1 /*
2  * Copyright (c) 2006 Intel Corporation.  All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/completion.h>
34 #include <linux/dma-mapping.h>
35 #include <linux/err.h>
36 #include <linux/interrupt.h>
37 #include <linux/export.h>
38 #include <linux/slab.h>
39 #include <linux/bitops.h>
40 #include <linux/random.h>
41 
42 #include <rdma/ib_cache.h>
43 #include "sa.h"
44 
45 static void mcast_add_one(struct ib_device *device);
46 static void mcast_remove_one(struct ib_device *device);
47 
48 static struct ib_client mcast_client = {
49 	.name   = "ib_multicast",
50 	.add    = mcast_add_one,
51 	.remove = mcast_remove_one
52 };
53 
54 static struct ib_sa_client	sa_client;
55 static struct workqueue_struct	*mcast_wq;
56 static union ib_gid mgid0;
57 
58 struct mcast_device;
59 
60 struct mcast_port {
61 	struct mcast_device	*dev;
62 	spinlock_t		lock;
63 	struct rb_root		table;
64 	atomic_t		refcount;
65 	struct completion	comp;
66 	u8			port_num;
67 };
68 
69 struct mcast_device {
70 	struct ib_device	*device;
71 	struct ib_event_handler	event_handler;
72 	int			start_port;
73 	int			end_port;
74 	struct mcast_port	port[0];
75 };
76 
77 enum mcast_state {
78 	MCAST_JOINING,
79 	MCAST_MEMBER,
80 	MCAST_ERROR,
81 };
82 
83 enum mcast_group_state {
84 	MCAST_IDLE,
85 	MCAST_BUSY,
86 	MCAST_GROUP_ERROR,
87 	MCAST_PKEY_EVENT
88 };
89 
90 enum {
91 	MCAST_INVALID_PKEY_INDEX = 0xFFFF
92 };
93 
94 struct mcast_member;
95 
96 struct mcast_group {
97 	struct ib_sa_mcmember_rec rec;
98 	struct rb_node		node;
99 	struct mcast_port	*port;
100 	spinlock_t		lock;
101 	struct work_struct	work;
102 	struct list_head	pending_list;
103 	struct list_head	active_list;
104 	struct mcast_member	*last_join;
105 	int			members[3];
106 	atomic_t		refcount;
107 	enum mcast_group_state	state;
108 	struct ib_sa_query	*query;
109 	int			query_id;
110 	u16			pkey_index;
111 	u8			leave_state;
112 	int			retries;
113 };
114 
115 struct mcast_member {
116 	struct ib_sa_multicast	multicast;
117 	struct ib_sa_client	*client;
118 	struct mcast_group	*group;
119 	struct list_head	list;
120 	enum mcast_state	state;
121 	atomic_t		refcount;
122 	struct completion	comp;
123 };
124 
125 static void join_handler(int status, struct ib_sa_mcmember_rec *rec,
126 			 void *context);
127 static void leave_handler(int status, struct ib_sa_mcmember_rec *rec,
128 			  void *context);
129 
130 static struct mcast_group *mcast_find(struct mcast_port *port,
131 				      union ib_gid *mgid)
132 {
133 	struct rb_node *node = port->table.rb_node;
134 	struct mcast_group *group;
135 	int ret;
136 
137 	while (node) {
138 		group = rb_entry(node, struct mcast_group, node);
139 		ret = memcmp(mgid->raw, group->rec.mgid.raw, sizeof *mgid);
140 		if (!ret)
141 			return group;
142 
143 		if (ret < 0)
144 			node = node->rb_left;
145 		else
146 			node = node->rb_right;
147 	}
148 	return NULL;
149 }
150 
151 static struct mcast_group *mcast_insert(struct mcast_port *port,
152 					struct mcast_group *group,
153 					int allow_duplicates)
154 {
155 	struct rb_node **link = &port->table.rb_node;
156 	struct rb_node *parent = NULL;
157 	struct mcast_group *cur_group;
158 	int ret;
159 
160 	while (*link) {
161 		parent = *link;
162 		cur_group = rb_entry(parent, struct mcast_group, node);
163 
164 		ret = memcmp(group->rec.mgid.raw, cur_group->rec.mgid.raw,
165 			     sizeof group->rec.mgid);
166 		if (ret < 0)
167 			link = &(*link)->rb_left;
168 		else if (ret > 0)
169 			link = &(*link)->rb_right;
170 		else if (allow_duplicates)
171 			link = &(*link)->rb_left;
172 		else
173 			return cur_group;
174 	}
175 	rb_link_node(&group->node, parent, link);
176 	rb_insert_color(&group->node, &port->table);
177 	return NULL;
178 }
179 
180 static void deref_port(struct mcast_port *port)
181 {
182 	if (atomic_dec_and_test(&port->refcount))
183 		complete(&port->comp);
184 }
185 
186 static void release_group(struct mcast_group *group)
187 {
188 	struct mcast_port *port = group->port;
189 	unsigned long flags;
190 
191 	spin_lock_irqsave(&port->lock, flags);
192 	if (atomic_dec_and_test(&group->refcount)) {
193 		rb_erase(&group->node, &port->table);
194 		spin_unlock_irqrestore(&port->lock, flags);
195 		kfree(group);
196 		deref_port(port);
197 	} else
198 		spin_unlock_irqrestore(&port->lock, flags);
199 }
200 
201 static void deref_member(struct mcast_member *member)
202 {
203 	if (atomic_dec_and_test(&member->refcount))
204 		complete(&member->comp);
205 }
206 
207 static void queue_join(struct mcast_member *member)
208 {
209 	struct mcast_group *group = member->group;
210 	unsigned long flags;
211 
212 	spin_lock_irqsave(&group->lock, flags);
213 	list_add_tail(&member->list, &group->pending_list);
214 	if (group->state == MCAST_IDLE) {
215 		group->state = MCAST_BUSY;
216 		atomic_inc(&group->refcount);
217 		queue_work(mcast_wq, &group->work);
218 	}
219 	spin_unlock_irqrestore(&group->lock, flags);
220 }
221 
222 /*
223  * A multicast group has three types of members: full member, non member, and
224  * send only member.  We need to keep track of the number of members of each
225  * type based on their join state.  Adjust the number of members the belong to
226  * the specified join states.
227  */
228 static void adjust_membership(struct mcast_group *group, u8 join_state, int inc)
229 {
230 	int i;
231 
232 	for (i = 0; i < 3; i++, join_state >>= 1)
233 		if (join_state & 0x1)
234 			group->members[i] += inc;
235 }
236 
237 /*
238  * If a multicast group has zero members left for a particular join state, but
239  * the group is still a member with the SA, we need to leave that join state.
240  * Determine which join states we still belong to, but that do not have any
241  * active members.
242  */
243 static u8 get_leave_state(struct mcast_group *group)
244 {
245 	u8 leave_state = 0;
246 	int i;
247 
248 	for (i = 0; i < 3; i++)
249 		if (!group->members[i])
250 			leave_state |= (0x1 << i);
251 
252 	return leave_state & group->rec.join_state;
253 }
254 
255 static int check_selector(ib_sa_comp_mask comp_mask,
256 			  ib_sa_comp_mask selector_mask,
257 			  ib_sa_comp_mask value_mask,
258 			  u8 selector, u8 src_value, u8 dst_value)
259 {
260 	int err;
261 
262 	if (!(comp_mask & selector_mask) || !(comp_mask & value_mask))
263 		return 0;
264 
265 	switch (selector) {
266 	case IB_SA_GT:
267 		err = (src_value <= dst_value);
268 		break;
269 	case IB_SA_LT:
270 		err = (src_value >= dst_value);
271 		break;
272 	case IB_SA_EQ:
273 		err = (src_value != dst_value);
274 		break;
275 	default:
276 		err = 0;
277 		break;
278 	}
279 
280 	return err;
281 }
282 
283 static int cmp_rec(struct ib_sa_mcmember_rec *src,
284 		   struct ib_sa_mcmember_rec *dst, ib_sa_comp_mask comp_mask)
285 {
286 	/* MGID must already match */
287 
288 	if (comp_mask & IB_SA_MCMEMBER_REC_PORT_GID &&
289 	    memcmp(&src->port_gid, &dst->port_gid, sizeof src->port_gid))
290 		return -EINVAL;
291 	if (comp_mask & IB_SA_MCMEMBER_REC_QKEY && src->qkey != dst->qkey)
292 		return -EINVAL;
293 	if (comp_mask & IB_SA_MCMEMBER_REC_MLID && src->mlid != dst->mlid)
294 		return -EINVAL;
295 	if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_MTU_SELECTOR,
296 			   IB_SA_MCMEMBER_REC_MTU, dst->mtu_selector,
297 			   src->mtu, dst->mtu))
298 		return -EINVAL;
299 	if (comp_mask & IB_SA_MCMEMBER_REC_TRAFFIC_CLASS &&
300 	    src->traffic_class != dst->traffic_class)
301 		return -EINVAL;
302 	if (comp_mask & IB_SA_MCMEMBER_REC_PKEY && src->pkey != dst->pkey)
303 		return -EINVAL;
304 	if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_RATE_SELECTOR,
305 			   IB_SA_MCMEMBER_REC_RATE, dst->rate_selector,
306 			   src->rate, dst->rate))
307 		return -EINVAL;
308 	if (check_selector(comp_mask,
309 			   IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME_SELECTOR,
310 			   IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME,
311 			   dst->packet_life_time_selector,
312 			   src->packet_life_time, dst->packet_life_time))
313 		return -EINVAL;
314 	if (comp_mask & IB_SA_MCMEMBER_REC_SL && src->sl != dst->sl)
315 		return -EINVAL;
316 	if (comp_mask & IB_SA_MCMEMBER_REC_FLOW_LABEL &&
317 	    src->flow_label != dst->flow_label)
318 		return -EINVAL;
319 	if (comp_mask & IB_SA_MCMEMBER_REC_HOP_LIMIT &&
320 	    src->hop_limit != dst->hop_limit)
321 		return -EINVAL;
322 	if (comp_mask & IB_SA_MCMEMBER_REC_SCOPE && src->scope != dst->scope)
323 		return -EINVAL;
324 
325 	/* join_state checked separately, proxy_join ignored */
326 
327 	return 0;
328 }
329 
330 static int send_join(struct mcast_group *group, struct mcast_member *member)
331 {
332 	struct mcast_port *port = group->port;
333 	int ret;
334 
335 	group->last_join = member;
336 	ret = ib_sa_mcmember_rec_query(&sa_client, port->dev->device,
337 				       port->port_num, IB_MGMT_METHOD_SET,
338 				       &member->multicast.rec,
339 				       member->multicast.comp_mask,
340 				       3000, GFP_KERNEL, join_handler, group,
341 				       &group->query);
342 	if (ret >= 0) {
343 		group->query_id = ret;
344 		ret = 0;
345 	}
346 	return ret;
347 }
348 
349 static int send_leave(struct mcast_group *group, u8 leave_state)
350 {
351 	struct mcast_port *port = group->port;
352 	struct ib_sa_mcmember_rec rec;
353 	int ret;
354 
355 	rec = group->rec;
356 	rec.join_state = leave_state;
357 	group->leave_state = leave_state;
358 
359 	ret = ib_sa_mcmember_rec_query(&sa_client, port->dev->device,
360 				       port->port_num, IB_SA_METHOD_DELETE, &rec,
361 				       IB_SA_MCMEMBER_REC_MGID     |
362 				       IB_SA_MCMEMBER_REC_PORT_GID |
363 				       IB_SA_MCMEMBER_REC_JOIN_STATE,
364 				       3000, GFP_KERNEL, leave_handler,
365 				       group, &group->query);
366 	if (ret >= 0) {
367 		group->query_id = ret;
368 		ret = 0;
369 	}
370 	return ret;
371 }
372 
373 static void join_group(struct mcast_group *group, struct mcast_member *member,
374 		       u8 join_state)
375 {
376 	member->state = MCAST_MEMBER;
377 	adjust_membership(group, join_state, 1);
378 	group->rec.join_state |= join_state;
379 	member->multicast.rec = group->rec;
380 	member->multicast.rec.join_state = join_state;
381 	list_move(&member->list, &group->active_list);
382 }
383 
384 static int fail_join(struct mcast_group *group, struct mcast_member *member,
385 		     int status)
386 {
387 	spin_lock_irq(&group->lock);
388 	list_del_init(&member->list);
389 	spin_unlock_irq(&group->lock);
390 	return member->multicast.callback(status, &member->multicast);
391 }
392 
393 static void process_group_error(struct mcast_group *group)
394 {
395 	struct mcast_member *member;
396 	int ret = 0;
397 	u16 pkey_index;
398 
399 	if (group->state == MCAST_PKEY_EVENT)
400 		ret = ib_find_pkey(group->port->dev->device,
401 				   group->port->port_num,
402 				   be16_to_cpu(group->rec.pkey), &pkey_index);
403 
404 	spin_lock_irq(&group->lock);
405 	if (group->state == MCAST_PKEY_EVENT && !ret &&
406 	    group->pkey_index == pkey_index)
407 		goto out;
408 
409 	while (!list_empty(&group->active_list)) {
410 		member = list_entry(group->active_list.next,
411 				    struct mcast_member, list);
412 		atomic_inc(&member->refcount);
413 		list_del_init(&member->list);
414 		adjust_membership(group, member->multicast.rec.join_state, -1);
415 		member->state = MCAST_ERROR;
416 		spin_unlock_irq(&group->lock);
417 
418 		ret = member->multicast.callback(-ENETRESET,
419 						 &member->multicast);
420 		deref_member(member);
421 		if (ret)
422 			ib_sa_free_multicast(&member->multicast);
423 		spin_lock_irq(&group->lock);
424 	}
425 
426 	group->rec.join_state = 0;
427 out:
428 	group->state = MCAST_BUSY;
429 	spin_unlock_irq(&group->lock);
430 }
431 
432 static void mcast_work_handler(struct work_struct *work)
433 {
434 	struct mcast_group *group;
435 	struct mcast_member *member;
436 	struct ib_sa_multicast *multicast;
437 	int status, ret;
438 	u8 join_state;
439 
440 	group = container_of(work, typeof(*group), work);
441 retest:
442 	spin_lock_irq(&group->lock);
443 	while (!list_empty(&group->pending_list) ||
444 	       (group->state != MCAST_BUSY)) {
445 
446 		if (group->state != MCAST_BUSY) {
447 			spin_unlock_irq(&group->lock);
448 			process_group_error(group);
449 			goto retest;
450 		}
451 
452 		member = list_entry(group->pending_list.next,
453 				    struct mcast_member, list);
454 		multicast = &member->multicast;
455 		join_state = multicast->rec.join_state;
456 		atomic_inc(&member->refcount);
457 
458 		if (join_state == (group->rec.join_state & join_state)) {
459 			status = cmp_rec(&group->rec, &multicast->rec,
460 					 multicast->comp_mask);
461 			if (!status)
462 				join_group(group, member, join_state);
463 			else
464 				list_del_init(&member->list);
465 			spin_unlock_irq(&group->lock);
466 			ret = multicast->callback(status, multicast);
467 		} else {
468 			spin_unlock_irq(&group->lock);
469 			status = send_join(group, member);
470 			if (!status) {
471 				deref_member(member);
472 				return;
473 			}
474 			ret = fail_join(group, member, status);
475 		}
476 
477 		deref_member(member);
478 		if (ret)
479 			ib_sa_free_multicast(&member->multicast);
480 		spin_lock_irq(&group->lock);
481 	}
482 
483 	join_state = get_leave_state(group);
484 	if (join_state) {
485 		group->rec.join_state &= ~join_state;
486 		spin_unlock_irq(&group->lock);
487 		if (send_leave(group, join_state))
488 			goto retest;
489 	} else {
490 		group->state = MCAST_IDLE;
491 		spin_unlock_irq(&group->lock);
492 		release_group(group);
493 	}
494 }
495 
496 /*
497  * Fail a join request if it is still active - at the head of the pending queue.
498  */
499 static void process_join_error(struct mcast_group *group, int status)
500 {
501 	struct mcast_member *member;
502 	int ret;
503 
504 	spin_lock_irq(&group->lock);
505 	member = list_entry(group->pending_list.next,
506 			    struct mcast_member, list);
507 	if (group->last_join == member) {
508 		atomic_inc(&member->refcount);
509 		list_del_init(&member->list);
510 		spin_unlock_irq(&group->lock);
511 		ret = member->multicast.callback(status, &member->multicast);
512 		deref_member(member);
513 		if (ret)
514 			ib_sa_free_multicast(&member->multicast);
515 	} else
516 		spin_unlock_irq(&group->lock);
517 }
518 
519 static void join_handler(int status, struct ib_sa_mcmember_rec *rec,
520 			 void *context)
521 {
522 	struct mcast_group *group = context;
523 	u16 pkey_index = MCAST_INVALID_PKEY_INDEX;
524 
525 	if (status)
526 		process_join_error(group, status);
527 	else {
528 		int mgids_changed, is_mgid0;
529 		ib_find_pkey(group->port->dev->device, group->port->port_num,
530 			     be16_to_cpu(rec->pkey), &pkey_index);
531 
532 		spin_lock_irq(&group->port->lock);
533 		if (group->state == MCAST_BUSY &&
534 		    group->pkey_index == MCAST_INVALID_PKEY_INDEX)
535 			group->pkey_index = pkey_index;
536 		mgids_changed = memcmp(&rec->mgid, &group->rec.mgid,
537 				       sizeof(group->rec.mgid));
538 		group->rec = *rec;
539 		if (mgids_changed) {
540 			rb_erase(&group->node, &group->port->table);
541 			is_mgid0 = !memcmp(&mgid0, &group->rec.mgid,
542 					   sizeof(mgid0));
543 			mcast_insert(group->port, group, is_mgid0);
544 		}
545 		spin_unlock_irq(&group->port->lock);
546 	}
547 	mcast_work_handler(&group->work);
548 }
549 
550 static void leave_handler(int status, struct ib_sa_mcmember_rec *rec,
551 			  void *context)
552 {
553 	struct mcast_group *group = context;
554 
555 	if (status && group->retries > 0 &&
556 	    !send_leave(group, group->leave_state))
557 		group->retries--;
558 	else
559 		mcast_work_handler(&group->work);
560 }
561 
562 static struct mcast_group *acquire_group(struct mcast_port *port,
563 					 union ib_gid *mgid, gfp_t gfp_mask)
564 {
565 	struct mcast_group *group, *cur_group;
566 	unsigned long flags;
567 	int is_mgid0;
568 
569 	is_mgid0 = !memcmp(&mgid0, mgid, sizeof mgid0);
570 	if (!is_mgid0) {
571 		spin_lock_irqsave(&port->lock, flags);
572 		group = mcast_find(port, mgid);
573 		if (group)
574 			goto found;
575 		spin_unlock_irqrestore(&port->lock, flags);
576 	}
577 
578 	group = kzalloc(sizeof *group, gfp_mask);
579 	if (!group)
580 		return NULL;
581 
582 	group->retries = 3;
583 	group->port = port;
584 	group->rec.mgid = *mgid;
585 	group->pkey_index = MCAST_INVALID_PKEY_INDEX;
586 	INIT_LIST_HEAD(&group->pending_list);
587 	INIT_LIST_HEAD(&group->active_list);
588 	INIT_WORK(&group->work, mcast_work_handler);
589 	spin_lock_init(&group->lock);
590 
591 	spin_lock_irqsave(&port->lock, flags);
592 	cur_group = mcast_insert(port, group, is_mgid0);
593 	if (cur_group) {
594 		kfree(group);
595 		group = cur_group;
596 	} else
597 		atomic_inc(&port->refcount);
598 found:
599 	atomic_inc(&group->refcount);
600 	spin_unlock_irqrestore(&port->lock, flags);
601 	return group;
602 }
603 
604 /*
605  * We serialize all join requests to a single group to make our lives much
606  * easier.  Otherwise, two users could try to join the same group
607  * simultaneously, with different configurations, one could leave while the
608  * join is in progress, etc., which makes locking around error recovery
609  * difficult.
610  */
611 struct ib_sa_multicast *
612 ib_sa_join_multicast(struct ib_sa_client *client,
613 		     struct ib_device *device, u8 port_num,
614 		     struct ib_sa_mcmember_rec *rec,
615 		     ib_sa_comp_mask comp_mask, gfp_t gfp_mask,
616 		     int (*callback)(int status,
617 				     struct ib_sa_multicast *multicast),
618 		     void *context)
619 {
620 	struct mcast_device *dev;
621 	struct mcast_member *member;
622 	struct ib_sa_multicast *multicast;
623 	int ret;
624 
625 	dev = ib_get_client_data(device, &mcast_client);
626 	if (!dev)
627 		return ERR_PTR(-ENODEV);
628 
629 	member = kmalloc(sizeof *member, gfp_mask);
630 	if (!member)
631 		return ERR_PTR(-ENOMEM);
632 
633 	ib_sa_client_get(client);
634 	member->client = client;
635 	member->multicast.rec = *rec;
636 	member->multicast.comp_mask = comp_mask;
637 	member->multicast.callback = callback;
638 	member->multicast.context = context;
639 	init_completion(&member->comp);
640 	atomic_set(&member->refcount, 1);
641 	member->state = MCAST_JOINING;
642 
643 	member->group = acquire_group(&dev->port[port_num - dev->start_port],
644 				      &rec->mgid, gfp_mask);
645 	if (!member->group) {
646 		ret = -ENOMEM;
647 		goto err;
648 	}
649 
650 	/*
651 	 * The user will get the multicast structure in their callback.  They
652 	 * could then free the multicast structure before we can return from
653 	 * this routine.  So we save the pointer to return before queuing
654 	 * any callback.
655 	 */
656 	multicast = &member->multicast;
657 	queue_join(member);
658 	return multicast;
659 
660 err:
661 	ib_sa_client_put(client);
662 	kfree(member);
663 	return ERR_PTR(ret);
664 }
665 EXPORT_SYMBOL(ib_sa_join_multicast);
666 
667 void ib_sa_free_multicast(struct ib_sa_multicast *multicast)
668 {
669 	struct mcast_member *member;
670 	struct mcast_group *group;
671 
672 	member = container_of(multicast, struct mcast_member, multicast);
673 	group = member->group;
674 
675 	spin_lock_irq(&group->lock);
676 	if (member->state == MCAST_MEMBER)
677 		adjust_membership(group, multicast->rec.join_state, -1);
678 
679 	list_del_init(&member->list);
680 
681 	if (group->state == MCAST_IDLE) {
682 		group->state = MCAST_BUSY;
683 		spin_unlock_irq(&group->lock);
684 		/* Continue to hold reference on group until callback */
685 		queue_work(mcast_wq, &group->work);
686 	} else {
687 		spin_unlock_irq(&group->lock);
688 		release_group(group);
689 	}
690 
691 	deref_member(member);
692 	wait_for_completion(&member->comp);
693 	ib_sa_client_put(member->client);
694 	kfree(member);
695 }
696 EXPORT_SYMBOL(ib_sa_free_multicast);
697 
698 int ib_sa_get_mcmember_rec(struct ib_device *device, u8 port_num,
699 			   union ib_gid *mgid, struct ib_sa_mcmember_rec *rec)
700 {
701 	struct mcast_device *dev;
702 	struct mcast_port *port;
703 	struct mcast_group *group;
704 	unsigned long flags;
705 	int ret = 0;
706 
707 	dev = ib_get_client_data(device, &mcast_client);
708 	if (!dev)
709 		return -ENODEV;
710 
711 	port = &dev->port[port_num - dev->start_port];
712 	spin_lock_irqsave(&port->lock, flags);
713 	group = mcast_find(port, mgid);
714 	if (group)
715 		*rec = group->rec;
716 	else
717 		ret = -EADDRNOTAVAIL;
718 	spin_unlock_irqrestore(&port->lock, flags);
719 
720 	return ret;
721 }
722 EXPORT_SYMBOL(ib_sa_get_mcmember_rec);
723 
724 int ib_init_ah_from_mcmember(struct ib_device *device, u8 port_num,
725 			     struct ib_sa_mcmember_rec *rec,
726 			     struct ib_ah_attr *ah_attr)
727 {
728 	int ret;
729 	u16 gid_index;
730 	u8 p;
731 
732 	ret = ib_find_cached_gid(device, &rec->port_gid, &p, &gid_index);
733 	if (ret)
734 		return ret;
735 
736 	memset(ah_attr, 0, sizeof *ah_attr);
737 	ah_attr->dlid = be16_to_cpu(rec->mlid);
738 	ah_attr->sl = rec->sl;
739 	ah_attr->port_num = port_num;
740 	ah_attr->static_rate = rec->rate;
741 
742 	ah_attr->ah_flags = IB_AH_GRH;
743 	ah_attr->grh.dgid = rec->mgid;
744 
745 	ah_attr->grh.sgid_index = (u8) gid_index;
746 	ah_attr->grh.flow_label = be32_to_cpu(rec->flow_label);
747 	ah_attr->grh.hop_limit = rec->hop_limit;
748 	ah_attr->grh.traffic_class = rec->traffic_class;
749 
750 	return 0;
751 }
752 EXPORT_SYMBOL(ib_init_ah_from_mcmember);
753 
754 static void mcast_groups_event(struct mcast_port *port,
755 			       enum mcast_group_state state)
756 {
757 	struct mcast_group *group;
758 	struct rb_node *node;
759 	unsigned long flags;
760 
761 	spin_lock_irqsave(&port->lock, flags);
762 	for (node = rb_first(&port->table); node; node = rb_next(node)) {
763 		group = rb_entry(node, struct mcast_group, node);
764 		spin_lock(&group->lock);
765 		if (group->state == MCAST_IDLE) {
766 			atomic_inc(&group->refcount);
767 			queue_work(mcast_wq, &group->work);
768 		}
769 		if (group->state != MCAST_GROUP_ERROR)
770 			group->state = state;
771 		spin_unlock(&group->lock);
772 	}
773 	spin_unlock_irqrestore(&port->lock, flags);
774 }
775 
776 static void mcast_event_handler(struct ib_event_handler *handler,
777 				struct ib_event *event)
778 {
779 	struct mcast_device *dev;
780 	int index;
781 
782 	dev = container_of(handler, struct mcast_device, event_handler);
783 	if (rdma_port_get_link_layer(dev->device, event->element.port_num) !=
784 	    IB_LINK_LAYER_INFINIBAND)
785 		return;
786 
787 	index = event->element.port_num - dev->start_port;
788 
789 	switch (event->event) {
790 	case IB_EVENT_PORT_ERR:
791 	case IB_EVENT_LID_CHANGE:
792 	case IB_EVENT_SM_CHANGE:
793 	case IB_EVENT_CLIENT_REREGISTER:
794 		mcast_groups_event(&dev->port[index], MCAST_GROUP_ERROR);
795 		break;
796 	case IB_EVENT_PKEY_CHANGE:
797 		mcast_groups_event(&dev->port[index], MCAST_PKEY_EVENT);
798 		break;
799 	default:
800 		break;
801 	}
802 }
803 
804 static void mcast_add_one(struct ib_device *device)
805 {
806 	struct mcast_device *dev;
807 	struct mcast_port *port;
808 	int i;
809 	int count = 0;
810 
811 	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
812 		return;
813 
814 	dev = kmalloc(sizeof *dev + device->phys_port_cnt * sizeof *port,
815 		      GFP_KERNEL);
816 	if (!dev)
817 		return;
818 
819 	if (device->node_type == RDMA_NODE_IB_SWITCH)
820 		dev->start_port = dev->end_port = 0;
821 	else {
822 		dev->start_port = 1;
823 		dev->end_port = device->phys_port_cnt;
824 	}
825 
826 	for (i = 0; i <= dev->end_port - dev->start_port; i++) {
827 		if (rdma_port_get_link_layer(device, dev->start_port + i) !=
828 		    IB_LINK_LAYER_INFINIBAND)
829 			continue;
830 		port = &dev->port[i];
831 		port->dev = dev;
832 		port->port_num = dev->start_port + i;
833 		spin_lock_init(&port->lock);
834 		port->table = RB_ROOT;
835 		init_completion(&port->comp);
836 		atomic_set(&port->refcount, 1);
837 		++count;
838 	}
839 
840 	if (!count) {
841 		kfree(dev);
842 		return;
843 	}
844 
845 	dev->device = device;
846 	ib_set_client_data(device, &mcast_client, dev);
847 
848 	INIT_IB_EVENT_HANDLER(&dev->event_handler, device, mcast_event_handler);
849 	ib_register_event_handler(&dev->event_handler);
850 }
851 
852 static void mcast_remove_one(struct ib_device *device)
853 {
854 	struct mcast_device *dev;
855 	struct mcast_port *port;
856 	int i;
857 
858 	dev = ib_get_client_data(device, &mcast_client);
859 	if (!dev)
860 		return;
861 
862 	ib_unregister_event_handler(&dev->event_handler);
863 	flush_workqueue(mcast_wq);
864 
865 	for (i = 0; i <= dev->end_port - dev->start_port; i++) {
866 		if (rdma_port_get_link_layer(device, dev->start_port + i) ==
867 		    IB_LINK_LAYER_INFINIBAND) {
868 			port = &dev->port[i];
869 			deref_port(port);
870 			wait_for_completion(&port->comp);
871 		}
872 	}
873 
874 	kfree(dev);
875 }
876 
877 int mcast_init(void)
878 {
879 	int ret;
880 
881 	mcast_wq = create_singlethread_workqueue("ib_mcast");
882 	if (!mcast_wq)
883 		return -ENOMEM;
884 
885 	ib_sa_register_client(&sa_client);
886 
887 	ret = ib_register_client(&mcast_client);
888 	if (ret)
889 		goto err;
890 	return 0;
891 
892 err:
893 	ib_sa_unregister_client(&sa_client);
894 	destroy_workqueue(mcast_wq);
895 	return ret;
896 }
897 
898 void mcast_cleanup(void)
899 {
900 	ib_unregister_client(&mcast_client);
901 	ib_sa_unregister_client(&sa_client);
902 	destroy_workqueue(mcast_wq);
903 }
904