err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1513 			       req->service_id);
1514 	if (err)
1515 		return ERR_PTR(err);
1516 
1517 	if (rdma_protocol_roce(req->device, req->port))
1518 		net_dev = roce_get_net_dev_by_cm_event(ib_event);
1519 	else
1520 		net_dev = ib_get_net_dev_by_params(req->device, req->port,
1521 						   req->pkey,
1522 						   gid, listen_addr);
1523 	if (!net_dev)
1524 		return ERR_PTR(-ENODEV);
1525 
1526 	return net_dev;
1527 }
1528 
1529 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1530 {
1531 	return (be64_to_cpu(service_id) >> 16) & 0xffff;
1532 }
1533 
1534 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1535 				   const struct cma_hdr *hdr)
1536 {
1537 	struct sockaddr *addr = cma_src_addr(id_priv);
1538 	__be32 ip4_addr;
1539 	struct in6_addr ip6_addr;
1540 
1541 	if (cma_any_addr(addr) && !id_priv->afonly)
1542 		return true;
1543 
1544 	switch (addr->sa_family) {
1545 	case AF_INET:
1546 		ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1547 		if (cma_get_ip_ver(hdr) != 4)
1548 			return false;
1549 		if (!cma_any_addr(addr) &&
1550 		    hdr->dst_addr.ip4.addr != ip4_addr)
1551 			return false;
1552 		break;
1553 	case AF_INET6:
1554 		ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1555 		if (cma_get_ip_ver(hdr) != 6)
1556 			return false;
1557 		if (!cma_any_addr(addr) &&
1558 		    memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1559 			return false;
1560 		break;
1561 	case AF_IB:
1562 		return true;
1563 	default:
1564 		return false;
1565 	}
1566 
1567 	return true;
1568 }
1569 
1570 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1571 {
1572 	struct ib_device *device = id->device;
1573 	const int port_num = id->port_num ?: rdma_start_port(device);
1574 
1575 	return rdma_protocol_roce(device, port_num);
1576 }
1577 
1578 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1579 {
1580 	const struct sockaddr *daddr =
1581 			(const struct sockaddr *)&req->listen_addr_storage;
1582 	const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1583 
1584 	/* Returns true if the req is for IPv6 link local */
1585 	return (daddr->sa_family == AF_INET6 &&
1586 		(ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1587 }
1588 
1589 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1590 			      const struct net_device *net_dev,
1591 			      const struct cma_req_info *req)
1592 {
1593 	const struct rdma_addr *addr = &id->route.addr;
1594 
1595 	if (!net_dev)
1596 		/* This request is an AF_IB request */
1597 		return (!id->port_num || id->port_num == req->port) &&
1598 		       (addr->src_addr.ss_family == AF_IB);
1599 
1600 	/*
1601 	 * If the request is not for IPv6 link local, allow matching
1602 	 * request to any netdevice of the one or multiport rdma device.
1603 	 */
1604 	if (!cma_is_req_ipv6_ll(req))
1605 		return true;
1606 	/*
1607 	 * Net namespaces must match, and if the listner is listening
1608 	 * on a specific netdevice than netdevice must match as well.
1609 	 */
1610 	if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1611 	    (!!addr->dev_addr.bound_dev_if ==
1612 	     (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1613 		return true;
1614 	else
1615 		return false;
1616 }
1617 
1618 static struct rdma_id_private *cma_find_listener(
1619 		const struct rdma_bind_list *bind_list,
1620 		const struct ib_cm_id *cm_id,
1621 		const struct ib_cm_event *ib_event,
1622 		const struct cma_req_info *req,
1623 		const struct net_device *net_dev)
1624 {
1625 	struct rdma_id_private *id_priv, *id_priv_dev;
1626 
1627 	lockdep_assert_held(&lock);
1628 
1629 	if (!bind_list)
1630 		return ERR_PTR(-EINVAL);
1631 
1632 	hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1633 		if (cma_match_private_data(id_priv, ib_event->private_data)) {
1634 			if (id_priv->id.device == cm_id->device &&
1635 			    cma_match_net_dev(&id_priv->id, net_dev, req))
1636 				return id_priv;
1637 			list_for_each_entry(id_priv_dev,
1638 					    &id_priv->listen_list,
1639 					    listen_list) {
1640 				if (id_priv_dev->id.device == cm_id->device &&
1641 				    cma_match_net_dev(&id_priv_dev->id,
1642 						      net_dev, req))
1643 					return id_priv_dev;
1644 			}
1645 		}
1646 	}
1647 
1648 	return ERR_PTR(-EINVAL);
1649 }
1650 
1651 static struct rdma_id_private *
1652 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1653 		     const struct ib_cm_event *ib_event,
1654 		     struct cma_req_info *req,
1655 		     struct net_device **net_dev)
1656 {
1657 	struct rdma_bind_list *bind_list;
1658 	struct rdma_id_private *id_priv;
1659 	int err;
1660 
1661 	err = cma_save_req_info(ib_event, req);
1662 	if (err)
1663 		return ERR_PTR(err);
1664 
1665 	*net_dev = cma_get_net_dev(ib_event, req);
1666 	if (IS_ERR(*net_dev)) {
1667 		if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1668 			/* Assuming the protocol is AF_IB */
1669 			*net_dev = NULL;
1670 		} else {
1671 			return ERR_CAST(*net_dev);
1672 		}
1673 	}
1674 
1675 	mutex_lock(&lock);
1676 	/*
1677 	 * Net namespace might be getting deleted while route lookup,
1678 	 * cm_id lookup is in progress. Therefore, perform netdevice
1679 	 * validation, cm_id lookup under rcu lock.
1680 	 * RCU lock along with netdevice state check, synchronizes with
1681 	 * netdevice migrating to different net namespace and also avoids
1682 	 * case where net namespace doesn't get deleted while lookup is in
1683 	 * progress.
1684 	 * If the device state is not IFF_UP, its properties such as ifindex
1685 	 * and nd_net cannot be trusted to remain valid without rcu lock.
1686 	 * net/core/dev.c change_net_namespace() ensures to synchronize with
1687 	 * ongoing operations on net device after device is closed using
1688 	 * synchronize_net().
1689 	 */
1690 	rcu_read_lock();
1691 	if (*net_dev) {
1692 		/*
1693 		 * If netdevice is down, it is likely that it is administratively
1694 		 * down or it might be migrating to different namespace.
1695 		 * In that case avoid further processing, as the net namespace
1696 		 * or ifindex may change.
1697 		 */
1698 		if (((*net_dev)->flags & IFF_UP) == 0) {
1699 			id_priv = ERR_PTR(-EHOSTUNREACH);
1700 			goto err;
1701 		}
1702 
1703 		if (!validate_net_dev(*net_dev,
1704 				 (struct sockaddr *)&req->listen_addr_storage,
1705 				 (struct sockaddr *)&req->src_addr_storage)) {
1706 			id_priv = ERR_PTR(-EHOSTUNREACH);
1707 			goto err;
1708 		}
1709 	}
1710 
1711 	bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1712 				rdma_ps_from_service_id(req->service_id),
1713 				cma_port_from_service_id(req->service_id));
1714 	id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1715 err:
1716 	rcu_read_unlock();
1717 	mutex_unlock(&lock);
1718 	if (IS_ERR(id_priv) && *net_dev) {
1719 		dev_put(*net_dev);
1720 		*net_dev = NULL;
1721 	}
1722 	return id_priv;
1723 }
1724 
1725 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1726 {
1727 	return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1728 }
1729 
1730 static void cma_cancel_route(struct rdma_id_private *id_priv)
1731 {
1732 	if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1733 		if (id_priv->query)
1734 			ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1735 	}
1736 }
1737 
1738 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1739 {
1740 	struct rdma_id_private *dev_id_priv;
1741 
1742 	/*
1743 	 * Remove from listen_any_list to prevent added devices from spawning
1744 	 * additional listen requests.
1745 	 */
1746 	mutex_lock(&lock);
1747 	list_del(&id_priv->list);
1748 
1749 	while (!list_empty(&id_priv->listen_list)) {
1750 		dev_id_priv = list_entry(id_priv->listen_list.next,
1751 					 struct rdma_id_private, listen_list);
1752 		/* sync with device removal to avoid duplicate destruction */
1753 		list_del_init(&dev_id_priv->list);
1754 		list_del(&dev_id_priv->listen_list);
1755 		mutex_unlock(&lock);
1756 
1757 		rdma_destroy_id(&dev_id_priv->id);
1758 		mutex_lock(&lock);
1759 	}
1760 	mutex_unlock(&lock);
1761 }
1762 
1763 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1764 				 enum rdma_cm_state state)
1765 {
1766 	switch (state) {
1767 	case RDMA_CM_ADDR_QUERY:
1768 		rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1769 		break;
1770 	case RDMA_CM_ROUTE_QUERY:
1771 		cma_cancel_route(id_priv);
1772 		break;
1773 	case RDMA_CM_LISTEN:
1774 		if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1775 			cma_cancel_listens(id_priv);
1776 		break;
1777 	default:
1778 		break;
1779 	}
1780 }
1781 
1782 static void cma_release_port(struct rdma_id_private *id_priv)
1783 {
1784 	struct rdma_bind_list *bind_list = id_priv->bind_list;
1785 	struct net *net = id_priv->id.route.addr.dev_addr.net;
1786 
1787 	if (!bind_list)
1788 		return;
1789 
1790 	mutex_lock(&lock);
1791 	hlist_del(&id_priv->node);
1792 	if (hlist_empty(&bind_list->owners)) {
1793 		cma_ps_remove(net, bind_list->ps, bind_list->port);
1794 		kfree(bind_list);
1795 	}
1796 	mutex_unlock(&lock);
1797 }
1798 
1799 static void cma_leave_roce_mc_group(struct rdma_id_private *id_priv,
1800 				    struct cma_multicast *mc)
1801 {
1802 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1803 	struct net_device *ndev = NULL;
1804 
1805 	if (dev_addr->bound_dev_if)
1806 		ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
1807 	if (ndev) {
1808 		cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid, false);
1809 		dev_put(ndev);
1810 	}
1811 	kref_put(&mc->mcref, release_mc);
1812 }
1813 
1814 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1815 {
1816 	struct cma_multicast *mc;
1817 
1818 	while (!list_empty(&id_priv->mc_list)) {
1819 		mc = container_of(id_priv->mc_list.next,
1820 				  struct cma_multicast, list);
1821 		list_del(&mc->list);
1822 		if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1823 				      id_priv->id.port_num)) {
1824 			ib_sa_free_multicast(mc->multicast.ib);
1825 			kfree(mc);
1826 		} else {
1827 			cma_leave_roce_mc_group(id_priv, mc);
1828 		}
1829 	}
1830 }
1831 
1832 void rdma_destroy_id(struct rdma_cm_id *id)
1833 {
1834 	struct rdma_id_private *id_priv;
1835 	enum rdma_cm_state state;
1836 
1837 	id_priv = container_of(id, struct rdma_id_private, id);
1838 	trace_cm_id_destroy(id_priv);
1839 	state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1840 	cma_cancel_operation(id_priv, state);
1841 
1842 	/*
1843 	 * Wait for any active callback to finish.  New callbacks will find
1844 	 * the id_priv state set to destroying and abort.
1845 	 */
1846 	mutex_lock(&id_priv->handler_mutex);
1847 	mutex_unlock(&id_priv->handler_mutex);
1848 
1849 	rdma_restrack_del(&id_priv->res);
1850 	if (id_priv->cma_dev) {
1851 		if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1852 			if (id_priv->cm_id.ib)
1853 				ib_destroy_cm_id(id_priv->cm_id.ib);
1854 		} else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1855 			if (id_priv->cm_id.iw)
1856 				iw_destroy_cm_id(id_priv->cm_id.iw);
1857 		}
1858 		cma_leave_mc_groups(id_priv);
1859 		cma_release_dev(id_priv);
1860 	}
1861 
1862 	cma_release_port(id_priv);
1863 	cma_id_put(id_priv);
1864 	wait_for_completion(&id_priv->comp);
1865 
1866 	if (id_priv->internal_id)
1867 		cma_id_put(id_priv->id.context);
1868 
1869 	kfree(id_priv->id.route.path_rec);
1870 
1871 	if (id_priv->id.route.addr.dev_addr.sgid_attr)
1872 		rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
1873 
1874 	put_net(id_priv->id.route.addr.dev_addr.net);
1875 	kfree(id_priv);
1876 }
1877 EXPORT_SYMBOL(rdma_destroy_id);
1878 
1879 static int cma_rep_recv(struct rdma_id_private *id_priv)
1880 {
1881 	int ret;
1882 
1883 	ret = cma_modify_qp_rtr(id_priv, NULL);
1884 	if (ret)
1885 		goto reject;
1886 
1887 	ret = cma_modify_qp_rts(id_priv, NULL);
1888 	if (ret)
1889 		goto reject;
1890 
1891 	trace_cm_send_rtu(id_priv);
1892 	ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1893 	if (ret)
1894 		goto reject;
1895 
1896 	return 0;
1897 reject:
1898 	pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1899 	cma_modify_qp_err(id_priv);
1900 	trace_cm_send_rej(id_priv);
1901 	ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1902 		       NULL, 0, NULL, 0);
1903 	return ret;
1904 }
1905 
1906 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1907 				   const struct ib_cm_rep_event_param *rep_data,
1908 				   void *private_data)
1909 {
1910 	event->param.conn.private_data = private_data;
1911 	event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1912 	event->param.conn.responder_resources = rep_data->responder_resources;
1913 	event->param.conn.initiator_depth = rep_data->initiator_depth;
1914 	event->param.conn.flow_control = rep_data->flow_control;
1915 	event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1916 	event->param.conn.srq = rep_data->srq;
1917 	event->param.conn.qp_num = rep_data->remote_qpn;
1918 
1919 	event->ece.vendor_id = rep_data->ece.vendor_id;
1920 	event->ece.attr_mod = rep_data->ece.attr_mod;
1921 }
1922 
1923 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
1924 				struct rdma_cm_event *event)
1925 {
1926 	int ret;
1927 
1928 	trace_cm_event_handler(id_priv, event);
1929 	ret = id_priv->id.event_handler(&id_priv->id, event);
1930 	trace_cm_event_done(id_priv, event, ret);
1931 	return ret;
1932 }
1933 
1934 static int cma_ib_handler(struct ib_cm_id *cm_id,
1935 			  const struct ib_cm_event *ib_event)
1936 {
1937 	struct rdma_id_private *id_priv = cm_id->context;
1938 	struct rdma_cm_event event = {};
1939 	int ret = 0;
1940 
1941 	mutex_lock(&id_priv->handler_mutex);
1942 	if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1943 	     id_priv->state != RDMA_CM_CONNECT) ||
1944 	    (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1945 	     id_priv->state != RDMA_CM_DISCONNECT))
1946 		goto out;
1947 
1948 	switch (ib_event->event) {
1949 	case IB_CM_REQ_ERROR:
1950 	case IB_CM_REP_ERROR:
1951 		event.event = RDMA_CM_EVENT_UNREACHABLE;
1952 		event.status = -ETIMEDOUT;
1953 		break;
1954 	case IB_CM_REP_RECEIVED:
1955 		if (cma_comp(id_priv, RDMA_CM_CONNECT) &&
1956 		    (id_priv->id.qp_type != IB_QPT_UD)) {
1957 			trace_cm_send_mra(id_priv);
1958 			ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1959 		}
1960 		if (id_priv->id.qp) {
1961 			event.status = cma_rep_recv(id_priv);
1962 			event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1963 						     RDMA_CM_EVENT_ESTABLISHED;
1964 		} else {
1965 			event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1966 		}
1967 		cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1968 				       ib_event->private_data);
1969 		break;
1970 	case IB_CM_RTU_RECEIVED:
1971 	case IB_CM_USER_ESTABLISHED:
1972 		event.event = RDMA_CM_EVENT_ESTABLISHED;
1973 		break;
1974 	case IB_CM_DREQ_ERROR:
1975 		event.status = -ETIMEDOUT; /* fall through */
1976 	case IB_CM_DREQ_RECEIVED:
1977 	case IB_CM_DREP_RECEIVED:
1978 		if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1979 				   RDMA_CM_DISCONNECT))
1980 			goto out;
1981 		event.event = RDMA_CM_EVENT_DISCONNECTED;
1982 		break;
1983 	case IB_CM_TIMEWAIT_EXIT:
1984 		event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1985 		break;
1986 	case IB_CM_MRA_RECEIVED:
1987 		/* ignore event */
1988 		goto out;
1989 	case IB_CM_REJ_RECEIVED:
1990 		pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
1991 										ib_event->param.rej_rcvd.reason));
1992 		cma_modify_qp_err(id_priv);
1993 		event.status = ib_event->param.rej_rcvd.reason;
1994 		event.event = RDMA_CM_EVENT_REJECTED;
1995 		event.param.conn.private_data = ib_event->private_data;
1996 		event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1997 		break;
1998 	default:
1999 		pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2000 		       ib_event->event);
2001 		goto out;
2002 	}
2003 
2004 	ret = cma_cm_event_handler(id_priv, &event);
2005 	if (ret) {
2006 		/* Destroy the CM ID by returning a non-zero value. */
2007 		id_priv->cm_id.ib = NULL;
2008 		cma_exch(id_priv, RDMA_CM_DESTROYING);
2009 		mutex_unlock(&id_priv->handler_mutex);
2010 		rdma_destroy_id(&id_priv->id);
2011 		return ret;
2012 	}
2013 out:
2014 	mutex_unlock(&id_priv->handler_mutex);
2015 	return ret;
2016 }
2017 
2018 static struct rdma_id_private *
2019 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2020 		   const struct ib_cm_event *ib_event,
2021 		   struct net_device *net_dev)
2022 {
2023 	struct rdma_id_private *listen_id_priv;
2024 	struct rdma_id_private *id_priv;
2025 	struct rdma_cm_id *id;
2026 	struct rdma_route *rt;
2027 	const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2028 	struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2029 	const __be64 service_id =
2030 		ib_event->param.req_rcvd.primary_path->service_id;
2031 	int ret;
2032 
2033 	listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2034 	id = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2035 			    listen_id->event_handler, listen_id->context,
2036 			    listen_id->ps, ib_event->param.req_rcvd.qp_type,
2037 			    listen_id_priv->res.kern_name);
2038 	if (IS_ERR(id))
2039 		return NULL;
2040 
2041 	id_priv = container_of(id, struct rdma_id_private, id);
2042 	if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2043 			      (struct sockaddr *)&id->route.addr.dst_addr,
2044 			      listen_id, ib_event, ss_family, service_id))
2045 		goto err;
2046 
2047 	rt = &id->route;
2048 	rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2049 	rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
2050 				     GFP_KERNEL);
2051 	if (!rt->path_rec)
2052 		goto err;
2053 
2054 	rt->path_rec[0] = *path;
2055 	if (rt->num_paths == 2)
2056 		rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2057 
2058 	if (net_dev) {
2059 		rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2060 	} else {
2061 		if (!cma_protocol_roce(listen_id) &&
2062 		    cma_any_addr(cma_src_addr(id_priv))) {
2063 			rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2064 			rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2065 			ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2066 		} else if (!cma_any_addr(cma_src_addr(id_priv))) {
2067 			ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2068 			if (ret)
2069 				goto err;
2070 		}
2071 	}
2072 	rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2073 
2074 	id_priv->state = RDMA_CM_CONNECT;
2075 	return id_priv;
2076 
2077 err:
2078 	rdma_destroy_id(id);
2079 	return NULL;
2080 }
2081 
2082 static struct rdma_id_private *
2083 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2084 		  const struct ib_cm_event *ib_event,
2085 		  struct net_device *net_dev)
2086 {
2087 	const struct rdma_id_private *listen_id_priv;
2088 	struct rdma_id_private *id_priv;
2089 	struct rdma_cm_id *id;
2090 	const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2091 	struct net *net = listen_id->route.addr.dev_addr.net;
2092 	int ret;
2093 
2094 	listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2095 	id = __rdma_create_id(net, listen_id->event_handler, listen_id->context,
2096 			      listen_id->ps, IB_QPT_UD,
2097 			      listen_id_priv->res.kern_name);
2098 	if (IS_ERR(id))
2099 		return NULL;
2100 
2101 	id_priv = container_of(id, struct rdma_id_private, id);
2102 	if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2103 			      (struct sockaddr *)&id->route.addr.dst_addr,
2104 			      listen_id, ib_event, ss_family,
2105 			      ib_event->param.sidr_req_rcvd.service_id))
2106 		goto err;
2107 
2108 	if (net_dev) {
2109 		rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2110 	} else {
2111 		if (!cma_any_addr(cma_src_addr(id_priv))) {
2112 			ret = cma_translate_addr(cma_src_addr(id_priv),
2113 						 &id->route.addr.dev_addr);
2114 			if (ret)
2115 				goto err;
2116 		}
2117 	}
2118 
2119 	id_priv->state = RDMA_CM_CONNECT;
2120 	return id_priv;
2121 err:
2122 	rdma_destroy_id(id);
2123 	return NULL;
2124 }
2125 
2126 static void cma_set_req_event_data(struct rdma_cm_event *event,
2127 				   const struct ib_cm_req_event_param *req_data,
2128 				   void *private_data, int offset)
2129 {
2130 	event->param.conn.private_data = private_data + offset;
2131 	event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2132 	event->param.conn.responder_resources = req_data->responder_resources;
2133 	event->param.conn.initiator_depth = req_data->initiator_depth;
2134 	event->param.conn.flow_control = req_data->flow_control;
2135 	event->param.conn.retry_count = req_data->retry_count;
2136 	event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2137 	event->param.conn.srq = req_data->srq;
2138 	event->param.conn.qp_num = req_data->remote_qpn;
2139 
2140 	event->ece.vendor_id = req_data->ece.vendor_id;
2141 	event->ece.attr_mod = req_data->ece.attr_mod;
2142 }
2143 
2144 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2145 				    const struct ib_cm_event *ib_event)
2146 {
2147 	return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2148 		 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2149 		((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2150 		 (id->qp_type == IB_QPT_UD)) ||
2151 		(!id->qp_type));
2152 }
2153 
2154 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2155 			      const struct ib_cm_event *ib_event)
2156 {
2157 	struct rdma_id_private *listen_id, *conn_id = NULL;
2158 	struct rdma_cm_event event = {};
2159 	struct cma_req_info req = {};
2160 	struct net_device *net_dev;
2161 	u8 offset;
2162 	int ret;
2163 
2164 	listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2165 	if (IS_ERR(listen_id))
2166 		return PTR_ERR(listen_id);
2167 
2168 	trace_cm_req_handler(listen_id, ib_event->event);
2169 	if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2170 		ret = -EINVAL;
2171 		goto net_dev_put;
2172 	}
2173 
2174 	mutex_lock(&listen_id->handler_mutex);
2175 	if (listen_id->state != RDMA_CM_LISTEN) {
2176 		ret = -ECONNABORTED;
2177 		goto err1;
2178 	}
2179 
2180 	offset = cma_user_data_offset(listen_id);
2181 	event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2182 	if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2183 		conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2184 		event.param.ud.private_data = ib_event->private_data + offset;
2185 		event.param.ud.private_data_len =
2186 				IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2187 	} else {
2188 		conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2189 		cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2190 				       ib_event->private_data, offset);
2191 	}
2192 	if (!conn_id) {
2193 		ret = -ENOMEM;
2194 		goto err1;
2195 	}
2196 
2197 	mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2198 	ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2199 	if (ret)
2200 		goto err2;
2201 
2202 	conn_id->cm_id.ib = cm_id;
2203 	cm_id->context = conn_id;
2204 	cm_id->cm_handler = cma_ib_handler;
2205 
2206 	/*
2207 	 * Protect against the user destroying conn_id from another thread
2208 	 * until we're done accessing it.
2209 	 */
2210 	cma_id_get(conn_id);
2211 	ret = cma_cm_event_handler(conn_id, &event);
2212 	if (ret)
2213 		goto err3;
2214 	/*
2215 	 * Acquire mutex to prevent user executing rdma_destroy_id()
2216 	 * while we're accessing the cm_id.
2217 	 */
2218 	mutex_lock(&lock);
2219 	if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
2220 	    (conn_id->id.qp_type != IB_QPT_UD)) {
2221 		trace_cm_send_mra(cm_id->context);
2222 		ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2223 	}
2224 	mutex_unlock(&lock);
2225 	mutex_unlock(&conn_id->handler_mutex);
2226 	mutex_unlock(&listen_id->handler_mutex);
2227 	cma_id_put(conn_id);
2228 	if (net_dev)
2229 		dev_put(net_dev);
2230 	return 0;
2231 
2232 err3:
2233 	cma_id_put(conn_id);
2234 	/* Destroy the CM ID by returning a non-zero value. */
2235 	conn_id->cm_id.ib = NULL;
2236 err2:
2237 	cma_exch(conn_id, RDMA_CM_DESTROYING);
2238 	mutex_unlock(&conn_id->handler_mutex);
2239 err1:
2240 	mutex_unlock(&listen_id->handler_mutex);
2241 	if (conn_id)
2242 		rdma_destroy_id(&conn_id->id);
2243 
2244 net_dev_put:
2245 	if (net_dev)
2246 		dev_put(net_dev);
2247 
2248 	return ret;
2249 }
2250 
2251 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2252 {
2253 	if (addr->sa_family == AF_IB)
2254 		return ((struct sockaddr_ib *) addr)->sib_sid;
2255 
2256 	return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2257 }
2258 EXPORT_SYMBOL(rdma_get_service_id);
2259 
2260 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2261 		    union ib_gid *dgid)
2262 {
2263 	struct rdma_addr *addr = &cm_id->route.addr;
2264 
2265 	if (!cm_id->device) {
2266 		if (sgid)
2267 			memset(sgid, 0, sizeof(*sgid));
2268 		if (dgid)
2269 			memset(dgid, 0, sizeof(*dgid));
2270 		return;
2271 	}
2272 
2273 	if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2274 		if (sgid)
2275 			rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2276 		if (dgid)
2277 			rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2278 	} else {
2279 		if (sgid)
2280 			rdma_addr_get_sgid(&addr->dev_addr, sgid);
2281 		if (dgid)
2282 			rdma_addr_get_dgid(&addr->dev_addr, dgid);
2283 	}
2284 }
2285 EXPORT_SYMBOL(rdma_read_gids);
2286 
2287 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2288 {
2289 	struct rdma_id_private *id_priv = iw_id->context;
2290 	struct rdma_cm_event event = {};
2291 	int ret = 0;
2292 	struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2293 	struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2294 
2295 	mutex_lock(&id_priv->handler_mutex);
2296 	if (id_priv->state != RDMA_CM_CONNECT)
2297 		goto out;
2298 
2299 	switch (iw_event->event) {
2300 	case IW_CM_EVENT_CLOSE:
2301 		event.event = RDMA_CM_EVENT_DISCONNECTED;
2302 		break;
2303 	case IW_CM_EVENT_CONNECT_REPLY:
2304 		memcpy(cma_src_addr(id_priv), laddr,
2305 		       rdma_addr_size(laddr));
2306 		memcpy(cma_dst_addr(id_priv), raddr,
2307 		       rdma_addr_size(raddr));
2308 		switch (iw_event->status) {
2309 		case 0:
2310 			event.event = RDMA_CM_EVENT_ESTABLISHED;
2311 			event.param.conn.initiator_depth = iw_event->ird;
2312 			event.param.conn.responder_resources = iw_event->ord;
2313 			break;
2314 		case -ECONNRESET:
2315 		case -ECONNREFUSED:
2316 			event.event = RDMA_CM_EVENT_REJECTED;
2317 			break;
2318 		case -ETIMEDOUT:
2319 			event.event = RDMA_CM_EVENT_UNREACHABLE;
2320 			break;
2321 		default:
2322 			event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2323 			break;
2324 		}
2325 		break;
2326 	case IW_CM_EVENT_ESTABLISHED:
2327 		event.event = RDMA_CM_EVENT_ESTABLISHED;
2328 		event.param.conn.initiator_depth = iw_event->ird;
2329 		event.param.conn.responder_resources = iw_event->ord;
2330 		break;
2331 	default:
2332 		goto out;
2333 	}
2334 
2335 	event.status = iw_event->status;
2336 	event.param.conn.private_data = iw_event->private_data;
2337 	event.param.conn.private_data_len = iw_event->private_data_len;
2338 	ret = cma_cm_event_handler(id_priv, &event);
2339 	if (ret) {
2340 		/* Destroy the CM ID by returning a non-zero value. */
2341 		id_priv->cm_id.iw = NULL;
2342 		cma_exch(id_priv, RDMA_CM_DESTROYING);
2343 		mutex_unlock(&id_priv->handler_mutex);
2344 		rdma_destroy_id(&id_priv->id);
2345 		return ret;
2346 	}
2347 
2348 out:
2349 	mutex_unlock(&id_priv->handler_mutex);
2350 	return ret;
2351 }
2352 
2353 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2354 			       struct iw_cm_event *iw_event)
2355 {
2356 	struct rdma_cm_id *new_cm_id;
2357 	struct rdma_id_private *listen_id, *conn_id;
2358 	struct rdma_cm_event event = {};
2359 	int ret = -ECONNABORTED;
2360 	struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2361 	struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2362 
2363 	event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2364 	event.param.conn.private_data = iw_event->private_data;
2365 	event.param.conn.private_data_len = iw_event->private_data_len;
2366 	event.param.conn.initiator_depth = iw_event->ird;
2367 	event.param.conn.responder_resources = iw_event->ord;
2368 
2369 	listen_id = cm_id->context;
2370 
2371 	mutex_lock(&listen_id->handler_mutex);
2372 	if (listen_id->state != RDMA_CM_LISTEN)
2373 		goto out;
2374 
2375 	/* Create a new RDMA id for the new IW CM ID */
2376 	new_cm_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2377 				     listen_id->id.event_handler,
2378 				     listen_id->id.context,
2379 				     RDMA_PS_TCP, IB_QPT_RC,
2380 				     listen_id->res.kern_name);
2381 	if (IS_ERR(new_cm_id)) {
2382 		ret = -ENOMEM;
2383 		goto out;
2384 	}
2385 	conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2386 	mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2387 	conn_id->state = RDMA_CM_CONNECT;
2388 
2389 	ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2390 	if (ret) {
2391 		mutex_unlock(&conn_id->handler_mutex);
2392 		rdma_destroy_id(new_cm_id);
2393 		goto out;
2394 	}
2395 
2396 	ret = cma_iw_acquire_dev(conn_id, listen_id);
2397 	if (ret) {
2398 		mutex_unlock(&conn_id->handler_mutex);
2399 		rdma_destroy_id(new_cm_id);
2400 		goto out;
2401 	}
2402 
2403 	conn_id->cm_id.iw = cm_id;
2404 	cm_id->context = conn_id;
2405 	cm_id->cm_handler = cma_iw_handler;
2406 
2407 	memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2408 	memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2409 
2410 	/*
2411 	 * Protect against the user destroying conn_id from another thread
2412 	 * until we're done accessing it.
2413 	 */
2414 	cma_id_get(conn_id);
2415 	ret = cma_cm_event_handler(conn_id, &event);
2416 	if (ret) {
2417 		/* User wants to destroy the CM ID */
2418 		conn_id->cm_id.iw = NULL;
2419 		cma_exch(conn_id, RDMA_CM_DESTROYING);
2420 		mutex_unlock(&conn_id->handler_mutex);
2421 		mutex_unlock(&listen_id->handler_mutex);
2422 		cma_id_put(conn_id);
2423 		rdma_destroy_id(&conn_id->id);
2424 		return ret;
2425 	}
2426 
2427 	mutex_unlock(&conn_id->handler_mutex);
2428 	cma_id_put(conn_id);
2429 
2430 out:
2431 	mutex_unlock(&listen_id->handler_mutex);
2432 	return ret;
2433 }
2434 
2435 static int cma_ib_listen(struct rdma_id_private *id_priv)
2436 {
2437 	struct sockaddr *addr;
2438 	struct ib_cm_id	*id;
2439 	__be64 svc_id;
2440 
2441 	addr = cma_src_addr(id_priv);
2442 	svc_id = rdma_get_service_id(&id_priv->id, addr);
2443 	id = ib_cm_insert_listen(id_priv->id.device,
2444 				 cma_ib_req_handler, svc_id);
2445 	if (IS_ERR(id))
2446 		return PTR_ERR(id);
2447 	id_priv->cm_id.ib = id;
2448 
2449 	return 0;
2450 }
2451 
2452 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2453 {
2454 	int ret;
2455 	struct iw_cm_id	*id;
2456 
2457 	id = iw_create_cm_id(id_priv->id.device,
2458 			     iw_conn_req_handler,
2459 			     id_priv);
2460 	if (IS_ERR(id))
2461 		return PTR_ERR(id);
2462 
2463 	id->tos = id_priv->tos;
2464 	id->tos_set = id_priv->tos_set;
2465 	id_priv->cm_id.iw = id;
2466 
2467 	memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2468 	       rdma_addr_size(cma_src_addr(id_priv)));
2469 
2470 	ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2471 
2472 	if (ret) {
2473 		iw_destroy_cm_id(id_priv->cm_id.iw);
2474 		id_priv->cm_id.iw = NULL;
2475 	}
2476 
2477 	return ret;
2478 }
2479 
2480 static int cma_listen_handler(struct rdma_cm_id *id,
2481 			      struct rdma_cm_event *event)
2482 {
2483 	struct rdma_id_private *id_priv = id->context;
2484 
2485 	id->context = id_priv->id.context;
2486 	id->event_handler = id_priv->id.event_handler;
2487 	trace_cm_event_handler(id_priv, event);
2488 	return id_priv->id.event_handler(id, event);
2489 }
2490 
2491 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2492 			      struct cma_device *cma_dev)
2493 {
2494 	struct rdma_id_private *dev_id_priv;
2495 	struct rdma_cm_id *id;
2496 	struct net *net = id_priv->id.route.addr.dev_addr.net;
2497 	int ret;
2498 
2499 	lockdep_assert_held(&lock);
2500 
2501 	if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2502 		return;
2503 
2504 	id = __rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2505 			      id_priv->id.qp_type, id_priv->res.kern_name);
2506 	if (IS_ERR(id))
2507 		return;
2508 
2509 	dev_id_priv = container_of(id, struct rdma_id_private, id);
2510 
2511 	dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2512 	memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2513 	       rdma_addr_size(cma_src_addr(id_priv)));
2514 
2515 	_cma_attach_to_dev(dev_id_priv, cma_dev);
2516 	list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2517 	cma_id_get(id_priv);
2518 	dev_id_priv->internal_id = 1;
2519 	dev_id_priv->afonly = id_priv->afonly;
2520 	dev_id_priv->tos_set = id_priv->tos_set;
2521 	dev_id_priv->tos = id_priv->tos;
2522 
2523 	ret = rdma_listen(id, id_priv->backlog);
2524 	if (ret)
2525 		dev_warn(&cma_dev->device->dev,
2526 			 "RDMA CMA: cma_listen_on_dev, error %d\n", ret);
2527 }
2528 
2529 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2530 {
2531 	struct cma_device *cma_dev;
2532 
2533 	mutex_lock(&lock);
2534 	list_add_tail(&id_priv->list, &listen_any_list);
2535 	list_for_each_entry(cma_dev, &dev_list, list)
2536 		cma_listen_on_dev(id_priv, cma_dev);
2537 	mutex_unlock(&lock);
2538 }
2539 
2540 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2541 {
2542 	struct rdma_id_private *id_priv;
2543 
2544 	id_priv = container_of(id, struct rdma_id_private, id);
2545 	id_priv->tos = (u8) tos;
2546 	id_priv->tos_set = true;
2547 }
2548 EXPORT_SYMBOL(rdma_set_service_type);
2549 
2550 /**
2551  * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2552  *                          with a connection identifier.
2553  * @id: Communication identifier to associated with service type.
2554  * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2555  *
2556  * This function should be called before rdma_connect() on active side,
2557  * and on passive side before rdma_accept(). It is applicable to primary
2558  * path only. The timeout will affect the local side of the QP, it is not
2559  * negotiated with remote side and zero disables the timer. In case it is
2560  * set before rdma_resolve_route, the value will also be used to determine
2561  * PacketLifeTime for RoCE.
2562  *
2563  * Return: 0 for success
2564  */
2565 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2566 {
2567 	struct rdma_id_private *id_priv;
2568 
2569 	if (id->qp_type != IB_QPT_RC)
2570 		return -EINVAL;
2571 
2572 	id_priv = container_of(id, struct rdma_id_private, id);
2573 	id_priv->timeout = timeout;
2574 	id_priv->timeout_set = true;
2575 
2576 	return 0;
2577 }
2578 EXPORT_SYMBOL(rdma_set_ack_timeout);
2579 
2580 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2581 			      void *context)
2582 {
2583 	struct cma_work *work = context;
2584 	struct rdma_route *route;
2585 
2586 	route = &work->id->id.route;
2587 
2588 	if (!status) {
2589 		route->num_paths = 1;
2590 		*route->path_rec = *path_rec;
2591 	} else {
2592 		work->old_state = RDMA_CM_ROUTE_QUERY;
2593 		work->new_state = RDMA_CM_ADDR_RESOLVED;
2594 		work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2595 		work->event.status = status;
2596 		pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2597 				     status);
2598 	}
2599 
2600 	queue_work(cma_wq, &work->work);
2601 }
2602 
2603 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2604 			      unsigned long timeout_ms, struct cma_work *work)
2605 {
2606 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2607 	struct sa_path_rec path_rec;
2608 	ib_sa_comp_mask comp_mask;
2609 	struct sockaddr_in6 *sin6;
2610 	struct sockaddr_ib *sib;
2611 
2612 	memset(&path_rec, 0, sizeof path_rec);
2613 
2614 	if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2615 		path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2616 	else
2617 		path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2618 	rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2619 	rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2620 	path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2621 	path_rec.numb_path = 1;
2622 	path_rec.reversible = 1;
2623 	path_rec.service_id = rdma_get_service_id(&id_priv->id,
2624 						  cma_dst_addr(id_priv));
2625 
2626 	comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2627 		    IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2628 		    IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2629 
2630 	switch (cma_family(id_priv)) {
2631 	case AF_INET:
2632 		path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2633 		comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2634 		break;
2635 	case AF_INET6:
2636 		sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2637 		path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2638 		comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2639 		break;
2640 	case AF_IB:
2641 		sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2642 		path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2643 		comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2644 		break;
2645 	}
2646 
2647 	id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2648 					       id_priv->id.port_num, &path_rec,
2649 					       comp_mask, timeout_ms,
2650 					       GFP_KERNEL, cma_query_handler,
2651 					       work, &id_priv->query);
2652 
2653 	return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2654 }
2655 
2656 static void cma_work_handler(struct work_struct *_work)
2657 {
2658 	struct cma_work *work = container_of(_work, struct cma_work, work);
2659 	struct rdma_id_private *id_priv = work->id;
2660 	int destroy = 0;
2661 
2662 	mutex_lock(&id_priv->handler_mutex);
2663 	if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2664 		goto out;
2665 
2666 	if (cma_cm_event_handler(id_priv, &work->event)) {
2667 		cma_exch(id_priv, RDMA_CM_DESTROYING);
2668 		destroy = 1;
2669 	}
2670 out:
2671 	mutex_unlock(&id_priv->handler_mutex);
2672 	cma_id_put(id_priv);
2673 	if (destroy)
2674 		rdma_destroy_id(&id_priv->id);
2675 	kfree(work);
2676 }
2677 
2678 static void cma_ndev_work_handler(struct work_struct *_work)
2679 {
2680 	struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2681 	struct rdma_id_private *id_priv = work->id;
2682 	int destroy = 0;
2683 
2684 	mutex_lock(&id_priv->handler_mutex);
2685 	if (id_priv->state == RDMA_CM_DESTROYING ||
2686 	    id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2687 		goto out;
2688 
2689 	if (cma_cm_event_handler(id_priv, &work->event)) {
2690 		cma_exch(id_priv, RDMA_CM_DESTROYING);
2691 		destroy = 1;
2692 	}
2693 
2694 out:
2695 	mutex_unlock(&id_priv->handler_mutex);
2696 	cma_id_put(id_priv);
2697 	if (destroy)
2698 		rdma_destroy_id(&id_priv->id);
2699 	kfree(work);
2700 }
2701 
2702 static void cma_init_resolve_route_work(struct cma_work *work,
2703 					struct rdma_id_private *id_priv)
2704 {
2705 	work->id = id_priv;
2706 	INIT_WORK(&work->work, cma_work_handler);
2707 	work->old_state = RDMA_CM_ROUTE_QUERY;
2708 	work->new_state = RDMA_CM_ROUTE_RESOLVED;
2709 	work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2710 }
2711 
2712 static void enqueue_resolve_addr_work(struct cma_work *work,
2713 				      struct rdma_id_private *id_priv)
2714 {
2715 	/* Balances with cma_id_put() in cma_work_handler */
2716 	cma_id_get(id_priv);
2717 
2718 	work->id = id_priv;
2719 	INIT_WORK(&work->work, cma_work_handler);
2720 	work->old_state = RDMA_CM_ADDR_QUERY;
2721 	work->new_state = RDMA_CM_ADDR_RESOLVED;
2722 	work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2723 
2724 	queue_work(cma_wq, &work->work);
2725 }
2726 
2727 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2728 				unsigned long timeout_ms)
2729 {
2730 	struct rdma_route *route = &id_priv->id.route;
2731 	struct cma_work *work;
2732 	int ret;
2733 
2734 	work = kzalloc(sizeof *work, GFP_KERNEL);
2735 	if (!work)
2736 		return -ENOMEM;
2737 
2738 	cma_init_resolve_route_work(work, id_priv);
2739 
2740 	route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2741 	if (!route->path_rec) {
2742 		ret = -ENOMEM;
2743 		goto err1;
2744 	}
2745 
2746 	ret = cma_query_ib_route(id_priv, timeout_ms, work);
2747 	if (ret)
2748 		goto err2;
2749 
2750 	return 0;
2751 err2:
2752 	kfree(route->path_rec);
2753 	route->path_rec = NULL;
2754 err1:
2755 	kfree(work);
2756 	return ret;
2757 }
2758 
2759 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2760 					   unsigned long supported_gids,
2761 					   enum ib_gid_type default_gid)
2762 {
2763 	if ((network_type == RDMA_NETWORK_IPV4 ||
2764 	     network_type == RDMA_NETWORK_IPV6) &&
2765 	    test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2766 		return IB_GID_TYPE_ROCE_UDP_ENCAP;
2767 
2768 	return default_gid;
2769 }
2770 
2771 /*
2772  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2773  * path record type based on GID type.
2774  * It also sets up other L2 fields which includes destination mac address
2775  * netdev ifindex, of the path record.
2776  * It returns the netdev of the bound interface for this path record entry.
2777  */
2778 static struct net_device *
2779 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2780 {
2781 	struct rdma_route *route = &id_priv->id.route;
2782 	enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2783 	struct rdma_addr *addr = &route->addr;
2784 	unsigned long supported_gids;
2785 	struct net_device *ndev;
2786 
2787 	if (!addr->dev_addr.bound_dev_if)
2788 		return NULL;
2789 
2790 	ndev = dev_get_by_index(addr->dev_addr.net,
2791 				addr->dev_addr.bound_dev_if);
2792 	if (!ndev)
2793 		return NULL;
2794 
2795 	supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2796 						    id_priv->id.port_num);
2797 	gid_type = cma_route_gid_type(addr->dev_addr.network,
2798 				      supported_gids,
2799 				      id_priv->gid_type);
2800 	/* Use the hint from IP Stack to select GID Type */
2801 	if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2802 		gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2803 	route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2804 
2805 	route->path_rec->roce.route_resolved = true;
2806 	sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2807 	return ndev;
2808 }
2809 
2810 int rdma_set_ib_path(struct rdma_cm_id *id,
2811 		     struct sa_path_rec *path_rec)
2812 {
2813 	struct rdma_id_private *id_priv;
2814 	struct net_device *ndev;
2815 	int ret;
2816 
2817 	id_priv = container_of(id, struct rdma_id_private, id);
2818 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2819 			   RDMA_CM_ROUTE_RESOLVED))
2820 		return -EINVAL;
2821 
2822 	id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2823 				     GFP_KERNEL);
2824 	if (!id->route.path_rec) {
2825 		ret = -ENOMEM;
2826 		goto err;
2827 	}
2828 
2829 	if (rdma_protocol_roce(id->device, id->port_num)) {
2830 		ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2831 		if (!ndev) {
2832 			ret = -ENODEV;
2833 			goto err_free;
2834 		}
2835 		dev_put(ndev);
2836 	}
2837 
2838 	id->route.num_paths = 1;
2839 	return 0;
2840 
2841 err_free:
2842 	kfree(id->route.path_rec);
2843 	id->route.path_rec = NULL;
2844 err:
2845 	cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2846 	return ret;
2847 }
2848 EXPORT_SYMBOL(rdma_set_ib_path);
2849 
2850 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
2851 {
2852 	struct cma_work *work;
2853 
2854 	work = kzalloc(sizeof *work, GFP_KERNEL);
2855 	if (!work)
2856 		return -ENOMEM;
2857 
2858 	cma_init_resolve_route_work(work, id_priv);
2859 	queue_work(cma_wq, &work->work);
2860 	return 0;
2861 }
2862 
2863 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
2864 {
2865 	struct net_device *dev;
2866 
2867 	dev = vlan_dev_real_dev(vlan_ndev);
2868 	if (dev->num_tc)
2869 		return netdev_get_prio_tc_map(dev, prio);
2870 
2871 	return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
2872 		VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2873 }
2874 
2875 struct iboe_prio_tc_map {
2876 	int input_prio;
2877 	int output_tc;
2878 	bool found;
2879 };
2880 
2881 static int get_lower_vlan_dev_tc(struct net_device *dev, void *data)
2882 {
2883 	struct iboe_prio_tc_map *map = data;
2884 
2885 	if (is_vlan_dev(dev))
2886 		map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
2887 	else if (dev->num_tc)
2888 		map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
2889 	else
2890 		map->output_tc = 0;
2891 	/* We are interested only in first level VLAN device, so always
2892 	 * return 1 to stop iterating over next level devices.
2893 	 */
2894 	map->found = true;
2895 	return 1;
2896 }
2897 
2898 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2899 {
2900 	struct iboe_prio_tc_map prio_tc_map = {};
2901 	int prio = rt_tos2priority(tos);
2902 
2903 	/* If VLAN device, get it directly from the VLAN netdev */
2904 	if (is_vlan_dev(ndev))
2905 		return get_vlan_ndev_tc(ndev, prio);
2906 
2907 	prio_tc_map.input_prio = prio;
2908 	rcu_read_lock();
2909 	netdev_walk_all_lower_dev_rcu(ndev,
2910 				      get_lower_vlan_dev_tc,
2911 				      &prio_tc_map);
2912 	rcu_read_unlock();
2913 	/* If map is found from lower device, use it; Otherwise
2914 	 * continue with the current netdevice to get priority to tc map.
2915 	 */
2916 	if (prio_tc_map.found)
2917 		return prio_tc_map.output_tc;
2918 	else if (ndev->num_tc)
2919 		return netdev_get_prio_tc_map(ndev, prio);
2920 	else
2921 		return 0;
2922 }
2923 
2924 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
2925 {
2926 	struct sockaddr_in6 *addr6;
2927 	u16 dport, sport;
2928 	u32 hash, fl;
2929 
2930 	addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
2931 	fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
2932 	if ((cma_family(id_priv) != AF_INET6) || !fl) {
2933 		dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
2934 		sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
2935 		hash = (u32)sport * 31 + dport;
2936 		fl = hash & IB_GRH_FLOWLABEL_MASK;
2937 	}
2938 
2939 	return cpu_to_be32(fl);
2940 }
2941 
2942 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2943 {
2944 	struct rdma_route *route = &id_priv->id.route;
2945 	struct rdma_addr *addr = &route->addr;
2946 	struct cma_work *work;
2947 	int ret;
2948 	struct net_device *ndev;
2949 
2950 	u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
2951 					rdma_start_port(id_priv->cma_dev->device)];
2952 	u8 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
2953 
2954 
2955 	work = kzalloc(sizeof *work, GFP_KERNEL);
2956 	if (!work)
2957 		return -ENOMEM;
2958 
2959 	route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2960 	if (!route->path_rec) {
2961 		ret = -ENOMEM;
2962 		goto err1;
2963 	}
2964 
2965 	route->num_paths = 1;
2966 
2967 	ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2968 	if (!ndev) {
2969 		ret = -ENODEV;
2970 		goto err2;
2971 	}
2972 
2973 	rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2974 		    &route->path_rec->sgid);
2975 	rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2976 		    &route->path_rec->dgid);
2977 
2978 	if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2979 		/* TODO: get the hoplimit from the inet/inet6 device */
2980 		route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2981 	else
2982 		route->path_rec->hop_limit = 1;
2983 	route->path_rec->reversible = 1;
2984 	route->path_rec->pkey = cpu_to_be16(0xffff);
2985 	route->path_rec->mtu_selector = IB_SA_EQ;
2986 	route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
2987 	route->path_rec->traffic_class = tos;
2988 	route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2989 	route->path_rec->rate_selector = IB_SA_EQ;
2990 	route->path_rec->rate = iboe_get_rate(ndev);
2991 	dev_put(ndev);
2992 	route->path_rec->packet_life_time_selector = IB_SA_EQ;
2993 	/* In case ACK timeout is set, use this value to calculate
2994 	 * PacketLifeTime.  As per IBTA 12.7.34,
2995 	 * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
2996 	 * Assuming a negligible local ACK delay, we can use
2997 	 * PacketLifeTime = local ACK timeout/2
2998 	 * as a reasonable approximation for RoCE networks.
2999 	 */
3000 	route->path_rec->packet_life_time = id_priv->timeout_set ?
3001 		id_priv->timeout - 1 : CMA_IBOE_PACKET_LIFETIME;
3002 
3003 	if (!route->path_rec->mtu) {
3004 		ret = -EINVAL;
3005 		goto err2;
3006 	}
3007 
3008 	if (rdma_protocol_roce_udp_encap(id_priv->id.device,
3009 					 id_priv->id.port_num))
3010 		route->path_rec->flow_label =
3011 			cma_get_roce_udp_flow_label(id_priv);
3012 
3013 	cma_init_resolve_route_work(work, id_priv);
3014 	queue_work(cma_wq, &work->work);
3015 
3016 	return 0;
3017 
3018 err2:
3019 	kfree(route->path_rec);
3020 	route->path_rec = NULL;
3021 	route->num_paths = 0;
3022 err1:
3023 	kfree(work);
3024 	return ret;
3025 }
3026 
3027 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3028 {
3029 	struct rdma_id_private *id_priv;
3030 	int ret;
3031 
3032 	id_priv = container_of(id, struct rdma_id_private, id);
3033 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3034 		return -EINVAL;
3035 
3036 	cma_id_get(id_priv);
3037 	if (rdma_cap_ib_sa(id->device, id->port_num))
3038 		ret = cma_resolve_ib_route(id_priv, timeout_ms);
3039 	else if (rdma_protocol_roce(id->device, id->port_num))
3040 		ret = cma_resolve_iboe_route(id_priv);
3041 	else if (rdma_protocol_iwarp(id->device, id->port_num))
3042 		ret = cma_resolve_iw_route(id_priv);
3043 	else
3044 		ret = -ENOSYS;
3045 
3046 	if (ret)
3047 		goto err;
3048 
3049 	return 0;
3050 err:
3051 	cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3052 	cma_id_put(id_priv);
3053 	return ret;
3054 }
3055 EXPORT_SYMBOL(rdma_resolve_route);
3056 
3057 static void cma_set_loopback(struct sockaddr *addr)
3058 {
3059 	switch (addr->sa_family) {
3060 	case AF_INET:
3061 		((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3062 		break;
3063 	case AF_INET6:
3064 		ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3065 			      0, 0, 0, htonl(1));
3066 		break;
3067 	default:
3068 		ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3069 			    0, 0, 0, htonl(1));
3070 		break;
3071 	}
3072 }
3073 
3074 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3075 {
3076 	struct cma_device *cma_dev, *cur_dev;
3077 	union ib_gid gid;
3078 	enum ib_port_state port_state;
3079 	unsigned int p;
3080 	u16 pkey;
3081 	int ret;
3082 
3083 	cma_dev = NULL;
3084 	mutex_lock(&lock);
3085 	list_for_each_entry(cur_dev, &dev_list, list) {
3086 		if (cma_family(id_priv) == AF_IB &&
3087 		    !rdma_cap_ib_cm(cur_dev->device, 1))
3088 			continue;
3089 
3090 		if (!cma_dev)
3091 			cma_dev = cur_dev;
3092 
3093 		rdma_for_each_port (cur_dev->device, p) {
3094 			if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3095 			    port_state == IB_PORT_ACTIVE) {
3096 				cma_dev = cur_dev;
3097 				goto port_found;
3098 			}
3099 		}
3100 	}
3101 
3102 	if (!cma_dev) {
3103 		ret = -ENODEV;
3104 		goto out;
3105 	}
3106 
3107 	p = 1;
3108 
3109 port_found:
3110 	ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3111 	if (ret)
3112 		goto out;
3113 
3114 	ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3115 	if (ret)
3116 		goto out;
3117 
3118 	id_priv->id.route.addr.dev_addr.dev_type =
3119 		(rdma_protocol_ib(cma_dev->device, p)) ?
3120 		ARPHRD_INFINIBAND : ARPHRD_ETHER;
3121 
3122 	rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3123 	ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3124 	id_priv->id.port_num = p;
3125 	cma_attach_to_dev(id_priv, cma_dev);
3126 	cma_set_loopback(cma_src_addr(id_priv));
3127 out:
3128 	mutex_unlock(&lock);
3129 	return ret;
3130 }
3131 
3132 static void addr_handler(int status, struct sockaddr *src_addr,
3133 			 struct rdma_dev_addr *dev_addr, void *context)
3134 {
3135 	struct rdma_id_private *id_priv = context;
3136 	struct rdma_cm_event event = {};
3137 	struct sockaddr *addr;
3138 	struct sockaddr_storage old_addr;
3139 
3140 	mutex_lock(&id_priv->handler_mutex);
3141 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3142 			   RDMA_CM_ADDR_RESOLVED))
3143 		goto out;
3144 
3145 	/*
3146 	 * Store the previous src address, so that if we fail to acquire
3147 	 * matching rdma device, old address can be restored back, which helps
3148 	 * to cancel the cma listen operation correctly.
3149 	 */
3150 	addr = cma_src_addr(id_priv);
3151 	memcpy(&old_addr, addr, rdma_addr_size(addr));
3152 	memcpy(addr, src_addr, rdma_addr_size(src_addr));
3153 	if (!status && !id_priv->cma_dev) {
3154 		status = cma_acquire_dev_by_src_ip(id_priv);
3155 		if (status)
3156 			pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3157 					     status);
3158 	} else if (status) {
3159 		pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3160 	}
3161 
3162 	if (status) {
3163 		memcpy(addr, &old_addr,
3164 		       rdma_addr_size((struct sockaddr *)&old_addr));
3165 		if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3166 				   RDMA_CM_ADDR_BOUND))
3167 			goto out;
3168 		event.event = RDMA_CM_EVENT_ADDR_ERROR;
3169 		event.status = status;
3170 	} else
3171 		event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3172 
3173 	if (cma_cm_event_handler(id_priv, &event)) {
3174 		cma_exch(id_priv, RDMA_CM_DESTROYING);
3175 		mutex_unlock(&id_priv->handler_mutex);
3176 		rdma_destroy_id(&id_priv->id);
3177 		return;
3178 	}
3179 out:
3180 	mutex_unlock(&id_priv->handler_mutex);
3181 }
3182 
3183 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3184 {
3185 	struct cma_work *work;
3186 	union ib_gid gid;
3187 	int ret;
3188 
3189 	work = kzalloc(sizeof *work, GFP_KERNEL);
3190 	if (!work)
3191 		return -ENOMEM;
3192 
3193 	if (!id_priv->cma_dev) {
3194 		ret = cma_bind_loopback(id_priv);
3195 		if (ret)
3196 			goto err;
3197 	}
3198 
3199 	rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3200 	rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3201 
3202 	enqueue_resolve_addr_work(work, id_priv);
3203 	return 0;
3204 err:
3205 	kfree(work);
3206 	return ret;
3207 }
3208 
3209 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3210 {
3211 	struct cma_work *work;
3212 	int ret;
3213 
3214 	work = kzalloc(sizeof *work, GFP_KERNEL);
3215 	if (!work)
3216 		return -ENOMEM;
3217 
3218 	if (!id_priv->cma_dev) {
3219 		ret = cma_resolve_ib_dev(id_priv);
3220 		if (ret)
3221 			goto err;
3222 	}
3223 
3224 	rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3225 		&(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3226 
3227 	enqueue_resolve_addr_work(work, id_priv);
3228 	return 0;
3229 err:
3230 	kfree(work);
3231 	return ret;
3232 }
3233 
3234 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3235 			 const struct sockaddr *dst_addr)
3236 {
3237 	if (!src_addr || !src_addr->sa_family) {
3238 		src_addr = (struct sockaddr *) &id->route.addr.src_addr;
3239 		src_addr->sa_family = dst_addr->sa_family;
3240 		if (IS_ENABLED(CONFIG_IPV6) &&
3241 		    dst_addr->sa_family == AF_INET6) {
3242 			struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
3243 			struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
3244 			src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3245 			if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3246 				id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
3247 		} else if (dst_addr->sa_family == AF_IB) {
3248 			((struct sockaddr_ib *) src_addr)->sib_pkey =
3249 				((struct sockaddr_ib *) dst_addr)->sib_pkey;
3250 		}
3251 	}
3252 	return rdma_bind_addr(id, src_addr);
3253 }
3254 
3255 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3256 		      const struct sockaddr *dst_addr, unsigned long timeout_ms)
3257 {
3258 	struct rdma_id_private *id_priv;
3259 	int ret;
3260 
3261 	id_priv = container_of(id, struct rdma_id_private, id);
3262 	memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3263 	if (id_priv->state == RDMA_CM_IDLE) {
3264 		ret = cma_bind_addr(id, src_addr, dst_addr);
3265 		if (ret) {
3266 			memset(cma_dst_addr(id_priv), 0,
3267 			       rdma_addr_size(dst_addr));
3268 			return ret;
3269 		}
3270 	}
3271 
3272 	if (cma_family(id_priv) != dst_addr->sa_family) {
3273 		memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3274 		return -EINVAL;
3275 	}
3276 
3277 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3278 		memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3279 		return -EINVAL;
3280 	}
3281 
3282 	if (cma_any_addr(dst_addr)) {
3283 		ret = cma_resolve_loopback(id_priv);
3284 	} else {
3285 		if (dst_addr->sa_family == AF_IB) {
3286 			ret = cma_resolve_ib_addr(id_priv);
3287 		} else {
3288 			ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3289 					      &id->route.addr.dev_addr,
3290 					      timeout_ms, addr_handler,
3291 					      false, id_priv);
3292 		}
3293 	}
3294 	if (ret)
3295 		goto err;
3296 
3297 	return 0;
3298 err:
3299 	cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3300 	return ret;
3301 }
3302 EXPORT_SYMBOL(rdma_resolve_addr);
3303 
3304 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3305 {
3306 	struct rdma_id_private *id_priv;
3307 	unsigned long flags;
3308 	int ret;
3309 
3310 	id_priv = container_of(id, struct rdma_id_private, id);
3311 	spin_lock_irqsave(&id_priv->lock, flags);
3312 	if (reuse || id_priv->state == RDMA_CM_IDLE) {
3313 		id_priv->reuseaddr = reuse;
3314 		ret = 0;
3315 	} else {
3316 		ret = -EINVAL;
3317 	}
3318 	spin_unlock_irqrestore(&id_priv->lock, flags);
3319 	return ret;
3320 }
3321 EXPORT_SYMBOL(rdma_set_reuseaddr);
3322 
3323 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3324 {
3325 	struct rdma_id_private *id_priv;
3326 	unsigned long flags;
3327 	int ret;
3328 
3329 	id_priv = container_of(id, struct rdma_id_private, id);
3330 	spin_lock_irqsave(&id_priv->lock, flags);
3331 	if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3332 		id_priv->options |= (1 << CMA_OPTION_AFONLY);
3333 		id_priv->afonly = afonly;
3334 		ret = 0;
3335 	} else {
3336 		ret = -EINVAL;
3337 	}
3338 	spin_unlock_irqrestore(&id_priv->lock, flags);
3339 	return ret;
3340 }
3341 EXPORT_SYMBOL(rdma_set_afonly);
3342 
3343 static void cma_bind_port(struct rdma_bind_list *bind_list,
3344 			  struct rdma_id_private *id_priv)
3345 {
3346 	struct sockaddr *addr;
3347 	struct sockaddr_ib *sib;
3348 	u64 sid, mask;
3349 	__be16 port;
3350 
3351 	lockdep_assert_held(&lock);
3352 
3353 	addr = cma_src_addr(id_priv);
3354 	port = htons(bind_list->port);
3355 
3356 	switch (addr->sa_family) {
3357 	case AF_INET:
3358 		((struct sockaddr_in *) addr)->sin_port = port;
3359 		break;
3360 	case AF_INET6:
3361 		((struct sockaddr_in6 *) addr)->sin6_port = port;
3362 		break;
3363 	case AF_IB:
3364 		sib = (struct sockaddr_ib *) addr;
3365 		sid = be64_to_cpu(sib->sib_sid);
3366 		mask = be64_to_cpu(sib->sib_sid_mask);
3367 		sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3368 		sib->sib_sid_mask = cpu_to_be64(~0ULL);
3369 		break;
3370 	}
3371 	id_priv->bind_list = bind_list;
3372 	hlist_add_head(&id_priv->node, &bind_list->owners);
3373 }
3374 
3375 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3376 			  struct rdma_id_private *id_priv, unsigned short snum)
3377 {
3378 	struct rdma_bind_list *bind_list;
3379 	int ret;
3380 
3381 	lockdep_assert_held(&lock);
3382 
3383 	bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3384 	if (!bind_list)
3385 		return -ENOMEM;
3386 
3387 	ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3388 			   snum);
3389 	if (ret < 0)
3390 		goto err;
3391 
3392 	bind_list->ps = ps;
3393 	bind_list->port = snum;
3394 	cma_bind_port(bind_list, id_priv);
3395 	return 0;
3396 err:
3397 	kfree(bind_list);
3398 	return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3399 }
3400 
3401 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3402 			      struct rdma_id_private *id_priv)
3403 {
3404 	struct rdma_id_private *cur_id;
3405 	struct sockaddr  *daddr = cma_dst_addr(id_priv);
3406 	struct sockaddr  *saddr = cma_src_addr(id_priv);
3407 	__be16 dport = cma_port(daddr);
3408 
3409 	lockdep_assert_held(&lock);
3410 
3411 	hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3412 		struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3413 		struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3414 		__be16 cur_dport = cma_port(cur_daddr);
3415 
3416 		if (id_priv == cur_id)
3417 			continue;
3418 
3419 		/* different dest port -> unique */
3420 		if (!cma_any_port(daddr) &&
3421 		    !cma_any_port(cur_daddr) &&
3422 		    (dport != cur_dport))
3423 			continue;
3424 
3425 		/* different src address -> unique */
3426 		if (!cma_any_addr(saddr) &&
3427 		    !cma_any_addr(cur_saddr) &&
3428 		    cma_addr_cmp(saddr, cur_saddr))
3429 			continue;
3430 
3431 		/* different dst address -> unique */
3432 		if (!cma_any_addr(daddr) &&
3433 		    !cma_any_addr(cur_daddr) &&
3434 		    cma_addr_cmp(daddr, cur_daddr))
3435 			continue;
3436 
3437 		return -EADDRNOTAVAIL;
3438 	}
3439 	return 0;
3440 }
3441 
3442 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3443 			      struct rdma_id_private *id_priv)
3444 {
3445 	static unsigned int last_used_port;
3446 	int low, high, remaining;
3447 	unsigned int rover;
3448 	struct net *net = id_priv->id.route.addr.dev_addr.net;
3449 
3450 	lockdep_assert_held(&lock);
3451 
3452 	inet_get_local_port_range(net, &low, &high);
3453 	remaining = (high - low) + 1;
3454 	rover = prandom_u32() % remaining + low;
3455 retry:
3456 	if (last_used_port != rover) {
3457 		struct rdma_bind_list *bind_list;
3458 		int ret;
3459 
3460 		bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3461 
3462 		if (!bind_list) {
3463 			ret = cma_alloc_port(ps, id_priv, rover);
3464 		} else {
3465 			ret = cma_port_is_unique(bind_list, id_priv);
3466 			if (!ret)
3467 				cma_bind_port(bind_list, id_priv);
3468 		}
3469 		/*
3470 		 * Remember previously used port number in order to avoid
3471 		 * re-using same port immediately after it is closed.
3472 		 */
3473 		if (!ret)
3474 			last_used_port = rover;
3475 		if (ret != -EADDRNOTAVAIL)
3476 			return ret;
3477 	}
3478 	if (--remaining) {
3479 		rover++;
3480 		if ((rover < low) || (rover > high))
3481 			rover = low;
3482 		goto retry;
3483 	}
3484 	return -EADDRNOTAVAIL;
3485 }
3486 
3487 /*
3488  * Check that the requested port is available.  This is called when trying to
3489  * bind to a specific port, or when trying to listen on a bound port.  In
3490  * the latter case, the provided id_priv may already be on the bind_list, but
3491  * we still need to check that it's okay to start listening.
3492  */
3493 static int cma_check_port(struct rdma_bind_list *bind_list,
3494 			  struct rdma_id_private *id_priv, uint8_t reuseaddr)
3495 {
3496 	struct rdma_id_private *cur_id;
3497 	struct sockaddr *addr, *cur_addr;
3498 
3499 	lockdep_assert_held(&lock);
3500 
3501 	addr = cma_src_addr(id_priv);
3502 	hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3503 		if (id_priv == cur_id)
3504 			continue;
3505 
3506 		if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
3507 		    cur_id->reuseaddr)
3508 			continue;
3509 
3510 		cur_addr = cma_src_addr(cur_id);
3511 		if (id_priv->afonly && cur_id->afonly &&
3512 		    (addr->sa_family != cur_addr->sa_family))
3513 			continue;
3514 
3515 		if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3516 			return -EADDRNOTAVAIL;
3517 
3518 		if (!cma_addr_cmp(addr, cur_addr))
3519 			return -EADDRINUSE;
3520 	}
3521 	return 0;
3522 }
3523 
3524 static int cma_use_port(enum rdma_ucm_port_space ps,
3525 			struct rdma_id_private *id_priv)
3526 {
3527 	struct rdma_bind_list *bind_list;
3528 	unsigned short snum;
3529 	int ret;
3530 
3531 	lockdep_assert_held(&lock);
3532 
3533 	snum = ntohs(cma_port(cma_src_addr(id_priv)));
3534 	if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3535 		return -EACCES;
3536 
3537 	bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3538 	if (!bind_list) {
3539 		ret = cma_alloc_port(ps, id_priv, snum);
3540 	} else {
3541 		ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3542 		if (!ret)
3543 			cma_bind_port(bind_list, id_priv);
3544 	}
3545 	return ret;
3546 }
3547 
3548 static int cma_bind_listen(struct rdma_id_private *id_priv)
3549 {
3550 	struct rdma_bind_list *bind_list = id_priv->bind_list;
3551 	int ret = 0;
3552 
3553 	mutex_lock(&lock);
3554 	if (bind_list->owners.first->next)
3555 		ret = cma_check_port(bind_list, id_priv, 0);
3556 	mutex_unlock(&lock);
3557 	return ret;
3558 }
3559 
3560 static enum rdma_ucm_port_space
3561 cma_select_inet_ps(struct rdma_id_private *id_priv)
3562 {
3563 	switch (id_priv->id.ps) {
3564 	case RDMA_PS_TCP:
3565 	case RDMA_PS_UDP:
3566 	case RDMA_PS_IPOIB:
3567 	case RDMA_PS_IB:
3568 		return id_priv->id.ps;
3569 	default:
3570 
3571 		return 0;
3572 	}
3573 }
3574 
3575 static enum rdma_ucm_port_space
3576 cma_select_ib_ps(struct rdma_id_private *id_priv)
3577 {
3578 	enum rdma_ucm_port_space ps = 0;
3579 	struct sockaddr_ib *sib;
3580 	u64 sid_ps, mask, sid;
3581 
3582 	sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3583 	mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3584 	sid = be64_to_cpu(sib->sib_sid) & mask;
3585 
3586 	if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3587 		sid_ps = RDMA_IB_IP_PS_IB;
3588 		ps = RDMA_PS_IB;
3589 	} else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3590 		   (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3591 		sid_ps = RDMA_IB_IP_PS_TCP;
3592 		ps = RDMA_PS_TCP;
3593 	} else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3594 		   (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3595 		sid_ps = RDMA_IB_IP_PS_UDP;
3596 		ps = RDMA_PS_UDP;
3597 	}
3598 
3599 	if (ps) {
3600 		sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3601 		sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3602 						be64_to_cpu(sib->sib_sid_mask));
3603 	}
3604 	return ps;
3605 }
3606 
3607 static int cma_get_port(struct rdma_id_private *id_priv)
3608 {
3609 	enum rdma_ucm_port_space ps;
3610 	int ret;
3611 
3612 	if (cma_family(id_priv) != AF_IB)
3613 		ps = cma_select_inet_ps(id_priv);
3614 	else
3615 		ps = cma_select_ib_ps(id_priv);
3616 	if (!ps)
3617 		return -EPROTONOSUPPORT;
3618 
3619 	mutex_lock(&lock);
3620 	if (cma_any_port(cma_src_addr(id_priv)))
3621 		ret = cma_alloc_any_port(ps, id_priv);
3622 	else
3623 		ret = cma_use_port(ps, id_priv);
3624 	mutex_unlock(&lock);
3625 
3626 	return ret;
3627 }
3628 
3629 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3630 			       struct sockaddr *addr)
3631 {
3632 #if IS_ENABLED(CONFIG_IPV6)
3633 	struct sockaddr_in6 *sin6;
3634 
3635 	if (addr->sa_family != AF_INET6)
3636 		return 0;
3637 
3638 	sin6 = (struct sockaddr_in6 *) addr;
3639 
3640 	if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3641 		return 0;
3642 
3643 	if (!sin6->sin6_scope_id)
3644 			return -EINVAL;
3645 
3646 	dev_addr->bound_dev_if = sin6->sin6_scope_id;
3647 #endif
3648 	return 0;
3649 }
3650 
3651 int rdma_listen(struct rdma_cm_id *id, int backlog)
3652 {
3653 	struct rdma_id_private *id_priv;
3654 	int ret;
3655 
3656 	id_priv = container_of(id, struct rdma_id_private, id);
3657 	if (id_priv->state == RDMA_CM_IDLE) {
3658 		id->route.addr.src_addr.ss_family = AF_INET;
3659 		ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3660 		if (ret)
3661 			return ret;
3662 	}
3663 
3664 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3665 		return -EINVAL;
3666 
3667 	if (id_priv->reuseaddr) {
3668 		ret = cma_bind_listen(id_priv);
3669 		if (ret)
3670 			goto err;
3671 	}
3672 
3673 	id_priv->backlog = backlog;
3674 	if (id->device) {
3675 		if (rdma_cap_ib_cm(id->device, 1)) {
3676 			ret = cma_ib_listen(id_priv);
3677 			if (ret)
3678 				goto err;
3679 		} else if (rdma_cap_iw_cm(id->device, 1)) {
3680 			ret = cma_iw_listen(id_priv, backlog);
3681 			if (ret)
3682 				goto err;
3683 		} else {
3684 			ret = -ENOSYS;
3685 			goto err;
3686 		}
3687 	} else
3688 		cma_listen_on_all(id_priv);
3689 
3690 	return 0;
3691 err:
3692 	id_priv->backlog = 0;
3693 	cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3694 	return ret;
3695 }
3696 EXPORT_SYMBOL(rdma_listen);
3697 
3698 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3699 {
3700 	struct rdma_id_private *id_priv;
3701 	int ret;
3702 	struct sockaddr  *daddr;
3703 
3704 	if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3705 	    addr->sa_family != AF_IB)
3706 		return -EAFNOSUPPORT;
3707 
3708 	id_priv = container_of(id, struct rdma_id_private, id);
3709 	if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3710 		return -EINVAL;
3711 
3712 	ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3713 	if (ret)
3714 		goto err1;
3715 
3716 	memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3717 	if (!cma_any_addr(addr)) {
3718 		ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3719 		if (ret)
3720 			goto err1;
3721 
3722 		ret = cma_acquire_dev_by_src_ip(id_priv);
3723 		if (ret)
3724 			goto err1;
3725 	}
3726 
3727 	if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3728 		if (addr->sa_family == AF_INET)
3729 			id_priv->afonly = 1;
3730 #if IS_ENABLED(CONFIG_IPV6)
3731 		else if (addr->sa_family == AF_INET6) {
3732 			struct net *net = id_priv->id.route.addr.dev_addr.net;
3733 
3734 			id_priv->afonly = net->ipv6.sysctl.bindv6only;
3735 		}
3736 #endif
3737 	}
3738 	daddr = cma_dst_addr(id_priv);
3739 	daddr->sa_family = addr->sa_family;
3740 
3741 	ret = cma_get_port(id_priv);
3742 	if (ret)
3743 		goto err2;
3744 
3745 	return 0;
3746 err2:
3747 	rdma_restrack_del(&id_priv->res);
3748 	if (id_priv->cma_dev)
3749 		cma_release_dev(id_priv);
3750 err1:
3751 	cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3752 	return ret;
3753 }
3754 EXPORT_SYMBOL(rdma_bind_addr);
3755 
3756 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3757 {
3758 	struct cma_hdr *cma_hdr;
3759 
3760 	cma_hdr = hdr;
3761 	cma_hdr->cma_version = CMA_VERSION;
3762 	if (cma_family(id_priv) == AF_INET) {
3763 		struct sockaddr_in *src4, *dst4;
3764 
3765 		src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3766 		dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3767 
3768 		cma_set_ip_ver(cma_hdr, 4);
3769 		cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3770 		cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3771 		cma_hdr->port = src4->sin_port;
3772 	} else if (cma_family(id_priv) == AF_INET6) {
3773 		struct sockaddr_in6 *src6, *dst6;
3774 
3775 		src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3776 		dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3777 
3778 		cma_set_ip_ver(cma_hdr, 6);
3779 		cma_hdr->src_addr.ip6 = src6->sin6_addr;
3780 		cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3781 		cma_hdr->port = src6->sin6_port;
3782 	}
3783 	return 0;
3784 }
3785 
3786 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3787 				const struct ib_cm_event *ib_event)
3788 {
3789 	struct rdma_id_private *id_priv = cm_id->context;
3790 	struct rdma_cm_event event = {};
3791 	const struct ib_cm_sidr_rep_event_param *rep =
3792 				&ib_event->param.sidr_rep_rcvd;
3793 	int ret = 0;
3794 
3795 	mutex_lock(&id_priv->handler_mutex);
3796 	if (id_priv->state != RDMA_CM_CONNECT)
3797 		goto out;
3798 
3799 	switch (ib_event->event) {
3800 	case IB_CM_SIDR_REQ_ERROR:
3801 		event.event = RDMA_CM_EVENT_UNREACHABLE;
3802 		event.status = -ETIMEDOUT;
3803 		break;
3804 	case IB_CM_SIDR_REP_RECEIVED:
3805 		event.param.ud.private_data = ib_event->private_data;
3806 		event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3807 		if (rep->status != IB_SIDR_SUCCESS) {
3808 			event.event = RDMA_CM_EVENT_UNREACHABLE;
3809 			event.status = ib_event->param.sidr_rep_rcvd.status;
3810 			pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3811 					     event.status);
3812 			break;
3813 		}
3814 		ret = cma_set_qkey(id_priv, rep->qkey);
3815 		if (ret) {
3816 			pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3817 			event.event = RDMA_CM_EVENT_ADDR_ERROR;
3818 			event.status = ret;
3819 			break;
3820 		}
3821 		ib_init_ah_attr_from_path(id_priv->id.device,
3822 					  id_priv->id.port_num,
3823 					  id_priv->id.route.path_rec,
3824 					  &event.param.ud.ah_attr,
3825 					  rep->sgid_attr);
3826 		event.param.ud.qp_num = rep->qpn;
3827 		event.param.ud.qkey = rep->qkey;
3828 		event.event = RDMA_CM_EVENT_ESTABLISHED;
3829 		event.status = 0;
3830 		break;
3831 	default:
3832 		pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3833 		       ib_event->event);
3834 		goto out;
3835 	}
3836 
3837 	ret = cma_cm_event_handler(id_priv, &event);
3838 
3839 	rdma_destroy_ah_attr(&event.param.ud.ah_attr);
3840 	if (ret) {
3841 		/* Destroy the CM ID by returning a non-zero value. */
3842 		id_priv->cm_id.ib = NULL;
3843 		cma_exch(id_priv, RDMA_CM_DESTROYING);
3844 		mutex_unlock(&id_priv->handler_mutex);
3845 		rdma_destroy_id(&id_priv->id);
3846 		return ret;
3847 	}
3848 out:
3849 	mutex_unlock(&id_priv->handler_mutex);
3850 	return ret;
3851 }
3852 
3853 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3854 			      struct rdma_conn_param *conn_param)
3855 {
3856 	struct ib_cm_sidr_req_param req;
3857 	struct ib_cm_id	*id;
3858 	void *private_data;
3859 	u8 offset;
3860 	int ret;
3861 
3862 	memset(&req, 0, sizeof req);
3863 	offset = cma_user_data_offset(id_priv);
3864 	req.private_data_len = offset + conn_param->private_data_len;
3865 	if (req.private_data_len < conn_param->private_data_len)
3866 		return -EINVAL;
3867 
3868 	if (req.private_data_len) {
3869 		private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3870 		if (!private_data)
3871 			return -ENOMEM;
3872 	} else {
3873 		private_data = NULL;
3874 	}
3875 
3876 	if (conn_param->private_data && conn_param->private_data_len)
3877 		memcpy(private_data + offset, conn_param->private_data,
3878 		       conn_param->private_data_len);
3879 
3880 	if (private_data) {
3881 		ret = cma_format_hdr(private_data, id_priv);
3882 		if (ret)
3883 			goto out;
3884 		req.private_data = private_data;
3885 	}
3886 
3887 	id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3888 			     id_priv);
3889 	if (IS_ERR(id)) {
3890 		ret = PTR_ERR(id);
3891 		goto out;
3892 	}
3893 	id_priv->cm_id.ib = id;
3894 
3895 	req.path = id_priv->id.route.path_rec;
3896 	req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3897 	req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3898 	req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3899 	req.max_cm_retries = CMA_MAX_CM_RETRIES;
3900 
3901 	trace_cm_send_sidr_req(id_priv);
3902 	ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3903 	if (ret) {
3904 		ib_destroy_cm_id(id_priv->cm_id.ib);
3905 		id_priv->cm_id.ib = NULL;
3906 	}
3907 out:
3908 	kfree(private_data);
3909 	return ret;
3910 }
3911 
3912 static int cma_connect_ib(struct rdma_id_private *id_priv,
3913 			  struct rdma_conn_param *conn_param)
3914 {
3915 	struct ib_cm_req_param req;
3916 	struct rdma_route *route;
3917 	void *private_data;
3918 	struct ib_cm_id	*id;
3919 	u8 offset;
3920 	int ret;
3921 
3922 	memset(&req, 0, sizeof req);
3923 	offset = cma_user_data_offset(id_priv);
3924 	req.private_data_len = offset + conn_param->private_data_len;
3925 	if (req.private_data_len < conn_param->private_data_len)
3926 		return -EINVAL;
3927 
3928 	if (req.private_data_len) {
3929 		private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3930 		if (!private_data)
3931 			return -ENOMEM;
3932 	} else {
3933 		private_data = NULL;
3934 	}
3935 
3936 	if (conn_param->private_data && conn_param->private_data_len)
3937 		memcpy(private_data + offset, conn_param->private_data,
3938 		       conn_param->private_data_len);
3939 
3940 	id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3941 	if (IS_ERR(id)) {
3942 		ret = PTR_ERR(id);
3943 		goto out;
3944 	}
3945 	id_priv->cm_id.ib = id;
3946 
3947 	route = &id_priv->id.route;
3948 	if (private_data) {
3949 		ret = cma_format_hdr(private_data, id_priv);
3950 		if (ret)
3951 			goto out;
3952 		req.private_data = private_data;
3953 	}
3954 
3955 	req.primary_path = &route->path_rec[0];
3956 	if (route->num_paths == 2)
3957 		req.alternate_path = &route->path_rec[1];
3958 
3959 	req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3960 	/* Alternate path SGID attribute currently unsupported */
3961 	req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3962 	req.qp_num = id_priv->qp_num;
3963 	req.qp_type = id_priv->id.qp_type;
3964 	req.starting_psn = id_priv->seq_num;
3965 	req.responder_resources = conn_param->responder_resources;
3966 	req.initiator_depth = conn_param->initiator_depth;
3967 	req.flow_control = conn_param->flow_control;
3968 	req.retry_count = min_t(u8, 7, conn_param->retry_count);
3969 	req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3970 	req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3971 	req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3972 	req.max_cm_retries = CMA_MAX_CM_RETRIES;
3973 	req.srq = id_priv->srq ? 1 : 0;
3974 	req.ece.vendor_id = id_priv->ece.vendor_id;
3975 	req.ece.attr_mod = id_priv->ece.attr_mod;
3976 
3977 	trace_cm_send_req(id_priv);
3978 	ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3979 out:
3980 	if (ret && !IS_ERR(id)) {
3981 		ib_destroy_cm_id(id);
3982 		id_priv->cm_id.ib = NULL;
3983 	}
3984 
3985 	kfree(private_data);
3986 	return ret;
3987 }
3988 
3989 static int cma_connect_iw(struct rdma_id_private *id_priv,
3990 			  struct rdma_conn_param *conn_param)
3991 {
3992 	struct iw_cm_id *cm_id;
3993 	int ret;
3994 	struct iw_cm_conn_param iw_param;
3995 
3996 	cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3997 	if (IS_ERR(cm_id))
3998 		return PTR_ERR(cm_id);
3999 
4000 	cm_id->tos = id_priv->tos;
4001 	cm_id->tos_set = id_priv->tos_set;
4002 	id_priv->cm_id.iw = cm_id;
4003 
4004 	memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4005 	       rdma_addr_size(cma_src_addr(id_priv)));
4006 	memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4007 	       rdma_addr_size(cma_dst_addr(id_priv)));
4008 
4009 	ret = cma_modify_qp_rtr(id_priv, conn_param);
4010 	if (ret)
4011 		goto out;
4012 
4013 	if (conn_param) {
4014 		iw_param.ord = conn_param->initiator_depth;
4015 		iw_param.ird = conn_param->responder_resources;
4016 		iw_param.private_data = conn_param->private_data;
4017 		iw_param.private_data_len = conn_param->private_data_len;
4018 		iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4019 	} else {
4020 		memset(&iw_param, 0, sizeof iw_param);
4021 		iw_param.qpn = id_priv->qp_num;
4022 	}
4023 	ret = iw_cm_connect(cm_id, &iw_param);
4024 out:
4025 	if (ret) {
4026 		iw_destroy_cm_id(cm_id);
4027 		id_priv->cm_id.iw = NULL;
4028 	}
4029 	return ret;
4030 }
4031 
4032 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4033 {
4034 	struct rdma_id_private *id_priv;
4035 	int ret;
4036 
4037 	id_priv = container_of(id, struct rdma_id_private, id);
4038 	if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4039 		return -EINVAL;
4040 
4041 	if (!id->qp) {
4042 		id_priv->qp_num = conn_param->qp_num;
4043 		id_priv->srq = conn_param->srq;
4044 	}
4045 
4046 	if (rdma_cap_ib_cm(id->device, id->port_num)) {
4047 		if (id->qp_type == IB_QPT_UD)
4048 			ret = cma_resolve_ib_udp(id_priv, conn_param);
4049 		else
4050 			ret = cma_connect_ib(id_priv, conn_param);
4051 	} else if (rdma_cap_iw_cm(id->device, id->port_num))
4052 		ret = cma_connect_iw(id_priv, conn_param);
4053 	else
4054 		ret = -ENOSYS;
4055 	if (ret)
4056 		goto err;
4057 
4058 	return 0;
4059 err:
4060 	cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4061 	return ret;
4062 }
4063 EXPORT_SYMBOL(rdma_connect);
4064 
4065 /**
4066  * rdma_connect_ece - Initiate an active connection request with ECE data.
4067  * @id: Connection identifier to connect.
4068  * @conn_param: Connection information used for connected QPs.
4069  * @ece: ECE parameters
4070  *
4071  * See rdma_connect() explanation.
4072  */
4073 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4074 		     struct rdma_ucm_ece *ece)
4075 {
4076 	struct rdma_id_private *id_priv =
4077 		container_of(id, struct rdma_id_private, id);
4078 
4079 	id_priv->ece.vendor_id = ece->vendor_id;
4080 	id_priv->ece.attr_mod = ece->attr_mod;
4081 
4082 	return rdma_connect(id, conn_param);
4083 }
4084 EXPORT_SYMBOL(rdma_connect_ece);
4085 
4086 static int cma_accept_ib(struct rdma_id_private *id_priv,
4087 			 struct rdma_conn_param *conn_param)
4088 {
4089 	struct ib_cm_rep_param rep;
4090 	int ret;
4091 
4092 	ret = cma_modify_qp_rtr(id_priv, conn_param);
4093 	if (ret)
4094 		goto out;
4095 
4096 	ret = cma_modify_qp_rts(id_priv, conn_param);
4097 	if (ret)
4098 		goto out;
4099 
4100 	memset(&rep, 0, sizeof rep);
4101 	rep.qp_num = id_priv->qp_num;
4102 	rep.starting_psn = id_priv->seq_num;
4103 	rep.private_data = conn_param->private_data;
4104 	rep.private_data_len = conn_param->private_data_len;
4105 	rep.responder_resources = conn_param->responder_resources;
4106 	rep.initiator_depth = conn_param->initiator_depth;
4107 	rep.failover_accepted = 0;
4108 	rep.flow_control = conn_param->flow_control;
4109 	rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4110 	rep.srq = id_priv->srq ? 1 : 0;
4111 	rep.ece.vendor_id = id_priv->ece.vendor_id;
4112 	rep.ece.attr_mod = id_priv->ece.attr_mod;
4113 
4114 	trace_cm_send_rep(id_priv);
4115 	ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4116 out:
4117 	return ret;
4118 }
4119 
4120 static int cma_accept_iw(struct rdma_id_private *id_priv,
4121 		  struct rdma_conn_param *conn_param)
4122 {
4123 	struct iw_cm_conn_param iw_param;
4124 	int ret;
4125 
4126 	if (!conn_param)
4127 		return -EINVAL;
4128 
4129 	ret = cma_modify_qp_rtr(id_priv, conn_param);
4130 	if (ret)
4131 		return ret;
4132 
4133 	iw_param.ord = conn_param->initiator_depth;
4134 	iw_param.ird = conn_param->responder_resources;
4135 	iw_param.private_data = conn_param->private_data;
4136 	iw_param.private_data_len = conn_param->private_data_len;
4137 	if (id_priv->id.qp) {
4138 		iw_param.qpn = id_priv->qp_num;
4139 	} else
4140 		iw_param.qpn = conn_param->qp_num;
4141 
4142 	return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4143 }
4144 
4145 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4146 			     enum ib_cm_sidr_status status, u32 qkey,
4147 			     const void *private_data, int private_data_len)
4148 {
4149 	struct ib_cm_sidr_rep_param rep;
4150 	int ret;
4151 
4152 	memset(&rep, 0, sizeof rep);
4153 	rep.status = status;
4154 	if (status == IB_SIDR_SUCCESS) {
4155 		ret = cma_set_qkey(id_priv, qkey);
4156 		if (ret)
4157 			return ret;
4158 		rep.qp_num = id_priv->qp_num;
4159 		rep.qkey = id_priv->qkey;
4160 
4161 		rep.ece.vendor_id = id_priv->ece.vendor_id;
4162 		rep.ece.attr_mod = id_priv->ece.attr_mod;
4163 	}
4164 
4165 	rep.private_data = private_data;
4166 	rep.private_data_len = private_data_len;
4167 
4168 	trace_cm_send_sidr_rep(id_priv);
4169 	return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4170 }
4171 
4172 int __rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4173 		  const char *caller)
4174 {
4175 	struct rdma_id_private *id_priv;
4176 	int ret;
4177 
4178 	id_priv = container_of(id, struct rdma_id_private, id);
4179 
4180 	rdma_restrack_set_task(&id_priv->res, caller);
4181 
4182 	if (!cma_comp(id_priv, RDMA_CM_CONNECT))
4183 		return -EINVAL;
4184 
4185 	if (!id->qp && conn_param) {
4186 		id_priv->qp_num = conn_param->qp_num;
4187 		id_priv->srq = conn_param->srq;
4188 	}
4189 
4190 	if (rdma_cap_ib_cm(id->device, id->port_num)) {
4191 		if (id->qp_type == IB_QPT_UD) {
4192 			if (conn_param)
4193 				ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4194 							conn_param->qkey,
4195 							conn_param->private_data,
4196 							conn_param->private_data_len);
4197 			else
4198 				ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4199 							0, NULL, 0);
4200 		} else {
4201 			if (conn_param)
4202 				ret = cma_accept_ib(id_priv, conn_param);
4203 			else
4204 				ret = cma_rep_recv(id_priv);
4205 		}
4206 	} else if (rdma_cap_iw_cm(id->device, id->port_num))
4207 		ret = cma_accept_iw(id_priv, conn_param);
4208 	else
4209 		ret = -ENOSYS;
4210 
4211 	if (ret)
4212 		goto reject;
4213 
4214 	return 0;
4215 reject:
4216 	cma_modify_qp_err(id_priv);
4217 	rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4218 	return ret;
4219 }
4220 EXPORT_SYMBOL(__rdma_accept);
4221 
4222 int __rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4223 		      const char *caller, struct rdma_ucm_ece *ece)
4224 {
4225 	struct rdma_id_private *id_priv =
4226 		container_of(id, struct rdma_id_private, id);
4227 
4228 	id_priv->ece.vendor_id = ece->vendor_id;
4229 	id_priv->ece.attr_mod = ece->attr_mod;
4230 
4231 	return __rdma_accept(id, conn_param, caller);
4232 }
4233 EXPORT_SYMBOL(__rdma_accept_ece);
4234 
4235 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4236 {
4237 	struct rdma_id_private *id_priv;
4238 	int ret;
4239 
4240 	id_priv = container_of(id, struct rdma_id_private, id);
4241 	if (!id_priv->cm_id.ib)
4242 		return -EINVAL;
4243 
4244 	switch (id->device->node_type) {
4245 	case RDMA_NODE_IB_CA:
4246 		ret = ib_cm_notify(id_priv->cm_id.ib, event);
4247 		break;
4248 	default:
4249 		ret = 0;
4250 		break;
4251 	}
4252 	return ret;
4253 }
4254 EXPORT_SYMBOL(rdma_notify);
4255 
4256 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4257 		u8 private_data_len, u8 reason)
4258 {
4259 	struct rdma_id_private *id_priv;
4260 	int ret;
4261 
4262 	id_priv = container_of(id, struct rdma_id_private, id);
4263 	if (!id_priv->cm_id.ib)
4264 		return -EINVAL;
4265 
4266 	if (rdma_cap_ib_cm(id->device, id->port_num)) {
4267 		if (id->qp_type == IB_QPT_UD) {
4268 			ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4269 						private_data, private_data_len);
4270 		} else {
4271 			trace_cm_send_rej(id_priv);
4272 			ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4273 					     private_data, private_data_len);
4274 		}
4275 	} else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4276 		ret = iw_cm_reject(id_priv->cm_id.iw,
4277 				   private_data, private_data_len);
4278 	} else
4279 		ret = -ENOSYS;
4280 
4281 	return ret;
4282 }
4283 EXPORT_SYMBOL(rdma_reject);
4284 
4285 int rdma_disconnect(struct rdma_cm_id *id)
4286 {
4287 	struct rdma_id_private *id_priv;
4288 	int ret;
4289 
4290 	id_priv = container_of(id, struct rdma_id_private, id);
4291 	if (!id_priv->cm_id.ib)
4292 		return -EINVAL;
4293 
4294 	if (rdma_cap_ib_cm(id->device, id->port_num)) {
4295 		ret = cma_modify_qp_err(id_priv);
4296 		if (ret)
4297 			goto out;
4298 		/* Initiate or respond to a disconnect. */
4299 		trace_cm_disconnect(id_priv);
4300 		if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4301 			if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4302 				trace_cm_sent_drep(id_priv);
4303 		} else {
4304 			trace_cm_sent_dreq(id_priv);
4305 		}
4306 	} else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4307 		ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4308 	} else
4309 		ret = -EINVAL;
4310 
4311 out:
4312 	return ret;
4313 }
4314 EXPORT_SYMBOL(rdma_disconnect);
4315 
4316 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4317 {
4318 	struct rdma_id_private *id_priv;
4319 	struct cma_multicast *mc = multicast->context;
4320 	struct rdma_cm_event event = {};
4321 	int ret = 0;
4322 
4323 	id_priv = mc->id_priv;
4324 	mutex_lock(&id_priv->handler_mutex);
4325 	if (id_priv->state != RDMA_CM_ADDR_BOUND &&
4326 	    id_priv->state != RDMA_CM_ADDR_RESOLVED)
4327 		goto out;
4328 
4329 	if (!status)
4330 		status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4331 	else
4332 		pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4333 				     status);
4334 	mutex_lock(&id_priv->qp_mutex);
4335 	if (!status && id_priv->id.qp) {
4336 		status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
4337 					 be16_to_cpu(multicast->rec.mlid));
4338 		if (status)
4339 			pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to attach QP. status %d\n",
4340 					     status);
4341 	}
4342 	mutex_unlock(&id_priv->qp_mutex);
4343 
4344 	event.status = status;
4345 	event.param.ud.private_data = mc->context;
4346 	if (!status) {
4347 		struct rdma_dev_addr *dev_addr =
4348 			&id_priv->id.route.addr.dev_addr;
4349 		struct net_device *ndev =
4350 			dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4351 		enum ib_gid_type gid_type =
4352 			id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4353 			rdma_start_port(id_priv->cma_dev->device)];
4354 
4355 		event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
4356 		ret = ib_init_ah_from_mcmember(id_priv->id.device,
4357 					       id_priv->id.port_num,
4358 					       &multicast->rec,
4359 					       ndev, gid_type,
4360 					       &event.param.ud.ah_attr);
4361 		if (ret)
4362 			event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4363 
4364 		event.param.ud.qp_num = 0xFFFFFF;
4365 		event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4366 		if (ndev)
4367 			dev_put(ndev);
4368 	} else
4369 		event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4370 
4371 	ret = cma_cm_event_handler(id_priv, &event);
4372 
4373 	rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4374 	if (ret) {
4375 		cma_exch(id_priv, RDMA_CM_DESTROYING);
4376 		mutex_unlock(&id_priv->handler_mutex);
4377 		rdma_destroy_id(&id_priv->id);
4378 		return 0;
4379 	}
4380 
4381 out:
4382 	mutex_unlock(&id_priv->handler_mutex);
4383 	return 0;
4384 }
4385 
4386 static void cma_set_mgid(struct rdma_id_private *id_priv,
4387 			 struct sockaddr *addr, union ib_gid *mgid)
4388 {
4389 	unsigned char mc_map[MAX_ADDR_LEN];
4390 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4391 	struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4392 	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4393 
4394 	if (cma_any_addr(addr)) {
4395 		memset(mgid, 0, sizeof *mgid);
4396 	} else if ((addr->sa_family == AF_INET6) &&
4397 		   ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4398 								 0xFF10A01B)) {
4399 		/* IPv6 address is an SA assigned MGID. */
4400 		memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4401 	} else if (addr->sa_family == AF_IB) {
4402 		memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4403 	} else if (addr->sa_family == AF_INET6) {
4404 		ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4405 		if (id_priv->id.ps == RDMA_PS_UDP)
4406 			mc_map[7] = 0x01;	/* Use RDMA CM signature */
4407 		*mgid = *(union ib_gid *) (mc_map + 4);
4408 	} else {
4409 		ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4410 		if (id_priv->id.ps == RDMA_PS_UDP)
4411 			mc_map[7] = 0x01;	/* Use RDMA CM signature */
4412 		*mgid = *(union ib_gid *) (mc_map + 4);
4413 	}
4414 }
4415 
4416 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4417 				 struct cma_multicast *mc)
4418 {
4419 	struct ib_sa_mcmember_rec rec;
4420 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4421 	ib_sa_comp_mask comp_mask;
4422 	int ret;
4423 
4424 	ib_addr_get_mgid(dev_addr, &rec.mgid);
4425 	ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4426 				     &rec.mgid, &rec);
4427 	if (ret)
4428 		return ret;
4429 
4430 	ret = cma_set_qkey(id_priv, 0);
4431 	if (ret)
4432 		return ret;
4433 
4434 	cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4435 	rec.qkey = cpu_to_be32(id_priv->qkey);
4436 	rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4437 	rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4438 	rec.join_state = mc->join_state;
4439 
4440 	if ((rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) &&
4441 	    (!ib_sa_sendonly_fullmem_support(&sa_client,
4442 					     id_priv->id.device,
4443 					     id_priv->id.port_num))) {
4444 		dev_warn(
4445 			&id_priv->id.device->dev,
4446 			"RDMA CM: port %u Unable to multicast join: SM doesn't support Send Only Full Member option\n",
4447 			id_priv->id.port_num);
4448 		return -EOPNOTSUPP;
4449 	}
4450 
4451 	comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4452 		    IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4453 		    IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4454 		    IB_SA_MCMEMBER_REC_FLOW_LABEL |
4455 		    IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4456 
4457 	if (id_priv->id.ps == RDMA_PS_IPOIB)
4458 		comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4459 			     IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4460 			     IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4461 			     IB_SA_MCMEMBER_REC_MTU |
4462 			     IB_SA_MCMEMBER_REC_HOP_LIMIT;
4463 
4464 	mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4465 						id_priv->id.port_num, &rec,
4466 						comp_mask, GFP_KERNEL,
4467 						cma_ib_mc_handler, mc);
4468 	return PTR_ERR_OR_ZERO(mc->multicast.ib);
4469 }
4470 
4471 static void iboe_mcast_work_handler(struct work_struct *work)
4472 {
4473 	struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
4474 	struct cma_multicast *mc = mw->mc;
4475 	struct ib_sa_multicast *m = mc->multicast.ib;
4476 
4477 	mc->multicast.ib->context = mc;
4478 	cma_ib_mc_handler(0, m);
4479 	kref_put(&mc->mcref, release_mc);
4480 	kfree(mw);
4481 }
4482 
4483 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4484 			      enum ib_gid_type gid_type)
4485 {
4486 	struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4487 	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4488 
4489 	if (cma_any_addr(addr)) {
4490 		memset(mgid, 0, sizeof *mgid);
4491 	} else if (addr->sa_family == AF_INET6) {
4492 		memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4493 	} else {
4494 		mgid->raw[0] =
4495 			(gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4496 		mgid->raw[1] =
4497 			(gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4498 		mgid->raw[2] = 0;
4499 		mgid->raw[3] = 0;
4500 		mgid->raw[4] = 0;
4501 		mgid->raw[5] = 0;
4502 		mgid->raw[6] = 0;
4503 		mgid->raw[7] = 0;
4504 		mgid->raw[8] = 0;
4505 		mgid->raw[9] = 0;
4506 		mgid->raw[10] = 0xff;
4507 		mgid->raw[11] = 0xff;
4508 		*(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4509 	}
4510 }
4511 
4512 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4513 				   struct cma_multicast *mc)
4514 {
4515 	struct iboe_mcast_work *work;
4516 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4517 	int err = 0;
4518 	struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4519 	struct net_device *ndev = NULL;
4520 	enum ib_gid_type gid_type;
4521 	bool send_only;
4522 
4523 	send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4524 
4525 	if (cma_zero_addr((struct sockaddr *)&mc->addr))
4526 		return -EINVAL;
4527 
4528 	work = kzalloc(sizeof *work, GFP_KERNEL);
4529 	if (!work)
4530 		return -ENOMEM;
4531 
4532 	mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
4533 	if (!mc->multicast.ib) {
4534 		err = -ENOMEM;
4535 		goto out1;
4536 	}
4537 
4538 	gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4539 		   rdma_start_port(id_priv->cma_dev->device)];
4540 	cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid, gid_type);
4541 
4542 	mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
4543 	if (id_priv->id.ps == RDMA_PS_UDP)
4544 		mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4545 
4546 	if (dev_addr->bound_dev_if)
4547 		ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4548 	if (!ndev) {
4549 		err = -ENODEV;
4550 		goto out2;
4551 	}
4552 	mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
4553 	mc->multicast.ib->rec.hop_limit = 1;
4554 	mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
4555 
4556 	if (addr->sa_family == AF_INET) {
4557 		if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4558 			mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4559 			if (!send_only) {
4560 				err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
4561 						    true);
4562 			}
4563 		}
4564 	} else {
4565 		if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4566 			err = -ENOTSUPP;
4567 	}
4568 	dev_put(ndev);
4569 	if (err || !mc->multicast.ib->rec.mtu) {
4570 		if (!err)
4571 			err = -EINVAL;
4572 		goto out2;
4573 	}
4574 	rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4575 		    &mc->multicast.ib->rec.port_gid);
4576 	work->id = id_priv;
4577 	work->mc = mc;
4578 	INIT_WORK(&work->work, iboe_mcast_work_handler);
4579 	kref_get(&mc->mcref);
4580 	queue_work(cma_wq, &work->work);
4581 
4582 	return 0;
4583 
4584 out2:
4585 	kfree(mc->multicast.ib);
4586 out1:
4587 	kfree(work);
4588 	return err;
4589 }
4590 
4591 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4592 			u8 join_state, void *context)
4593 {
4594 	struct rdma_id_private *id_priv;
4595 	struct cma_multicast *mc;
4596 	int ret;
4597 
4598 	if (!id->device)
4599 		return -EINVAL;
4600 
4601 	id_priv = container_of(id, struct rdma_id_private, id);
4602 	if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4603 	    !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4604 		return -EINVAL;
4605 
4606 	mc = kmalloc(sizeof *mc, GFP_KERNEL);
4607 	if (!mc)
4608 		return -ENOMEM;
4609 
4610 	memcpy(&mc->addr, addr, rdma_addr_size(addr));
4611 	mc->context = context;
4612 	mc->id_priv = id_priv;
4613 	mc->join_state = join_state;
4614 
4615 	if (rdma_protocol_roce(id->device, id->port_num)) {
4616 		kref_init(&mc->mcref);
4617 		ret = cma_iboe_join_multicast(id_priv, mc);
4618 		if (ret)
4619 			goto out_err;
4620 	} else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4621 		ret = cma_join_ib_multicast(id_priv, mc);
4622 		if (ret)
4623 			goto out_err;
4624 	} else {
4625 		ret = -ENOSYS;
4626 		goto out_err;
4627 	}
4628 
4629 	spin_lock(&id_priv->lock);
4630 	list_add(&mc->list, &id_priv->mc_list);
4631 	spin_unlock(&id_priv->lock);
4632 
4633 	return 0;
4634 out_err:
4635 	kfree(mc);
4636 	return ret;
4637 }
4638 EXPORT_SYMBOL(rdma_join_multicast);
4639 
4640 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4641 {
4642 	struct rdma_id_private *id_priv;
4643 	struct cma_multicast *mc;
4644 
4645 	id_priv = container_of(id, struct rdma_id_private, id);
4646 	spin_lock_irq(&id_priv->lock);
4647 	list_for_each_entry(mc, &id_priv->mc_list, list) {
4648 		if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
4649 			list_del(&mc->list);
4650 			spin_unlock_irq(&id_priv->lock);
4651 
4652 			if (id->qp)
4653 				ib_detach_mcast(id->qp,
4654 						&mc->multicast.ib->rec.mgid,
4655 						be16_to_cpu(mc->multicast.ib->rec.mlid));
4656 
4657 			BUG_ON(id_priv->cma_dev->device != id->device);
4658 
4659 			if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4660 				ib_sa_free_multicast(mc->multicast.ib);
4661 				kfree(mc);
4662 			} else if (rdma_protocol_roce(id->device, id->port_num)) {
4663 				cma_leave_roce_mc_group(id_priv, mc);
4664 			}
4665 			return;
4666 		}
4667 	}
4668 	spin_unlock_irq(&id_priv->lock);
4669 }
4670 EXPORT_SYMBOL(rdma_leave_multicast);
4671 
4672 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4673 {
4674 	struct rdma_dev_addr *dev_addr;
4675 	struct cma_ndev_work *work;
4676 
4677 	dev_addr = &id_priv->id.route.addr.dev_addr;
4678 
4679 	if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4680 	    (net_eq(dev_net(ndev), dev_addr->net)) &&
4681 	    memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4682 		pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4683 			ndev->name, &id_priv->id);
4684 		work = kzalloc(sizeof *work, GFP_KERNEL);
4685 		if (!work)
4686 			return -ENOMEM;
4687 
4688 		INIT_WORK(&work->work, cma_ndev_work_handler);
4689 		work->id = id_priv;
4690 		work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4691 		cma_id_get(id_priv);
4692 		queue_work(cma_wq, &work->work);
4693 	}
4694 
4695 	return 0;
4696 }
4697 
4698 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4699 			       void *ptr)
4700 {
4701 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4702 	struct cma_device *cma_dev;
4703 	struct rdma_id_private *id_priv;
4704 	int ret = NOTIFY_DONE;
4705 
4706 	if (event != NETDEV_BONDING_FAILOVER)
4707 		return NOTIFY_DONE;
4708 
4709 	if (!netif_is_bond_master(ndev))
4710 		return NOTIFY_DONE;
4711 
4712 	mutex_lock(&lock);
4713 	list_for_each_entry(cma_dev, &dev_list, list)
4714 		list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4715 			ret = cma_netdev_change(ndev, id_priv);
4716 			if (ret)
4717 				goto out;
4718 		}
4719 
4720 out:
4721 	mutex_unlock(&lock);
4722 	return ret;
4723 }
4724 
4725 static struct notifier_block cma_nb = {
4726 	.notifier_call = cma_netdev_callback
4727 };
4728 
4729 static int cma_add_one(struct ib_device *device)
4730 {
4731 	struct cma_device *cma_dev;
4732 	struct rdma_id_private *id_priv;
4733 	unsigned int i;
4734 	unsigned long supported_gids = 0;
4735 	int ret;
4736 
4737 	cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4738 	if (!cma_dev)
4739 		return -ENOMEM;
4740 
4741 	cma_dev->device = device;
4742 	cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4743 					    sizeof(*cma_dev->default_gid_type),
4744 					    GFP_KERNEL);
4745 	if (!cma_dev->default_gid_type) {
4746 		ret = -ENOMEM;
4747 		goto free_cma_dev;
4748 	}
4749 
4750 	cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4751 					    sizeof(*cma_dev->default_roce_tos),
4752 					    GFP_KERNEL);
4753 	if (!cma_dev->default_roce_tos) {
4754 		ret = -ENOMEM;
4755 		goto free_gid_type;
4756 	}
4757 
4758 	rdma_for_each_port (device, i) {
4759 		supported_gids = roce_gid_type_mask_support(device, i);
4760 		WARN_ON(!supported_gids);
4761 		if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4762 			cma_dev->default_gid_type[i - rdma_start_port(device)] =
4763 				CMA_PREFERRED_ROCE_GID_TYPE;
4764 		else
4765 			cma_dev->default_gid_type[i - rdma_start_port(device)] =
4766 				find_first_bit(&supported_gids, BITS_PER_LONG);
4767 		cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4768 	}
4769 
4770 	init_completion(&cma_dev->comp);
4771 	refcount_set(&cma_dev->refcount, 1);
4772 	INIT_LIST_HEAD(&cma_dev->id_list);
4773 	ib_set_client_data(device, &cma_client, cma_dev);
4774 
4775 	mutex_lock(&lock);
4776 	list_add_tail(&cma_dev->list, &dev_list);
4777 	list_for_each_entry(id_priv, &listen_any_list, list)
4778 		cma_listen_on_dev(id_priv, cma_dev);
4779 	mutex_unlock(&lock);
4780 
4781 	trace_cm_add_one(device);
4782 	return 0;
4783 
4784 free_gid_type:
4785 	kfree(cma_dev->default_gid_type);
4786 
4787 free_cma_dev:
4788 	kfree(cma_dev);
4789 	return ret;
4790 }
4791 
4792 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4793 {
4794 	struct rdma_cm_event event = {};
4795 	enum rdma_cm_state state;
4796 	int ret = 0;
4797 
4798 	/* Record that we want to remove the device */
4799 	state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4800 	if (state == RDMA_CM_DESTROYING)
4801 		return 0;
4802 
4803 	cma_cancel_operation(id_priv, state);
4804 	mutex_lock(&id_priv->handler_mutex);
4805 
4806 	/* Check for destruction from another callback. */
4807 	if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4808 		goto out;
4809 
4810 	event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4811 	ret = cma_cm_event_handler(id_priv, &event);
4812 out:
4813 	mutex_unlock(&id_priv->handler_mutex);
4814 	return ret;
4815 }
4816 
4817 static void cma_process_remove(struct cma_device *cma_dev)
4818 {
4819 	struct rdma_id_private *id_priv;
4820 	int ret;
4821 
4822 	mutex_lock(&lock);
4823 	while (!list_empty(&cma_dev->id_list)) {
4824 		id_priv = list_entry(cma_dev->id_list.next,
4825 				     struct rdma_id_private, list);
4826 
4827 		list_del(&id_priv->listen_list);
4828 		list_del_init(&id_priv->list);
4829 		cma_id_get(id_priv);
4830 		mutex_unlock(&lock);
4831 
4832 		ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4833 		cma_id_put(id_priv);
4834 		if (ret)
4835 			rdma_destroy_id(&id_priv->id);
4836 
4837 		mutex_lock(&lock);
4838 	}
4839 	mutex_unlock(&lock);
4840 
4841 	cma_dev_put(cma_dev);
4842 	wait_for_completion(&cma_dev->comp);
4843 }
4844 
4845 static void cma_remove_one(struct ib_device *device, void *client_data)
4846 {
4847 	struct cma_device *cma_dev = client_data;
4848 
4849 	trace_cm_remove_one(device);
4850 
4851 	mutex_lock(&lock);
4852 	list_del(&cma_dev->list);
4853 	mutex_unlock(&lock);
4854 
4855 	cma_process_remove(cma_dev);
4856 	kfree(cma_dev->default_roce_tos);
4857 	kfree(cma_dev->default_gid_type);
4858 	kfree(cma_dev);
4859 }
4860 
4861 static int cma_init_net(struct net *net)
4862 {
4863 	struct cma_pernet *pernet = cma_pernet(net);
4864 
4865 	xa_init(&pernet->tcp_ps);
4866 	xa_init(&pernet->udp_ps);
4867 	xa_init(&pernet->ipoib_ps);
4868 	xa_init(&pernet->ib_ps);
4869 
4870 	return 0;
4871 }
4872 
4873 static void cma_exit_net(struct net *net)
4874 {
4875 	struct cma_pernet *pernet = cma_pernet(net);
4876 
4877 	WARN_ON(!xa_empty(&pernet->tcp_ps));
4878 	WARN_ON(!xa_empty(&pernet->udp_ps));
4879 	WARN_ON(!xa_empty(&pernet->ipoib_ps));
4880 	WARN_ON(!xa_empty(&pernet->ib_ps));
4881 }
4882 
4883 static struct pernet_operations cma_pernet_operations = {
4884 	.init = cma_init_net,
4885 	.exit = cma_exit_net,
4886 	.id = &cma_pernet_id,
4887 	.size = sizeof(struct cma_pernet),
4888 };
4889 
4890 static int __init cma_init(void)
4891 {
4892 	int ret;
4893 
4894 	/*
4895 	 * There is a rare lock ordering dependency in cma_netdev_callback()
4896 	 * that only happens when bonding is enabled. Teach lockdep that rtnl
4897 	 * must never be nested under lock so it can find these without having
4898 	 * to test with bonding.
4899 	 */
4900 	if (IS_ENABLED(CONFIG_LOCKDEP)) {
4901 		rtnl_lock();
4902 		mutex_lock(&lock);
4903 		mutex_unlock(&lock);
4904 		rtnl_unlock();
4905 	}
4906 
4907 	cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4908 	if (!cma_wq)
4909 		return -ENOMEM;
4910 
4911 	ret = register_pernet_subsys(&cma_pernet_operations);
4912 	if (ret)
4913 		goto err_wq;
4914 
4915 	ib_sa_register_client(&sa_client);
4916 	register_netdevice_notifier(&cma_nb);
4917 
4918 	ret = ib_register_client(&cma_client);
4919 	if (ret)
4920 		goto err;
4921 
4922 	ret = cma_configfs_init();
4923 	if (ret)
4924 		goto err_ib;
4925 
4926 	return 0;
4927 
4928 err_ib:
4929 	ib_unregister_client(&cma_client);
4930 err:
4931 	unregister_netdevice_notifier(&cma_nb);
4932 	ib_sa_unregister_client(&sa_client);
4933 	unregister_pernet_subsys(&cma_pernet_operations);
4934 err_wq:
4935 	destroy_workqueue(cma_wq);
4936 	return ret;
4937 }
4938 
4939 static void __exit cma_cleanup(void)
4940 {
4941 	cma_configfs_exit();
4942 	ib_unregister_client(&cma_client);
4943 	unregister_netdevice_notifier(&cma_nb);
4944 	ib_sa_unregister_client(&sa_client);
4945 	unregister_pernet_subsys(&cma_pernet_operations);
4946 	destroy_workqueue(cma_wq);
4947 }
4948 
4949 module_init(cma_init);
4950 module_exit(cma_cleanup);
4951