2 * Copyright (c) 2006 Intel Corporation. All rights reserved.
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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.
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
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>
42 #include <rdma/ib_cache.h>
45 static void mcast_add_one(struct ib_device *device);
46 static void mcast_remove_one(struct ib_device *device);
48 static struct ib_client mcast_client = {
49 .name = "ib_multicast",
51 .remove = mcast_remove_one
54 static struct ib_sa_client sa_client;
55 static struct workqueue_struct *mcast_wq;
56 static union ib_gid mgid0;
61 struct mcast_device *dev;
65 struct completion comp;
70 struct ib_device *device;
71 struct ib_event_handler event_handler;
74 struct mcast_port port[0];
83 enum mcast_group_state {
91 MCAST_INVALID_PKEY_INDEX = 0xFFFF
97 struct ib_sa_mcmember_rec rec;
99 struct mcast_port *port;
101 struct work_struct work;
102 struct list_head pending_list;
103 struct list_head active_list;
104 struct mcast_member *last_join;
107 enum mcast_group_state state;
108 struct ib_sa_query *query;
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;
122 struct completion comp;
125 static void join_handler(int status, struct ib_sa_mcmember_rec *rec,
127 static void leave_handler(int status, struct ib_sa_mcmember_rec *rec,
130 static struct mcast_group *mcast_find(struct mcast_port *port,
133 struct rb_node *node = port->table.rb_node;
134 struct mcast_group *group;
138 group = rb_entry(node, struct mcast_group, node);
139 ret = memcmp(mgid->raw, group->rec.mgid.raw, sizeof *mgid);
144 node = node->rb_left;
146 node = node->rb_right;
151 static struct mcast_group *mcast_insert(struct mcast_port *port,
152 struct mcast_group *group,
153 int allow_duplicates)
155 struct rb_node **link = &port->table.rb_node;
156 struct rb_node *parent = NULL;
157 struct mcast_group *cur_group;
162 cur_group = rb_entry(parent, struct mcast_group, node);
164 ret = memcmp(group->rec.mgid.raw, cur_group->rec.mgid.raw,
165 sizeof group->rec.mgid);
167 link = &(*link)->rb_left;
169 link = &(*link)->rb_right;
170 else if (allow_duplicates)
171 link = &(*link)->rb_left;
175 rb_link_node(&group->node, parent, link);
176 rb_insert_color(&group->node, &port->table);
180 static void deref_port(struct mcast_port *port)
182 if (atomic_dec_and_test(&port->refcount))
183 complete(&port->comp);
186 static void release_group(struct mcast_group *group)
188 struct mcast_port *port = group->port;
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);
198 spin_unlock_irqrestore(&port->lock, flags);
201 static void deref_member(struct mcast_member *member)
203 if (atomic_dec_and_test(&member->refcount))
204 complete(&member->comp);
207 static void queue_join(struct mcast_member *member)
209 struct mcast_group *group = member->group;
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);
219 spin_unlock_irqrestore(&group->lock, flags);
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.
228 static void adjust_membership(struct mcast_group *group, u8 join_state, int inc)
232 for (i = 0; i < 3; i++, join_state >>= 1)
233 if (join_state & 0x1)
234 group->members[i] += inc;
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
243 static u8 get_leave_state(struct mcast_group *group)
248 for (i = 0; i < 3; i++)
249 if (!group->members[i])
250 leave_state |= (0x1 << i);
252 return leave_state & group->rec.join_state;
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)
262 if (!(comp_mask & selector_mask) || !(comp_mask & value_mask))
267 err = (src_value <= dst_value);
270 err = (src_value >= dst_value);
273 err = (src_value != dst_value);
283 static int cmp_rec(struct ib_sa_mcmember_rec *src,
284 struct ib_sa_mcmember_rec *dst, ib_sa_comp_mask comp_mask)
286 /* MGID must already match */
288 if (comp_mask & IB_SA_MCMEMBER_REC_PORT_GID &&
289 memcmp(&src->port_gid, &dst->port_gid, sizeof src->port_gid))
291 if (comp_mask & IB_SA_MCMEMBER_REC_QKEY && src->qkey != dst->qkey)
293 if (comp_mask & IB_SA_MCMEMBER_REC_MLID && src->mlid != dst->mlid)
295 if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_MTU_SELECTOR,
296 IB_SA_MCMEMBER_REC_MTU, dst->mtu_selector,
299 if (comp_mask & IB_SA_MCMEMBER_REC_TRAFFIC_CLASS &&
300 src->traffic_class != dst->traffic_class)
302 if (comp_mask & IB_SA_MCMEMBER_REC_PKEY && src->pkey != dst->pkey)
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))
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))
314 if (comp_mask & IB_SA_MCMEMBER_REC_SL && src->sl != dst->sl)
316 if (comp_mask & IB_SA_MCMEMBER_REC_FLOW_LABEL &&
317 src->flow_label != dst->flow_label)
319 if (comp_mask & IB_SA_MCMEMBER_REC_HOP_LIMIT &&
320 src->hop_limit != dst->hop_limit)
322 if (comp_mask & IB_SA_MCMEMBER_REC_SCOPE && src->scope != dst->scope)
325 /* join_state checked separately, proxy_join ignored */
330 static int send_join(struct mcast_group *group, struct mcast_member *member)
332 struct mcast_port *port = group->port;
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,
343 group->query_id = ret;
349 static int send_leave(struct mcast_group *group, u8 leave_state)
351 struct mcast_port *port = group->port;
352 struct ib_sa_mcmember_rec rec;
356 rec.join_state = leave_state;
357 group->leave_state = leave_state;
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);
367 group->query_id = ret;
373 static void join_group(struct mcast_group *group, struct mcast_member *member,
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);
384 static int fail_join(struct mcast_group *group, struct mcast_member *member,
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);
393 static void process_group_error(struct mcast_group *group)
395 struct mcast_member *member;
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);
404 spin_lock_irq(&group->lock);
405 if (group->state == MCAST_PKEY_EVENT && !ret &&
406 group->pkey_index == pkey_index)
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);
418 ret = member->multicast.callback(-ENETRESET,
420 deref_member(member);
422 ib_sa_free_multicast(&member->multicast);
423 spin_lock_irq(&group->lock);
426 group->rec.join_state = 0;
428 group->state = MCAST_BUSY;
429 spin_unlock_irq(&group->lock);
432 static void mcast_work_handler(struct work_struct *work)
434 struct mcast_group *group;
435 struct mcast_member *member;
436 struct ib_sa_multicast *multicast;
440 group = container_of(work, typeof(*group), work);
442 spin_lock_irq(&group->lock);
443 while (!list_empty(&group->pending_list) ||
444 (group->state != MCAST_BUSY)) {
446 if (group->state != MCAST_BUSY) {
447 spin_unlock_irq(&group->lock);
448 process_group_error(group);
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);
458 if (join_state == (group->rec.join_state & join_state)) {
459 status = cmp_rec(&group->rec, &multicast->rec,
460 multicast->comp_mask);
462 join_group(group, member, join_state);
464 list_del_init(&member->list);
465 spin_unlock_irq(&group->lock);
466 ret = multicast->callback(status, multicast);
468 spin_unlock_irq(&group->lock);
469 status = send_join(group, member);
471 deref_member(member);
474 ret = fail_join(group, member, status);
477 deref_member(member);
479 ib_sa_free_multicast(&member->multicast);
480 spin_lock_irq(&group->lock);
483 join_state = get_leave_state(group);
485 group->rec.join_state &= ~join_state;
486 spin_unlock_irq(&group->lock);
487 if (send_leave(group, join_state))
490 group->state = MCAST_IDLE;
491 spin_unlock_irq(&group->lock);
492 release_group(group);
497 * Fail a join request if it is still active - at the head of the pending queue.
499 static void process_join_error(struct mcast_group *group, int status)
501 struct mcast_member *member;
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);
514 ib_sa_free_multicast(&member->multicast);
516 spin_unlock_irq(&group->lock);
519 static void join_handler(int status, struct ib_sa_mcmember_rec *rec,
522 struct mcast_group *group = context;
523 u16 pkey_index = MCAST_INVALID_PKEY_INDEX;
526 process_join_error(group, status);
528 ib_find_pkey(group->port->dev->device, group->port->port_num,
529 be16_to_cpu(rec->pkey), &pkey_index);
531 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 if (!memcmp(&mgid0, &group->rec.mgid, sizeof mgid0)) {
537 rb_erase(&group->node, &group->port->table);
538 mcast_insert(group->port, group, 1);
540 spin_unlock_irq(&group->port->lock);
542 mcast_work_handler(&group->work);
545 static void leave_handler(int status, struct ib_sa_mcmember_rec *rec,
548 struct mcast_group *group = context;
550 if (status && group->retries > 0 &&
551 !send_leave(group, group->leave_state))
554 mcast_work_handler(&group->work);
557 static struct mcast_group *acquire_group(struct mcast_port *port,
558 union ib_gid *mgid, gfp_t gfp_mask)
560 struct mcast_group *group, *cur_group;
564 is_mgid0 = !memcmp(&mgid0, mgid, sizeof mgid0);
566 spin_lock_irqsave(&port->lock, flags);
567 group = mcast_find(port, mgid);
570 spin_unlock_irqrestore(&port->lock, flags);
573 group = kzalloc(sizeof *group, gfp_mask);
579 group->rec.mgid = *mgid;
580 group->pkey_index = MCAST_INVALID_PKEY_INDEX;
581 INIT_LIST_HEAD(&group->pending_list);
582 INIT_LIST_HEAD(&group->active_list);
583 INIT_WORK(&group->work, mcast_work_handler);
584 spin_lock_init(&group->lock);
586 spin_lock_irqsave(&port->lock, flags);
587 cur_group = mcast_insert(port, group, is_mgid0);
592 atomic_inc(&port->refcount);
594 atomic_inc(&group->refcount);
595 spin_unlock_irqrestore(&port->lock, flags);
600 * We serialize all join requests to a single group to make our lives much
601 * easier. Otherwise, two users could try to join the same group
602 * simultaneously, with different configurations, one could leave while the
603 * join is in progress, etc., which makes locking around error recovery
606 struct ib_sa_multicast *
607 ib_sa_join_multicast(struct ib_sa_client *client,
608 struct ib_device *device, u8 port_num,
609 struct ib_sa_mcmember_rec *rec,
610 ib_sa_comp_mask comp_mask, gfp_t gfp_mask,
611 int (*callback)(int status,
612 struct ib_sa_multicast *multicast),
615 struct mcast_device *dev;
616 struct mcast_member *member;
617 struct ib_sa_multicast *multicast;
620 dev = ib_get_client_data(device, &mcast_client);
622 return ERR_PTR(-ENODEV);
624 member = kmalloc(sizeof *member, gfp_mask);
626 return ERR_PTR(-ENOMEM);
628 ib_sa_client_get(client);
629 member->client = client;
630 member->multicast.rec = *rec;
631 member->multicast.comp_mask = comp_mask;
632 member->multicast.callback = callback;
633 member->multicast.context = context;
634 init_completion(&member->comp);
635 atomic_set(&member->refcount, 1);
636 member->state = MCAST_JOINING;
638 member->group = acquire_group(&dev->port[port_num - dev->start_port],
639 &rec->mgid, gfp_mask);
640 if (!member->group) {
646 * The user will get the multicast structure in their callback. They
647 * could then free the multicast structure before we can return from
648 * this routine. So we save the pointer to return before queuing
651 multicast = &member->multicast;
656 ib_sa_client_put(client);
660 EXPORT_SYMBOL(ib_sa_join_multicast);
662 void ib_sa_free_multicast(struct ib_sa_multicast *multicast)
664 struct mcast_member *member;
665 struct mcast_group *group;
667 member = container_of(multicast, struct mcast_member, multicast);
668 group = member->group;
670 spin_lock_irq(&group->lock);
671 if (member->state == MCAST_MEMBER)
672 adjust_membership(group, multicast->rec.join_state, -1);
674 list_del_init(&member->list);
676 if (group->state == MCAST_IDLE) {
677 group->state = MCAST_BUSY;
678 spin_unlock_irq(&group->lock);
679 /* Continue to hold reference on group until callback */
680 queue_work(mcast_wq, &group->work);
682 spin_unlock_irq(&group->lock);
683 release_group(group);
686 deref_member(member);
687 wait_for_completion(&member->comp);
688 ib_sa_client_put(member->client);
691 EXPORT_SYMBOL(ib_sa_free_multicast);
693 int ib_sa_get_mcmember_rec(struct ib_device *device, u8 port_num,
694 union ib_gid *mgid, struct ib_sa_mcmember_rec *rec)
696 struct mcast_device *dev;
697 struct mcast_port *port;
698 struct mcast_group *group;
702 dev = ib_get_client_data(device, &mcast_client);
706 port = &dev->port[port_num - dev->start_port];
707 spin_lock_irqsave(&port->lock, flags);
708 group = mcast_find(port, mgid);
712 ret = -EADDRNOTAVAIL;
713 spin_unlock_irqrestore(&port->lock, flags);
717 EXPORT_SYMBOL(ib_sa_get_mcmember_rec);
719 int ib_init_ah_from_mcmember(struct ib_device *device, u8 port_num,
720 struct ib_sa_mcmember_rec *rec,
721 struct ib_ah_attr *ah_attr)
727 ret = ib_find_cached_gid(device, &rec->port_gid, &p, &gid_index);
731 memset(ah_attr, 0, sizeof *ah_attr);
732 ah_attr->dlid = be16_to_cpu(rec->mlid);
733 ah_attr->sl = rec->sl;
734 ah_attr->port_num = port_num;
735 ah_attr->static_rate = rec->rate;
737 ah_attr->ah_flags = IB_AH_GRH;
738 ah_attr->grh.dgid = rec->mgid;
740 ah_attr->grh.sgid_index = (u8) gid_index;
741 ah_attr->grh.flow_label = be32_to_cpu(rec->flow_label);
742 ah_attr->grh.hop_limit = rec->hop_limit;
743 ah_attr->grh.traffic_class = rec->traffic_class;
747 EXPORT_SYMBOL(ib_init_ah_from_mcmember);
749 static void mcast_groups_event(struct mcast_port *port,
750 enum mcast_group_state state)
752 struct mcast_group *group;
753 struct rb_node *node;
756 spin_lock_irqsave(&port->lock, flags);
757 for (node = rb_first(&port->table); node; node = rb_next(node)) {
758 group = rb_entry(node, struct mcast_group, node);
759 spin_lock(&group->lock);
760 if (group->state == MCAST_IDLE) {
761 atomic_inc(&group->refcount);
762 queue_work(mcast_wq, &group->work);
764 if (group->state != MCAST_GROUP_ERROR)
765 group->state = state;
766 spin_unlock(&group->lock);
768 spin_unlock_irqrestore(&port->lock, flags);
771 static void mcast_event_handler(struct ib_event_handler *handler,
772 struct ib_event *event)
774 struct mcast_device *dev;
777 dev = container_of(handler, struct mcast_device, event_handler);
778 if (rdma_port_get_link_layer(dev->device, event->element.port_num) !=
779 IB_LINK_LAYER_INFINIBAND)
782 index = event->element.port_num - dev->start_port;
784 switch (event->event) {
785 case IB_EVENT_PORT_ERR:
786 case IB_EVENT_LID_CHANGE:
787 case IB_EVENT_SM_CHANGE:
788 case IB_EVENT_CLIENT_REREGISTER:
789 mcast_groups_event(&dev->port[index], MCAST_GROUP_ERROR);
791 case IB_EVENT_PKEY_CHANGE:
792 mcast_groups_event(&dev->port[index], MCAST_PKEY_EVENT);
799 static void mcast_add_one(struct ib_device *device)
801 struct mcast_device *dev;
802 struct mcast_port *port;
806 if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
809 dev = kmalloc(sizeof *dev + device->phys_port_cnt * sizeof *port,
814 if (device->node_type == RDMA_NODE_IB_SWITCH)
815 dev->start_port = dev->end_port = 0;
818 dev->end_port = device->phys_port_cnt;
821 for (i = 0; i <= dev->end_port - dev->start_port; i++) {
822 if (rdma_port_get_link_layer(device, dev->start_port + i) !=
823 IB_LINK_LAYER_INFINIBAND)
825 port = &dev->port[i];
827 port->port_num = dev->start_port + i;
828 spin_lock_init(&port->lock);
829 port->table = RB_ROOT;
830 init_completion(&port->comp);
831 atomic_set(&port->refcount, 1);
840 dev->device = device;
841 ib_set_client_data(device, &mcast_client, dev);
843 INIT_IB_EVENT_HANDLER(&dev->event_handler, device, mcast_event_handler);
844 ib_register_event_handler(&dev->event_handler);
847 static void mcast_remove_one(struct ib_device *device)
849 struct mcast_device *dev;
850 struct mcast_port *port;
853 dev = ib_get_client_data(device, &mcast_client);
857 ib_unregister_event_handler(&dev->event_handler);
858 flush_workqueue(mcast_wq);
860 for (i = 0; i <= dev->end_port - dev->start_port; i++) {
861 if (rdma_port_get_link_layer(device, dev->start_port + i) ==
862 IB_LINK_LAYER_INFINIBAND) {
863 port = &dev->port[i];
865 wait_for_completion(&port->comp);
876 mcast_wq = create_singlethread_workqueue("ib_mcast");
880 ib_sa_register_client(&sa_client);
882 ret = ib_register_client(&mcast_client);
888 ib_sa_unregister_client(&sa_client);
889 destroy_workqueue(mcast_wq);
893 void mcast_cleanup(void)
895 ib_unregister_client(&mcast_client);
896 ib_sa_unregister_client(&sa_client);
897 destroy_workqueue(mcast_wq);