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/rbtree.h>
15 #include <linux/igmp.h>
16 #include <linux/xarray.h>
17 #include <linux/inetdevice.h>
18 #include <linux/slab.h>
19 #include <linux/module.h>
20 #include <net/route.h>
21 
22 #include <net/net_namespace.h>
23 #include <net/netns/generic.h>
24 #include <net/netevent.h>
25 #include <net/tcp.h>
26 #include <net/ipv6.h>
27 #include <net/ip_fib.h>
28 #include <net/ip6_route.h>
29 
30 #include <rdma/rdma_cm.h>
31 #include <rdma/rdma_cm_ib.h>
32 #include <rdma/rdma_netlink.h>
33 #include <rdma/ib.h>
34 #include <rdma/ib_cache.h>
35 #include <rdma/ib_cm.h>
36 #include <rdma/ib_sa.h>
37 #include <rdma/iw_cm.h>
38 
39 #include "core_priv.h"
40 #include "cma_priv.h"
41 #include "cma_trace.h"
42 
43 MODULE_AUTHOR("Sean Hefty");
44 MODULE_DESCRIPTION("Generic RDMA CM Agent");
45 MODULE_LICENSE("Dual BSD/GPL");
46 
47 #define CMA_CM_RESPONSE_TIMEOUT 20
48 #define CMA_MAX_CM_RETRIES 15
49 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
50 #define CMA_IBOE_PACKET_LIFETIME 18
51 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
52 
53 static const char * const cma_events[] = {
54 	[RDMA_CM_EVENT_ADDR_RESOLVED]	 = "address resolved",
55 	[RDMA_CM_EVENT_ADDR_ERROR]	 = "address error",
56 	[RDMA_CM_EVENT_ROUTE_RESOLVED]	 = "route resolved ",
57 	[RDMA_CM_EVENT_ROUTE_ERROR]	 = "route error",
58 	[RDMA_CM_EVENT_CONNECT_REQUEST]	 = "connect request",
59 	[RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
60 	[RDMA_CM_EVENT_CONNECT_ERROR]	 = "connect error",
61 	[RDMA_CM_EVENT_UNREACHABLE]	 = "unreachable",
62 	[RDMA_CM_EVENT_REJECTED]	 = "rejected",
63 	[RDMA_CM_EVENT_ESTABLISHED]	 = "established",
64 	[RDMA_CM_EVENT_DISCONNECTED]	 = "disconnected",
65 	[RDMA_CM_EVENT_DEVICE_REMOVAL]	 = "device removal",
66 	[RDMA_CM_EVENT_MULTICAST_JOIN]	 = "multicast join",
67 	[RDMA_CM_EVENT_MULTICAST_ERROR]	 = "multicast error",
68 	[RDMA_CM_EVENT_ADDR_CHANGE]	 = "address change",
69 	[RDMA_CM_EVENT_TIMEWAIT_EXIT]	 = "timewait exit",
70 };
71 
72 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
73 			      enum ib_gid_type gid_type);
74 
rdma_event_msg(enum rdma_cm_event_type event)75 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
76 {
77 	size_t index = event;
78 
79 	return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
80 			cma_events[index] : "unrecognized event";
81 }
82 EXPORT_SYMBOL(rdma_event_msg);
83 
rdma_reject_msg(struct rdma_cm_id * id,int reason)84 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
85 						int reason)
86 {
87 	if (rdma_ib_or_roce(id->device, id->port_num))
88 		return ibcm_reject_msg(reason);
89 
90 	if (rdma_protocol_iwarp(id->device, id->port_num))
91 		return iwcm_reject_msg(reason);
92 
93 	WARN_ON_ONCE(1);
94 	return "unrecognized transport";
95 }
96 EXPORT_SYMBOL(rdma_reject_msg);
97 
98 /**
99  * rdma_is_consumer_reject - return true if the consumer rejected the connect
100  *                           request.
101  * @id: Communication identifier that received the REJECT event.
102  * @reason: Value returned in the REJECT event status field.
103  */
rdma_is_consumer_reject(struct rdma_cm_id * id,int reason)104 static bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
105 {
106 	if (rdma_ib_or_roce(id->device, id->port_num))
107 		return reason == IB_CM_REJ_CONSUMER_DEFINED;
108 
109 	if (rdma_protocol_iwarp(id->device, id->port_num))
110 		return reason == -ECONNREFUSED;
111 
112 	WARN_ON_ONCE(1);
113 	return false;
114 }
115 
rdma_consumer_reject_data(struct rdma_cm_id * id,struct rdma_cm_event * ev,u8 * data_len)116 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
117 				      struct rdma_cm_event *ev, u8 *data_len)
118 {
119 	const void *p;
120 
121 	if (rdma_is_consumer_reject(id, ev->status)) {
122 		*data_len = ev->param.conn.private_data_len;
123 		p = ev->param.conn.private_data;
124 	} else {
125 		*data_len = 0;
126 		p = NULL;
127 	}
128 	return p;
129 }
130 EXPORT_SYMBOL(rdma_consumer_reject_data);
131 
132 /**
133  * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
134  * @id: Communication Identifier
135  */
rdma_iw_cm_id(struct rdma_cm_id * id)136 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
137 {
138 	struct rdma_id_private *id_priv;
139 
140 	id_priv = container_of(id, struct rdma_id_private, id);
141 	if (id->device->node_type == RDMA_NODE_RNIC)
142 		return id_priv->cm_id.iw;
143 	return NULL;
144 }
145 EXPORT_SYMBOL(rdma_iw_cm_id);
146 
147 /**
148  * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
149  * @res: rdma resource tracking entry pointer
150  */
rdma_res_to_id(struct rdma_restrack_entry * res)151 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
152 {
153 	struct rdma_id_private *id_priv =
154 		container_of(res, struct rdma_id_private, res);
155 
156 	return &id_priv->id;
157 }
158 EXPORT_SYMBOL(rdma_res_to_id);
159 
160 static int cma_add_one(struct ib_device *device);
161 static void cma_remove_one(struct ib_device *device, void *client_data);
162 
163 static struct ib_client cma_client = {
164 	.name   = "cma",
165 	.add    = cma_add_one,
166 	.remove = cma_remove_one
167 };
168 
169 static struct ib_sa_client sa_client;
170 static LIST_HEAD(dev_list);
171 static LIST_HEAD(listen_any_list);
172 static DEFINE_MUTEX(lock);
173 static struct rb_root id_table = RB_ROOT;
174 /* Serialize operations of id_table tree */
175 static DEFINE_SPINLOCK(id_table_lock);
176 static struct workqueue_struct *cma_wq;
177 static unsigned int cma_pernet_id;
178 
179 struct cma_pernet {
180 	struct xarray tcp_ps;
181 	struct xarray udp_ps;
182 	struct xarray ipoib_ps;
183 	struct xarray ib_ps;
184 };
185 
cma_pernet(struct net * net)186 static struct cma_pernet *cma_pernet(struct net *net)
187 {
188 	return net_generic(net, cma_pernet_id);
189 }
190 
191 static
cma_pernet_xa(struct net * net,enum rdma_ucm_port_space ps)192 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
193 {
194 	struct cma_pernet *pernet = cma_pernet(net);
195 
196 	switch (ps) {
197 	case RDMA_PS_TCP:
198 		return &pernet->tcp_ps;
199 	case RDMA_PS_UDP:
200 		return &pernet->udp_ps;
201 	case RDMA_PS_IPOIB:
202 		return &pernet->ipoib_ps;
203 	case RDMA_PS_IB:
204 		return &pernet->ib_ps;
205 	default:
206 		return NULL;
207 	}
208 }
209 
210 struct id_table_entry {
211 	struct list_head id_list;
212 	struct rb_node rb_node;
213 };
214 
215 struct cma_device {
216 	struct list_head	list;
217 	struct ib_device	*device;
218 	struct completion	comp;
219 	refcount_t refcount;
220 	struct list_head	id_list;
221 	enum ib_gid_type	*default_gid_type;
222 	u8			*default_roce_tos;
223 };
224 
225 struct rdma_bind_list {
226 	enum rdma_ucm_port_space ps;
227 	struct hlist_head	owners;
228 	unsigned short		port;
229 };
230 
cma_ps_alloc(struct net * net,enum rdma_ucm_port_space ps,struct rdma_bind_list * bind_list,int snum)231 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
232 			struct rdma_bind_list *bind_list, int snum)
233 {
234 	struct xarray *xa = cma_pernet_xa(net, ps);
235 
236 	return xa_insert(xa, snum, bind_list, GFP_KERNEL);
237 }
238 
cma_ps_find(struct net * net,enum rdma_ucm_port_space ps,int snum)239 static struct rdma_bind_list *cma_ps_find(struct net *net,
240 					  enum rdma_ucm_port_space ps, int snum)
241 {
242 	struct xarray *xa = cma_pernet_xa(net, ps);
243 
244 	return xa_load(xa, snum);
245 }
246 
cma_ps_remove(struct net * net,enum rdma_ucm_port_space ps,int snum)247 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
248 			  int snum)
249 {
250 	struct xarray *xa = cma_pernet_xa(net, ps);
251 
252 	xa_erase(xa, snum);
253 }
254 
255 enum {
256 	CMA_OPTION_AFONLY,
257 };
258 
cma_dev_get(struct cma_device * cma_dev)259 void cma_dev_get(struct cma_device *cma_dev)
260 {
261 	refcount_inc(&cma_dev->refcount);
262 }
263 
cma_dev_put(struct cma_device * cma_dev)264 void cma_dev_put(struct cma_device *cma_dev)
265 {
266 	if (refcount_dec_and_test(&cma_dev->refcount))
267 		complete(&cma_dev->comp);
268 }
269 
cma_enum_devices_by_ibdev(cma_device_filter filter,void * cookie)270 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter	filter,
271 					     void		*cookie)
272 {
273 	struct cma_device *cma_dev;
274 	struct cma_device *found_cma_dev = NULL;
275 
276 	mutex_lock(&lock);
277 
278 	list_for_each_entry(cma_dev, &dev_list, list)
279 		if (filter(cma_dev->device, cookie)) {
280 			found_cma_dev = cma_dev;
281 			break;
282 		}
283 
284 	if (found_cma_dev)
285 		cma_dev_get(found_cma_dev);
286 	mutex_unlock(&lock);
287 	return found_cma_dev;
288 }
289 
cma_get_default_gid_type(struct cma_device * cma_dev,u32 port)290 int cma_get_default_gid_type(struct cma_device *cma_dev,
291 			     u32 port)
292 {
293 	if (!rdma_is_port_valid(cma_dev->device, port))
294 		return -EINVAL;
295 
296 	return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
297 }
298 
cma_set_default_gid_type(struct cma_device * cma_dev,u32 port,enum ib_gid_type default_gid_type)299 int cma_set_default_gid_type(struct cma_device *cma_dev,
300 			     u32 port,
301 			     enum ib_gid_type default_gid_type)
302 {
303 	unsigned long supported_gids;
304 
305 	if (!rdma_is_port_valid(cma_dev->device, port))
306 		return -EINVAL;
307 
308 	if (default_gid_type == IB_GID_TYPE_IB &&
309 	    rdma_protocol_roce_eth_encap(cma_dev->device, port))
310 		default_gid_type = IB_GID_TYPE_ROCE;
311 
312 	supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
313 
314 	if (!(supported_gids & 1 << default_gid_type))
315 		return -EINVAL;
316 
317 	cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
318 		default_gid_type;
319 
320 	return 0;
321 }
322 
cma_get_default_roce_tos(struct cma_device * cma_dev,u32 port)323 int cma_get_default_roce_tos(struct cma_device *cma_dev, u32 port)
324 {
325 	if (!rdma_is_port_valid(cma_dev->device, port))
326 		return -EINVAL;
327 
328 	return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
329 }
330 
cma_set_default_roce_tos(struct cma_device * cma_dev,u32 port,u8 default_roce_tos)331 int cma_set_default_roce_tos(struct cma_device *cma_dev, u32 port,
332 			     u8 default_roce_tos)
333 {
334 	if (!rdma_is_port_valid(cma_dev->device, port))
335 		return -EINVAL;
336 
337 	cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
338 		 default_roce_tos;
339 
340 	return 0;
341 }
cma_get_ib_dev(struct cma_device * cma_dev)342 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
343 {
344 	return cma_dev->device;
345 }
346 
347 /*
348  * Device removal can occur at anytime, so we need extra handling to
349  * serialize notifying the user of device removal with other callbacks.
350  * We do this by disabling removal notification while a callback is in process,
351  * and reporting it after the callback completes.
352  */
353 
354 struct cma_multicast {
355 	struct rdma_id_private *id_priv;
356 	union {
357 		struct ib_sa_multicast *sa_mc;
358 		struct {
359 			struct work_struct work;
360 			struct rdma_cm_event event;
361 		} iboe_join;
362 	};
363 	struct list_head	list;
364 	void			*context;
365 	struct sockaddr_storage	addr;
366 	u8			join_state;
367 };
368 
369 struct cma_work {
370 	struct work_struct	work;
371 	struct rdma_id_private	*id;
372 	enum rdma_cm_state	old_state;
373 	enum rdma_cm_state	new_state;
374 	struct rdma_cm_event	event;
375 };
376 
377 union cma_ip_addr {
378 	struct in6_addr ip6;
379 	struct {
380 		__be32 pad[3];
381 		__be32 addr;
382 	} ip4;
383 };
384 
385 struct cma_hdr {
386 	u8 cma_version;
387 	u8 ip_version;	/* IP version: 7:4 */
388 	__be16 port;
389 	union cma_ip_addr src_addr;
390 	union cma_ip_addr dst_addr;
391 };
392 
393 #define CMA_VERSION 0x00
394 
395 struct cma_req_info {
396 	struct sockaddr_storage listen_addr_storage;
397 	struct sockaddr_storage src_addr_storage;
398 	struct ib_device *device;
399 	union ib_gid local_gid;
400 	__be64 service_id;
401 	int port;
402 	bool has_gid;
403 	u16 pkey;
404 };
405 
cma_comp_exch(struct rdma_id_private * id_priv,enum rdma_cm_state comp,enum rdma_cm_state exch)406 static int cma_comp_exch(struct rdma_id_private *id_priv,
407 			 enum rdma_cm_state comp, enum rdma_cm_state exch)
408 {
409 	unsigned long flags;
410 	int ret;
411 
412 	/*
413 	 * The FSM uses a funny double locking where state is protected by both
414 	 * the handler_mutex and the spinlock. State is not allowed to change
415 	 * to/from a handler_mutex protected value without also holding
416 	 * handler_mutex.
417 	 */
418 	if (comp == RDMA_CM_CONNECT || exch == RDMA_CM_CONNECT)
419 		lockdep_assert_held(&id_priv->handler_mutex);
420 
421 	spin_lock_irqsave(&id_priv->lock, flags);
422 	if ((ret = (id_priv->state == comp)))
423 		id_priv->state = exch;
424 	spin_unlock_irqrestore(&id_priv->lock, flags);
425 	return ret;
426 }
427 
cma_get_ip_ver(const struct cma_hdr * hdr)428 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
429 {
430 	return hdr->ip_version >> 4;
431 }
432 
cma_set_ip_ver(struct cma_hdr * hdr,u8 ip_ver)433 static void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
434 {
435 	hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
436 }
437 
cma_src_addr(struct rdma_id_private * id_priv)438 static struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
439 {
440 	return (struct sockaddr *)&id_priv->id.route.addr.src_addr;
441 }
442 
cma_dst_addr(struct rdma_id_private * id_priv)443 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
444 {
445 	return (struct sockaddr *)&id_priv->id.route.addr.dst_addr;
446 }
447 
cma_igmp_send(struct net_device * ndev,union ib_gid * mgid,bool join)448 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
449 {
450 	struct in_device *in_dev = NULL;
451 
452 	if (ndev) {
453 		rtnl_lock();
454 		in_dev = __in_dev_get_rtnl(ndev);
455 		if (in_dev) {
456 			if (join)
457 				ip_mc_inc_group(in_dev,
458 						*(__be32 *)(mgid->raw + 12));
459 			else
460 				ip_mc_dec_group(in_dev,
461 						*(__be32 *)(mgid->raw + 12));
462 		}
463 		rtnl_unlock();
464 	}
465 	return (in_dev) ? 0 : -ENODEV;
466 }
467 
compare_netdev_and_ip(int ifindex_a,struct sockaddr * sa,struct id_table_entry * entry_b)468 static int compare_netdev_and_ip(int ifindex_a, struct sockaddr *sa,
469 				 struct id_table_entry *entry_b)
470 {
471 	struct rdma_id_private *id_priv = list_first_entry(
472 		&entry_b->id_list, struct rdma_id_private, id_list_entry);
473 	int ifindex_b = id_priv->id.route.addr.dev_addr.bound_dev_if;
474 	struct sockaddr *sb = cma_dst_addr(id_priv);
475 
476 	if (ifindex_a != ifindex_b)
477 		return (ifindex_a > ifindex_b) ? 1 : -1;
478 
479 	if (sa->sa_family != sb->sa_family)
480 		return sa->sa_family - sb->sa_family;
481 
482 	if (sa->sa_family == AF_INET)
483 		return memcmp((char *)&((struct sockaddr_in *)sa)->sin_addr,
484 			      (char *)&((struct sockaddr_in *)sb)->sin_addr,
485 			      sizeof(((struct sockaddr_in *)sa)->sin_addr));
486 
487 	return ipv6_addr_cmp(&((struct sockaddr_in6 *)sa)->sin6_addr,
488 			     &((struct sockaddr_in6 *)sb)->sin6_addr);
489 }
490 
cma_add_id_to_tree(struct rdma_id_private * node_id_priv)491 static int cma_add_id_to_tree(struct rdma_id_private *node_id_priv)
492 {
493 	struct rb_node **new, *parent = NULL;
494 	struct id_table_entry *this, *node;
495 	unsigned long flags;
496 	int result;
497 
498 	node = kzalloc(sizeof(*node), GFP_KERNEL);
499 	if (!node)
500 		return -ENOMEM;
501 
502 	spin_lock_irqsave(&id_table_lock, flags);
503 	new = &id_table.rb_node;
504 	while (*new) {
505 		this = container_of(*new, struct id_table_entry, rb_node);
506 		result = compare_netdev_and_ip(
507 			node_id_priv->id.route.addr.dev_addr.bound_dev_if,
508 			cma_dst_addr(node_id_priv), this);
509 
510 		parent = *new;
511 		if (result < 0)
512 			new = &((*new)->rb_left);
513 		else if (result > 0)
514 			new = &((*new)->rb_right);
515 		else {
516 			list_add_tail(&node_id_priv->id_list_entry,
517 				      &this->id_list);
518 			kfree(node);
519 			goto unlock;
520 		}
521 	}
522 
523 	INIT_LIST_HEAD(&node->id_list);
524 	list_add_tail(&node_id_priv->id_list_entry, &node->id_list);
525 
526 	rb_link_node(&node->rb_node, parent, new);
527 	rb_insert_color(&node->rb_node, &id_table);
528 
529 unlock:
530 	spin_unlock_irqrestore(&id_table_lock, flags);
531 	return 0;
532 }
533 
534 static struct id_table_entry *
node_from_ndev_ip(struct rb_root * root,int ifindex,struct sockaddr * sa)535 node_from_ndev_ip(struct rb_root *root, int ifindex, struct sockaddr *sa)
536 {
537 	struct rb_node *node = root->rb_node;
538 	struct id_table_entry *data;
539 	int result;
540 
541 	while (node) {
542 		data = container_of(node, struct id_table_entry, rb_node);
543 		result = compare_netdev_and_ip(ifindex, sa, data);
544 		if (result < 0)
545 			node = node->rb_left;
546 		else if (result > 0)
547 			node = node->rb_right;
548 		else
549 			return data;
550 	}
551 
552 	return NULL;
553 }
554 
cma_remove_id_from_tree(struct rdma_id_private * id_priv)555 static void cma_remove_id_from_tree(struct rdma_id_private *id_priv)
556 {
557 	struct id_table_entry *data;
558 	unsigned long flags;
559 
560 	spin_lock_irqsave(&id_table_lock, flags);
561 	if (list_empty(&id_priv->id_list_entry))
562 		goto out;
563 
564 	data = node_from_ndev_ip(&id_table,
565 				 id_priv->id.route.addr.dev_addr.bound_dev_if,
566 				 cma_dst_addr(id_priv));
567 	if (!data)
568 		goto out;
569 
570 	list_del_init(&id_priv->id_list_entry);
571 	if (list_empty(&data->id_list)) {
572 		rb_erase(&data->rb_node, &id_table);
573 		kfree(data);
574 	}
575 out:
576 	spin_unlock_irqrestore(&id_table_lock, flags);
577 }
578 
_cma_attach_to_dev(struct rdma_id_private * id_priv,struct cma_device * cma_dev)579 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
580 			       struct cma_device *cma_dev)
581 {
582 	cma_dev_get(cma_dev);
583 	id_priv->cma_dev = cma_dev;
584 	id_priv->id.device = cma_dev->device;
585 	id_priv->id.route.addr.dev_addr.transport =
586 		rdma_node_get_transport(cma_dev->device->node_type);
587 	list_add_tail(&id_priv->device_item, &cma_dev->id_list);
588 
589 	trace_cm_id_attach(id_priv, cma_dev->device);
590 }
591 
cma_attach_to_dev(struct rdma_id_private * id_priv,struct cma_device * cma_dev)592 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
593 			      struct cma_device *cma_dev)
594 {
595 	_cma_attach_to_dev(id_priv, cma_dev);
596 	id_priv->gid_type =
597 		cma_dev->default_gid_type[id_priv->id.port_num -
598 					  rdma_start_port(cma_dev->device)];
599 }
600 
cma_release_dev(struct rdma_id_private * id_priv)601 static void cma_release_dev(struct rdma_id_private *id_priv)
602 {
603 	mutex_lock(&lock);
604 	list_del_init(&id_priv->device_item);
605 	cma_dev_put(id_priv->cma_dev);
606 	id_priv->cma_dev = NULL;
607 	id_priv->id.device = NULL;
608 	if (id_priv->id.route.addr.dev_addr.sgid_attr) {
609 		rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
610 		id_priv->id.route.addr.dev_addr.sgid_attr = NULL;
611 	}
612 	mutex_unlock(&lock);
613 }
614 
cma_family(struct rdma_id_private * id_priv)615 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
616 {
617 	return id_priv->id.route.addr.src_addr.ss_family;
618 }
619 
cma_set_qkey(struct rdma_id_private * id_priv,u32 qkey)620 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
621 {
622 	struct ib_sa_mcmember_rec rec;
623 	int ret = 0;
624 
625 	if (id_priv->qkey) {
626 		if (qkey && id_priv->qkey != qkey)
627 			return -EINVAL;
628 		return 0;
629 	}
630 
631 	if (qkey) {
632 		id_priv->qkey = qkey;
633 		return 0;
634 	}
635 
636 	switch (id_priv->id.ps) {
637 	case RDMA_PS_UDP:
638 	case RDMA_PS_IB:
639 		id_priv->qkey = RDMA_UDP_QKEY;
640 		break;
641 	case RDMA_PS_IPOIB:
642 		ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
643 		ret = ib_sa_get_mcmember_rec(id_priv->id.device,
644 					     id_priv->id.port_num, &rec.mgid,
645 					     &rec);
646 		if (!ret)
647 			id_priv->qkey = be32_to_cpu(rec.qkey);
648 		break;
649 	default:
650 		break;
651 	}
652 	return ret;
653 }
654 
cma_translate_ib(struct sockaddr_ib * sib,struct rdma_dev_addr * dev_addr)655 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
656 {
657 	dev_addr->dev_type = ARPHRD_INFINIBAND;
658 	rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
659 	ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
660 }
661 
cma_translate_addr(struct sockaddr * addr,struct rdma_dev_addr * dev_addr)662 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
663 {
664 	int ret;
665 
666 	if (addr->sa_family != AF_IB) {
667 		ret = rdma_translate_ip(addr, dev_addr);
668 	} else {
669 		cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
670 		ret = 0;
671 	}
672 
673 	return ret;
674 }
675 
676 static const struct ib_gid_attr *
cma_validate_port(struct ib_device * device,u32 port,enum ib_gid_type gid_type,union ib_gid * gid,struct rdma_id_private * id_priv)677 cma_validate_port(struct ib_device *device, u32 port,
678 		  enum ib_gid_type gid_type,
679 		  union ib_gid *gid,
680 		  struct rdma_id_private *id_priv)
681 {
682 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
683 	int bound_if_index = dev_addr->bound_dev_if;
684 	const struct ib_gid_attr *sgid_attr;
685 	int dev_type = dev_addr->dev_type;
686 	struct net_device *ndev = NULL;
687 
688 	if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
689 		return ERR_PTR(-ENODEV);
690 
691 	if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
692 		return ERR_PTR(-ENODEV);
693 
694 	if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
695 		return ERR_PTR(-ENODEV);
696 
697 	if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
698 		ndev = dev_get_by_index(dev_addr->net, bound_if_index);
699 		if (!ndev)
700 			return ERR_PTR(-ENODEV);
701 	} else {
702 		gid_type = IB_GID_TYPE_IB;
703 	}
704 
705 	sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
706 	if (ndev)
707 		dev_put(ndev);
708 	return sgid_attr;
709 }
710 
cma_bind_sgid_attr(struct rdma_id_private * id_priv,const struct ib_gid_attr * sgid_attr)711 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
712 			       const struct ib_gid_attr *sgid_attr)
713 {
714 	WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
715 	id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
716 }
717 
718 /**
719  * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
720  * based on source ip address.
721  * @id_priv:	cm_id which should be bound to cma device
722  *
723  * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
724  * based on source IP address. It returns 0 on success or error code otherwise.
725  * It is applicable to active and passive side cm_id.
726  */
cma_acquire_dev_by_src_ip(struct rdma_id_private * id_priv)727 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
728 {
729 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
730 	const struct ib_gid_attr *sgid_attr;
731 	union ib_gid gid, iboe_gid, *gidp;
732 	struct cma_device *cma_dev;
733 	enum ib_gid_type gid_type;
734 	int ret = -ENODEV;
735 	u32 port;
736 
737 	if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
738 	    id_priv->id.ps == RDMA_PS_IPOIB)
739 		return -EINVAL;
740 
741 	rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
742 		    &iboe_gid);
743 
744 	memcpy(&gid, dev_addr->src_dev_addr +
745 	       rdma_addr_gid_offset(dev_addr), sizeof(gid));
746 
747 	mutex_lock(&lock);
748 	list_for_each_entry(cma_dev, &dev_list, list) {
749 		rdma_for_each_port (cma_dev->device, port) {
750 			gidp = rdma_protocol_roce(cma_dev->device, port) ?
751 			       &iboe_gid : &gid;
752 			gid_type = cma_dev->default_gid_type[port - 1];
753 			sgid_attr = cma_validate_port(cma_dev->device, port,
754 						      gid_type, gidp, id_priv);
755 			if (!IS_ERR(sgid_attr)) {
756 				id_priv->id.port_num = port;
757 				cma_bind_sgid_attr(id_priv, sgid_attr);
758 				cma_attach_to_dev(id_priv, cma_dev);
759 				ret = 0;
760 				goto out;
761 			}
762 		}
763 	}
764 out:
765 	mutex_unlock(&lock);
766 	return ret;
767 }
768 
769 /**
770  * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
771  * @id_priv:		cm id to bind to cma device
772  * @listen_id_priv:	listener cm id to match against
773  * @req:		Pointer to req structure containaining incoming
774  *			request information
775  * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
776  * rdma device matches for listen_id and incoming request. It also verifies
777  * that a GID table entry is present for the source address.
778  * Returns 0 on success, or returns error code otherwise.
779  */
cma_ib_acquire_dev(struct rdma_id_private * id_priv,const struct rdma_id_private * listen_id_priv,struct cma_req_info * req)780 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
781 			      const struct rdma_id_private *listen_id_priv,
782 			      struct cma_req_info *req)
783 {
784 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
785 	const struct ib_gid_attr *sgid_attr;
786 	enum ib_gid_type gid_type;
787 	union ib_gid gid;
788 
789 	if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
790 	    id_priv->id.ps == RDMA_PS_IPOIB)
791 		return -EINVAL;
792 
793 	if (rdma_protocol_roce(req->device, req->port))
794 		rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
795 			    &gid);
796 	else
797 		memcpy(&gid, dev_addr->src_dev_addr +
798 		       rdma_addr_gid_offset(dev_addr), sizeof(gid));
799 
800 	gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
801 	sgid_attr = cma_validate_port(req->device, req->port,
802 				      gid_type, &gid, id_priv);
803 	if (IS_ERR(sgid_attr))
804 		return PTR_ERR(sgid_attr);
805 
806 	id_priv->id.port_num = req->port;
807 	cma_bind_sgid_attr(id_priv, sgid_attr);
808 	/* Need to acquire lock to protect against reader
809 	 * of cma_dev->id_list such as cma_netdev_callback() and
810 	 * cma_process_remove().
811 	 */
812 	mutex_lock(&lock);
813 	cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
814 	mutex_unlock(&lock);
815 	rdma_restrack_add(&id_priv->res);
816 	return 0;
817 }
818 
cma_iw_acquire_dev(struct rdma_id_private * id_priv,const struct rdma_id_private * listen_id_priv)819 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
820 			      const struct rdma_id_private *listen_id_priv)
821 {
822 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
823 	const struct ib_gid_attr *sgid_attr;
824 	struct cma_device *cma_dev;
825 	enum ib_gid_type gid_type;
826 	int ret = -ENODEV;
827 	union ib_gid gid;
828 	u32 port;
829 
830 	if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
831 	    id_priv->id.ps == RDMA_PS_IPOIB)
832 		return -EINVAL;
833 
834 	memcpy(&gid, dev_addr->src_dev_addr +
835 	       rdma_addr_gid_offset(dev_addr), sizeof(gid));
836 
837 	mutex_lock(&lock);
838 
839 	cma_dev = listen_id_priv->cma_dev;
840 	port = listen_id_priv->id.port_num;
841 	gid_type = listen_id_priv->gid_type;
842 	sgid_attr = cma_validate_port(cma_dev->device, port,
843 				      gid_type, &gid, id_priv);
844 	if (!IS_ERR(sgid_attr)) {
845 		id_priv->id.port_num = port;
846 		cma_bind_sgid_attr(id_priv, sgid_attr);
847 		ret = 0;
848 		goto out;
849 	}
850 
851 	list_for_each_entry(cma_dev, &dev_list, list) {
852 		rdma_for_each_port (cma_dev->device, port) {
853 			if (listen_id_priv->cma_dev == cma_dev &&
854 			    listen_id_priv->id.port_num == port)
855 				continue;
856 
857 			gid_type = cma_dev->default_gid_type[port - 1];
858 			sgid_attr = cma_validate_port(cma_dev->device, port,
859 						      gid_type, &gid, id_priv);
860 			if (!IS_ERR(sgid_attr)) {
861 				id_priv->id.port_num = port;
862 				cma_bind_sgid_attr(id_priv, sgid_attr);
863 				ret = 0;
864 				goto out;
865 			}
866 		}
867 	}
868 
869 out:
870 	if (!ret) {
871 		cma_attach_to_dev(id_priv, cma_dev);
872 		rdma_restrack_add(&id_priv->res);
873 	}
874 
875 	mutex_unlock(&lock);
876 	return ret;
877 }
878 
879 /*
880  * Select the source IB device and address to reach the destination IB address.
881  */
cma_resolve_ib_dev(struct rdma_id_private * id_priv)882 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
883 {
884 	struct cma_device *cma_dev, *cur_dev;
885 	struct sockaddr_ib *addr;
886 	union ib_gid gid, sgid, *dgid;
887 	unsigned int p;
888 	u16 pkey, index;
889 	enum ib_port_state port_state;
890 	int ret;
891 	int i;
892 
893 	cma_dev = NULL;
894 	addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
895 	dgid = (union ib_gid *) &addr->sib_addr;
896 	pkey = ntohs(addr->sib_pkey);
897 
898 	mutex_lock(&lock);
899 	list_for_each_entry(cur_dev, &dev_list, list) {
900 		rdma_for_each_port (cur_dev->device, p) {
901 			if (!rdma_cap_af_ib(cur_dev->device, p))
902 				continue;
903 
904 			if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
905 				continue;
906 
907 			if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
908 				continue;
909 
910 			for (i = 0; i < cur_dev->device->port_data[p].immutable.gid_tbl_len;
911 			     ++i) {
912 				ret = rdma_query_gid(cur_dev->device, p, i,
913 						     &gid);
914 				if (ret)
915 					continue;
916 
917 				if (!memcmp(&gid, dgid, sizeof(gid))) {
918 					cma_dev = cur_dev;
919 					sgid = gid;
920 					id_priv->id.port_num = p;
921 					goto found;
922 				}
923 
924 				if (!cma_dev && (gid.global.subnet_prefix ==
925 				    dgid->global.subnet_prefix) &&
926 				    port_state == IB_PORT_ACTIVE) {
927 					cma_dev = cur_dev;
928 					sgid = gid;
929 					id_priv->id.port_num = p;
930 					goto found;
931 				}
932 			}
933 		}
934 	}
935 	mutex_unlock(&lock);
936 	return -ENODEV;
937 
938 found:
939 	cma_attach_to_dev(id_priv, cma_dev);
940 	rdma_restrack_add(&id_priv->res);
941 	mutex_unlock(&lock);
942 	addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
943 	memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
944 	cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
945 	return 0;
946 }
947 
cma_id_get(struct rdma_id_private * id_priv)948 static void cma_id_get(struct rdma_id_private *id_priv)
949 {
950 	refcount_inc(&id_priv->refcount);
951 }
952 
cma_id_put(struct rdma_id_private * id_priv)953 static void cma_id_put(struct rdma_id_private *id_priv)
954 {
955 	if (refcount_dec_and_test(&id_priv->refcount))
956 		complete(&id_priv->comp);
957 }
958 
959 static struct rdma_id_private *
__rdma_create_id(struct net * net,rdma_cm_event_handler event_handler,void * context,enum rdma_ucm_port_space ps,enum ib_qp_type qp_type,const struct rdma_id_private * parent)960 __rdma_create_id(struct net *net, rdma_cm_event_handler event_handler,
961 		 void *context, enum rdma_ucm_port_space ps,
962 		 enum ib_qp_type qp_type, const struct rdma_id_private *parent)
963 {
964 	struct rdma_id_private *id_priv;
965 
966 	id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
967 	if (!id_priv)
968 		return ERR_PTR(-ENOMEM);
969 
970 	id_priv->state = RDMA_CM_IDLE;
971 	id_priv->id.context = context;
972 	id_priv->id.event_handler = event_handler;
973 	id_priv->id.ps = ps;
974 	id_priv->id.qp_type = qp_type;
975 	id_priv->tos_set = false;
976 	id_priv->timeout_set = false;
977 	id_priv->min_rnr_timer_set = false;
978 	id_priv->gid_type = IB_GID_TYPE_IB;
979 	spin_lock_init(&id_priv->lock);
980 	mutex_init(&id_priv->qp_mutex);
981 	init_completion(&id_priv->comp);
982 	refcount_set(&id_priv->refcount, 1);
983 	mutex_init(&id_priv->handler_mutex);
984 	INIT_LIST_HEAD(&id_priv->device_item);
985 	INIT_LIST_HEAD(&id_priv->id_list_entry);
986 	INIT_LIST_HEAD(&id_priv->listen_list);
987 	INIT_LIST_HEAD(&id_priv->mc_list);
988 	get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
989 	id_priv->id.route.addr.dev_addr.net = get_net(net);
990 	id_priv->seq_num &= 0x00ffffff;
991 
992 	rdma_restrack_new(&id_priv->res, RDMA_RESTRACK_CM_ID);
993 	if (parent)
994 		rdma_restrack_parent_name(&id_priv->res, &parent->res);
995 
996 	return id_priv;
997 }
998 
999 struct rdma_cm_id *
__rdma_create_kernel_id(struct net * net,rdma_cm_event_handler event_handler,void * context,enum rdma_ucm_port_space ps,enum ib_qp_type qp_type,const char * caller)1000 __rdma_create_kernel_id(struct net *net, rdma_cm_event_handler event_handler,
1001 			void *context, enum rdma_ucm_port_space ps,
1002 			enum ib_qp_type qp_type, const char *caller)
1003 {
1004 	struct rdma_id_private *ret;
1005 
1006 	ret = __rdma_create_id(net, event_handler, context, ps, qp_type, NULL);
1007 	if (IS_ERR(ret))
1008 		return ERR_CAST(ret);
1009 
1010 	rdma_restrack_set_name(&ret->res, caller);
1011 	return &ret->id;
1012 }
1013 EXPORT_SYMBOL(__rdma_create_kernel_id);
1014 
rdma_create_user_id(rdma_cm_event_handler event_handler,void * context,enum rdma_ucm_port_space ps,enum ib_qp_type qp_type)1015 struct rdma_cm_id *rdma_create_user_id(rdma_cm_event_handler event_handler,
1016 				       void *context,
1017 				       enum rdma_ucm_port_space ps,
1018 				       enum ib_qp_type qp_type)
1019 {
1020 	struct rdma_id_private *ret;
1021 
1022 	ret = __rdma_create_id(current->nsproxy->net_ns, event_handler, context,
1023 			       ps, qp_type, NULL);
1024 	if (IS_ERR(ret))
1025 		return ERR_CAST(ret);
1026 
1027 	rdma_restrack_set_name(&ret->res, NULL);
1028 	return &ret->id;
1029 }
1030 EXPORT_SYMBOL(rdma_create_user_id);
1031 
cma_init_ud_qp(struct rdma_id_private * id_priv,struct ib_qp * qp)1032 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
1033 {
1034 	struct ib_qp_attr qp_attr;
1035 	int qp_attr_mask, ret;
1036 
1037 	qp_attr.qp_state = IB_QPS_INIT;
1038 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1039 	if (ret)
1040 		return ret;
1041 
1042 	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1043 	if (ret)
1044 		return ret;
1045 
1046 	qp_attr.qp_state = IB_QPS_RTR;
1047 	ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
1048 	if (ret)
1049 		return ret;
1050 
1051 	qp_attr.qp_state = IB_QPS_RTS;
1052 	qp_attr.sq_psn = 0;
1053 	ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
1054 
1055 	return ret;
1056 }
1057 
cma_init_conn_qp(struct rdma_id_private * id_priv,struct ib_qp * qp)1058 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
1059 {
1060 	struct ib_qp_attr qp_attr;
1061 	int qp_attr_mask, ret;
1062 
1063 	qp_attr.qp_state = IB_QPS_INIT;
1064 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1065 	if (ret)
1066 		return ret;
1067 
1068 	return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1069 }
1070 
rdma_create_qp(struct rdma_cm_id * id,struct ib_pd * pd,struct ib_qp_init_attr * qp_init_attr)1071 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
1072 		   struct ib_qp_init_attr *qp_init_attr)
1073 {
1074 	struct rdma_id_private *id_priv;
1075 	struct ib_qp *qp;
1076 	int ret;
1077 
1078 	id_priv = container_of(id, struct rdma_id_private, id);
1079 	if (id->device != pd->device) {
1080 		ret = -EINVAL;
1081 		goto out_err;
1082 	}
1083 
1084 	qp_init_attr->port_num = id->port_num;
1085 	qp = ib_create_qp(pd, qp_init_attr);
1086 	if (IS_ERR(qp)) {
1087 		ret = PTR_ERR(qp);
1088 		goto out_err;
1089 	}
1090 
1091 	if (id->qp_type == IB_QPT_UD)
1092 		ret = cma_init_ud_qp(id_priv, qp);
1093 	else
1094 		ret = cma_init_conn_qp(id_priv, qp);
1095 	if (ret)
1096 		goto out_destroy;
1097 
1098 	id->qp = qp;
1099 	id_priv->qp_num = qp->qp_num;
1100 	id_priv->srq = (qp->srq != NULL);
1101 	trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
1102 	return 0;
1103 out_destroy:
1104 	ib_destroy_qp(qp);
1105 out_err:
1106 	trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
1107 	return ret;
1108 }
1109 EXPORT_SYMBOL(rdma_create_qp);
1110 
rdma_destroy_qp(struct rdma_cm_id * id)1111 void rdma_destroy_qp(struct rdma_cm_id *id)
1112 {
1113 	struct rdma_id_private *id_priv;
1114 
1115 	id_priv = container_of(id, struct rdma_id_private, id);
1116 	trace_cm_qp_destroy(id_priv);
1117 	mutex_lock(&id_priv->qp_mutex);
1118 	ib_destroy_qp(id_priv->id.qp);
1119 	id_priv->id.qp = NULL;
1120 	mutex_unlock(&id_priv->qp_mutex);
1121 }
1122 EXPORT_SYMBOL(rdma_destroy_qp);
1123 
cma_modify_qp_rtr(struct rdma_id_private * id_priv,struct rdma_conn_param * conn_param)1124 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
1125 			     struct rdma_conn_param *conn_param)
1126 {
1127 	struct ib_qp_attr qp_attr;
1128 	int qp_attr_mask, ret;
1129 
1130 	mutex_lock(&id_priv->qp_mutex);
1131 	if (!id_priv->id.qp) {
1132 		ret = 0;
1133 		goto out;
1134 	}
1135 
1136 	/* Need to update QP attributes from default values. */
1137 	qp_attr.qp_state = IB_QPS_INIT;
1138 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1139 	if (ret)
1140 		goto out;
1141 
1142 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1143 	if (ret)
1144 		goto out;
1145 
1146 	qp_attr.qp_state = IB_QPS_RTR;
1147 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1148 	if (ret)
1149 		goto out;
1150 
1151 	BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1152 
1153 	if (conn_param)
1154 		qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1155 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1156 out:
1157 	mutex_unlock(&id_priv->qp_mutex);
1158 	return ret;
1159 }
1160 
cma_modify_qp_rts(struct rdma_id_private * id_priv,struct rdma_conn_param * conn_param)1161 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1162 			     struct rdma_conn_param *conn_param)
1163 {
1164 	struct ib_qp_attr qp_attr;
1165 	int qp_attr_mask, ret;
1166 
1167 	mutex_lock(&id_priv->qp_mutex);
1168 	if (!id_priv->id.qp) {
1169 		ret = 0;
1170 		goto out;
1171 	}
1172 
1173 	qp_attr.qp_state = IB_QPS_RTS;
1174 	ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1175 	if (ret)
1176 		goto out;
1177 
1178 	if (conn_param)
1179 		qp_attr.max_rd_atomic = conn_param->initiator_depth;
1180 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1181 out:
1182 	mutex_unlock(&id_priv->qp_mutex);
1183 	return ret;
1184 }
1185 
cma_modify_qp_err(struct rdma_id_private * id_priv)1186 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1187 {
1188 	struct ib_qp_attr qp_attr;
1189 	int ret;
1190 
1191 	mutex_lock(&id_priv->qp_mutex);
1192 	if (!id_priv->id.qp) {
1193 		ret = 0;
1194 		goto out;
1195 	}
1196 
1197 	qp_attr.qp_state = IB_QPS_ERR;
1198 	ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1199 out:
1200 	mutex_unlock(&id_priv->qp_mutex);
1201 	return ret;
1202 }
1203 
cma_ib_init_qp_attr(struct rdma_id_private * id_priv,struct ib_qp_attr * qp_attr,int * qp_attr_mask)1204 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1205 			       struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1206 {
1207 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1208 	int ret;
1209 	u16 pkey;
1210 
1211 	if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1212 		pkey = 0xffff;
1213 	else
1214 		pkey = ib_addr_get_pkey(dev_addr);
1215 
1216 	ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1217 				  pkey, &qp_attr->pkey_index);
1218 	if (ret)
1219 		return ret;
1220 
1221 	qp_attr->port_num = id_priv->id.port_num;
1222 	*qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1223 
1224 	if (id_priv->id.qp_type == IB_QPT_UD) {
1225 		ret = cma_set_qkey(id_priv, 0);
1226 		if (ret)
1227 			return ret;
1228 
1229 		qp_attr->qkey = id_priv->qkey;
1230 		*qp_attr_mask |= IB_QP_QKEY;
1231 	} else {
1232 		qp_attr->qp_access_flags = 0;
1233 		*qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1234 	}
1235 	return 0;
1236 }
1237 
rdma_init_qp_attr(struct rdma_cm_id * id,struct ib_qp_attr * qp_attr,int * qp_attr_mask)1238 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1239 		       int *qp_attr_mask)
1240 {
1241 	struct rdma_id_private *id_priv;
1242 	int ret = 0;
1243 
1244 	id_priv = container_of(id, struct rdma_id_private, id);
1245 	if (rdma_cap_ib_cm(id->device, id->port_num)) {
1246 		if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1247 			ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1248 		else
1249 			ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1250 						 qp_attr_mask);
1251 
1252 		if (qp_attr->qp_state == IB_QPS_RTR)
1253 			qp_attr->rq_psn = id_priv->seq_num;
1254 	} else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1255 		if (!id_priv->cm_id.iw) {
1256 			qp_attr->qp_access_flags = 0;
1257 			*qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1258 		} else
1259 			ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1260 						 qp_attr_mask);
1261 		qp_attr->port_num = id_priv->id.port_num;
1262 		*qp_attr_mask |= IB_QP_PORT;
1263 	} else {
1264 		ret = -ENOSYS;
1265 	}
1266 
1267 	if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1268 		qp_attr->timeout = id_priv->timeout;
1269 
1270 	if ((*qp_attr_mask & IB_QP_MIN_RNR_TIMER) && id_priv->min_rnr_timer_set)
1271 		qp_attr->min_rnr_timer = id_priv->min_rnr_timer;
1272 
1273 	return ret;
1274 }
1275 EXPORT_SYMBOL(rdma_init_qp_attr);
1276 
cma_zero_addr(const struct sockaddr * addr)1277 static inline bool cma_zero_addr(const struct sockaddr *addr)
1278 {
1279 	switch (addr->sa_family) {
1280 	case AF_INET:
1281 		return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1282 	case AF_INET6:
1283 		return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1284 	case AF_IB:
1285 		return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1286 	default:
1287 		return false;
1288 	}
1289 }
1290 
cma_loopback_addr(const struct sockaddr * addr)1291 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1292 {
1293 	switch (addr->sa_family) {
1294 	case AF_INET:
1295 		return ipv4_is_loopback(
1296 			((struct sockaddr_in *)addr)->sin_addr.s_addr);
1297 	case AF_INET6:
1298 		return ipv6_addr_loopback(
1299 			&((struct sockaddr_in6 *)addr)->sin6_addr);
1300 	case AF_IB:
1301 		return ib_addr_loopback(
1302 			&((struct sockaddr_ib *)addr)->sib_addr);
1303 	default:
1304 		return false;
1305 	}
1306 }
1307 
cma_any_addr(const struct sockaddr * addr)1308 static inline bool cma_any_addr(const struct sockaddr *addr)
1309 {
1310 	return cma_zero_addr(addr) || cma_loopback_addr(addr);
1311 }
1312 
cma_addr_cmp(const struct sockaddr * src,const struct sockaddr * dst)1313 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1314 {
1315 	if (src->sa_family != dst->sa_family)
1316 		return -1;
1317 
1318 	switch (src->sa_family) {
1319 	case AF_INET:
1320 		return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1321 		       ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1322 	case AF_INET6: {
1323 		struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1324 		struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1325 		bool link_local;
1326 
1327 		if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1328 					  &dst_addr6->sin6_addr))
1329 			return 1;
1330 		link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1331 			     IPV6_ADDR_LINKLOCAL;
1332 		/* Link local must match their scope_ids */
1333 		return link_local ? (src_addr6->sin6_scope_id !=
1334 				     dst_addr6->sin6_scope_id) :
1335 				    0;
1336 	}
1337 
1338 	default:
1339 		return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1340 				   &((struct sockaddr_ib *) dst)->sib_addr);
1341 	}
1342 }
1343 
cma_port(const struct sockaddr * addr)1344 static __be16 cma_port(const struct sockaddr *addr)
1345 {
1346 	struct sockaddr_ib *sib;
1347 
1348 	switch (addr->sa_family) {
1349 	case AF_INET:
1350 		return ((struct sockaddr_in *) addr)->sin_port;
1351 	case AF_INET6:
1352 		return ((struct sockaddr_in6 *) addr)->sin6_port;
1353 	case AF_IB:
1354 		sib = (struct sockaddr_ib *) addr;
1355 		return htons((u16) (be64_to_cpu(sib->sib_sid) &
1356 				    be64_to_cpu(sib->sib_sid_mask)));
1357 	default:
1358 		return 0;
1359 	}
1360 }
1361 
cma_any_port(const struct sockaddr * addr)1362 static inline int cma_any_port(const struct sockaddr *addr)
1363 {
1364 	return !cma_port(addr);
1365 }
1366 
cma_save_ib_info(struct sockaddr * src_addr,struct sockaddr * dst_addr,const struct rdma_cm_id * listen_id,const struct sa_path_rec * path)1367 static void cma_save_ib_info(struct sockaddr *src_addr,
1368 			     struct sockaddr *dst_addr,
1369 			     const struct rdma_cm_id *listen_id,
1370 			     const struct sa_path_rec *path)
1371 {
1372 	struct sockaddr_ib *listen_ib, *ib;
1373 
1374 	listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1375 	if (src_addr) {
1376 		ib = (struct sockaddr_ib *)src_addr;
1377 		ib->sib_family = AF_IB;
1378 		if (path) {
1379 			ib->sib_pkey = path->pkey;
1380 			ib->sib_flowinfo = path->flow_label;
1381 			memcpy(&ib->sib_addr, &path->sgid, 16);
1382 			ib->sib_sid = path->service_id;
1383 			ib->sib_scope_id = 0;
1384 		} else {
1385 			ib->sib_pkey = listen_ib->sib_pkey;
1386 			ib->sib_flowinfo = listen_ib->sib_flowinfo;
1387 			ib->sib_addr = listen_ib->sib_addr;
1388 			ib->sib_sid = listen_ib->sib_sid;
1389 			ib->sib_scope_id = listen_ib->sib_scope_id;
1390 		}
1391 		ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1392 	}
1393 	if (dst_addr) {
1394 		ib = (struct sockaddr_ib *)dst_addr;
1395 		ib->sib_family = AF_IB;
1396 		if (path) {
1397 			ib->sib_pkey = path->pkey;
1398 			ib->sib_flowinfo = path->flow_label;
1399 			memcpy(&ib->sib_addr, &path->dgid, 16);
1400 		}
1401 	}
1402 }
1403 
cma_save_ip4_info(struct sockaddr_in * src_addr,struct sockaddr_in * dst_addr,struct cma_hdr * hdr,__be16 local_port)1404 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1405 			      struct sockaddr_in *dst_addr,
1406 			      struct cma_hdr *hdr,
1407 			      __be16 local_port)
1408 {
1409 	if (src_addr) {
1410 		*src_addr = (struct sockaddr_in) {
1411 			.sin_family = AF_INET,
1412 			.sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1413 			.sin_port = local_port,
1414 		};
1415 	}
1416 
1417 	if (dst_addr) {
1418 		*dst_addr = (struct sockaddr_in) {
1419 			.sin_family = AF_INET,
1420 			.sin_addr.s_addr = hdr->src_addr.ip4.addr,
1421 			.sin_port = hdr->port,
1422 		};
1423 	}
1424 }
1425 
cma_save_ip6_info(struct sockaddr_in6 * src_addr,struct sockaddr_in6 * dst_addr,struct cma_hdr * hdr,__be16 local_port)1426 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1427 			      struct sockaddr_in6 *dst_addr,
1428 			      struct cma_hdr *hdr,
1429 			      __be16 local_port)
1430 {
1431 	if (src_addr) {
1432 		*src_addr = (struct sockaddr_in6) {
1433 			.sin6_family = AF_INET6,
1434 			.sin6_addr = hdr->dst_addr.ip6,
1435 			.sin6_port = local_port,
1436 		};
1437 	}
1438 
1439 	if (dst_addr) {
1440 		*dst_addr = (struct sockaddr_in6) {
1441 			.sin6_family = AF_INET6,
1442 			.sin6_addr = hdr->src_addr.ip6,
1443 			.sin6_port = hdr->port,
1444 		};
1445 	}
1446 }
1447 
cma_port_from_service_id(__be64 service_id)1448 static u16 cma_port_from_service_id(__be64 service_id)
1449 {
1450 	return (u16)be64_to_cpu(service_id);
1451 }
1452 
cma_save_ip_info(struct sockaddr * src_addr,struct sockaddr * dst_addr,const struct ib_cm_event * ib_event,__be64 service_id)1453 static int cma_save_ip_info(struct sockaddr *src_addr,
1454 			    struct sockaddr *dst_addr,
1455 			    const struct ib_cm_event *ib_event,
1456 			    __be64 service_id)
1457 {
1458 	struct cma_hdr *hdr;
1459 	__be16 port;
1460 
1461 	hdr = ib_event->private_data;
1462 	if (hdr->cma_version != CMA_VERSION)
1463 		return -EINVAL;
1464 
1465 	port = htons(cma_port_from_service_id(service_id));
1466 
1467 	switch (cma_get_ip_ver(hdr)) {
1468 	case 4:
1469 		cma_save_ip4_info((struct sockaddr_in *)src_addr,
1470 				  (struct sockaddr_in *)dst_addr, hdr, port);
1471 		break;
1472 	case 6:
1473 		cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1474 				  (struct sockaddr_in6 *)dst_addr, hdr, port);
1475 		break;
1476 	default:
1477 		return -EAFNOSUPPORT;
1478 	}
1479 
1480 	return 0;
1481 }
1482 
cma_save_net_info(struct sockaddr * src_addr,struct sockaddr * dst_addr,const struct rdma_cm_id * listen_id,const struct ib_cm_event * ib_event,sa_family_t sa_family,__be64 service_id)1483 static int cma_save_net_info(struct sockaddr *src_addr,
1484 			     struct sockaddr *dst_addr,
1485 			     const struct rdma_cm_id *listen_id,
1486 			     const struct ib_cm_event *ib_event,
1487 			     sa_family_t sa_family, __be64 service_id)
1488 {
1489 	if (sa_family == AF_IB) {
1490 		if (ib_event->event == IB_CM_REQ_RECEIVED)
1491 			cma_save_ib_info(src_addr, dst_addr, listen_id,
1492 					 ib_event->param.req_rcvd.primary_path);
1493 		else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1494 			cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1495 		return 0;
1496 	}
1497 
1498 	return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1499 }
1500 
cma_save_req_info(const struct ib_cm_event * ib_event,struct cma_req_info * req)1501 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1502 			     struct cma_req_info *req)
1503 {
1504 	const struct ib_cm_req_event_param *req_param =
1505 		&ib_event->param.req_rcvd;
1506 	const struct ib_cm_sidr_req_event_param *sidr_param =
1507 		&ib_event->param.sidr_req_rcvd;
1508 
1509 	switch (ib_event->event) {
1510 	case IB_CM_REQ_RECEIVED:
1511 		req->device	= req_param->listen_id->device;
1512 		req->port	= req_param->port;
1513 		memcpy(&req->local_gid, &req_param->primary_path->sgid,
1514 		       sizeof(req->local_gid));
1515 		req->has_gid	= true;
1516 		req->service_id = req_param->primary_path->service_id;
1517 		req->pkey	= be16_to_cpu(req_param->primary_path->pkey);
1518 		if (req->pkey != req_param->bth_pkey)
1519 			pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1520 					    "RDMA CMA: in the future this may cause the request to be dropped\n",
1521 					    req_param->bth_pkey, req->pkey);
1522 		break;
1523 	case IB_CM_SIDR_REQ_RECEIVED:
1524 		req->device	= sidr_param->listen_id->device;
1525 		req->port	= sidr_param->port;
1526 		req->has_gid	= false;
1527 		req->service_id	= sidr_param->service_id;
1528 		req->pkey	= sidr_param->pkey;
1529 		if (req->pkey != sidr_param->bth_pkey)
1530 			pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1531 					    "RDMA CMA: in the future this may cause the request to be dropped\n",
1532 					    sidr_param->bth_pkey, req->pkey);
1533 		break;
1534 	default:
1535 		return -EINVAL;
1536 	}
1537 
1538 	return 0;
1539 }
1540 
validate_ipv4_net_dev(struct net_device * net_dev,const struct sockaddr_in * dst_addr,const struct sockaddr_in * src_addr)1541 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1542 				  const struct sockaddr_in *dst_addr,
1543 				  const struct sockaddr_in *src_addr)
1544 {
1545 	__be32 daddr = dst_addr->sin_addr.s_addr,
1546 	       saddr = src_addr->sin_addr.s_addr;
1547 	struct fib_result res;
1548 	struct flowi4 fl4;
1549 	int err;
1550 	bool ret;
1551 
1552 	if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1553 	    ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1554 	    ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1555 	    ipv4_is_loopback(saddr))
1556 		return false;
1557 
1558 	memset(&fl4, 0, sizeof(fl4));
1559 	fl4.flowi4_oif = net_dev->ifindex;
1560 	fl4.daddr = daddr;
1561 	fl4.saddr = saddr;
1562 
1563 	rcu_read_lock();
1564 	err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1565 	ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1566 	rcu_read_unlock();
1567 
1568 	return ret;
1569 }
1570 
validate_ipv6_net_dev(struct net_device * net_dev,const struct sockaddr_in6 * dst_addr,const struct sockaddr_in6 * src_addr)1571 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1572 				  const struct sockaddr_in6 *dst_addr,
1573 				  const struct sockaddr_in6 *src_addr)
1574 {
1575 #if IS_ENABLED(CONFIG_IPV6)
1576 	const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1577 			   IPV6_ADDR_LINKLOCAL;
1578 	struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1579 					 &src_addr->sin6_addr, net_dev->ifindex,
1580 					 NULL, strict);
1581 	bool ret;
1582 
1583 	if (!rt)
1584 		return false;
1585 
1586 	ret = rt->rt6i_idev->dev == net_dev;
1587 	ip6_rt_put(rt);
1588 
1589 	return ret;
1590 #else
1591 	return false;
1592 #endif
1593 }
1594 
validate_net_dev(struct net_device * net_dev,const struct sockaddr * daddr,const struct sockaddr * saddr)1595 static bool validate_net_dev(struct net_device *net_dev,
1596 			     const struct sockaddr *daddr,
1597 			     const struct sockaddr *saddr)
1598 {
1599 	const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1600 	const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1601 	const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1602 	const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1603 
1604 	switch (daddr->sa_family) {
1605 	case AF_INET:
1606 		return saddr->sa_family == AF_INET &&
1607 		       validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1608 
1609 	case AF_INET6:
1610 		return saddr->sa_family == AF_INET6 &&
1611 		       validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1612 
1613 	default:
1614 		return false;
1615 	}
1616 }
1617 
1618 static struct net_device *
roce_get_net_dev_by_cm_event(const struct ib_cm_event * ib_event)1619 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1620 {
1621 	const struct ib_gid_attr *sgid_attr = NULL;
1622 	struct net_device *ndev;
1623 
1624 	if (ib_event->event == IB_CM_REQ_RECEIVED)
1625 		sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1626 	else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1627 		sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1628 
1629 	if (!sgid_attr)
1630 		return NULL;
1631 
1632 	rcu_read_lock();
1633 	ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1634 	if (IS_ERR(ndev))
1635 		ndev = NULL;
1636 	else
1637 		dev_hold(ndev);
1638 	rcu_read_unlock();
1639 	return ndev;
1640 }
1641 
cma_get_net_dev(const struct ib_cm_event * ib_event,struct cma_req_info * req)1642 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1643 					  struct cma_req_info *req)
1644 {
1645 	struct sockaddr *listen_addr =
1646 			(struct sockaddr *)&req->listen_addr_storage;
1647 	struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1648 	struct net_device *net_dev;
1649 	const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1650 	int err;
1651 
1652 	err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1653 			       req->service_id);
1654 	if (err)
1655 		return ERR_PTR(err);
1656 
1657 	if (rdma_protocol_roce(req->device, req->port))
1658 		net_dev = roce_get_net_dev_by_cm_event(ib_event);
1659 	else
1660 		net_dev = ib_get_net_dev_by_params(req->device, req->port,
1661 						   req->pkey,
1662 						   gid, listen_addr);
1663 	if (!net_dev)
1664 		return ERR_PTR(-ENODEV);
1665 
1666 	return net_dev;
1667 }
1668 
rdma_ps_from_service_id(__be64 service_id)1669 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1670 {
1671 	return (be64_to_cpu(service_id) >> 16) & 0xffff;
1672 }
1673 
cma_match_private_data(struct rdma_id_private * id_priv,const struct cma_hdr * hdr)1674 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1675 				   const struct cma_hdr *hdr)
1676 {
1677 	struct sockaddr *addr = cma_src_addr(id_priv);
1678 	__be32 ip4_addr;
1679 	struct in6_addr ip6_addr;
1680 
1681 	if (cma_any_addr(addr) && !id_priv->afonly)
1682 		return true;
1683 
1684 	switch (addr->sa_family) {
1685 	case AF_INET:
1686 		ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1687 		if (cma_get_ip_ver(hdr) != 4)
1688 			return false;
1689 		if (!cma_any_addr(addr) &&
1690 		    hdr->dst_addr.ip4.addr != ip4_addr)
1691 			return false;
1692 		break;
1693 	case AF_INET6:
1694 		ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1695 		if (cma_get_ip_ver(hdr) != 6)
1696 			return false;
1697 		if (!cma_any_addr(addr) &&
1698 		    memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1699 			return false;
1700 		break;
1701 	case AF_IB:
1702 		return true;
1703 	default:
1704 		return false;
1705 	}
1706 
1707 	return true;
1708 }
1709 
cma_protocol_roce(const struct rdma_cm_id * id)1710 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1711 {
1712 	struct ib_device *device = id->device;
1713 	const u32 port_num = id->port_num ?: rdma_start_port(device);
1714 
1715 	return rdma_protocol_roce(device, port_num);
1716 }
1717 
cma_is_req_ipv6_ll(const struct cma_req_info * req)1718 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1719 {
1720 	const struct sockaddr *daddr =
1721 			(const struct sockaddr *)&req->listen_addr_storage;
1722 	const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1723 
1724 	/* Returns true if the req is for IPv6 link local */
1725 	return (daddr->sa_family == AF_INET6 &&
1726 		(ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1727 }
1728 
cma_match_net_dev(const struct rdma_cm_id * id,const struct net_device * net_dev,const struct cma_req_info * req)1729 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1730 			      const struct net_device *net_dev,
1731 			      const struct cma_req_info *req)
1732 {
1733 	const struct rdma_addr *addr = &id->route.addr;
1734 
1735 	if (!net_dev)
1736 		/* This request is an AF_IB request */
1737 		return (!id->port_num || id->port_num == req->port) &&
1738 		       (addr->src_addr.ss_family == AF_IB);
1739 
1740 	/*
1741 	 * If the request is not for IPv6 link local, allow matching
1742 	 * request to any netdevice of the one or multiport rdma device.
1743 	 */
1744 	if (!cma_is_req_ipv6_ll(req))
1745 		return true;
1746 	/*
1747 	 * Net namespaces must match, and if the listner is listening
1748 	 * on a specific netdevice than netdevice must match as well.
1749 	 */
1750 	if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1751 	    (!!addr->dev_addr.bound_dev_if ==
1752 	     (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1753 		return true;
1754 	else
1755 		return false;
1756 }
1757 
cma_find_listener(const struct rdma_bind_list * bind_list,const struct ib_cm_id * cm_id,const struct ib_cm_event * ib_event,const struct cma_req_info * req,const struct net_device * net_dev)1758 static struct rdma_id_private *cma_find_listener(
1759 		const struct rdma_bind_list *bind_list,
1760 		const struct ib_cm_id *cm_id,
1761 		const struct ib_cm_event *ib_event,
1762 		const struct cma_req_info *req,
1763 		const struct net_device *net_dev)
1764 {
1765 	struct rdma_id_private *id_priv, *id_priv_dev;
1766 
1767 	lockdep_assert_held(&lock);
1768 
1769 	if (!bind_list)
1770 		return ERR_PTR(-EINVAL);
1771 
1772 	hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1773 		if (cma_match_private_data(id_priv, ib_event->private_data)) {
1774 			if (id_priv->id.device == cm_id->device &&
1775 			    cma_match_net_dev(&id_priv->id, net_dev, req))
1776 				return id_priv;
1777 			list_for_each_entry(id_priv_dev,
1778 					    &id_priv->listen_list,
1779 					    listen_item) {
1780 				if (id_priv_dev->id.device == cm_id->device &&
1781 				    cma_match_net_dev(&id_priv_dev->id,
1782 						      net_dev, req))
1783 					return id_priv_dev;
1784 			}
1785 		}
1786 	}
1787 
1788 	return ERR_PTR(-EINVAL);
1789 }
1790 
1791 static struct rdma_id_private *
cma_ib_id_from_event(struct ib_cm_id * cm_id,const struct ib_cm_event * ib_event,struct cma_req_info * req,struct net_device ** net_dev)1792 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1793 		     const struct ib_cm_event *ib_event,
1794 		     struct cma_req_info *req,
1795 		     struct net_device **net_dev)
1796 {
1797 	struct rdma_bind_list *bind_list;
1798 	struct rdma_id_private *id_priv;
1799 	int err;
1800 
1801 	err = cma_save_req_info(ib_event, req);
1802 	if (err)
1803 		return ERR_PTR(err);
1804 
1805 	*net_dev = cma_get_net_dev(ib_event, req);
1806 	if (IS_ERR(*net_dev)) {
1807 		if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1808 			/* Assuming the protocol is AF_IB */
1809 			*net_dev = NULL;
1810 		} else {
1811 			return ERR_CAST(*net_dev);
1812 		}
1813 	}
1814 
1815 	mutex_lock(&lock);
1816 	/*
1817 	 * Net namespace might be getting deleted while route lookup,
1818 	 * cm_id lookup is in progress. Therefore, perform netdevice
1819 	 * validation, cm_id lookup under rcu lock.
1820 	 * RCU lock along with netdevice state check, synchronizes with
1821 	 * netdevice migrating to different net namespace and also avoids
1822 	 * case where net namespace doesn't get deleted while lookup is in
1823 	 * progress.
1824 	 * If the device state is not IFF_UP, its properties such as ifindex
1825 	 * and nd_net cannot be trusted to remain valid without rcu lock.
1826 	 * net/core/dev.c change_net_namespace() ensures to synchronize with
1827 	 * ongoing operations on net device after device is closed using
1828 	 * synchronize_net().
1829 	 */
1830 	rcu_read_lock();
1831 	if (*net_dev) {
1832 		/*
1833 		 * If netdevice is down, it is likely that it is administratively
1834 		 * down or it might be migrating to different namespace.
1835 		 * In that case avoid further processing, as the net namespace
1836 		 * or ifindex may change.
1837 		 */
1838 		if (((*net_dev)->flags & IFF_UP) == 0) {
1839 			id_priv = ERR_PTR(-EHOSTUNREACH);
1840 			goto err;
1841 		}
1842 
1843 		if (!validate_net_dev(*net_dev,
1844 				 (struct sockaddr *)&req->src_addr_storage,
1845 				 (struct sockaddr *)&req->listen_addr_storage)) {
1846 			id_priv = ERR_PTR(-EHOSTUNREACH);
1847 			goto err;
1848 		}
1849 	}
1850 
1851 	bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1852 				rdma_ps_from_service_id(req->service_id),
1853 				cma_port_from_service_id(req->service_id));
1854 	id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1855 err:
1856 	rcu_read_unlock();
1857 	mutex_unlock(&lock);
1858 	if (IS_ERR(id_priv) && *net_dev) {
1859 		dev_put(*net_dev);
1860 		*net_dev = NULL;
1861 	}
1862 	return id_priv;
1863 }
1864 
cma_user_data_offset(struct rdma_id_private * id_priv)1865 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1866 {
1867 	return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1868 }
1869 
cma_cancel_route(struct rdma_id_private * id_priv)1870 static void cma_cancel_route(struct rdma_id_private *id_priv)
1871 {
1872 	if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1873 		if (id_priv->query)
1874 			ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1875 	}
1876 }
1877 
_cma_cancel_listens(struct rdma_id_private * id_priv)1878 static void _cma_cancel_listens(struct rdma_id_private *id_priv)
1879 {
1880 	struct rdma_id_private *dev_id_priv;
1881 
1882 	lockdep_assert_held(&lock);
1883 
1884 	/*
1885 	 * Remove from listen_any_list to prevent added devices from spawning
1886 	 * additional listen requests.
1887 	 */
1888 	list_del_init(&id_priv->listen_any_item);
1889 
1890 	while (!list_empty(&id_priv->listen_list)) {
1891 		dev_id_priv =
1892 			list_first_entry(&id_priv->listen_list,
1893 					 struct rdma_id_private, listen_item);
1894 		/* sync with device removal to avoid duplicate destruction */
1895 		list_del_init(&dev_id_priv->device_item);
1896 		list_del_init(&dev_id_priv->listen_item);
1897 		mutex_unlock(&lock);
1898 
1899 		rdma_destroy_id(&dev_id_priv->id);
1900 		mutex_lock(&lock);
1901 	}
1902 }
1903 
cma_cancel_listens(struct rdma_id_private * id_priv)1904 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1905 {
1906 	mutex_lock(&lock);
1907 	_cma_cancel_listens(id_priv);
1908 	mutex_unlock(&lock);
1909 }
1910 
cma_cancel_operation(struct rdma_id_private * id_priv,enum rdma_cm_state state)1911 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1912 				 enum rdma_cm_state state)
1913 {
1914 	switch (state) {
1915 	case RDMA_CM_ADDR_QUERY:
1916 		/*
1917 		 * We can avoid doing the rdma_addr_cancel() based on state,
1918 		 * only RDMA_CM_ADDR_QUERY has a work that could still execute.
1919 		 * Notice that the addr_handler work could still be exiting
1920 		 * outside this state, however due to the interaction with the
1921 		 * handler_mutex the work is guaranteed not to touch id_priv
1922 		 * during exit.
1923 		 */
1924 		rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1925 		break;
1926 	case RDMA_CM_ROUTE_QUERY:
1927 		cma_cancel_route(id_priv);
1928 		break;
1929 	case RDMA_CM_LISTEN:
1930 		if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1931 			cma_cancel_listens(id_priv);
1932 		break;
1933 	default:
1934 		break;
1935 	}
1936 }
1937 
cma_release_port(struct rdma_id_private * id_priv)1938 static void cma_release_port(struct rdma_id_private *id_priv)
1939 {
1940 	struct rdma_bind_list *bind_list = id_priv->bind_list;
1941 	struct net *net = id_priv->id.route.addr.dev_addr.net;
1942 
1943 	if (!bind_list)
1944 		return;
1945 
1946 	mutex_lock(&lock);
1947 	hlist_del(&id_priv->node);
1948 	if (hlist_empty(&bind_list->owners)) {
1949 		cma_ps_remove(net, bind_list->ps, bind_list->port);
1950 		kfree(bind_list);
1951 	}
1952 	mutex_unlock(&lock);
1953 }
1954 
destroy_mc(struct rdma_id_private * id_priv,struct cma_multicast * mc)1955 static void destroy_mc(struct rdma_id_private *id_priv,
1956 		       struct cma_multicast *mc)
1957 {
1958 	bool send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
1959 
1960 	if (rdma_cap_ib_mcast(id_priv->id.device, id_priv->id.port_num))
1961 		ib_sa_free_multicast(mc->sa_mc);
1962 
1963 	if (rdma_protocol_roce(id_priv->id.device, id_priv->id.port_num)) {
1964 		struct rdma_dev_addr *dev_addr =
1965 			&id_priv->id.route.addr.dev_addr;
1966 		struct net_device *ndev = NULL;
1967 
1968 		if (dev_addr->bound_dev_if)
1969 			ndev = dev_get_by_index(dev_addr->net,
1970 						dev_addr->bound_dev_if);
1971 		if (ndev && !send_only) {
1972 			enum ib_gid_type gid_type;
1973 			union ib_gid mgid;
1974 
1975 			gid_type = id_priv->cma_dev->default_gid_type
1976 					   [id_priv->id.port_num -
1977 					    rdma_start_port(
1978 						    id_priv->cma_dev->device)];
1979 			cma_iboe_set_mgid((struct sockaddr *)&mc->addr, &mgid,
1980 					  gid_type);
1981 			cma_igmp_send(ndev, &mgid, false);
1982 		}
1983 		dev_put(ndev);
1984 
1985 		cancel_work_sync(&mc->iboe_join.work);
1986 	}
1987 	kfree(mc);
1988 }
1989 
cma_leave_mc_groups(struct rdma_id_private * id_priv)1990 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1991 {
1992 	struct cma_multicast *mc;
1993 
1994 	while (!list_empty(&id_priv->mc_list)) {
1995 		mc = list_first_entry(&id_priv->mc_list, struct cma_multicast,
1996 				      list);
1997 		list_del(&mc->list);
1998 		destroy_mc(id_priv, mc);
1999 	}
2000 }
2001 
_destroy_id(struct rdma_id_private * id_priv,enum rdma_cm_state state)2002 static void _destroy_id(struct rdma_id_private *id_priv,
2003 			enum rdma_cm_state state)
2004 {
2005 	cma_cancel_operation(id_priv, state);
2006 
2007 	rdma_restrack_del(&id_priv->res);
2008 	cma_remove_id_from_tree(id_priv);
2009 	if (id_priv->cma_dev) {
2010 		if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
2011 			if (id_priv->cm_id.ib)
2012 				ib_destroy_cm_id(id_priv->cm_id.ib);
2013 		} else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
2014 			if (id_priv->cm_id.iw)
2015 				iw_destroy_cm_id(id_priv->cm_id.iw);
2016 		}
2017 		cma_leave_mc_groups(id_priv);
2018 		cma_release_dev(id_priv);
2019 	}
2020 
2021 	cma_release_port(id_priv);
2022 	cma_id_put(id_priv);
2023 	wait_for_completion(&id_priv->comp);
2024 
2025 	if (id_priv->internal_id)
2026 		cma_id_put(id_priv->id.context);
2027 
2028 	kfree(id_priv->id.route.path_rec);
2029 	kfree(id_priv->id.route.path_rec_inbound);
2030 	kfree(id_priv->id.route.path_rec_outbound);
2031 
2032 	put_net(id_priv->id.route.addr.dev_addr.net);
2033 	kfree(id_priv);
2034 }
2035 
2036 /*
2037  * destroy an ID from within the handler_mutex. This ensures that no other
2038  * handlers can start running concurrently.
2039  */
destroy_id_handler_unlock(struct rdma_id_private * id_priv)2040 static void destroy_id_handler_unlock(struct rdma_id_private *id_priv)
2041 	__releases(&idprv->handler_mutex)
2042 {
2043 	enum rdma_cm_state state;
2044 	unsigned long flags;
2045 
2046 	trace_cm_id_destroy(id_priv);
2047 
2048 	/*
2049 	 * Setting the state to destroyed under the handler mutex provides a
2050 	 * fence against calling handler callbacks. If this is invoked due to
2051 	 * the failure of a handler callback then it guarentees that no future
2052 	 * handlers will be called.
2053 	 */
2054 	lockdep_assert_held(&id_priv->handler_mutex);
2055 	spin_lock_irqsave(&id_priv->lock, flags);
2056 	state = id_priv->state;
2057 	id_priv->state = RDMA_CM_DESTROYING;
2058 	spin_unlock_irqrestore(&id_priv->lock, flags);
2059 	mutex_unlock(&id_priv->handler_mutex);
2060 	_destroy_id(id_priv, state);
2061 }
2062 
rdma_destroy_id(struct rdma_cm_id * id)2063 void rdma_destroy_id(struct rdma_cm_id *id)
2064 {
2065 	struct rdma_id_private *id_priv =
2066 		container_of(id, struct rdma_id_private, id);
2067 
2068 	mutex_lock(&id_priv->handler_mutex);
2069 	destroy_id_handler_unlock(id_priv);
2070 }
2071 EXPORT_SYMBOL(rdma_destroy_id);
2072 
cma_rep_recv(struct rdma_id_private * id_priv)2073 static int cma_rep_recv(struct rdma_id_private *id_priv)
2074 {
2075 	int ret;
2076 
2077 	ret = cma_modify_qp_rtr(id_priv, NULL);
2078 	if (ret)
2079 		goto reject;
2080 
2081 	ret = cma_modify_qp_rts(id_priv, NULL);
2082 	if (ret)
2083 		goto reject;
2084 
2085 	trace_cm_send_rtu(id_priv);
2086 	ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
2087 	if (ret)
2088 		goto reject;
2089 
2090 	return 0;
2091 reject:
2092 	pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
2093 	cma_modify_qp_err(id_priv);
2094 	trace_cm_send_rej(id_priv);
2095 	ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
2096 		       NULL, 0, NULL, 0);
2097 	return ret;
2098 }
2099 
cma_set_rep_event_data(struct rdma_cm_event * event,const struct ib_cm_rep_event_param * rep_data,void * private_data)2100 static void cma_set_rep_event_data(struct rdma_cm_event *event,
2101 				   const struct ib_cm_rep_event_param *rep_data,
2102 				   void *private_data)
2103 {
2104 	event->param.conn.private_data = private_data;
2105 	event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
2106 	event->param.conn.responder_resources = rep_data->responder_resources;
2107 	event->param.conn.initiator_depth = rep_data->initiator_depth;
2108 	event->param.conn.flow_control = rep_data->flow_control;
2109 	event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
2110 	event->param.conn.srq = rep_data->srq;
2111 	event->param.conn.qp_num = rep_data->remote_qpn;
2112 
2113 	event->ece.vendor_id = rep_data->ece.vendor_id;
2114 	event->ece.attr_mod = rep_data->ece.attr_mod;
2115 }
2116 
cma_cm_event_handler(struct rdma_id_private * id_priv,struct rdma_cm_event * event)2117 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
2118 				struct rdma_cm_event *event)
2119 {
2120 	int ret;
2121 
2122 	lockdep_assert_held(&id_priv->handler_mutex);
2123 
2124 	trace_cm_event_handler(id_priv, event);
2125 	ret = id_priv->id.event_handler(&id_priv->id, event);
2126 	trace_cm_event_done(id_priv, event, ret);
2127 	return ret;
2128 }
2129 
cma_ib_handler(struct ib_cm_id * cm_id,const struct ib_cm_event * ib_event)2130 static int cma_ib_handler(struct ib_cm_id *cm_id,
2131 			  const struct ib_cm_event *ib_event)
2132 {
2133 	struct rdma_id_private *id_priv = cm_id->context;
2134 	struct rdma_cm_event event = {};
2135 	enum rdma_cm_state state;
2136 	int ret;
2137 
2138 	mutex_lock(&id_priv->handler_mutex);
2139 	state = READ_ONCE(id_priv->state);
2140 	if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
2141 	     state != RDMA_CM_CONNECT) ||
2142 	    (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
2143 	     state != RDMA_CM_DISCONNECT))
2144 		goto out;
2145 
2146 	switch (ib_event->event) {
2147 	case IB_CM_REQ_ERROR:
2148 	case IB_CM_REP_ERROR:
2149 		event.event = RDMA_CM_EVENT_UNREACHABLE;
2150 		event.status = -ETIMEDOUT;
2151 		break;
2152 	case IB_CM_REP_RECEIVED:
2153 		if (state == RDMA_CM_CONNECT &&
2154 		    (id_priv->id.qp_type != IB_QPT_UD)) {
2155 			trace_cm_send_mra(id_priv);
2156 			ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2157 		}
2158 		if (id_priv->id.qp) {
2159 			event.status = cma_rep_recv(id_priv);
2160 			event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
2161 						     RDMA_CM_EVENT_ESTABLISHED;
2162 		} else {
2163 			event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
2164 		}
2165 		cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
2166 				       ib_event->private_data);
2167 		break;
2168 	case IB_CM_RTU_RECEIVED:
2169 	case IB_CM_USER_ESTABLISHED:
2170 		event.event = RDMA_CM_EVENT_ESTABLISHED;
2171 		break;
2172 	case IB_CM_DREQ_ERROR:
2173 		event.status = -ETIMEDOUT;
2174 		fallthrough;
2175 	case IB_CM_DREQ_RECEIVED:
2176 	case IB_CM_DREP_RECEIVED:
2177 		if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
2178 				   RDMA_CM_DISCONNECT))
2179 			goto out;
2180 		event.event = RDMA_CM_EVENT_DISCONNECTED;
2181 		break;
2182 	case IB_CM_TIMEWAIT_EXIT:
2183 		event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
2184 		break;
2185 	case IB_CM_MRA_RECEIVED:
2186 		/* ignore event */
2187 		goto out;
2188 	case IB_CM_REJ_RECEIVED:
2189 		pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
2190 										ib_event->param.rej_rcvd.reason));
2191 		cma_modify_qp_err(id_priv);
2192 		event.status = ib_event->param.rej_rcvd.reason;
2193 		event.event = RDMA_CM_EVENT_REJECTED;
2194 		event.param.conn.private_data = ib_event->private_data;
2195 		event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
2196 		break;
2197 	default:
2198 		pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2199 		       ib_event->event);
2200 		goto out;
2201 	}
2202 
2203 	ret = cma_cm_event_handler(id_priv, &event);
2204 	if (ret) {
2205 		/* Destroy the CM ID by returning a non-zero value. */
2206 		id_priv->cm_id.ib = NULL;
2207 		destroy_id_handler_unlock(id_priv);
2208 		return ret;
2209 	}
2210 out:
2211 	mutex_unlock(&id_priv->handler_mutex);
2212 	return 0;
2213 }
2214 
2215 static struct rdma_id_private *
cma_ib_new_conn_id(const struct rdma_cm_id * listen_id,const struct ib_cm_event * ib_event,struct net_device * net_dev)2216 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2217 		   const struct ib_cm_event *ib_event,
2218 		   struct net_device *net_dev)
2219 {
2220 	struct rdma_id_private *listen_id_priv;
2221 	struct rdma_id_private *id_priv;
2222 	struct rdma_cm_id *id;
2223 	struct rdma_route *rt;
2224 	const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2225 	struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2226 	const __be64 service_id =
2227 		ib_event->param.req_rcvd.primary_path->service_id;
2228 	int ret;
2229 
2230 	listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2231 	id_priv = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2232 				   listen_id->event_handler, listen_id->context,
2233 				   listen_id->ps,
2234 				   ib_event->param.req_rcvd.qp_type,
2235 				   listen_id_priv);
2236 	if (IS_ERR(id_priv))
2237 		return NULL;
2238 
2239 	id = &id_priv->id;
2240 	if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2241 			      (struct sockaddr *)&id->route.addr.dst_addr,
2242 			      listen_id, ib_event, ss_family, service_id))
2243 		goto err;
2244 
2245 	rt = &id->route;
2246 	rt->num_pri_alt_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2247 	rt->path_rec = kmalloc_array(rt->num_pri_alt_paths,
2248 				     sizeof(*rt->path_rec), GFP_KERNEL);
2249 	if (!rt->path_rec)
2250 		goto err;
2251 
2252 	rt->path_rec[0] = *path;
2253 	if (rt->num_pri_alt_paths == 2)
2254 		rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2255 
2256 	if (net_dev) {
2257 		rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2258 	} else {
2259 		if (!cma_protocol_roce(listen_id) &&
2260 		    cma_any_addr(cma_src_addr(id_priv))) {
2261 			rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2262 			rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2263 			ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2264 		} else if (!cma_any_addr(cma_src_addr(id_priv))) {
2265 			ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2266 			if (ret)
2267 				goto err;
2268 		}
2269 	}
2270 	rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2271 
2272 	id_priv->state = RDMA_CM_CONNECT;
2273 	return id_priv;
2274 
2275 err:
2276 	rdma_destroy_id(id);
2277 	return NULL;
2278 }
2279 
2280 static struct rdma_id_private *
cma_ib_new_udp_id(const struct rdma_cm_id * listen_id,const struct ib_cm_event * ib_event,struct net_device * net_dev)2281 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2282 		  const struct ib_cm_event *ib_event,
2283 		  struct net_device *net_dev)
2284 {
2285 	const struct rdma_id_private *listen_id_priv;
2286 	struct rdma_id_private *id_priv;
2287 	struct rdma_cm_id *id;
2288 	const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2289 	struct net *net = listen_id->route.addr.dev_addr.net;
2290 	int ret;
2291 
2292 	listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2293 	id_priv = __rdma_create_id(net, listen_id->event_handler,
2294 				   listen_id->context, listen_id->ps, IB_QPT_UD,
2295 				   listen_id_priv);
2296 	if (IS_ERR(id_priv))
2297 		return NULL;
2298 
2299 	id = &id_priv->id;
2300 	if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2301 			      (struct sockaddr *)&id->route.addr.dst_addr,
2302 			      listen_id, ib_event, ss_family,
2303 			      ib_event->param.sidr_req_rcvd.service_id))
2304 		goto err;
2305 
2306 	if (net_dev) {
2307 		rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2308 	} else {
2309 		if (!cma_any_addr(cma_src_addr(id_priv))) {
2310 			ret = cma_translate_addr(cma_src_addr(id_priv),
2311 						 &id->route.addr.dev_addr);
2312 			if (ret)
2313 				goto err;
2314 		}
2315 	}
2316 
2317 	id_priv->state = RDMA_CM_CONNECT;
2318 	return id_priv;
2319 err:
2320 	rdma_destroy_id(id);
2321 	return NULL;
2322 }
2323 
cma_set_req_event_data(struct rdma_cm_event * event,const struct ib_cm_req_event_param * req_data,void * private_data,int offset)2324 static void cma_set_req_event_data(struct rdma_cm_event *event,
2325 				   const struct ib_cm_req_event_param *req_data,
2326 				   void *private_data, int offset)
2327 {
2328 	event->param.conn.private_data = private_data + offset;
2329 	event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2330 	event->param.conn.responder_resources = req_data->responder_resources;
2331 	event->param.conn.initiator_depth = req_data->initiator_depth;
2332 	event->param.conn.flow_control = req_data->flow_control;
2333 	event->param.conn.retry_count = req_data->retry_count;
2334 	event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2335 	event->param.conn.srq = req_data->srq;
2336 	event->param.conn.qp_num = req_data->remote_qpn;
2337 
2338 	event->ece.vendor_id = req_data->ece.vendor_id;
2339 	event->ece.attr_mod = req_data->ece.attr_mod;
2340 }
2341 
cma_ib_check_req_qp_type(const struct rdma_cm_id * id,const struct ib_cm_event * ib_event)2342 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2343 				    const struct ib_cm_event *ib_event)
2344 {
2345 	return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2346 		 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2347 		((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2348 		 (id->qp_type == IB_QPT_UD)) ||
2349 		(!id->qp_type));
2350 }
2351 
cma_ib_req_handler(struct ib_cm_id * cm_id,const struct ib_cm_event * ib_event)2352 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2353 			      const struct ib_cm_event *ib_event)
2354 {
2355 	struct rdma_id_private *listen_id, *conn_id = NULL;
2356 	struct rdma_cm_event event = {};
2357 	struct cma_req_info req = {};
2358 	struct net_device *net_dev;
2359 	u8 offset;
2360 	int ret;
2361 
2362 	listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2363 	if (IS_ERR(listen_id))
2364 		return PTR_ERR(listen_id);
2365 
2366 	trace_cm_req_handler(listen_id, ib_event->event);
2367 	if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2368 		ret = -EINVAL;
2369 		goto net_dev_put;
2370 	}
2371 
2372 	mutex_lock(&listen_id->handler_mutex);
2373 	if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN) {
2374 		ret = -ECONNABORTED;
2375 		goto err_unlock;
2376 	}
2377 
2378 	offset = cma_user_data_offset(listen_id);
2379 	event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2380 	if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2381 		conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2382 		event.param.ud.private_data = ib_event->private_data + offset;
2383 		event.param.ud.private_data_len =
2384 				IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2385 	} else {
2386 		conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2387 		cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2388 				       ib_event->private_data, offset);
2389 	}
2390 	if (!conn_id) {
2391 		ret = -ENOMEM;
2392 		goto err_unlock;
2393 	}
2394 
2395 	mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2396 	ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2397 	if (ret) {
2398 		destroy_id_handler_unlock(conn_id);
2399 		goto err_unlock;
2400 	}
2401 
2402 	conn_id->cm_id.ib = cm_id;
2403 	cm_id->context = conn_id;
2404 	cm_id->cm_handler = cma_ib_handler;
2405 
2406 	ret = cma_cm_event_handler(conn_id, &event);
2407 	if (ret) {
2408 		/* Destroy the CM ID by returning a non-zero value. */
2409 		conn_id->cm_id.ib = NULL;
2410 		mutex_unlock(&listen_id->handler_mutex);
2411 		destroy_id_handler_unlock(conn_id);
2412 		goto net_dev_put;
2413 	}
2414 
2415 	if (READ_ONCE(conn_id->state) == RDMA_CM_CONNECT &&
2416 	    conn_id->id.qp_type != IB_QPT_UD) {
2417 		trace_cm_send_mra(cm_id->context);
2418 		ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2419 	}
2420 	mutex_unlock(&conn_id->handler_mutex);
2421 
2422 err_unlock:
2423 	mutex_unlock(&listen_id->handler_mutex);
2424 
2425 net_dev_put:
2426 	if (net_dev)
2427 		dev_put(net_dev);
2428 
2429 	return ret;
2430 }
2431 
rdma_get_service_id(struct rdma_cm_id * id,struct sockaddr * addr)2432 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2433 {
2434 	if (addr->sa_family == AF_IB)
2435 		return ((struct sockaddr_ib *) addr)->sib_sid;
2436 
2437 	return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2438 }
2439 EXPORT_SYMBOL(rdma_get_service_id);
2440 
rdma_read_gids(struct rdma_cm_id * cm_id,union ib_gid * sgid,union ib_gid * dgid)2441 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2442 		    union ib_gid *dgid)
2443 {
2444 	struct rdma_addr *addr = &cm_id->route.addr;
2445 
2446 	if (!cm_id->device) {
2447 		if (sgid)
2448 			memset(sgid, 0, sizeof(*sgid));
2449 		if (dgid)
2450 			memset(dgid, 0, sizeof(*dgid));
2451 		return;
2452 	}
2453 
2454 	if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2455 		if (sgid)
2456 			rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2457 		if (dgid)
2458 			rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2459 	} else {
2460 		if (sgid)
2461 			rdma_addr_get_sgid(&addr->dev_addr, sgid);
2462 		if (dgid)
2463 			rdma_addr_get_dgid(&addr->dev_addr, dgid);
2464 	}
2465 }
2466 EXPORT_SYMBOL(rdma_read_gids);
2467 
cma_iw_handler(struct iw_cm_id * iw_id,struct iw_cm_event * iw_event)2468 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2469 {
2470 	struct rdma_id_private *id_priv = iw_id->context;
2471 	struct rdma_cm_event event = {};
2472 	int ret = 0;
2473 	struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2474 	struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2475 
2476 	mutex_lock(&id_priv->handler_mutex);
2477 	if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
2478 		goto out;
2479 
2480 	switch (iw_event->event) {
2481 	case IW_CM_EVENT_CLOSE:
2482 		event.event = RDMA_CM_EVENT_DISCONNECTED;
2483 		break;
2484 	case IW_CM_EVENT_CONNECT_REPLY:
2485 		memcpy(cma_src_addr(id_priv), laddr,
2486 		       rdma_addr_size(laddr));
2487 		memcpy(cma_dst_addr(id_priv), raddr,
2488 		       rdma_addr_size(raddr));
2489 		switch (iw_event->status) {
2490 		case 0:
2491 			event.event = RDMA_CM_EVENT_ESTABLISHED;
2492 			event.param.conn.initiator_depth = iw_event->ird;
2493 			event.param.conn.responder_resources = iw_event->ord;
2494 			break;
2495 		case -ECONNRESET:
2496 		case -ECONNREFUSED:
2497 			event.event = RDMA_CM_EVENT_REJECTED;
2498 			break;
2499 		case -ETIMEDOUT:
2500 			event.event = RDMA_CM_EVENT_UNREACHABLE;
2501 			break;
2502 		default:
2503 			event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2504 			break;
2505 		}
2506 		break;
2507 	case IW_CM_EVENT_ESTABLISHED:
2508 		event.event = RDMA_CM_EVENT_ESTABLISHED;
2509 		event.param.conn.initiator_depth = iw_event->ird;
2510 		event.param.conn.responder_resources = iw_event->ord;
2511 		break;
2512 	default:
2513 		goto out;
2514 	}
2515 
2516 	event.status = iw_event->status;
2517 	event.param.conn.private_data = iw_event->private_data;
2518 	event.param.conn.private_data_len = iw_event->private_data_len;
2519 	ret = cma_cm_event_handler(id_priv, &event);
2520 	if (ret) {
2521 		/* Destroy the CM ID by returning a non-zero value. */
2522 		id_priv->cm_id.iw = NULL;
2523 		destroy_id_handler_unlock(id_priv);
2524 		return ret;
2525 	}
2526 
2527 out:
2528 	mutex_unlock(&id_priv->handler_mutex);
2529 	return ret;
2530 }
2531 
iw_conn_req_handler(struct iw_cm_id * cm_id,struct iw_cm_event * iw_event)2532 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2533 			       struct iw_cm_event *iw_event)
2534 {
2535 	struct rdma_id_private *listen_id, *conn_id;
2536 	struct rdma_cm_event event = {};
2537 	int ret = -ECONNABORTED;
2538 	struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2539 	struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2540 
2541 	event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2542 	event.param.conn.private_data = iw_event->private_data;
2543 	event.param.conn.private_data_len = iw_event->private_data_len;
2544 	event.param.conn.initiator_depth = iw_event->ird;
2545 	event.param.conn.responder_resources = iw_event->ord;
2546 
2547 	listen_id = cm_id->context;
2548 
2549 	mutex_lock(&listen_id->handler_mutex);
2550 	if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN)
2551 		goto out;
2552 
2553 	/* Create a new RDMA id for the new IW CM ID */
2554 	conn_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2555 				   listen_id->id.event_handler,
2556 				   listen_id->id.context, RDMA_PS_TCP,
2557 				   IB_QPT_RC, listen_id);
2558 	if (IS_ERR(conn_id)) {
2559 		ret = -ENOMEM;
2560 		goto out;
2561 	}
2562 	mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2563 	conn_id->state = RDMA_CM_CONNECT;
2564 
2565 	ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2566 	if (ret) {
2567 		mutex_unlock(&listen_id->handler_mutex);
2568 		destroy_id_handler_unlock(conn_id);
2569 		return ret;
2570 	}
2571 
2572 	ret = cma_iw_acquire_dev(conn_id, listen_id);
2573 	if (ret) {
2574 		mutex_unlock(&listen_id->handler_mutex);
2575 		destroy_id_handler_unlock(conn_id);
2576 		return ret;
2577 	}
2578 
2579 	conn_id->cm_id.iw = cm_id;
2580 	cm_id->context = conn_id;
2581 	cm_id->cm_handler = cma_iw_handler;
2582 
2583 	memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2584 	memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2585 
2586 	ret = cma_cm_event_handler(conn_id, &event);
2587 	if (ret) {
2588 		/* User wants to destroy the CM ID */
2589 		conn_id->cm_id.iw = NULL;
2590 		mutex_unlock(&listen_id->handler_mutex);
2591 		destroy_id_handler_unlock(conn_id);
2592 		return ret;
2593 	}
2594 
2595 	mutex_unlock(&conn_id->handler_mutex);
2596 
2597 out:
2598 	mutex_unlock(&listen_id->handler_mutex);
2599 	return ret;
2600 }
2601 
cma_ib_listen(struct rdma_id_private * id_priv)2602 static int cma_ib_listen(struct rdma_id_private *id_priv)
2603 {
2604 	struct sockaddr *addr;
2605 	struct ib_cm_id	*id;
2606 	__be64 svc_id;
2607 
2608 	addr = cma_src_addr(id_priv);
2609 	svc_id = rdma_get_service_id(&id_priv->id, addr);
2610 	id = ib_cm_insert_listen(id_priv->id.device,
2611 				 cma_ib_req_handler, svc_id);
2612 	if (IS_ERR(id))
2613 		return PTR_ERR(id);
2614 	id_priv->cm_id.ib = id;
2615 
2616 	return 0;
2617 }
2618 
cma_iw_listen(struct rdma_id_private * id_priv,int backlog)2619 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2620 {
2621 	int ret;
2622 	struct iw_cm_id	*id;
2623 
2624 	id = iw_create_cm_id(id_priv->id.device,
2625 			     iw_conn_req_handler,
2626 			     id_priv);
2627 	if (IS_ERR(id))
2628 		return PTR_ERR(id);
2629 
2630 	mutex_lock(&id_priv->qp_mutex);
2631 	id->tos = id_priv->tos;
2632 	id->tos_set = id_priv->tos_set;
2633 	mutex_unlock(&id_priv->qp_mutex);
2634 	id->afonly = id_priv->afonly;
2635 	id_priv->cm_id.iw = id;
2636 
2637 	memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2638 	       rdma_addr_size(cma_src_addr(id_priv)));
2639 
2640 	ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2641 
2642 	if (ret) {
2643 		iw_destroy_cm_id(id_priv->cm_id.iw);
2644 		id_priv->cm_id.iw = NULL;
2645 	}
2646 
2647 	return ret;
2648 }
2649 
cma_listen_handler(struct rdma_cm_id * id,struct rdma_cm_event * event)2650 static int cma_listen_handler(struct rdma_cm_id *id,
2651 			      struct rdma_cm_event *event)
2652 {
2653 	struct rdma_id_private *id_priv = id->context;
2654 
2655 	/* Listening IDs are always destroyed on removal */
2656 	if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2657 		return -1;
2658 
2659 	id->context = id_priv->id.context;
2660 	id->event_handler = id_priv->id.event_handler;
2661 	trace_cm_event_handler(id_priv, event);
2662 	return id_priv->id.event_handler(id, event);
2663 }
2664 
cma_listen_on_dev(struct rdma_id_private * id_priv,struct cma_device * cma_dev,struct rdma_id_private ** to_destroy)2665 static int cma_listen_on_dev(struct rdma_id_private *id_priv,
2666 			     struct cma_device *cma_dev,
2667 			     struct rdma_id_private **to_destroy)
2668 {
2669 	struct rdma_id_private *dev_id_priv;
2670 	struct net *net = id_priv->id.route.addr.dev_addr.net;
2671 	int ret;
2672 
2673 	lockdep_assert_held(&lock);
2674 
2675 	*to_destroy = NULL;
2676 	if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2677 		return 0;
2678 
2679 	dev_id_priv =
2680 		__rdma_create_id(net, cma_listen_handler, id_priv,
2681 				 id_priv->id.ps, id_priv->id.qp_type, id_priv);
2682 	if (IS_ERR(dev_id_priv))
2683 		return PTR_ERR(dev_id_priv);
2684 
2685 	dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2686 	memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2687 	       rdma_addr_size(cma_src_addr(id_priv)));
2688 
2689 	_cma_attach_to_dev(dev_id_priv, cma_dev);
2690 	rdma_restrack_add(&dev_id_priv->res);
2691 	cma_id_get(id_priv);
2692 	dev_id_priv->internal_id = 1;
2693 	dev_id_priv->afonly = id_priv->afonly;
2694 	mutex_lock(&id_priv->qp_mutex);
2695 	dev_id_priv->tos_set = id_priv->tos_set;
2696 	dev_id_priv->tos = id_priv->tos;
2697 	mutex_unlock(&id_priv->qp_mutex);
2698 
2699 	ret = rdma_listen(&dev_id_priv->id, id_priv->backlog);
2700 	if (ret)
2701 		goto err_listen;
2702 	list_add_tail(&dev_id_priv->listen_item, &id_priv->listen_list);
2703 	return 0;
2704 err_listen:
2705 	/* Caller must destroy this after releasing lock */
2706 	*to_destroy = dev_id_priv;
2707 	dev_warn(&cma_dev->device->dev, "RDMA CMA: %s, error %d\n", __func__, ret);
2708 	return ret;
2709 }
2710 
cma_listen_on_all(struct rdma_id_private * id_priv)2711 static int cma_listen_on_all(struct rdma_id_private *id_priv)
2712 {
2713 	struct rdma_id_private *to_destroy;
2714 	struct cma_device *cma_dev;
2715 	int ret;
2716 
2717 	mutex_lock(&lock);
2718 	list_add_tail(&id_priv->listen_any_item, &listen_any_list);
2719 	list_for_each_entry(cma_dev, &dev_list, list) {
2720 		ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
2721 		if (ret) {
2722 			/* Prevent racing with cma_process_remove() */
2723 			if (to_destroy)
2724 				list_del_init(&to_destroy->device_item);
2725 			goto err_listen;
2726 		}
2727 	}
2728 	mutex_unlock(&lock);
2729 	return 0;
2730 
2731 err_listen:
2732 	_cma_cancel_listens(id_priv);
2733 	mutex_unlock(&lock);
2734 	if (to_destroy)
2735 		rdma_destroy_id(&to_destroy->id);
2736 	return ret;
2737 }
2738 
rdma_set_service_type(struct rdma_cm_id * id,int tos)2739 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2740 {
2741 	struct rdma_id_private *id_priv;
2742 
2743 	id_priv = container_of(id, struct rdma_id_private, id);
2744 	mutex_lock(&id_priv->qp_mutex);
2745 	id_priv->tos = (u8) tos;
2746 	id_priv->tos_set = true;
2747 	mutex_unlock(&id_priv->qp_mutex);
2748 }
2749 EXPORT_SYMBOL(rdma_set_service_type);
2750 
2751 /**
2752  * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2753  *                          with a connection identifier.
2754  * @id: Communication identifier to associated with service type.
2755  * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2756  *
2757  * This function should be called before rdma_connect() on active side,
2758  * and on passive side before rdma_accept(). It is applicable to primary
2759  * path only. The timeout will affect the local side of the QP, it is not
2760  * negotiated with remote side and zero disables the timer. In case it is
2761  * set before rdma_resolve_route, the value will also be used to determine
2762  * PacketLifeTime for RoCE.
2763  *
2764  * Return: 0 for success
2765  */
rdma_set_ack_timeout(struct rdma_cm_id * id,u8 timeout)2766 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2767 {
2768 	struct rdma_id_private *id_priv;
2769 
2770 	if (id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_INI)
2771 		return -EINVAL;
2772 
2773 	id_priv = container_of(id, struct rdma_id_private, id);
2774 	mutex_lock(&id_priv->qp_mutex);
2775 	id_priv->timeout = timeout;
2776 	id_priv->timeout_set = true;
2777 	mutex_unlock(&id_priv->qp_mutex);
2778 
2779 	return 0;
2780 }
2781 EXPORT_SYMBOL(rdma_set_ack_timeout);
2782 
2783 /**
2784  * rdma_set_min_rnr_timer() - Set the minimum RNR Retry timer of the
2785  *			      QP associated with a connection identifier.
2786  * @id: Communication identifier to associated with service type.
2787  * @min_rnr_timer: 5-bit value encoded as Table 45: "Encoding for RNR NAK
2788  *		   Timer Field" in the IBTA specification.
2789  *
2790  * This function should be called before rdma_connect() on active
2791  * side, and on passive side before rdma_accept(). The timer value
2792  * will be associated with the local QP. When it receives a send it is
2793  * not read to handle, typically if the receive queue is empty, an RNR
2794  * Retry NAK is returned to the requester with the min_rnr_timer
2795  * encoded. The requester will then wait at least the time specified
2796  * in the NAK before retrying. The default is zero, which translates
2797  * to a minimum RNR Timer value of 655 ms.
2798  *
2799  * Return: 0 for success
2800  */
rdma_set_min_rnr_timer(struct rdma_cm_id * id,u8 min_rnr_timer)2801 int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer)
2802 {
2803 	struct rdma_id_private *id_priv;
2804 
2805 	/* It is a five-bit value */
2806 	if (min_rnr_timer & 0xe0)
2807 		return -EINVAL;
2808 
2809 	if (WARN_ON(id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_TGT))
2810 		return -EINVAL;
2811 
2812 	id_priv = container_of(id, struct rdma_id_private, id);
2813 	mutex_lock(&id_priv->qp_mutex);
2814 	id_priv->min_rnr_timer = min_rnr_timer;
2815 	id_priv->min_rnr_timer_set = true;
2816 	mutex_unlock(&id_priv->qp_mutex);
2817 
2818 	return 0;
2819 }
2820 EXPORT_SYMBOL(rdma_set_min_rnr_timer);
2821 
route_set_path_rec_inbound(struct cma_work * work,struct sa_path_rec * path_rec)2822 static void route_set_path_rec_inbound(struct cma_work *work,
2823 				       struct sa_path_rec *path_rec)
2824 {
2825 	struct rdma_route *route = &work->id->id.route;
2826 
2827 	if (!route->path_rec_inbound) {
2828 		route->path_rec_inbound =
2829 			kzalloc(sizeof(*route->path_rec_inbound), GFP_KERNEL);
2830 		if (!route->path_rec_inbound)
2831 			return;
2832 	}
2833 
2834 	*route->path_rec_inbound = *path_rec;
2835 }
2836 
route_set_path_rec_outbound(struct cma_work * work,struct sa_path_rec * path_rec)2837 static void route_set_path_rec_outbound(struct cma_work *work,
2838 					struct sa_path_rec *path_rec)
2839 {
2840 	struct rdma_route *route = &work->id->id.route;
2841 
2842 	if (!route->path_rec_outbound) {
2843 		route->path_rec_outbound =
2844 			kzalloc(sizeof(*route->path_rec_outbound), GFP_KERNEL);
2845 		if (!route->path_rec_outbound)
2846 			return;
2847 	}
2848 
2849 	*route->path_rec_outbound = *path_rec;
2850 }
2851 
cma_query_handler(int status,struct sa_path_rec * path_rec,int num_prs,void * context)2852 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2853 			      int num_prs, void *context)
2854 {
2855 	struct cma_work *work = context;
2856 	struct rdma_route *route;
2857 	int i;
2858 
2859 	route = &work->id->id.route;
2860 
2861 	if (status)
2862 		goto fail;
2863 
2864 	for (i = 0; i < num_prs; i++) {
2865 		if (!path_rec[i].flags || (path_rec[i].flags & IB_PATH_GMP))
2866 			*route->path_rec = path_rec[i];
2867 		else if (path_rec[i].flags & IB_PATH_INBOUND)
2868 			route_set_path_rec_inbound(work, &path_rec[i]);
2869 		else if (path_rec[i].flags & IB_PATH_OUTBOUND)
2870 			route_set_path_rec_outbound(work, &path_rec[i]);
2871 	}
2872 	if (!route->path_rec) {
2873 		status = -EINVAL;
2874 		goto fail;
2875 	}
2876 
2877 	route->num_pri_alt_paths = 1;
2878 	queue_work(cma_wq, &work->work);
2879 	return;
2880 
2881 fail:
2882 	work->old_state = RDMA_CM_ROUTE_QUERY;
2883 	work->new_state = RDMA_CM_ADDR_RESOLVED;
2884 	work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2885 	work->event.status = status;
2886 	pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2887 			     status);
2888 	queue_work(cma_wq, &work->work);
2889 }
2890 
cma_query_ib_route(struct rdma_id_private * id_priv,unsigned long timeout_ms,struct cma_work * work)2891 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2892 			      unsigned long timeout_ms, struct cma_work *work)
2893 {
2894 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2895 	struct sa_path_rec path_rec;
2896 	ib_sa_comp_mask comp_mask;
2897 	struct sockaddr_in6 *sin6;
2898 	struct sockaddr_ib *sib;
2899 
2900 	memset(&path_rec, 0, sizeof path_rec);
2901 
2902 	if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2903 		path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2904 	else
2905 		path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2906 	rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2907 	rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2908 	path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2909 	path_rec.numb_path = 1;
2910 	path_rec.reversible = 1;
2911 	path_rec.service_id = rdma_get_service_id(&id_priv->id,
2912 						  cma_dst_addr(id_priv));
2913 
2914 	comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2915 		    IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2916 		    IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2917 
2918 	switch (cma_family(id_priv)) {
2919 	case AF_INET:
2920 		path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2921 		comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2922 		break;
2923 	case AF_INET6:
2924 		sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2925 		path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2926 		comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2927 		break;
2928 	case AF_IB:
2929 		sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2930 		path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2931 		comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2932 		break;
2933 	}
2934 
2935 	id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2936 					       id_priv->id.port_num, &path_rec,
2937 					       comp_mask, timeout_ms,
2938 					       GFP_KERNEL, cma_query_handler,
2939 					       work, &id_priv->query);
2940 
2941 	return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2942 }
2943 
cma_iboe_join_work_handler(struct work_struct * work)2944 static void cma_iboe_join_work_handler(struct work_struct *work)
2945 {
2946 	struct cma_multicast *mc =
2947 		container_of(work, struct cma_multicast, iboe_join.work);
2948 	struct rdma_cm_event *event = &mc->iboe_join.event;
2949 	struct rdma_id_private *id_priv = mc->id_priv;
2950 	int ret;
2951 
2952 	mutex_lock(&id_priv->handler_mutex);
2953 	if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2954 	    READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2955 		goto out_unlock;
2956 
2957 	ret = cma_cm_event_handler(id_priv, event);
2958 	WARN_ON(ret);
2959 
2960 out_unlock:
2961 	mutex_unlock(&id_priv->handler_mutex);
2962 	if (event->event == RDMA_CM_EVENT_MULTICAST_JOIN)
2963 		rdma_destroy_ah_attr(&event->param.ud.ah_attr);
2964 }
2965 
cma_work_handler(struct work_struct * _work)2966 static void cma_work_handler(struct work_struct *_work)
2967 {
2968 	struct cma_work *work = container_of(_work, struct cma_work, work);
2969 	struct rdma_id_private *id_priv = work->id;
2970 
2971 	mutex_lock(&id_priv->handler_mutex);
2972 	if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2973 	    READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2974 		goto out_unlock;
2975 	if (work->old_state != 0 || work->new_state != 0) {
2976 		if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2977 			goto out_unlock;
2978 	}
2979 
2980 	if (cma_cm_event_handler(id_priv, &work->event)) {
2981 		cma_id_put(id_priv);
2982 		destroy_id_handler_unlock(id_priv);
2983 		goto out_free;
2984 	}
2985 
2986 out_unlock:
2987 	mutex_unlock(&id_priv->handler_mutex);
2988 	cma_id_put(id_priv);
2989 out_free:
2990 	if (work->event.event == RDMA_CM_EVENT_MULTICAST_JOIN)
2991 		rdma_destroy_ah_attr(&work->event.param.ud.ah_attr);
2992 	kfree(work);
2993 }
2994 
cma_init_resolve_route_work(struct cma_work * work,struct rdma_id_private * id_priv)2995 static void cma_init_resolve_route_work(struct cma_work *work,
2996 					struct rdma_id_private *id_priv)
2997 {
2998 	work->id = id_priv;
2999 	INIT_WORK(&work->work, cma_work_handler);
3000 	work->old_state = RDMA_CM_ROUTE_QUERY;
3001 	work->new_state = RDMA_CM_ROUTE_RESOLVED;
3002 	work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
3003 }
3004 
enqueue_resolve_addr_work(struct cma_work * work,struct rdma_id_private * id_priv)3005 static void enqueue_resolve_addr_work(struct cma_work *work,
3006 				      struct rdma_id_private *id_priv)
3007 {
3008 	/* Balances with cma_id_put() in cma_work_handler */
3009 	cma_id_get(id_priv);
3010 
3011 	work->id = id_priv;
3012 	INIT_WORK(&work->work, cma_work_handler);
3013 	work->old_state = RDMA_CM_ADDR_QUERY;
3014 	work->new_state = RDMA_CM_ADDR_RESOLVED;
3015 	work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3016 
3017 	queue_work(cma_wq, &work->work);
3018 }
3019 
cma_resolve_ib_route(struct rdma_id_private * id_priv,unsigned long timeout_ms)3020 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
3021 				unsigned long timeout_ms)
3022 {
3023 	struct rdma_route *route = &id_priv->id.route;
3024 	struct cma_work *work;
3025 	int ret;
3026 
3027 	work = kzalloc(sizeof *work, GFP_KERNEL);
3028 	if (!work)
3029 		return -ENOMEM;
3030 
3031 	cma_init_resolve_route_work(work, id_priv);
3032 
3033 	if (!route->path_rec)
3034 		route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
3035 	if (!route->path_rec) {
3036 		ret = -ENOMEM;
3037 		goto err1;
3038 	}
3039 
3040 	ret = cma_query_ib_route(id_priv, timeout_ms, work);
3041 	if (ret)
3042 		goto err2;
3043 
3044 	return 0;
3045 err2:
3046 	kfree(route->path_rec);
3047 	route->path_rec = NULL;
3048 err1:
3049 	kfree(work);
3050 	return ret;
3051 }
3052 
cma_route_gid_type(enum rdma_network_type network_type,unsigned long supported_gids,enum ib_gid_type default_gid)3053 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
3054 					   unsigned long supported_gids,
3055 					   enum ib_gid_type default_gid)
3056 {
3057 	if ((network_type == RDMA_NETWORK_IPV4 ||
3058 	     network_type == RDMA_NETWORK_IPV6) &&
3059 	    test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
3060 		return IB_GID_TYPE_ROCE_UDP_ENCAP;
3061 
3062 	return default_gid;
3063 }
3064 
3065 /*
3066  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
3067  * path record type based on GID type.
3068  * It also sets up other L2 fields which includes destination mac address
3069  * netdev ifindex, of the path record.
3070  * It returns the netdev of the bound interface for this path record entry.
3071  */
3072 static struct net_device *
cma_iboe_set_path_rec_l2_fields(struct rdma_id_private * id_priv)3073 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
3074 {
3075 	struct rdma_route *route = &id_priv->id.route;
3076 	enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
3077 	struct rdma_addr *addr = &route->addr;
3078 	unsigned long supported_gids;
3079 	struct net_device *ndev;
3080 
3081 	if (!addr->dev_addr.bound_dev_if)
3082 		return NULL;
3083 
3084 	ndev = dev_get_by_index(addr->dev_addr.net,
3085 				addr->dev_addr.bound_dev_if);
3086 	if (!ndev)
3087 		return NULL;
3088 
3089 	supported_gids = roce_gid_type_mask_support(id_priv->id.device,
3090 						    id_priv->id.port_num);
3091 	gid_type = cma_route_gid_type(addr->dev_addr.network,
3092 				      supported_gids,
3093 				      id_priv->gid_type);
3094 	/* Use the hint from IP Stack to select GID Type */
3095 	if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
3096 		gid_type = ib_network_to_gid_type(addr->dev_addr.network);
3097 	route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
3098 
3099 	route->path_rec->roce.route_resolved = true;
3100 	sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
3101 	return ndev;
3102 }
3103 
rdma_set_ib_path(struct rdma_cm_id * id,struct sa_path_rec * path_rec)3104 int rdma_set_ib_path(struct rdma_cm_id *id,
3105 		     struct sa_path_rec *path_rec)
3106 {
3107 	struct rdma_id_private *id_priv;
3108 	struct net_device *ndev;
3109 	int ret;
3110 
3111 	id_priv = container_of(id, struct rdma_id_private, id);
3112 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3113 			   RDMA_CM_ROUTE_RESOLVED))
3114 		return -EINVAL;
3115 
3116 	id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
3117 				     GFP_KERNEL);
3118 	if (!id->route.path_rec) {
3119 		ret = -ENOMEM;
3120 		goto err;
3121 	}
3122 
3123 	if (rdma_protocol_roce(id->device, id->port_num)) {
3124 		ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3125 		if (!ndev) {
3126 			ret = -ENODEV;
3127 			goto err_free;
3128 		}
3129 		dev_put(ndev);
3130 	}
3131 
3132 	id->route.num_pri_alt_paths = 1;
3133 	return 0;
3134 
3135 err_free:
3136 	kfree(id->route.path_rec);
3137 	id->route.path_rec = NULL;
3138 err:
3139 	cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
3140 	return ret;
3141 }
3142 EXPORT_SYMBOL(rdma_set_ib_path);
3143 
cma_resolve_iw_route(struct rdma_id_private * id_priv)3144 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
3145 {
3146 	struct cma_work *work;
3147 
3148 	work = kzalloc(sizeof *work, GFP_KERNEL);
3149 	if (!work)
3150 		return -ENOMEM;
3151 
3152 	cma_init_resolve_route_work(work, id_priv);
3153 	queue_work(cma_wq, &work->work);
3154 	return 0;
3155 }
3156 
get_vlan_ndev_tc(struct net_device * vlan_ndev,int prio)3157 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
3158 {
3159 	struct net_device *dev;
3160 
3161 	dev = vlan_dev_real_dev(vlan_ndev);
3162 	if (dev->num_tc)
3163 		return netdev_get_prio_tc_map(dev, prio);
3164 
3165 	return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
3166 		VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
3167 }
3168 
3169 struct iboe_prio_tc_map {
3170 	int input_prio;
3171 	int output_tc;
3172 	bool found;
3173 };
3174 
get_lower_vlan_dev_tc(struct net_device * dev,struct netdev_nested_priv * priv)3175 static int get_lower_vlan_dev_tc(struct net_device *dev,
3176 				 struct netdev_nested_priv *priv)
3177 {
3178 	struct iboe_prio_tc_map *map = (struct iboe_prio_tc_map *)priv->data;
3179 
3180 	if (is_vlan_dev(dev))
3181 		map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
3182 	else if (dev->num_tc)
3183 		map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
3184 	else
3185 		map->output_tc = 0;
3186 	/* We are interested only in first level VLAN device, so always
3187 	 * return 1 to stop iterating over next level devices.
3188 	 */
3189 	map->found = true;
3190 	return 1;
3191 }
3192 
iboe_tos_to_sl(struct net_device * ndev,int tos)3193 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
3194 {
3195 	struct iboe_prio_tc_map prio_tc_map = {};
3196 	int prio = rt_tos2priority(tos);
3197 	struct netdev_nested_priv priv;
3198 
3199 	/* If VLAN device, get it directly from the VLAN netdev */
3200 	if (is_vlan_dev(ndev))
3201 		return get_vlan_ndev_tc(ndev, prio);
3202 
3203 	prio_tc_map.input_prio = prio;
3204 	priv.data = (void *)&prio_tc_map;
3205 	rcu_read_lock();
3206 	netdev_walk_all_lower_dev_rcu(ndev,
3207 				      get_lower_vlan_dev_tc,
3208 				      &priv);
3209 	rcu_read_unlock();
3210 	/* If map is found from lower device, use it; Otherwise
3211 	 * continue with the current netdevice to get priority to tc map.
3212 	 */
3213 	if (prio_tc_map.found)
3214 		return prio_tc_map.output_tc;
3215 	else if (ndev->num_tc)
3216 		return netdev_get_prio_tc_map(ndev, prio);
3217 	else
3218 		return 0;
3219 }
3220 
cma_get_roce_udp_flow_label(struct rdma_id_private * id_priv)3221 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
3222 {
3223 	struct sockaddr_in6 *addr6;
3224 	u16 dport, sport;
3225 	u32 hash, fl;
3226 
3227 	addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
3228 	fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
3229 	if ((cma_family(id_priv) != AF_INET6) || !fl) {
3230 		dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
3231 		sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
3232 		hash = (u32)sport * 31 + dport;
3233 		fl = hash & IB_GRH_FLOWLABEL_MASK;
3234 	}
3235 
3236 	return cpu_to_be32(fl);
3237 }
3238 
cma_resolve_iboe_route(struct rdma_id_private * id_priv)3239 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
3240 {
3241 	struct rdma_route *route = &id_priv->id.route;
3242 	struct rdma_addr *addr = &route->addr;
3243 	struct cma_work *work;
3244 	int ret;
3245 	struct net_device *ndev;
3246 
3247 	u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
3248 					rdma_start_port(id_priv->cma_dev->device)];
3249 	u8 tos;
3250 
3251 	mutex_lock(&id_priv->qp_mutex);
3252 	tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
3253 	mutex_unlock(&id_priv->qp_mutex);
3254 
3255 	work = kzalloc(sizeof *work, GFP_KERNEL);
3256 	if (!work)
3257 		return -ENOMEM;
3258 
3259 	route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
3260 	if (!route->path_rec) {
3261 		ret = -ENOMEM;
3262 		goto err1;
3263 	}
3264 
3265 	route->num_pri_alt_paths = 1;
3266 
3267 	ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3268 	if (!ndev) {
3269 		ret = -ENODEV;
3270 		goto err2;
3271 	}
3272 
3273 	rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3274 		    &route->path_rec->sgid);
3275 	rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
3276 		    &route->path_rec->dgid);
3277 
3278 	if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
3279 		/* TODO: get the hoplimit from the inet/inet6 device */
3280 		route->path_rec->hop_limit = addr->dev_addr.hoplimit;
3281 	else
3282 		route->path_rec->hop_limit = 1;
3283 	route->path_rec->reversible = 1;
3284 	route->path_rec->pkey = cpu_to_be16(0xffff);
3285 	route->path_rec->mtu_selector = IB_SA_EQ;
3286 	route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
3287 	route->path_rec->traffic_class = tos;
3288 	route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
3289 	route->path_rec->rate_selector = IB_SA_EQ;
3290 	route->path_rec->rate = iboe_get_rate(ndev);
3291 	dev_put(ndev);
3292 	route->path_rec->packet_life_time_selector = IB_SA_EQ;
3293 	/* In case ACK timeout is set, use this value to calculate
3294 	 * PacketLifeTime.  As per IBTA 12.7.34,
3295 	 * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
3296 	 * Assuming a negligible local ACK delay, we can use
3297 	 * PacketLifeTime = local ACK timeout/2
3298 	 * as a reasonable approximation for RoCE networks.
3299 	 */
3300 	mutex_lock(&id_priv->qp_mutex);
3301 	if (id_priv->timeout_set && id_priv->timeout)
3302 		route->path_rec->packet_life_time = id_priv->timeout - 1;
3303 	else
3304 		route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
3305 	mutex_unlock(&id_priv->qp_mutex);
3306 
3307 	if (!route->path_rec->mtu) {
3308 		ret = -EINVAL;
3309 		goto err2;
3310 	}
3311 
3312 	if (rdma_protocol_roce_udp_encap(id_priv->id.device,
3313 					 id_priv->id.port_num))
3314 		route->path_rec->flow_label =
3315 			cma_get_roce_udp_flow_label(id_priv);
3316 
3317 	cma_init_resolve_route_work(work, id_priv);
3318 	queue_work(cma_wq, &work->work);
3319 
3320 	return 0;
3321 
3322 err2:
3323 	kfree(route->path_rec);
3324 	route->path_rec = NULL;
3325 	route->num_pri_alt_paths = 0;
3326 err1:
3327 	kfree(work);
3328 	return ret;
3329 }
3330 
rdma_resolve_route(struct rdma_cm_id * id,unsigned long timeout_ms)3331 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3332 {
3333 	struct rdma_id_private *id_priv;
3334 	int ret;
3335 
3336 	if (!timeout_ms)
3337 		return -EINVAL;
3338 
3339 	id_priv = container_of(id, struct rdma_id_private, id);
3340 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3341 		return -EINVAL;
3342 
3343 	cma_id_get(id_priv);
3344 	if (rdma_cap_ib_sa(id->device, id->port_num))
3345 		ret = cma_resolve_ib_route(id_priv, timeout_ms);
3346 	else if (rdma_protocol_roce(id->device, id->port_num)) {
3347 		ret = cma_resolve_iboe_route(id_priv);
3348 		if (!ret)
3349 			cma_add_id_to_tree(id_priv);
3350 	}
3351 	else if (rdma_protocol_iwarp(id->device, id->port_num))
3352 		ret = cma_resolve_iw_route(id_priv);
3353 	else
3354 		ret = -ENOSYS;
3355 
3356 	if (ret)
3357 		goto err;
3358 
3359 	return 0;
3360 err:
3361 	cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3362 	cma_id_put(id_priv);
3363 	return ret;
3364 }
3365 EXPORT_SYMBOL(rdma_resolve_route);
3366 
cma_set_loopback(struct sockaddr * addr)3367 static void cma_set_loopback(struct sockaddr *addr)
3368 {
3369 	switch (addr->sa_family) {
3370 	case AF_INET:
3371 		((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3372 		break;
3373 	case AF_INET6:
3374 		ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3375 			      0, 0, 0, htonl(1));
3376 		break;
3377 	default:
3378 		ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3379 			    0, 0, 0, htonl(1));
3380 		break;
3381 	}
3382 }
3383 
cma_bind_loopback(struct rdma_id_private * id_priv)3384 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3385 {
3386 	struct cma_device *cma_dev, *cur_dev;
3387 	union ib_gid gid;
3388 	enum ib_port_state port_state;
3389 	unsigned int p;
3390 	u16 pkey;
3391 	int ret;
3392 
3393 	cma_dev = NULL;
3394 	mutex_lock(&lock);
3395 	list_for_each_entry(cur_dev, &dev_list, list) {
3396 		if (cma_family(id_priv) == AF_IB &&
3397 		    !rdma_cap_ib_cm(cur_dev->device, 1))
3398 			continue;
3399 
3400 		if (!cma_dev)
3401 			cma_dev = cur_dev;
3402 
3403 		rdma_for_each_port (cur_dev->device, p) {
3404 			if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3405 			    port_state == IB_PORT_ACTIVE) {
3406 				cma_dev = cur_dev;
3407 				goto port_found;
3408 			}
3409 		}
3410 	}
3411 
3412 	if (!cma_dev) {
3413 		ret = -ENODEV;
3414 		goto out;
3415 	}
3416 
3417 	p = 1;
3418 
3419 port_found:
3420 	ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3421 	if (ret)
3422 		goto out;
3423 
3424 	ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3425 	if (ret)
3426 		goto out;
3427 
3428 	id_priv->id.route.addr.dev_addr.dev_type =
3429 		(rdma_protocol_ib(cma_dev->device, p)) ?
3430 		ARPHRD_INFINIBAND : ARPHRD_ETHER;
3431 
3432 	rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3433 	ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3434 	id_priv->id.port_num = p;
3435 	cma_attach_to_dev(id_priv, cma_dev);
3436 	rdma_restrack_add(&id_priv->res);
3437 	cma_set_loopback(cma_src_addr(id_priv));
3438 out:
3439 	mutex_unlock(&lock);
3440 	return ret;
3441 }
3442 
addr_handler(int status,struct sockaddr * src_addr,struct rdma_dev_addr * dev_addr,void * context)3443 static void addr_handler(int status, struct sockaddr *src_addr,
3444 			 struct rdma_dev_addr *dev_addr, void *context)
3445 {
3446 	struct rdma_id_private *id_priv = context;
3447 	struct rdma_cm_event event = {};
3448 	struct sockaddr *addr;
3449 	struct sockaddr_storage old_addr;
3450 
3451 	mutex_lock(&id_priv->handler_mutex);
3452 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3453 			   RDMA_CM_ADDR_RESOLVED))
3454 		goto out;
3455 
3456 	/*
3457 	 * Store the previous src address, so that if we fail to acquire
3458 	 * matching rdma device, old address can be restored back, which helps
3459 	 * to cancel the cma listen operation correctly.
3460 	 */
3461 	addr = cma_src_addr(id_priv);
3462 	memcpy(&old_addr, addr, rdma_addr_size(addr));
3463 	memcpy(addr, src_addr, rdma_addr_size(src_addr));
3464 	if (!status && !id_priv->cma_dev) {
3465 		status = cma_acquire_dev_by_src_ip(id_priv);
3466 		if (status)
3467 			pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3468 					     status);
3469 		rdma_restrack_add(&id_priv->res);
3470 	} else if (status) {
3471 		pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3472 	}
3473 
3474 	if (status) {
3475 		memcpy(addr, &old_addr,
3476 		       rdma_addr_size((struct sockaddr *)&old_addr));
3477 		if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3478 				   RDMA_CM_ADDR_BOUND))
3479 			goto out;
3480 		event.event = RDMA_CM_EVENT_ADDR_ERROR;
3481 		event.status = status;
3482 	} else
3483 		event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3484 
3485 	if (cma_cm_event_handler(id_priv, &event)) {
3486 		destroy_id_handler_unlock(id_priv);
3487 		return;
3488 	}
3489 out:
3490 	mutex_unlock(&id_priv->handler_mutex);
3491 }
3492 
cma_resolve_loopback(struct rdma_id_private * id_priv)3493 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3494 {
3495 	struct cma_work *work;
3496 	union ib_gid gid;
3497 	int ret;
3498 
3499 	work = kzalloc(sizeof *work, GFP_KERNEL);
3500 	if (!work)
3501 		return -ENOMEM;
3502 
3503 	if (!id_priv->cma_dev) {
3504 		ret = cma_bind_loopback(id_priv);
3505 		if (ret)
3506 			goto err;
3507 	}
3508 
3509 	rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3510 	rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3511 
3512 	enqueue_resolve_addr_work(work, id_priv);
3513 	return 0;
3514 err:
3515 	kfree(work);
3516 	return ret;
3517 }
3518 
cma_resolve_ib_addr(struct rdma_id_private * id_priv)3519 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3520 {
3521 	struct cma_work *work;
3522 	int ret;
3523 
3524 	work = kzalloc(sizeof *work, GFP_KERNEL);
3525 	if (!work)
3526 		return -ENOMEM;
3527 
3528 	if (!id_priv->cma_dev) {
3529 		ret = cma_resolve_ib_dev(id_priv);
3530 		if (ret)
3531 			goto err;
3532 	}
3533 
3534 	rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3535 		&(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3536 
3537 	enqueue_resolve_addr_work(work, id_priv);
3538 	return 0;
3539 err:
3540 	kfree(work);
3541 	return ret;
3542 }
3543 
cma_bind_addr(struct rdma_cm_id * id,struct sockaddr * src_addr,const struct sockaddr * dst_addr)3544 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3545 			 const struct sockaddr *dst_addr)
3546 {
3547 	struct sockaddr_storage zero_sock = {};
3548 
3549 	if (src_addr && src_addr->sa_family)
3550 		return rdma_bind_addr(id, src_addr);
3551 
3552 	/*
3553 	 * When the src_addr is not specified, automatically supply an any addr
3554 	 */
3555 	zero_sock.ss_family = dst_addr->sa_family;
3556 	if (IS_ENABLED(CONFIG_IPV6) && dst_addr->sa_family == AF_INET6) {
3557 		struct sockaddr_in6 *src_addr6 =
3558 			(struct sockaddr_in6 *)&zero_sock;
3559 		struct sockaddr_in6 *dst_addr6 =
3560 			(struct sockaddr_in6 *)dst_addr;
3561 
3562 		src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3563 		if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3564 			id->route.addr.dev_addr.bound_dev_if =
3565 				dst_addr6->sin6_scope_id;
3566 	} else if (dst_addr->sa_family == AF_IB) {
3567 		((struct sockaddr_ib *)&zero_sock)->sib_pkey =
3568 			((struct sockaddr_ib *)dst_addr)->sib_pkey;
3569 	}
3570 	return rdma_bind_addr(id, (struct sockaddr *)&zero_sock);
3571 }
3572 
3573 /*
3574  * If required, resolve the source address for bind and leave the id_priv in
3575  * state RDMA_CM_ADDR_BOUND. This oddly uses the state to determine the prior
3576  * calls made by ULP, a previously bound ID will not be re-bound and src_addr is
3577  * ignored.
3578  */
resolve_prepare_src(struct rdma_id_private * id_priv,struct sockaddr * src_addr,const struct sockaddr * dst_addr)3579 static int resolve_prepare_src(struct rdma_id_private *id_priv,
3580 			       struct sockaddr *src_addr,
3581 			       const struct sockaddr *dst_addr)
3582 {
3583 	int ret;
3584 
3585 	memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3586 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3587 		/* For a well behaved ULP state will be RDMA_CM_IDLE */
3588 		ret = cma_bind_addr(&id_priv->id, src_addr, dst_addr);
3589 		if (ret)
3590 			goto err_dst;
3591 		if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3592 					   RDMA_CM_ADDR_QUERY))) {
3593 			ret = -EINVAL;
3594 			goto err_dst;
3595 		}
3596 	}
3597 
3598 	if (cma_family(id_priv) != dst_addr->sa_family) {
3599 		ret = -EINVAL;
3600 		goto err_state;
3601 	}
3602 	return 0;
3603 
3604 err_state:
3605 	cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3606 err_dst:
3607 	memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3608 	return ret;
3609 }
3610 
rdma_resolve_addr(struct rdma_cm_id * id,struct sockaddr * src_addr,const struct sockaddr * dst_addr,unsigned long timeout_ms)3611 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3612 		      const struct sockaddr *dst_addr, unsigned long timeout_ms)
3613 {
3614 	struct rdma_id_private *id_priv =
3615 		container_of(id, struct rdma_id_private, id);
3616 	int ret;
3617 
3618 	ret = resolve_prepare_src(id_priv, src_addr, dst_addr);
3619 	if (ret)
3620 		return ret;
3621 
3622 	if (cma_any_addr(dst_addr)) {
3623 		ret = cma_resolve_loopback(id_priv);
3624 	} else {
3625 		if (dst_addr->sa_family == AF_IB) {
3626 			ret = cma_resolve_ib_addr(id_priv);
3627 		} else {
3628 			/*
3629 			 * The FSM can return back to RDMA_CM_ADDR_BOUND after
3630 			 * rdma_resolve_ip() is called, eg through the error
3631 			 * path in addr_handler(). If this happens the existing
3632 			 * request must be canceled before issuing a new one.
3633 			 * Since canceling a request is a bit slow and this
3634 			 * oddball path is rare, keep track once a request has
3635 			 * been issued. The track turns out to be a permanent
3636 			 * state since this is the only cancel as it is
3637 			 * immediately before rdma_resolve_ip().
3638 			 */
3639 			if (id_priv->used_resolve_ip)
3640 				rdma_addr_cancel(&id->route.addr.dev_addr);
3641 			else
3642 				id_priv->used_resolve_ip = 1;
3643 			ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3644 					      &id->route.addr.dev_addr,
3645 					      timeout_ms, addr_handler,
3646 					      false, id_priv);
3647 		}
3648 	}
3649 	if (ret)
3650 		goto err;
3651 
3652 	return 0;
3653 err:
3654 	cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3655 	return ret;
3656 }
3657 EXPORT_SYMBOL(rdma_resolve_addr);
3658 
rdma_set_reuseaddr(struct rdma_cm_id * id,int reuse)3659 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3660 {
3661 	struct rdma_id_private *id_priv;
3662 	unsigned long flags;
3663 	int ret;
3664 
3665 	id_priv = container_of(id, struct rdma_id_private, id);
3666 	spin_lock_irqsave(&id_priv->lock, flags);
3667 	if ((reuse && id_priv->state != RDMA_CM_LISTEN) ||
3668 	    id_priv->state == RDMA_CM_IDLE) {
3669 		id_priv->reuseaddr = reuse;
3670 		ret = 0;
3671 	} else {
3672 		ret = -EINVAL;
3673 	}
3674 	spin_unlock_irqrestore(&id_priv->lock, flags);
3675 	return ret;
3676 }
3677 EXPORT_SYMBOL(rdma_set_reuseaddr);
3678 
rdma_set_afonly(struct rdma_cm_id * id,int afonly)3679 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3680 {
3681 	struct rdma_id_private *id_priv;
3682 	unsigned long flags;
3683 	int ret;
3684 
3685 	id_priv = container_of(id, struct rdma_id_private, id);
3686 	spin_lock_irqsave(&id_priv->lock, flags);
3687 	if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3688 		id_priv->options |= (1 << CMA_OPTION_AFONLY);
3689 		id_priv->afonly = afonly;
3690 		ret = 0;
3691 	} else {
3692 		ret = -EINVAL;
3693 	}
3694 	spin_unlock_irqrestore(&id_priv->lock, flags);
3695 	return ret;
3696 }
3697 EXPORT_SYMBOL(rdma_set_afonly);
3698 
cma_bind_port(struct rdma_bind_list * bind_list,struct rdma_id_private * id_priv)3699 static void cma_bind_port(struct rdma_bind_list *bind_list,
3700 			  struct rdma_id_private *id_priv)
3701 {
3702 	struct sockaddr *addr;
3703 	struct sockaddr_ib *sib;
3704 	u64 sid, mask;
3705 	__be16 port;
3706 
3707 	lockdep_assert_held(&lock);
3708 
3709 	addr = cma_src_addr(id_priv);
3710 	port = htons(bind_list->port);
3711 
3712 	switch (addr->sa_family) {
3713 	case AF_INET:
3714 		((struct sockaddr_in *) addr)->sin_port = port;
3715 		break;
3716 	case AF_INET6:
3717 		((struct sockaddr_in6 *) addr)->sin6_port = port;
3718 		break;
3719 	case AF_IB:
3720 		sib = (struct sockaddr_ib *) addr;
3721 		sid = be64_to_cpu(sib->sib_sid);
3722 		mask = be64_to_cpu(sib->sib_sid_mask);
3723 		sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3724 		sib->sib_sid_mask = cpu_to_be64(~0ULL);
3725 		break;
3726 	}
3727 	id_priv->bind_list = bind_list;
3728 	hlist_add_head(&id_priv->node, &bind_list->owners);
3729 }
3730 
cma_alloc_port(enum rdma_ucm_port_space ps,struct rdma_id_private * id_priv,unsigned short snum)3731 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3732 			  struct rdma_id_private *id_priv, unsigned short snum)
3733 {
3734 	struct rdma_bind_list *bind_list;
3735 	int ret;
3736 
3737 	lockdep_assert_held(&lock);
3738 
3739 	bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3740 	if (!bind_list)
3741 		return -ENOMEM;
3742 
3743 	ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3744 			   snum);
3745 	if (ret < 0)
3746 		goto err;
3747 
3748 	bind_list->ps = ps;
3749 	bind_list->port = snum;
3750 	cma_bind_port(bind_list, id_priv);
3751 	return 0;
3752 err:
3753 	kfree(bind_list);
3754 	return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3755 }
3756 
cma_port_is_unique(struct rdma_bind_list * bind_list,struct rdma_id_private * id_priv)3757 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3758 			      struct rdma_id_private *id_priv)
3759 {
3760 	struct rdma_id_private *cur_id;
3761 	struct sockaddr  *daddr = cma_dst_addr(id_priv);
3762 	struct sockaddr  *saddr = cma_src_addr(id_priv);
3763 	__be16 dport = cma_port(daddr);
3764 
3765 	lockdep_assert_held(&lock);
3766 
3767 	hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3768 		struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3769 		struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3770 		__be16 cur_dport = cma_port(cur_daddr);
3771 
3772 		if (id_priv == cur_id)
3773 			continue;
3774 
3775 		/* different dest port -> unique */
3776 		if (!cma_any_port(daddr) &&
3777 		    !cma_any_port(cur_daddr) &&
3778 		    (dport != cur_dport))
3779 			continue;
3780 
3781 		/* different src address -> unique */
3782 		if (!cma_any_addr(saddr) &&
3783 		    !cma_any_addr(cur_saddr) &&
3784 		    cma_addr_cmp(saddr, cur_saddr))
3785 			continue;
3786 
3787 		/* different dst address -> unique */
3788 		if (!cma_any_addr(daddr) &&
3789 		    !cma_any_addr(cur_daddr) &&
3790 		    cma_addr_cmp(daddr, cur_daddr))
3791 			continue;
3792 
3793 		return -EADDRNOTAVAIL;
3794 	}
3795 	return 0;
3796 }
3797 
cma_alloc_any_port(enum rdma_ucm_port_space ps,struct rdma_id_private * id_priv)3798 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3799 			      struct rdma_id_private *id_priv)
3800 {
3801 	static unsigned int last_used_port;
3802 	int low, high, remaining;
3803 	unsigned int rover;
3804 	struct net *net = id_priv->id.route.addr.dev_addr.net;
3805 
3806 	lockdep_assert_held(&lock);
3807 
3808 	inet_get_local_port_range(net, &low, &high);
3809 	remaining = (high - low) + 1;
3810 	rover = prandom_u32_max(remaining) + low;
3811 retry:
3812 	if (last_used_port != rover) {
3813 		struct rdma_bind_list *bind_list;
3814 		int ret;
3815 
3816 		bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3817 
3818 		if (!bind_list) {
3819 			ret = cma_alloc_port(ps, id_priv, rover);
3820 		} else {
3821 			ret = cma_port_is_unique(bind_list, id_priv);
3822 			if (!ret)
3823 				cma_bind_port(bind_list, id_priv);
3824 		}
3825 		/*
3826 		 * Remember previously used port number in order to avoid
3827 		 * re-using same port immediately after it is closed.
3828 		 */
3829 		if (!ret)
3830 			last_used_port = rover;
3831 		if (ret != -EADDRNOTAVAIL)
3832 			return ret;
3833 	}
3834 	if (--remaining) {
3835 		rover++;
3836 		if ((rover < low) || (rover > high))
3837 			rover = low;
3838 		goto retry;
3839 	}
3840 	return -EADDRNOTAVAIL;
3841 }
3842 
3843 /*
3844  * Check that the requested port is available.  This is called when trying to
3845  * bind to a specific port, or when trying to listen on a bound port.  In
3846  * the latter case, the provided id_priv may already be on the bind_list, but
3847  * we still need to check that it's okay to start listening.
3848  */
cma_check_port(struct rdma_bind_list * bind_list,struct rdma_id_private * id_priv,uint8_t reuseaddr)3849 static int cma_check_port(struct rdma_bind_list *bind_list,
3850 			  struct rdma_id_private *id_priv, uint8_t reuseaddr)
3851 {
3852 	struct rdma_id_private *cur_id;
3853 	struct sockaddr *addr, *cur_addr;
3854 
3855 	lockdep_assert_held(&lock);
3856 
3857 	addr = cma_src_addr(id_priv);
3858 	hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3859 		if (id_priv == cur_id)
3860 			continue;
3861 
3862 		if (reuseaddr && cur_id->reuseaddr)
3863 			continue;
3864 
3865 		cur_addr = cma_src_addr(cur_id);
3866 		if (id_priv->afonly && cur_id->afonly &&
3867 		    (addr->sa_family != cur_addr->sa_family))
3868 			continue;
3869 
3870 		if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3871 			return -EADDRNOTAVAIL;
3872 
3873 		if (!cma_addr_cmp(addr, cur_addr))
3874 			return -EADDRINUSE;
3875 	}
3876 	return 0;
3877 }
3878 
cma_use_port(enum rdma_ucm_port_space ps,struct rdma_id_private * id_priv)3879 static int cma_use_port(enum rdma_ucm_port_space ps,
3880 			struct rdma_id_private *id_priv)
3881 {
3882 	struct rdma_bind_list *bind_list;
3883 	unsigned short snum;
3884 	int ret;
3885 
3886 	lockdep_assert_held(&lock);
3887 
3888 	snum = ntohs(cma_port(cma_src_addr(id_priv)));
3889 	if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3890 		return -EACCES;
3891 
3892 	bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3893 	if (!bind_list) {
3894 		ret = cma_alloc_port(ps, id_priv, snum);
3895 	} else {
3896 		ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3897 		if (!ret)
3898 			cma_bind_port(bind_list, id_priv);
3899 	}
3900 	return ret;
3901 }
3902 
3903 static enum rdma_ucm_port_space
cma_select_inet_ps(struct rdma_id_private * id_priv)3904 cma_select_inet_ps(struct rdma_id_private *id_priv)
3905 {
3906 	switch (id_priv->id.ps) {
3907 	case RDMA_PS_TCP:
3908 	case RDMA_PS_UDP:
3909 	case RDMA_PS_IPOIB:
3910 	case RDMA_PS_IB:
3911 		return id_priv->id.ps;
3912 	default:
3913 
3914 		return 0;
3915 	}
3916 }
3917 
3918 static enum rdma_ucm_port_space
cma_select_ib_ps(struct rdma_id_private * id_priv)3919 cma_select_ib_ps(struct rdma_id_private *id_priv)
3920 {
3921 	enum rdma_ucm_port_space ps = 0;
3922 	struct sockaddr_ib *sib;
3923 	u64 sid_ps, mask, sid;
3924 
3925 	sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3926 	mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3927 	sid = be64_to_cpu(sib->sib_sid) & mask;
3928 
3929 	if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3930 		sid_ps = RDMA_IB_IP_PS_IB;
3931 		ps = RDMA_PS_IB;
3932 	} else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3933 		   (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3934 		sid_ps = RDMA_IB_IP_PS_TCP;
3935 		ps = RDMA_PS_TCP;
3936 	} else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3937 		   (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3938 		sid_ps = RDMA_IB_IP_PS_UDP;
3939 		ps = RDMA_PS_UDP;
3940 	}
3941 
3942 	if (ps) {
3943 		sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3944 		sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3945 						be64_to_cpu(sib->sib_sid_mask));
3946 	}
3947 	return ps;
3948 }
3949 
cma_get_port(struct rdma_id_private * id_priv)3950 static int cma_get_port(struct rdma_id_private *id_priv)
3951 {
3952 	enum rdma_ucm_port_space ps;
3953 	int ret;
3954 
3955 	if (cma_family(id_priv) != AF_IB)
3956 		ps = cma_select_inet_ps(id_priv);
3957 	else
3958 		ps = cma_select_ib_ps(id_priv);
3959 	if (!ps)
3960 		return -EPROTONOSUPPORT;
3961 
3962 	mutex_lock(&lock);
3963 	if (cma_any_port(cma_src_addr(id_priv)))
3964 		ret = cma_alloc_any_port(ps, id_priv);
3965 	else
3966 		ret = cma_use_port(ps, id_priv);
3967 	mutex_unlock(&lock);
3968 
3969 	return ret;
3970 }
3971 
cma_check_linklocal(struct rdma_dev_addr * dev_addr,struct sockaddr * addr)3972 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3973 			       struct sockaddr *addr)
3974 {
3975 #if IS_ENABLED(CONFIG_IPV6)
3976 	struct sockaddr_in6 *sin6;
3977 
3978 	if (addr->sa_family != AF_INET6)
3979 		return 0;
3980 
3981 	sin6 = (struct sockaddr_in6 *) addr;
3982 
3983 	if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3984 		return 0;
3985 
3986 	if (!sin6->sin6_scope_id)
3987 			return -EINVAL;
3988 
3989 	dev_addr->bound_dev_if = sin6->sin6_scope_id;
3990 #endif
3991 	return 0;
3992 }
3993 
rdma_listen(struct rdma_cm_id * id,int backlog)3994 int rdma_listen(struct rdma_cm_id *id, int backlog)
3995 {
3996 	struct rdma_id_private *id_priv =
3997 		container_of(id, struct rdma_id_private, id);
3998 	int ret;
3999 
4000 	if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) {
4001 		struct sockaddr_in any_in = {
4002 			.sin_family = AF_INET,
4003 			.sin_addr.s_addr = htonl(INADDR_ANY),
4004 		};
4005 
4006 		/* For a well behaved ULP state will be RDMA_CM_IDLE */
4007 		ret = rdma_bind_addr(id, (struct sockaddr *)&any_in);
4008 		if (ret)
4009 			return ret;
4010 		if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
4011 					   RDMA_CM_LISTEN)))
4012 			return -EINVAL;
4013 	}
4014 
4015 	/*
4016 	 * Once the ID reaches RDMA_CM_LISTEN it is not allowed to be reusable
4017 	 * any more, and has to be unique in the bind list.
4018 	 */
4019 	if (id_priv->reuseaddr) {
4020 		mutex_lock(&lock);
4021 		ret = cma_check_port(id_priv->bind_list, id_priv, 0);
4022 		if (!ret)
4023 			id_priv->reuseaddr = 0;
4024 		mutex_unlock(&lock);
4025 		if (ret)
4026 			goto err;
4027 	}
4028 
4029 	id_priv->backlog = backlog;
4030 	if (id_priv->cma_dev) {
4031 		if (rdma_cap_ib_cm(id->device, 1)) {
4032 			ret = cma_ib_listen(id_priv);
4033 			if (ret)
4034 				goto err;
4035 		} else if (rdma_cap_iw_cm(id->device, 1)) {
4036 			ret = cma_iw_listen(id_priv, backlog);
4037 			if (ret)
4038 				goto err;
4039 		} else {
4040 			ret = -ENOSYS;
4041 			goto err;
4042 		}
4043 	} else {
4044 		ret = cma_listen_on_all(id_priv);
4045 		if (ret)
4046 			goto err;
4047 	}
4048 
4049 	return 0;
4050 err:
4051 	id_priv->backlog = 0;
4052 	/*
4053 	 * All the failure paths that lead here will not allow the req_handler's
4054 	 * to have run.
4055 	 */
4056 	cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
4057 	return ret;
4058 }
4059 EXPORT_SYMBOL(rdma_listen);
4060 
rdma_bind_addr(struct rdma_cm_id * id,struct sockaddr * addr)4061 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
4062 {
4063 	struct rdma_id_private *id_priv;
4064 	int ret;
4065 	struct sockaddr  *daddr;
4066 
4067 	if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
4068 	    addr->sa_family != AF_IB)
4069 		return -EAFNOSUPPORT;
4070 
4071 	id_priv = container_of(id, struct rdma_id_private, id);
4072 	if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
4073 		return -EINVAL;
4074 
4075 	ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
4076 	if (ret)
4077 		goto err1;
4078 
4079 	memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
4080 	if (!cma_any_addr(addr)) {
4081 		ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
4082 		if (ret)
4083 			goto err1;
4084 
4085 		ret = cma_acquire_dev_by_src_ip(id_priv);
4086 		if (ret)
4087 			goto err1;
4088 	}
4089 
4090 	if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
4091 		if (addr->sa_family == AF_INET)
4092 			id_priv->afonly = 1;
4093 #if IS_ENABLED(CONFIG_IPV6)
4094 		else if (addr->sa_family == AF_INET6) {
4095 			struct net *net = id_priv->id.route.addr.dev_addr.net;
4096 
4097 			id_priv->afonly = net->ipv6.sysctl.bindv6only;
4098 		}
4099 #endif
4100 	}
4101 	daddr = cma_dst_addr(id_priv);
4102 	daddr->sa_family = addr->sa_family;
4103 
4104 	ret = cma_get_port(id_priv);
4105 	if (ret)
4106 		goto err2;
4107 
4108 	if (!cma_any_addr(addr))
4109 		rdma_restrack_add(&id_priv->res);
4110 	return 0;
4111 err2:
4112 	if (id_priv->cma_dev)
4113 		cma_release_dev(id_priv);
4114 err1:
4115 	cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
4116 	return ret;
4117 }
4118 EXPORT_SYMBOL(rdma_bind_addr);
4119 
cma_format_hdr(void * hdr,struct rdma_id_private * id_priv)4120 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
4121 {
4122 	struct cma_hdr *cma_hdr;
4123 
4124 	cma_hdr = hdr;
4125 	cma_hdr->cma_version = CMA_VERSION;
4126 	if (cma_family(id_priv) == AF_INET) {
4127 		struct sockaddr_in *src4, *dst4;
4128 
4129 		src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
4130 		dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
4131 
4132 		cma_set_ip_ver(cma_hdr, 4);
4133 		cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
4134 		cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
4135 		cma_hdr->port = src4->sin_port;
4136 	} else if (cma_family(id_priv) == AF_INET6) {
4137 		struct sockaddr_in6 *src6, *dst6;
4138 
4139 		src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
4140 		dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
4141 
4142 		cma_set_ip_ver(cma_hdr, 6);
4143 		cma_hdr->src_addr.ip6 = src6->sin6_addr;
4144 		cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
4145 		cma_hdr->port = src6->sin6_port;
4146 	}
4147 	return 0;
4148 }
4149 
cma_sidr_rep_handler(struct ib_cm_id * cm_id,const struct ib_cm_event * ib_event)4150 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
4151 				const struct ib_cm_event *ib_event)
4152 {
4153 	struct rdma_id_private *id_priv = cm_id->context;
4154 	struct rdma_cm_event event = {};
4155 	const struct ib_cm_sidr_rep_event_param *rep =
4156 				&ib_event->param.sidr_rep_rcvd;
4157 	int ret;
4158 
4159 	mutex_lock(&id_priv->handler_mutex);
4160 	if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4161 		goto out;
4162 
4163 	switch (ib_event->event) {
4164 	case IB_CM_SIDR_REQ_ERROR:
4165 		event.event = RDMA_CM_EVENT_UNREACHABLE;
4166 		event.status = -ETIMEDOUT;
4167 		break;
4168 	case IB_CM_SIDR_REP_RECEIVED:
4169 		event.param.ud.private_data = ib_event->private_data;
4170 		event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
4171 		if (rep->status != IB_SIDR_SUCCESS) {
4172 			event.event = RDMA_CM_EVENT_UNREACHABLE;
4173 			event.status = ib_event->param.sidr_rep_rcvd.status;
4174 			pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
4175 					     event.status);
4176 			break;
4177 		}
4178 		ret = cma_set_qkey(id_priv, rep->qkey);
4179 		if (ret) {
4180 			pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
4181 			event.event = RDMA_CM_EVENT_ADDR_ERROR;
4182 			event.status = ret;
4183 			break;
4184 		}
4185 		ib_init_ah_attr_from_path(id_priv->id.device,
4186 					  id_priv->id.port_num,
4187 					  id_priv->id.route.path_rec,
4188 					  &event.param.ud.ah_attr,
4189 					  rep->sgid_attr);
4190 		event.param.ud.qp_num = rep->qpn;
4191 		event.param.ud.qkey = rep->qkey;
4192 		event.event = RDMA_CM_EVENT_ESTABLISHED;
4193 		event.status = 0;
4194 		break;
4195 	default:
4196 		pr_err("RDMA CMA: unexpected IB CM event: %d\n",
4197 		       ib_event->event);
4198 		goto out;
4199 	}
4200 
4201 	ret = cma_cm_event_handler(id_priv, &event);
4202 
4203 	rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4204 	if (ret) {
4205 		/* Destroy the CM ID by returning a non-zero value. */
4206 		id_priv->cm_id.ib = NULL;
4207 		destroy_id_handler_unlock(id_priv);
4208 		return ret;
4209 	}
4210 out:
4211 	mutex_unlock(&id_priv->handler_mutex);
4212 	return 0;
4213 }
4214 
cma_resolve_ib_udp(struct rdma_id_private * id_priv,struct rdma_conn_param * conn_param)4215 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
4216 			      struct rdma_conn_param *conn_param)
4217 {
4218 	struct ib_cm_sidr_req_param req;
4219 	struct ib_cm_id	*id;
4220 	void *private_data;
4221 	u8 offset;
4222 	int ret;
4223 
4224 	memset(&req, 0, sizeof req);
4225 	offset = cma_user_data_offset(id_priv);
4226 	if (check_add_overflow(offset, conn_param->private_data_len, &req.private_data_len))
4227 		return -EINVAL;
4228 
4229 	if (req.private_data_len) {
4230 		private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4231 		if (!private_data)
4232 			return -ENOMEM;
4233 	} else {
4234 		private_data = NULL;
4235 	}
4236 
4237 	if (conn_param->private_data && conn_param->private_data_len)
4238 		memcpy(private_data + offset, conn_param->private_data,
4239 		       conn_param->private_data_len);
4240 
4241 	if (private_data) {
4242 		ret = cma_format_hdr(private_data, id_priv);
4243 		if (ret)
4244 			goto out;
4245 		req.private_data = private_data;
4246 	}
4247 
4248 	id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
4249 			     id_priv);
4250 	if (IS_ERR(id)) {
4251 		ret = PTR_ERR(id);
4252 		goto out;
4253 	}
4254 	id_priv->cm_id.ib = id;
4255 
4256 	req.path = id_priv->id.route.path_rec;
4257 	req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4258 	req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4259 	req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
4260 	req.max_cm_retries = CMA_MAX_CM_RETRIES;
4261 
4262 	trace_cm_send_sidr_req(id_priv);
4263 	ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
4264 	if (ret) {
4265 		ib_destroy_cm_id(id_priv->cm_id.ib);
4266 		id_priv->cm_id.ib = NULL;
4267 	}
4268 out:
4269 	kfree(private_data);
4270 	return ret;
4271 }
4272 
cma_connect_ib(struct rdma_id_private * id_priv,struct rdma_conn_param * conn_param)4273 static int cma_connect_ib(struct rdma_id_private *id_priv,
4274 			  struct rdma_conn_param *conn_param)
4275 {
4276 	struct ib_cm_req_param req;
4277 	struct rdma_route *route;
4278 	void *private_data;
4279 	struct ib_cm_id	*id;
4280 	u8 offset;
4281 	int ret;
4282 
4283 	memset(&req, 0, sizeof req);
4284 	offset = cma_user_data_offset(id_priv);
4285 	if (check_add_overflow(offset, conn_param->private_data_len, &req.private_data_len))
4286 		return -EINVAL;
4287 
4288 	if (req.private_data_len) {
4289 		private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4290 		if (!private_data)
4291 			return -ENOMEM;
4292 	} else {
4293 		private_data = NULL;
4294 	}
4295 
4296 	if (conn_param->private_data && conn_param->private_data_len)
4297 		memcpy(private_data + offset, conn_param->private_data,
4298 		       conn_param->private_data_len);
4299 
4300 	id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
4301 	if (IS_ERR(id)) {
4302 		ret = PTR_ERR(id);
4303 		goto out;
4304 	}
4305 	id_priv->cm_id.ib = id;
4306 
4307 	route = &id_priv->id.route;
4308 	if (private_data) {
4309 		ret = cma_format_hdr(private_data, id_priv);
4310 		if (ret)
4311 			goto out;
4312 		req.private_data = private_data;
4313 	}
4314 
4315 	req.primary_path = &route->path_rec[0];
4316 	req.primary_path_inbound = route->path_rec_inbound;
4317 	req.primary_path_outbound = route->path_rec_outbound;
4318 	if (route->num_pri_alt_paths == 2)
4319 		req.alternate_path = &route->path_rec[1];
4320 
4321 	req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4322 	/* Alternate path SGID attribute currently unsupported */
4323 	req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4324 	req.qp_num = id_priv->qp_num;
4325 	req.qp_type = id_priv->id.qp_type;
4326 	req.starting_psn = id_priv->seq_num;
4327 	req.responder_resources = conn_param->responder_resources;
4328 	req.initiator_depth = conn_param->initiator_depth;
4329 	req.flow_control = conn_param->flow_control;
4330 	req.retry_count = min_t(u8, 7, conn_param->retry_count);
4331 	req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4332 	req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4333 	req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4334 	req.max_cm_retries = CMA_MAX_CM_RETRIES;
4335 	req.srq = id_priv->srq ? 1 : 0;
4336 	req.ece.vendor_id = id_priv->ece.vendor_id;
4337 	req.ece.attr_mod = id_priv->ece.attr_mod;
4338 
4339 	trace_cm_send_req(id_priv);
4340 	ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
4341 out:
4342 	if (ret && !IS_ERR(id)) {
4343 		ib_destroy_cm_id(id);
4344 		id_priv->cm_id.ib = NULL;
4345 	}
4346 
4347 	kfree(private_data);
4348 	return ret;
4349 }
4350 
cma_connect_iw(struct rdma_id_private * id_priv,struct rdma_conn_param * conn_param)4351 static int cma_connect_iw(struct rdma_id_private *id_priv,
4352 			  struct rdma_conn_param *conn_param)
4353 {
4354 	struct iw_cm_id *cm_id;
4355 	int ret;
4356 	struct iw_cm_conn_param iw_param;
4357 
4358 	cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
4359 	if (IS_ERR(cm_id))
4360 		return PTR_ERR(cm_id);
4361 
4362 	mutex_lock(&id_priv->qp_mutex);
4363 	cm_id->tos = id_priv->tos;
4364 	cm_id->tos_set = id_priv->tos_set;
4365 	mutex_unlock(&id_priv->qp_mutex);
4366 
4367 	id_priv->cm_id.iw = cm_id;
4368 
4369 	memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4370 	       rdma_addr_size(cma_src_addr(id_priv)));
4371 	memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4372 	       rdma_addr_size(cma_dst_addr(id_priv)));
4373 
4374 	ret = cma_modify_qp_rtr(id_priv, conn_param);
4375 	if (ret)
4376 		goto out;
4377 
4378 	if (conn_param) {
4379 		iw_param.ord = conn_param->initiator_depth;
4380 		iw_param.ird = conn_param->responder_resources;
4381 		iw_param.private_data = conn_param->private_data;
4382 		iw_param.private_data_len = conn_param->private_data_len;
4383 		iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4384 	} else {
4385 		memset(&iw_param, 0, sizeof iw_param);
4386 		iw_param.qpn = id_priv->qp_num;
4387 	}
4388 	ret = iw_cm_connect(cm_id, &iw_param);
4389 out:
4390 	if (ret) {
4391 		iw_destroy_cm_id(cm_id);
4392 		id_priv->cm_id.iw = NULL;
4393 	}
4394 	return ret;
4395 }
4396 
4397 /**
4398  * rdma_connect_locked - Initiate an active connection request.
4399  * @id: Connection identifier to connect.
4400  * @conn_param: Connection information used for connected QPs.
4401  *
4402  * Same as rdma_connect() but can only be called from the
4403  * RDMA_CM_EVENT_ROUTE_RESOLVED handler callback.
4404  */
rdma_connect_locked(struct rdma_cm_id * id,struct rdma_conn_param * conn_param)4405 int rdma_connect_locked(struct rdma_cm_id *id,
4406 			struct rdma_conn_param *conn_param)
4407 {
4408 	struct rdma_id_private *id_priv =
4409 		container_of(id, struct rdma_id_private, id);
4410 	int ret;
4411 
4412 	if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4413 		return -EINVAL;
4414 
4415 	if (!id->qp) {
4416 		id_priv->qp_num = conn_param->qp_num;
4417 		id_priv->srq = conn_param->srq;
4418 	}
4419 
4420 	if (rdma_cap_ib_cm(id->device, id->port_num)) {
4421 		if (id->qp_type == IB_QPT_UD)
4422 			ret = cma_resolve_ib_udp(id_priv, conn_param);
4423 		else
4424 			ret = cma_connect_ib(id_priv, conn_param);
4425 	} else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4426 		ret = cma_connect_iw(id_priv, conn_param);
4427 	} else {
4428 		ret = -ENOSYS;
4429 	}
4430 	if (ret)
4431 		goto err_state;
4432 	return 0;
4433 err_state:
4434 	cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4435 	return ret;
4436 }
4437 EXPORT_SYMBOL(rdma_connect_locked);
4438 
4439 /**
4440  * rdma_connect - Initiate an active connection request.
4441  * @id: Connection identifier to connect.
4442  * @conn_param: Connection information used for connected QPs.
4443  *
4444  * Users must have resolved a route for the rdma_cm_id to connect with by having
4445  * called rdma_resolve_route before calling this routine.
4446  *
4447  * This call will either connect to a remote QP or obtain remote QP information
4448  * for unconnected rdma_cm_id's.  The actual operation is based on the
4449  * rdma_cm_id's port space.
4450  */
rdma_connect(struct rdma_cm_id * id,struct rdma_conn_param * conn_param)4451 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4452 {
4453 	struct rdma_id_private *id_priv =
4454 		container_of(id, struct rdma_id_private, id);
4455 	int ret;
4456 
4457 	mutex_lock(&id_priv->handler_mutex);
4458 	ret = rdma_connect_locked(id, conn_param);
4459 	mutex_unlock(&id_priv->handler_mutex);
4460 	return ret;
4461 }
4462 EXPORT_SYMBOL(rdma_connect);
4463 
4464 /**
4465  * rdma_connect_ece - Initiate an active connection request with ECE data.
4466  * @id: Connection identifier to connect.
4467  * @conn_param: Connection information used for connected QPs.
4468  * @ece: ECE parameters
4469  *
4470  * See rdma_connect() explanation.
4471  */
rdma_connect_ece(struct rdma_cm_id * id,struct rdma_conn_param * conn_param,struct rdma_ucm_ece * ece)4472 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4473 		     struct rdma_ucm_ece *ece)
4474 {
4475 	struct rdma_id_private *id_priv =
4476 		container_of(id, struct rdma_id_private, id);
4477 
4478 	id_priv->ece.vendor_id = ece->vendor_id;
4479 	id_priv->ece.attr_mod = ece->attr_mod;
4480 
4481 	return rdma_connect(id, conn_param);
4482 }
4483 EXPORT_SYMBOL(rdma_connect_ece);
4484 
cma_accept_ib(struct rdma_id_private * id_priv,struct rdma_conn_param * conn_param)4485 static int cma_accept_ib(struct rdma_id_private *id_priv,
4486 			 struct rdma_conn_param *conn_param)
4487 {
4488 	struct ib_cm_rep_param rep;
4489 	int ret;
4490 
4491 	ret = cma_modify_qp_rtr(id_priv, conn_param);
4492 	if (ret)
4493 		goto out;
4494 
4495 	ret = cma_modify_qp_rts(id_priv, conn_param);
4496 	if (ret)
4497 		goto out;
4498 
4499 	memset(&rep, 0, sizeof rep);
4500 	rep.qp_num = id_priv->qp_num;
4501 	rep.starting_psn = id_priv->seq_num;
4502 	rep.private_data = conn_param->private_data;
4503 	rep.private_data_len = conn_param->private_data_len;
4504 	rep.responder_resources = conn_param->responder_resources;
4505 	rep.initiator_depth = conn_param->initiator_depth;
4506 	rep.failover_accepted = 0;
4507 	rep.flow_control = conn_param->flow_control;
4508 	rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4509 	rep.srq = id_priv->srq ? 1 : 0;
4510 	rep.ece.vendor_id = id_priv->ece.vendor_id;
4511 	rep.ece.attr_mod = id_priv->ece.attr_mod;
4512 
4513 	trace_cm_send_rep(id_priv);
4514 	ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4515 out:
4516 	return ret;
4517 }
4518 
cma_accept_iw(struct rdma_id_private * id_priv,struct rdma_conn_param * conn_param)4519 static int cma_accept_iw(struct rdma_id_private *id_priv,
4520 		  struct rdma_conn_param *conn_param)
4521 {
4522 	struct iw_cm_conn_param iw_param;
4523 	int ret;
4524 
4525 	if (!conn_param)
4526 		return -EINVAL;
4527 
4528 	ret = cma_modify_qp_rtr(id_priv, conn_param);
4529 	if (ret)
4530 		return ret;
4531 
4532 	iw_param.ord = conn_param->initiator_depth;
4533 	iw_param.ird = conn_param->responder_resources;
4534 	iw_param.private_data = conn_param->private_data;
4535 	iw_param.private_data_len = conn_param->private_data_len;
4536 	if (id_priv->id.qp)
4537 		iw_param.qpn = id_priv->qp_num;
4538 	else
4539 		iw_param.qpn = conn_param->qp_num;
4540 
4541 	return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4542 }
4543 
cma_send_sidr_rep(struct rdma_id_private * id_priv,enum ib_cm_sidr_status status,u32 qkey,const void * private_data,int private_data_len)4544 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4545 			     enum ib_cm_sidr_status status, u32 qkey,
4546 			     const void *private_data, int private_data_len)
4547 {
4548 	struct ib_cm_sidr_rep_param rep;
4549 	int ret;
4550 
4551 	memset(&rep, 0, sizeof rep);
4552 	rep.status = status;
4553 	if (status == IB_SIDR_SUCCESS) {
4554 		ret = cma_set_qkey(id_priv, qkey);
4555 		if (ret)
4556 			return ret;
4557 		rep.qp_num = id_priv->qp_num;
4558 		rep.qkey = id_priv->qkey;
4559 
4560 		rep.ece.vendor_id = id_priv->ece.vendor_id;
4561 		rep.ece.attr_mod = id_priv->ece.attr_mod;
4562 	}
4563 
4564 	rep.private_data = private_data;
4565 	rep.private_data_len = private_data_len;
4566 
4567 	trace_cm_send_sidr_rep(id_priv);
4568 	return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4569 }
4570 
4571 /**
4572  * rdma_accept - Called to accept a connection request or response.
4573  * @id: Connection identifier associated with the request.
4574  * @conn_param: Information needed to establish the connection.  This must be
4575  *   provided if accepting a connection request.  If accepting a connection
4576  *   response, this parameter must be NULL.
4577  *
4578  * Typically, this routine is only called by the listener to accept a connection
4579  * request.  It must also be called on the active side of a connection if the
4580  * user is performing their own QP transitions.
4581  *
4582  * In the case of error, a reject message is sent to the remote side and the
4583  * state of the qp associated with the id is modified to error, such that any
4584  * previously posted receive buffers would be flushed.
4585  *
4586  * This function is for use by kernel ULPs and must be called from under the
4587  * handler callback.
4588  */
rdma_accept(struct rdma_cm_id * id,struct rdma_conn_param * conn_param)4589 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4590 {
4591 	struct rdma_id_private *id_priv =
4592 		container_of(id, struct rdma_id_private, id);
4593 	int ret;
4594 
4595 	lockdep_assert_held(&id_priv->handler_mutex);
4596 
4597 	if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4598 		return -EINVAL;
4599 
4600 	if (!id->qp && conn_param) {
4601 		id_priv->qp_num = conn_param->qp_num;
4602 		id_priv->srq = conn_param->srq;
4603 	}
4604 
4605 	if (rdma_cap_ib_cm(id->device, id->port_num)) {
4606 		if (id->qp_type == IB_QPT_UD) {
4607 			if (conn_param)
4608 				ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4609 							conn_param->qkey,
4610 							conn_param->private_data,
4611 							conn_param->private_data_len);
4612 			else
4613 				ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4614 							0, NULL, 0);
4615 		} else {
4616 			if (conn_param)
4617 				ret = cma_accept_ib(id_priv, conn_param);
4618 			else
4619 				ret = cma_rep_recv(id_priv);
4620 		}
4621 	} else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4622 		ret = cma_accept_iw(id_priv, conn_param);
4623 	} else {
4624 		ret = -ENOSYS;
4625 	}
4626 	if (ret)
4627 		goto reject;
4628 
4629 	return 0;
4630 reject:
4631 	cma_modify_qp_err(id_priv);
4632 	rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4633 	return ret;
4634 }
4635 EXPORT_SYMBOL(rdma_accept);
4636 
rdma_accept_ece(struct rdma_cm_id * id,struct rdma_conn_param * conn_param,struct rdma_ucm_ece * ece)4637 int rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4638 		    struct rdma_ucm_ece *ece)
4639 {
4640 	struct rdma_id_private *id_priv =
4641 		container_of(id, struct rdma_id_private, id);
4642 
4643 	id_priv->ece.vendor_id = ece->vendor_id;
4644 	id_priv->ece.attr_mod = ece->attr_mod;
4645 
4646 	return rdma_accept(id, conn_param);
4647 }
4648 EXPORT_SYMBOL(rdma_accept_ece);
4649 
rdma_lock_handler(struct rdma_cm_id * id)4650 void rdma_lock_handler(struct rdma_cm_id *id)
4651 {
4652 	struct rdma_id_private *id_priv =
4653 		container_of(id, struct rdma_id_private, id);
4654 
4655 	mutex_lock(&id_priv->handler_mutex);
4656 }
4657 EXPORT_SYMBOL(rdma_lock_handler);
4658 
rdma_unlock_handler(struct rdma_cm_id * id)4659 void rdma_unlock_handler(struct rdma_cm_id *id)
4660 {
4661 	struct rdma_id_private *id_priv =
4662 		container_of(id, struct rdma_id_private, id);
4663 
4664 	mutex_unlock(&id_priv->handler_mutex);
4665 }
4666 EXPORT_SYMBOL(rdma_unlock_handler);
4667 
rdma_notify(struct rdma_cm_id * id,enum ib_event_type event)4668 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4669 {
4670 	struct rdma_id_private *id_priv;
4671 	int ret;
4672 
4673 	id_priv = container_of(id, struct rdma_id_private, id);
4674 	if (!id_priv->cm_id.ib)
4675 		return -EINVAL;
4676 
4677 	switch (id->device->node_type) {
4678 	case RDMA_NODE_IB_CA:
4679 		ret = ib_cm_notify(id_priv->cm_id.ib, event);
4680 		break;
4681 	default:
4682 		ret = 0;
4683 		break;
4684 	}
4685 	return ret;
4686 }
4687 EXPORT_SYMBOL(rdma_notify);
4688 
rdma_reject(struct rdma_cm_id * id,const void * private_data,u8 private_data_len,u8 reason)4689 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4690 		u8 private_data_len, u8 reason)
4691 {
4692 	struct rdma_id_private *id_priv;
4693 	int ret;
4694 
4695 	id_priv = container_of(id, struct rdma_id_private, id);
4696 	if (!id_priv->cm_id.ib)
4697 		return -EINVAL;
4698 
4699 	if (rdma_cap_ib_cm(id->device, id->port_num)) {
4700 		if (id->qp_type == IB_QPT_UD) {
4701 			ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4702 						private_data, private_data_len);
4703 		} else {
4704 			trace_cm_send_rej(id_priv);
4705 			ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4706 					     private_data, private_data_len);
4707 		}
4708 	} else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4709 		ret = iw_cm_reject(id_priv->cm_id.iw,
4710 				   private_data, private_data_len);
4711 	} else {
4712 		ret = -ENOSYS;
4713 	}
4714 
4715 	return ret;
4716 }
4717 EXPORT_SYMBOL(rdma_reject);
4718 
rdma_disconnect(struct rdma_cm_id * id)4719 int rdma_disconnect(struct rdma_cm_id *id)
4720 {
4721 	struct rdma_id_private *id_priv;
4722 	int ret;
4723 
4724 	id_priv = container_of(id, struct rdma_id_private, id);
4725 	if (!id_priv->cm_id.ib)
4726 		return -EINVAL;
4727 
4728 	if (rdma_cap_ib_cm(id->device, id->port_num)) {
4729 		ret = cma_modify_qp_err(id_priv);
4730 		if (ret)
4731 			goto out;
4732 		/* Initiate or respond to a disconnect. */
4733 		trace_cm_disconnect(id_priv);
4734 		if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4735 			if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4736 				trace_cm_sent_drep(id_priv);
4737 		} else {
4738 			trace_cm_sent_dreq(id_priv);
4739 		}
4740 	} else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4741 		ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4742 	} else
4743 		ret = -EINVAL;
4744 
4745 out:
4746 	return ret;
4747 }
4748 EXPORT_SYMBOL(rdma_disconnect);
4749 
cma_make_mc_event(int status,struct rdma_id_private * id_priv,struct ib_sa_multicast * multicast,struct rdma_cm_event * event,struct cma_multicast * mc)4750 static void cma_make_mc_event(int status, struct rdma_id_private *id_priv,
4751 			      struct ib_sa_multicast *multicast,
4752 			      struct rdma_cm_event *event,
4753 			      struct cma_multicast *mc)
4754 {
4755 	struct rdma_dev_addr *dev_addr;
4756 	enum ib_gid_type gid_type;
4757 	struct net_device *ndev;
4758 
4759 	if (!status)
4760 		status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4761 	else
4762 		pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4763 				     status);
4764 
4765 	event->status = status;
4766 	event->param.ud.private_data = mc->context;
4767 	if (status) {
4768 		event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4769 		return;
4770 	}
4771 
4772 	dev_addr = &id_priv->id.route.addr.dev_addr;
4773 	ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4774 	gid_type =
4775 		id_priv->cma_dev
4776 			->default_gid_type[id_priv->id.port_num -
4777 					   rdma_start_port(
4778 						   id_priv->cma_dev->device)];
4779 
4780 	event->event = RDMA_CM_EVENT_MULTICAST_JOIN;
4781 	if (ib_init_ah_from_mcmember(id_priv->id.device, id_priv->id.port_num,
4782 				     &multicast->rec, ndev, gid_type,
4783 				     &event->param.ud.ah_attr)) {
4784 		event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4785 		goto out;
4786 	}
4787 
4788 	event->param.ud.qp_num = 0xFFFFFF;
4789 	event->param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4790 
4791 out:
4792 	if (ndev)
4793 		dev_put(ndev);
4794 }
4795 
cma_ib_mc_handler(int status,struct ib_sa_multicast * multicast)4796 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4797 {
4798 	struct cma_multicast *mc = multicast->context;
4799 	struct rdma_id_private *id_priv = mc->id_priv;
4800 	struct rdma_cm_event event = {};
4801 	int ret = 0;
4802 
4803 	mutex_lock(&id_priv->handler_mutex);
4804 	if (READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL ||
4805 	    READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING)
4806 		goto out;
4807 
4808 	cma_make_mc_event(status, id_priv, multicast, &event, mc);
4809 	ret = cma_cm_event_handler(id_priv, &event);
4810 	rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4811 	WARN_ON(ret);
4812 
4813 out:
4814 	mutex_unlock(&id_priv->handler_mutex);
4815 	return 0;
4816 }
4817 
cma_set_mgid(struct rdma_id_private * id_priv,struct sockaddr * addr,union ib_gid * mgid)4818 static void cma_set_mgid(struct rdma_id_private *id_priv,
4819 			 struct sockaddr *addr, union ib_gid *mgid)
4820 {
4821 	unsigned char mc_map[MAX_ADDR_LEN];
4822 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4823 	struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4824 	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4825 
4826 	if (cma_any_addr(addr)) {
4827 		memset(mgid, 0, sizeof *mgid);
4828 	} else if ((addr->sa_family == AF_INET6) &&
4829 		   ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4830 								 0xFF10A01B)) {
4831 		/* IPv6 address is an SA assigned MGID. */
4832 		memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4833 	} else if (addr->sa_family == AF_IB) {
4834 		memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4835 	} else if (addr->sa_family == AF_INET6) {
4836 		ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4837 		if (id_priv->id.ps == RDMA_PS_UDP)
4838 			mc_map[7] = 0x01;	/* Use RDMA CM signature */
4839 		*mgid = *(union ib_gid *) (mc_map + 4);
4840 	} else {
4841 		ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4842 		if (id_priv->id.ps == RDMA_PS_UDP)
4843 			mc_map[7] = 0x01;	/* Use RDMA CM signature */
4844 		*mgid = *(union ib_gid *) (mc_map + 4);
4845 	}
4846 }
4847 
cma_join_ib_multicast(struct rdma_id_private * id_priv,struct cma_multicast * mc)4848 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4849 				 struct cma_multicast *mc)
4850 {
4851 	struct ib_sa_mcmember_rec rec;
4852 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4853 	ib_sa_comp_mask comp_mask;
4854 	int ret;
4855 
4856 	ib_addr_get_mgid(dev_addr, &rec.mgid);
4857 	ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4858 				     &rec.mgid, &rec);
4859 	if (ret)
4860 		return ret;
4861 
4862 	ret = cma_set_qkey(id_priv, 0);
4863 	if (ret)
4864 		return ret;
4865 
4866 	cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4867 	rec.qkey = cpu_to_be32(id_priv->qkey);
4868 	rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4869 	rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4870 	rec.join_state = mc->join_state;
4871 
4872 	comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4873 		    IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4874 		    IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4875 		    IB_SA_MCMEMBER_REC_FLOW_LABEL |
4876 		    IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4877 
4878 	if (id_priv->id.ps == RDMA_PS_IPOIB)
4879 		comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4880 			     IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4881 			     IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4882 			     IB_SA_MCMEMBER_REC_MTU |
4883 			     IB_SA_MCMEMBER_REC_HOP_LIMIT;
4884 
4885 	mc->sa_mc = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4886 					 id_priv->id.port_num, &rec, comp_mask,
4887 					 GFP_KERNEL, cma_ib_mc_handler, mc);
4888 	return PTR_ERR_OR_ZERO(mc->sa_mc);
4889 }
4890 
cma_iboe_set_mgid(struct sockaddr * addr,union ib_gid * mgid,enum ib_gid_type gid_type)4891 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4892 			      enum ib_gid_type gid_type)
4893 {
4894 	struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4895 	struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4896 
4897 	if (cma_any_addr(addr)) {
4898 		memset(mgid, 0, sizeof *mgid);
4899 	} else if (addr->sa_family == AF_INET6) {
4900 		memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4901 	} else {
4902 		mgid->raw[0] =
4903 			(gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4904 		mgid->raw[1] =
4905 			(gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4906 		mgid->raw[2] = 0;
4907 		mgid->raw[3] = 0;
4908 		mgid->raw[4] = 0;
4909 		mgid->raw[5] = 0;
4910 		mgid->raw[6] = 0;
4911 		mgid->raw[7] = 0;
4912 		mgid->raw[8] = 0;
4913 		mgid->raw[9] = 0;
4914 		mgid->raw[10] = 0xff;
4915 		mgid->raw[11] = 0xff;
4916 		*(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4917 	}
4918 }
4919 
cma_iboe_join_multicast(struct rdma_id_private * id_priv,struct cma_multicast * mc)4920 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4921 				   struct cma_multicast *mc)
4922 {
4923 	struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4924 	int err = 0;
4925 	struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4926 	struct net_device *ndev = NULL;
4927 	struct ib_sa_multicast ib;
4928 	enum ib_gid_type gid_type;
4929 	bool send_only;
4930 
4931 	send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4932 
4933 	if (cma_zero_addr(addr))
4934 		return -EINVAL;
4935 
4936 	gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4937 		   rdma_start_port(id_priv->cma_dev->device)];
4938 	cma_iboe_set_mgid(addr, &ib.rec.mgid, gid_type);
4939 
4940 	ib.rec.pkey = cpu_to_be16(0xffff);
4941 	if (id_priv->id.ps == RDMA_PS_UDP)
4942 		ib.rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4943 
4944 	if (dev_addr->bound_dev_if)
4945 		ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4946 	if (!ndev)
4947 		return -ENODEV;
4948 
4949 	ib.rec.rate = iboe_get_rate(ndev);
4950 	ib.rec.hop_limit = 1;
4951 	ib.rec.mtu = iboe_get_mtu(ndev->mtu);
4952 
4953 	if (addr->sa_family == AF_INET) {
4954 		if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4955 			ib.rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4956 			if (!send_only) {
4957 				err = cma_igmp_send(ndev, &ib.rec.mgid,
4958 						    true);
4959 			}
4960 		}
4961 	} else {
4962 		if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4963 			err = -ENOTSUPP;
4964 	}
4965 	dev_put(ndev);
4966 	if (err || !ib.rec.mtu)
4967 		return err ?: -EINVAL;
4968 
4969 	rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4970 		    &ib.rec.port_gid);
4971 	INIT_WORK(&mc->iboe_join.work, cma_iboe_join_work_handler);
4972 	cma_make_mc_event(0, id_priv, &ib, &mc->iboe_join.event, mc);
4973 	queue_work(cma_wq, &mc->iboe_join.work);
4974 	return 0;
4975 }
4976 
rdma_join_multicast(struct rdma_cm_id * id,struct sockaddr * addr,u8 join_state,void * context)4977 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4978 			u8 join_state, void *context)
4979 {
4980 	struct rdma_id_private *id_priv =
4981 		container_of(id, struct rdma_id_private, id);
4982 	struct cma_multicast *mc;
4983 	int ret;
4984 
4985 	/* Not supported for kernel QPs */
4986 	if (WARN_ON(id->qp))
4987 		return -EINVAL;
4988 
4989 	/* ULP is calling this wrong. */
4990 	if (!id->device || (READ_ONCE(id_priv->state) != RDMA_CM_ADDR_BOUND &&
4991 			    READ_ONCE(id_priv->state) != RDMA_CM_ADDR_RESOLVED))
4992 		return -EINVAL;
4993 
4994 	mc = kzalloc(sizeof(*mc), GFP_KERNEL);
4995 	if (!mc)
4996 		return -ENOMEM;
4997 
4998 	memcpy(&mc->addr, addr, rdma_addr_size(addr));
4999 	mc->context = context;
5000 	mc->id_priv = id_priv;
5001 	mc->join_state = join_state;
5002 
5003 	if (rdma_protocol_roce(id->device, id->port_num)) {
5004 		ret = cma_iboe_join_multicast(id_priv, mc);
5005 		if (ret)
5006 			goto out_err;
5007 	} else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
5008 		ret = cma_join_ib_multicast(id_priv, mc);
5009 		if (ret)
5010 			goto out_err;
5011 	} else {
5012 		ret = -ENOSYS;
5013 		goto out_err;
5014 	}
5015 
5016 	spin_lock(&id_priv->lock);
5017 	list_add(&mc->list, &id_priv->mc_list);
5018 	spin_unlock(&id_priv->lock);
5019 
5020 	return 0;
5021 out_err:
5022 	kfree(mc);
5023 	return ret;
5024 }
5025 EXPORT_SYMBOL(rdma_join_multicast);
5026 
rdma_leave_multicast(struct rdma_cm_id * id,struct sockaddr * addr)5027 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
5028 {
5029 	struct rdma_id_private *id_priv;
5030 	struct cma_multicast *mc;
5031 
5032 	id_priv = container_of(id, struct rdma_id_private, id);
5033 	spin_lock_irq(&id_priv->lock);
5034 	list_for_each_entry(mc, &id_priv->mc_list, list) {
5035 		if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
5036 			continue;
5037 		list_del(&mc->list);
5038 		spin_unlock_irq(&id_priv->lock);
5039 
5040 		WARN_ON(id_priv->cma_dev->device != id->device);
5041 		destroy_mc(id_priv, mc);
5042 		return;
5043 	}
5044 	spin_unlock_irq(&id_priv->lock);
5045 }
5046 EXPORT_SYMBOL(rdma_leave_multicast);
5047 
cma_netdev_change(struct net_device * ndev,struct rdma_id_private * id_priv)5048 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
5049 {
5050 	struct rdma_dev_addr *dev_addr;
5051 	struct cma_work *work;
5052 
5053 	dev_addr = &id_priv->id.route.addr.dev_addr;
5054 
5055 	if ((dev_addr->bound_dev_if == ndev->ifindex) &&
5056 	    (net_eq(dev_net(ndev), dev_addr->net)) &&
5057 	    memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
5058 		pr_info("RDMA CM addr change for ndev %s used by id %p\n",
5059 			ndev->name, &id_priv->id);
5060 		work = kzalloc(sizeof *work, GFP_KERNEL);
5061 		if (!work)
5062 			return -ENOMEM;
5063 
5064 		INIT_WORK(&work->work, cma_work_handler);
5065 		work->id = id_priv;
5066 		work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
5067 		cma_id_get(id_priv);
5068 		queue_work(cma_wq, &work->work);
5069 	}
5070 
5071 	return 0;
5072 }
5073 
cma_netdev_callback(struct notifier_block * self,unsigned long event,void * ptr)5074 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
5075 			       void *ptr)
5076 {
5077 	struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
5078 	struct cma_device *cma_dev;
5079 	struct rdma_id_private *id_priv;
5080 	int ret = NOTIFY_DONE;
5081 
5082 	if (event != NETDEV_BONDING_FAILOVER)
5083 		return NOTIFY_DONE;
5084 
5085 	if (!netif_is_bond_master(ndev))
5086 		return NOTIFY_DONE;
5087 
5088 	mutex_lock(&lock);
5089 	list_for_each_entry(cma_dev, &dev_list, list)
5090 		list_for_each_entry(id_priv, &cma_dev->id_list, device_item) {
5091 			ret = cma_netdev_change(ndev, id_priv);
5092 			if (ret)
5093 				goto out;
5094 		}
5095 
5096 out:
5097 	mutex_unlock(&lock);
5098 	return ret;
5099 }
5100 
cma_netevent_work_handler(struct work_struct * _work)5101 static void cma_netevent_work_handler(struct work_struct *_work)
5102 {
5103 	struct rdma_id_private *id_priv =
5104 		container_of(_work, struct rdma_id_private, id.net_work);
5105 	struct rdma_cm_event event = {};
5106 
5107 	mutex_lock(&id_priv->handler_mutex);
5108 
5109 	if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
5110 	    READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
5111 		goto out_unlock;
5112 
5113 	event.event = RDMA_CM_EVENT_UNREACHABLE;
5114 	event.status = -ETIMEDOUT;
5115 
5116 	if (cma_cm_event_handler(id_priv, &event)) {
5117 		__acquire(&id_priv->handler_mutex);
5118 		id_priv->cm_id.ib = NULL;
5119 		cma_id_put(id_priv);
5120 		destroy_id_handler_unlock(id_priv);
5121 		return;
5122 	}
5123 
5124 out_unlock:
5125 	mutex_unlock(&id_priv->handler_mutex);
5126 	cma_id_put(id_priv);
5127 }
5128 
cma_netevent_callback(struct notifier_block * self,unsigned long event,void * ctx)5129 static int cma_netevent_callback(struct notifier_block *self,
5130 				 unsigned long event, void *ctx)
5131 {
5132 	struct id_table_entry *ips_node = NULL;
5133 	struct rdma_id_private *current_id;
5134 	struct neighbour *neigh = ctx;
5135 	unsigned long flags;
5136 
5137 	if (event != NETEVENT_NEIGH_UPDATE)
5138 		return NOTIFY_DONE;
5139 
5140 	spin_lock_irqsave(&id_table_lock, flags);
5141 	if (neigh->tbl->family == AF_INET6) {
5142 		struct sockaddr_in6 neigh_sock_6;
5143 
5144 		neigh_sock_6.sin6_family = AF_INET6;
5145 		neigh_sock_6.sin6_addr = *(struct in6_addr *)neigh->primary_key;
5146 		ips_node = node_from_ndev_ip(&id_table, neigh->dev->ifindex,
5147 					     (struct sockaddr *)&neigh_sock_6);
5148 	} else if (neigh->tbl->family == AF_INET) {
5149 		struct sockaddr_in neigh_sock_4;
5150 
5151 		neigh_sock_4.sin_family = AF_INET;
5152 		neigh_sock_4.sin_addr.s_addr = *(__be32 *)(neigh->primary_key);
5153 		ips_node = node_from_ndev_ip(&id_table, neigh->dev->ifindex,
5154 					     (struct sockaddr *)&neigh_sock_4);
5155 	} else
5156 		goto out;
5157 
5158 	if (!ips_node)
5159 		goto out;
5160 
5161 	list_for_each_entry(current_id, &ips_node->id_list, id_list_entry) {
5162 		if (!memcmp(current_id->id.route.addr.dev_addr.dst_dev_addr,
5163 			   neigh->ha, ETH_ALEN))
5164 			continue;
5165 		INIT_WORK(&current_id->id.net_work, cma_netevent_work_handler);
5166 		cma_id_get(current_id);
5167 		queue_work(cma_wq, &current_id->id.net_work);
5168 	}
5169 out:
5170 	spin_unlock_irqrestore(&id_table_lock, flags);
5171 	return NOTIFY_DONE;
5172 }
5173 
5174 static struct notifier_block cma_nb = {
5175 	.notifier_call = cma_netdev_callback
5176 };
5177 
5178 static struct notifier_block cma_netevent_cb = {
5179 	.notifier_call = cma_netevent_callback
5180 };
5181 
cma_send_device_removal_put(struct rdma_id_private * id_priv)5182 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
5183 {
5184 	struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
5185 	enum rdma_cm_state state;
5186 	unsigned long flags;
5187 
5188 	mutex_lock(&id_priv->handler_mutex);
5189 	/* Record that we want to remove the device */
5190 	spin_lock_irqsave(&id_priv->lock, flags);
5191 	state = id_priv->state;
5192 	if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
5193 		spin_unlock_irqrestore(&id_priv->lock, flags);
5194 		mutex_unlock(&id_priv->handler_mutex);
5195 		cma_id_put(id_priv);
5196 		return;
5197 	}
5198 	id_priv->state = RDMA_CM_DEVICE_REMOVAL;
5199 	spin_unlock_irqrestore(&id_priv->lock, flags);
5200 
5201 	if (cma_cm_event_handler(id_priv, &event)) {
5202 		/*
5203 		 * At this point the ULP promises it won't call
5204 		 * rdma_destroy_id() concurrently
5205 		 */
5206 		cma_id_put(id_priv);
5207 		mutex_unlock(&id_priv->handler_mutex);
5208 		trace_cm_id_destroy(id_priv);
5209 		_destroy_id(id_priv, state);
5210 		return;
5211 	}
5212 	mutex_unlock(&id_priv->handler_mutex);
5213 
5214 	/*
5215 	 * If this races with destroy then the thread that first assigns state
5216 	 * to a destroying does the cancel.
5217 	 */
5218 	cma_cancel_operation(id_priv, state);
5219 	cma_id_put(id_priv);
5220 }
5221 
cma_process_remove(struct cma_device * cma_dev)5222 static void cma_process_remove(struct cma_device *cma_dev)
5223 {
5224 	mutex_lock(&lock);
5225 	while (!list_empty(&cma_dev->id_list)) {
5226 		struct rdma_id_private *id_priv = list_first_entry(
5227 			&cma_dev->id_list, struct rdma_id_private, device_item);
5228 
5229 		list_del_init(&id_priv->listen_item);
5230 		list_del_init(&id_priv->device_item);
5231 		cma_id_get(id_priv);
5232 		mutex_unlock(&lock);
5233 
5234 		cma_send_device_removal_put(id_priv);
5235 
5236 		mutex_lock(&lock);
5237 	}
5238 	mutex_unlock(&lock);
5239 
5240 	cma_dev_put(cma_dev);
5241 	wait_for_completion(&cma_dev->comp);
5242 }
5243 
cma_supported(struct ib_device * device)5244 static bool cma_supported(struct ib_device *device)
5245 {
5246 	u32 i;
5247 
5248 	rdma_for_each_port(device, i) {
5249 		if (rdma_cap_ib_cm(device, i) || rdma_cap_iw_cm(device, i))
5250 			return true;
5251 	}
5252 	return false;
5253 }
5254 
cma_add_one(struct ib_device * device)5255 static int cma_add_one(struct ib_device *device)
5256 {
5257 	struct rdma_id_private *to_destroy;
5258 	struct cma_device *cma_dev;
5259 	struct rdma_id_private *id_priv;
5260 	unsigned long supported_gids = 0;
5261 	int ret;
5262 	u32 i;
5263 
5264 	if (!cma_supported(device))
5265 		return -EOPNOTSUPP;
5266 
5267 	cma_dev = kmalloc(sizeof(*cma_dev), GFP_KERNEL);
5268 	if (!cma_dev)
5269 		return -ENOMEM;
5270 
5271 	cma_dev->device = device;
5272 	cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
5273 					    sizeof(*cma_dev->default_gid_type),
5274 					    GFP_KERNEL);
5275 	if (!cma_dev->default_gid_type) {
5276 		ret = -ENOMEM;
5277 		goto free_cma_dev;
5278 	}
5279 
5280 	cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
5281 					    sizeof(*cma_dev->default_roce_tos),
5282 					    GFP_KERNEL);
5283 	if (!cma_dev->default_roce_tos) {
5284 		ret = -ENOMEM;
5285 		goto free_gid_type;
5286 	}
5287 
5288 	rdma_for_each_port (device, i) {
5289 		supported_gids = roce_gid_type_mask_support(device, i);
5290 		WARN_ON(!supported_gids);
5291 		if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
5292 			cma_dev->default_gid_type[i - rdma_start_port(device)] =
5293 				CMA_PREFERRED_ROCE_GID_TYPE;
5294 		else
5295 			cma_dev->default_gid_type[i - rdma_start_port(device)] =
5296 				find_first_bit(&supported_gids, BITS_PER_LONG);
5297 		cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
5298 	}
5299 
5300 	init_completion(&cma_dev->comp);
5301 	refcount_set(&cma_dev->refcount, 1);
5302 	INIT_LIST_HEAD(&cma_dev->id_list);
5303 	ib_set_client_data(device, &cma_client, cma_dev);
5304 
5305 	mutex_lock(&lock);
5306 	list_add_tail(&cma_dev->list, &dev_list);
5307 	list_for_each_entry(id_priv, &listen_any_list, listen_any_item) {
5308 		ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
5309 		if (ret)
5310 			goto free_listen;
5311 	}
5312 	mutex_unlock(&lock);
5313 
5314 	trace_cm_add_one(device);
5315 	return 0;
5316 
5317 free_listen:
5318 	list_del(&cma_dev->list);
5319 	mutex_unlock(&lock);
5320 
5321 	/* cma_process_remove() will delete to_destroy */
5322 	cma_process_remove(cma_dev);
5323 	kfree(cma_dev->default_roce_tos);
5324 free_gid_type:
5325 	kfree(cma_dev->default_gid_type);
5326 
5327 free_cma_dev:
5328 	kfree(cma_dev);
5329 	return ret;
5330 }
5331 
cma_remove_one(struct ib_device * device,void * client_data)5332 static void cma_remove_one(struct ib_device *device, void *client_data)
5333 {
5334 	struct cma_device *cma_dev = client_data;
5335 
5336 	trace_cm_remove_one(device);
5337 
5338 	mutex_lock(&lock);
5339 	list_del(&cma_dev->list);
5340 	mutex_unlock(&lock);
5341 
5342 	cma_process_remove(cma_dev);
5343 	kfree(cma_dev->default_roce_tos);
5344 	kfree(cma_dev->default_gid_type);
5345 	kfree(cma_dev);
5346 }
5347 
cma_init_net(struct net * net)5348 static int cma_init_net(struct net *net)
5349 {
5350 	struct cma_pernet *pernet = cma_pernet(net);
5351 
5352 	xa_init(&pernet->tcp_ps);
5353 	xa_init(&pernet->udp_ps);
5354 	xa_init(&pernet->ipoib_ps);
5355 	xa_init(&pernet->ib_ps);
5356 
5357 	return 0;
5358 }
5359 
cma_exit_net(struct net * net)5360 static void cma_exit_net(struct net *net)
5361 {
5362 	struct cma_pernet *pernet = cma_pernet(net);
5363 
5364 	WARN_ON(!xa_empty(&pernet->tcp_ps));
5365 	WARN_ON(!xa_empty(&pernet->udp_ps));
5366 	WARN_ON(!xa_empty(&pernet->ipoib_ps));
5367 	WARN_ON(!xa_empty(&pernet->ib_ps));
5368 }
5369 
5370 static struct pernet_operations cma_pernet_operations = {
5371 	.init = cma_init_net,
5372 	.exit = cma_exit_net,
5373 	.id = &cma_pernet_id,
5374 	.size = sizeof(struct cma_pernet),
5375 };
5376 
cma_init(void)5377 static int __init cma_init(void)
5378 {
5379 	int ret;
5380 
5381 	/*
5382 	 * There is a rare lock ordering dependency in cma_netdev_callback()
5383 	 * that only happens when bonding is enabled. Teach lockdep that rtnl
5384 	 * must never be nested under lock so it can find these without having
5385 	 * to test with bonding.
5386 	 */
5387 	if (IS_ENABLED(CONFIG_LOCKDEP)) {
5388 		rtnl_lock();
5389 		mutex_lock(&lock);
5390 		mutex_unlock(&lock);
5391 		rtnl_unlock();
5392 	}
5393 
5394 	cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
5395 	if (!cma_wq)
5396 		return -ENOMEM;
5397 
5398 	ret = register_pernet_subsys(&cma_pernet_operations);
5399 	if (ret)
5400 		goto err_wq;
5401 
5402 	ib_sa_register_client(&sa_client);
5403 	register_netdevice_notifier(&cma_nb);
5404 	register_netevent_notifier(&cma_netevent_cb);
5405 
5406 	ret = ib_register_client(&cma_client);
5407 	if (ret)
5408 		goto err;
5409 
5410 	ret = cma_configfs_init();
5411 	if (ret)
5412 		goto err_ib;
5413 
5414 	return 0;
5415 
5416 err_ib:
5417 	ib_unregister_client(&cma_client);
5418 err:
5419 	unregister_netevent_notifier(&cma_netevent_cb);
5420 	unregister_netdevice_notifier(&cma_nb);
5421 	ib_sa_unregister_client(&sa_client);
5422 	unregister_pernet_subsys(&cma_pernet_operations);
5423 err_wq:
5424 	destroy_workqueue(cma_wq);
5425 	return ret;
5426 }
5427 
cma_cleanup(void)5428 static void __exit cma_cleanup(void)
5429 {
5430 	cma_configfs_exit();
5431 	ib_unregister_client(&cma_client);
5432 	unregister_netevent_notifier(&cma_netevent_cb);
5433 	unregister_netdevice_notifier(&cma_nb);
5434 	ib_sa_unregister_client(&sa_client);
5435 	unregister_pernet_subsys(&cma_pernet_operations);
5436 	destroy_workqueue(cma_wq);
5437 }
5438 
5439 module_init(cma_init);
5440 module_exit(cma_cleanup);
5441