IB/sa: Track multicast join/leave requests

The IB SA tracks multicast join/leave requests on a per port basis and
does not do any reference counting: if two users of the same port join
the same group, and one leaves that group, then the SA will remove the
port from the group even though there is one user who wants to stay a
member left.  Therefore, in order to support multiple users of the
same multicast group from the same port, we need to perform reference
counting locally.

To do this, add an multicast submodule to ib_sa to perform reference
counting of multicast join/leave operations.  Modify ib_ipoib (the
only in-kernel user of multicast) to use the new interface.

Signed-off-by: Roland Dreier <rolandd@cisco.com>

4 files changed, 919 insertions, 15 deletions

diff --git a/drivers/infiniband/core/Makefile b/drivers/infiniband/core/Makefile
index 50fb1cd447b..189e5d4b9b1 100644
--- a/drivers/infiniband/core/Makefile
+++ b/drivers/infiniband/core/Makefile
@@ -12,7 +12,7 @@ ib_core-y :=			packer.o ud_header.o verbs.o sysfs.o \
 
 ib_mad-y :=			mad.o smi.o agent.o mad_rmpp.o
 
-ib_sa-y :=			sa_query.o
+ib_sa-y :=			sa_query.o multicast.o
 
 ib_cm-y :=			cm.o
 
diff --git a/drivers/infiniband/core/multicast.c b/drivers/infiniband/core/multicast.c
new file mode 100644
index 00000000000..4a579b3a1c9
--- /dev/null
+++ b/drivers/infiniband/core/multicast.c
@@ -0,0 +1,837 @@
+/*
+ * Copyright (c) 2006 Intel Corporation.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/completion.h>
+#include <linux/dma-mapping.h>
+#include <linux/err.h>
+#include <linux/interrupt.h>
+#include <linux/pci.h>
+#include <linux/bitops.h>
+#include <linux/random.h>
+
+#include <rdma/ib_cache.h>
+#include "sa.h"
+
+static void mcast_add_one(struct ib_device *device);
+static void mcast_remove_one(struct ib_device *device);
+
+static struct ib_client mcast_client = {
+	.name   = "ib_multicast",
+	.add    = mcast_add_one,
+	.remove = mcast_remove_one
+};
+
+static struct ib_sa_client	sa_client;
+static struct workqueue_struct	*mcast_wq;
+static union ib_gid mgid0;
+
+struct mcast_device;
+
+struct mcast_port {
+	struct mcast_device	*dev;
+	spinlock_t		lock;
+	struct rb_root		table;
+	atomic_t		refcount;
+	struct completion	comp;
+	u8			port_num;
+};
+
+struct mcast_device {
+	struct ib_device	*device;
+	struct ib_event_handler	event_handler;
+	int			start_port;
+	int			end_port;
+	struct mcast_port	port[0];
+};
+
+enum mcast_state {
+	MCAST_IDLE,
+	MCAST_JOINING,
+	MCAST_MEMBER,
+	MCAST_BUSY,
+	MCAST_ERROR
+};
+
+struct mcast_member;
+
+struct mcast_group {
+	struct ib_sa_mcmember_rec rec;
+	struct rb_node		node;
+	struct mcast_port	*port;
+	spinlock_t		lock;
+	struct work_struct	work;
+	struct list_head	pending_list;
+	struct list_head	active_list;
+	struct mcast_member	*last_join;
+	int			members[3];
+	atomic_t		refcount;
+	enum mcast_state	state;
+	struct ib_sa_query	*query;
+	int			query_id;
+};
+
+struct mcast_member {
+	struct ib_sa_multicast	multicast;
+	struct ib_sa_client	*client;
+	struct mcast_group	*group;
+	struct list_head	list;
+	enum mcast_state	state;
+	atomic_t		refcount;
+	struct completion	comp;
+};
+
+static void join_handler(int status, struct ib_sa_mcmember_rec *rec,
+			 void *context);
+static void leave_handler(int status, struct ib_sa_mcmember_rec *rec,
+			  void *context);
+
+static struct mcast_group *mcast_find(struct mcast_port *port,
+				      union ib_gid *mgid)
+{
+	struct rb_node *node = port->table.rb_node;
+	struct mcast_group *group;
+	int ret;
+
+	while (node) {
+		group = rb_entry(node, struct mcast_group, node);
+		ret = memcmp(mgid->raw, group->rec.mgid.raw, sizeof *mgid);
+		if (!ret)
+			return group;
+
+		if (ret < 0)
+			node = node->rb_left;
+		else
+			node = node->rb_right;
+	}
+	return NULL;
+}
+
+static struct mcast_group *mcast_insert(struct mcast_port *port,
+					struct mcast_group *group,
+					int allow_duplicates)
+{
+	struct rb_node **link = &port->table.rb_node;
+	struct rb_node *parent = NULL;
+	struct mcast_group *cur_group;
+	int ret;
+
+	while (*link) {
+		parent = *link;
+		cur_group = rb_entry(parent, struct mcast_group, node);
+
+		ret = memcmp(group->rec.mgid.raw, cur_group->rec.mgid.raw,
+			     sizeof group->rec.mgid);
+		if (ret < 0)
+			link = &(*link)->rb_left;
+		else if (ret > 0)
+			link = &(*link)->rb_right;
+		else if (allow_duplicates)
+			link = &(*link)->rb_left;
+		else
+			return cur_group;
+	}
+	rb_link_node(&group->node, parent, link);
+	rb_insert_color(&group->node, &port->table);
+	return NULL;
+}
+
+static void deref_port(struct mcast_port *port)
+{
+	if (atomic_dec_and_test(&port->refcount))
+		complete(&port->comp);
+}
+
+static void release_group(struct mcast_group *group)
+{
+	struct mcast_port *port = group->port;
+	unsigned long flags;
+
+	spin_lock_irqsave(&port->lock, flags);
+	if (atomic_dec_and_test(&group->refcount)) {
+		rb_erase(&group->node, &port->table);
+		spin_unlock_irqrestore(&port->lock, flags);
+		kfree(group);
+		deref_port(port);
+	} else
+		spin_unlock_irqrestore(&port->lock, flags);
+}
+
+static void deref_member(struct mcast_member *member)
+{
+	if (atomic_dec_and_test(&member->refcount))
+		complete(&member->comp);
+}
+
+static void queue_join(struct mcast_member *member)
+{
+	struct mcast_group *group = member->group;
+	unsigned long flags;
+
+	spin_lock_irqsave(&group->lock, flags);
+	list_add(&member->list, &group->pending_list);
+	if (group->state == MCAST_IDLE) {
+		group->state = MCAST_BUSY;
+		atomic_inc(&group->refcount);
+		queue_work(mcast_wq, &group->work);
+	}
+	spin_unlock_irqrestore(&group->lock, flags);
+}
+
+/*
+ * A multicast group has three types of members: full member, non member, and
+ * send only member.  We need to keep track of the number of members of each
+ * type based on their join state.  Adjust the number of members the belong to
+ * the specified join states.
+ */
+static void adjust_membership(struct mcast_group *group, u8 join_state, int inc)
+{
+	int i;
+
+	for (i = 0; i < 3; i++, join_state >>= 1)
+		if (join_state & 0x1)
+			group->members[i] += inc;
+}
+
+/*
+ * If a multicast group has zero members left for a particular join state, but
+ * the group is still a member with the SA, we need to leave that join state.
+ * Determine which join states we still belong to, but that do not have any
+ * active members.
+ */
+static u8 get_leave_state(struct mcast_group *group)
+{
+	u8 leave_state = 0;
+	int i;
+
+	for (i = 0; i < 3; i++)
+		if (!group->members[i])
+			leave_state |= (0x1 << i);
+
+	return leave_state & group->rec.join_state;
+}
+
+static int check_selector(ib_sa_comp_mask comp_mask,
+			  ib_sa_comp_mask selector_mask,
+			  ib_sa_comp_mask value_mask,
+			  u8 selector, u8 src_value, u8 dst_value)
+{
+	int err;
+
+	if (!(comp_mask & selector_mask) || !(comp_mask & value_mask))
+		return 0;
+
+	switch (selector) {
+	case IB_SA_GT:
+		err = (src_value <= dst_value);
+		break;
+	case IB_SA_LT:
+		err = (src_value >= dst_value);
+		break;
+	case IB_SA_EQ:
+		err = (src_value != dst_value);
+		break;
+	default:
+		err = 0;
+		break;
+	}
+
+	return err;
+}
+
+static int cmp_rec(struct ib_sa_mcmember_rec *src,
+		   struct ib_sa_mcmember_rec *dst, ib_sa_comp_mask comp_mask)
+{
+	/* MGID must already match */
+
+	if (comp_mask & IB_SA_MCMEMBER_REC_PORT_GID &&
+	    memcmp(&src->port_gid, &dst->port_gid, sizeof src->port_gid))
+		return -EINVAL;
+	if (comp_mask & IB_SA_MCMEMBER_REC_QKEY && src->qkey != dst->qkey)
+		return -EINVAL;
+	if (comp_mask & IB_SA_MCMEMBER_REC_MLID && src->mlid != dst->mlid)
+		return -EINVAL;
+	if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_MTU_SELECTOR,
+			   IB_SA_MCMEMBER_REC_MTU, dst->mtu_selector,
+			   src->mtu, dst->mtu))
+		return -EINVAL;
+	if (comp_mask & IB_SA_MCMEMBER_REC_TRAFFIC_CLASS &&
+	    src->traffic_class != dst->traffic_class)
+		return -EINVAL;
+	if (comp_mask & IB_SA_MCMEMBER_REC_PKEY && src->pkey != dst->pkey)
+		return -EINVAL;
+	if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_RATE_SELECTOR,
+			   IB_SA_MCMEMBER_REC_RATE, dst->rate_selector,
+			   src->rate, dst->rate))
+		return -EINVAL;
+	if (check_selector(comp_mask,
+			   IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME_SELECTOR,
+			   IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME,
+			   dst->packet_life_time_selector,
+			   src->packet_life_time, dst->packet_life_time))
+		return -EINVAL;
+	if (comp_mask & IB_SA_MCMEMBER_REC_SL && src->sl != dst->sl)
+		return -EINVAL;
+	if (comp_mask & IB_SA_MCMEMBER_REC_FLOW_LABEL &&
+	    src->flow_label != dst->flow_label)
+		return -EINVAL;
+	if (comp_mask & IB_SA_MCMEMBER_REC_HOP_LIMIT &&
+	    src->hop_limit != dst->hop_limit)
+		return -EINVAL;
+	if (comp_mask & IB_SA_MCMEMBER_REC_SCOPE && src->scope != dst->scope)
+		return -EINVAL;
+
+	/* join_state checked separately, proxy_join ignored */
+
+	return 0;
+}
+
+static int send_join(struct mcast_group *group, struct mcast_member *member)
+{
+	struct mcast_port *port = group->port;
+	int ret;
+
+	group->last_join = member;
+	ret = ib_sa_mcmember_rec_query(&sa_client, port->dev->device,
+				       port->port_num, IB_MGMT_METHOD_SET,
+				       &member->multicast.rec,
+				       member->multicast.comp_mask,
+				       3000, GFP_KERNEL, join_handler, group,
+				       &group->query);
+	if (ret >= 0) {
+		group->query_id = ret;
+		ret = 0;
+	}
+	return ret;
+}
+
+static int send_leave(struct mcast_group *group, u8 leave_state)
+{
+	struct mcast_port *port = group->port;
+	struct ib_sa_mcmember_rec rec;
+	int ret;
+
+	rec = group->rec;
+	rec.join_state = leave_state;
+
+	ret = ib_sa_mcmember_rec_query(&sa_client, port->dev->device,
+				       port->port_num, IB_SA_METHOD_DELETE, &rec,
+				       IB_SA_MCMEMBER_REC_MGID     |
+				       IB_SA_MCMEMBER_REC_PORT_GID |
+				       IB_SA_MCMEMBER_REC_JOIN_STATE,
+				       3000, GFP_KERNEL, leave_handler,
+				       group, &group->query);
+	if (ret >= 0) {
+		group->query_id = ret;
+		ret = 0;
+	}
+	return ret;
+}
+
+static void join_group(struct mcast_group *group, struct mcast_member *member,
+		       u8 join_state)
+{
+	member->state = MCAST_MEMBER;
+	adjust_membership(group, join_state, 1);
+	group->rec.join_state |= join_state;
+	member->multicast.rec = group->rec;
+	member->multicast.rec.join_state = join_state;
+	list_move(&member->list, &group->active_list);
+}
+
+static int fail_join(struct mcast_group *group, struct mcast_member *member,
+		     int status)
+{
+	spin_lock_irq(&group->lock);
+	list_del_init(&member->list);
+	spin_unlock_irq(&group->lock);
+	return member->multicast.callback(status, &member->multicast);
+}
+
+static void process_group_error(struct mcast_group *group)
+{
+	struct mcast_member *member;
+	int ret;
+
+	spin_lock_irq(&group->lock);
+	while (!list_empty(&group->active_list)) {
+		member = list_entry(group->active_list.next,
+				    struct mcast_member, list);
+		atomic_inc(&member->refcount);
+		list_del_init(&member->list);
+		adjust_membership(group, member->multicast.rec.join_state, -1);
+		member->state = MCAST_ERROR;
+		spin_unlock_irq(&group->lock);
+
+		ret = member->multicast.callback(-ENETRESET,
+						 &member->multicast);
+		deref_member(member);
+		if (ret)
+			ib_sa_free_multicast(&member->multicast);
+		spin_lock_irq(&group->lock);
+	}
+
+	group->rec.join_state = 0;
+	group->state = MCAST_BUSY;
+	spin_unlock_irq(&group->lock);
+}
+
+static void mcast_work_handler(struct work_struct *work)
+{
+	struct mcast_group *group;
+	struct mcast_member *member;
+	struct ib_sa_multicast *multicast;
+	int status, ret;
+	u8 join_state;
+
+	group = container_of(work, typeof(*group), work);
+retest:
+	spin_lock_irq(&group->lock);
+	while (!list_empty(&group->pending_list) ||
+	       (group->state == MCAST_ERROR)) {
+
+		if (group->state == MCAST_ERROR) {
+			spin_unlock_irq(&group->lock);
+			process_group_error(group);
+			goto retest;
+		}
+
+		member = list_entry(group->pending_list.next,
+				    struct mcast_member, list);
+		multicast = &member->multicast;
+		join_state = multicast->rec.join_state;
+		atomic_inc(&member->refcount);
+
+		if (join_state == (group->rec.join_state & join_state)) {
+			status = cmp_rec(&group->rec, &multicast->rec,
+					 multicast->comp_mask);
+			if (!status)
+				join_group(group, member, join_state);
+			else
+				list_del_init(&member->list);
+			spin_unlock_irq(&group->lock);
+			ret = multicast->callback(status, multicast);
+		} else {
+			spin_unlock_irq(&group->lock);
+			status = send_join(group, member);
+			if (!status) {
+				deref_member(member);
+				return;
+			}
+			ret = fail_join(group, member, status);
+		}
+
+		deref_member(member);
+		if (ret)
+			ib_sa_free_multicast(&member->multicast);
+		spin_lock_irq(&group->lock);
+	}
+
+	join_state = get_leave_state(group);
+	if (join_state) {
+		group->rec.join_state &= ~join_state;
+		spin_unlock_irq(&group->lock);
+		if (send_leave(group, join_state))
+			goto retest;
+	} else {
+		group->state = MCAST_IDLE;
+		spin_unlock_irq(&group->lock);
+		release_group(group);
+	}
+}
+
+/*
+ * Fail a join request if it is still active - at the head of the pending queue.
+ */
+static void process_join_error(struct mcast_group *group, int status)
+{
+	struct mcast_member *member;
+	int ret;
+
+	spin_lock_irq(&group->lock);
+	member = list_entry(group->pending_list.next,
+			    struct mcast_member, list);
+	if (group->last_join == member) {
+		atomic_inc(&member->refcount);
+		list_del_init(&member->list);
+		spin_unlock_irq(&group->lock);
+		ret = member->multicast.callback(status, &member->multicast);
+		deref_member(member);
+		if (ret)
+			ib_sa_free_multicast(&member->multicast);
+	} else
+		spin_unlock_irq(&group->lock);
+}
+
+static void join_handler(int status, struct ib_sa_mcmember_rec *rec,
+			 void *context)
+{
+	struct mcast_group *group = context;
+
+	if (status)
+		process_join_error(group, status);
+	else {
+		spin_lock_irq(&group->port->lock);
+		group->rec = *rec;
+		if (!memcmp(&mgid0, &group->rec.mgid, sizeof mgid0)) {
+			rb_erase(&group->node, &group->port->table);
+			mcast_insert(group->port, group, 1);
+		}
+		spin_unlock_irq(&group->port->lock);
+	}
+	mcast_work_handler(&group->work);
+}
+
+static void leave_handler(int status, struct ib_sa_mcmember_rec *rec,
+			  void *context)
+{
+	struct mcast_group *group = context;
+
+	mcast_work_handler(&group->work);
+}
+
+static struct mcast_group *acquire_group(struct mcast_port *port,
+					 union ib_gid *mgid, gfp_t gfp_mask)
+{
+	struct mcast_group *group, *cur_group;
+	unsigned long flags;
+	int is_mgid0;
+
+	is_mgid0 = !memcmp(&mgid0, mgid, sizeof mgid0);
+	if (!is_mgid0) {
+		spin_lock_irqsave(&port->lock, flags);
+		group = mcast_find(port, mgid);
+		if (group)
+			goto found;
+		spin_unlock_irqrestore(&port->lock, flags);
+	}
+
+	group = kzalloc(sizeof *group, gfp_mask);
+	if (!group)
+		return NULL;
+
+	group->port = port;
+	group->rec.mgid = *mgid;
+	INIT_LIST_HEAD(&group->pending_list);
+	INIT_LIST_HEAD(&group->active_list);
+	INIT_WORK(&group->work, mcast_work_handler);
+	spin_lock_init(&group->lock);
+
+	spin_lock_irqsave(&port->lock, flags);
+	cur_group = mcast_insert(port, group, is_mgid0);
+	if (cur_group) {
+		kfree(group);
+		group = cur_group;
+	} else
+		atomic_inc(&port->refcount);
+found:
+	atomic_inc(&group->refcount);
+	spin_unlock_irqrestore(&port->lock, flags);
+	return group;
+}
+
+/*
+ * We serialize all join requests to a single group to make our lives much
+ * easier.  Otherwise, two users could try to join the same group
+ * simultaneously, with different configurations, one could leave while the
+ * join is in progress, etc., which makes locking around error recovery
+ * difficult.
+ */
+struct ib_sa_multicast *
+ib_sa_join_multicast(struct ib_sa_client *client,
+		     struct ib_device *device, u8 port_num,
+		     struct ib_sa_mcmember_rec *rec,
+		     ib_sa_comp_mask comp_mask, gfp_t gfp_mask,
+		     int (*callback)(int status,
+				     struct ib_sa_multicast *multicast),
+		     void *context)
+{
+	struct mcast_device *dev;
+	struct mcast_member *member;
+	struct ib_sa_multicast *multicast;
+	int ret;
+
+	dev = ib_get_client_data(device, &mcast_client);
+	if (!dev)
+		return ERR_PTR(-ENODEV);
+
+	member = kmalloc(sizeof *member, gfp_mask);
+	if (!member)
+		return ERR_PTR(-ENOMEM);
+
+	ib_sa_client_get(client);
+	member->client = client;
+	member->multicast.rec = *rec;
+	member->multicast.comp_mask = comp_mask;
+	member->multicast.callback = callback;
+	member->multicast.context = context;
+	init_completion(&member->comp);
+	atomic_set(&member->refcount, 1);
+	member->state = MCAST_JOINING;
+
+	member->group = acquire_group(&dev->port[port_num - dev->start_port],
+				      &rec->mgid, gfp_mask);
+	if (!member->group) {
+		ret = -ENOMEM;
+		goto err;
+	}
+
+	/*
+	 * The user will get the multicast structure in their callback.  They
+	 * could then free the multicast structure before we can return from
+	 * this routine.  So we save the pointer to return before queuing
+	 * any callback.
+	 */
+	multicast = &member->multicast;
+	queue_join(member);
+	return multicast;
+
+err:
+	ib_sa_client_put(client);
+	kfree(member);
+	return ERR_PTR(ret);
+}
+EXPORT_SYMBOL(ib_sa_join_multicast);
+
+void ib_sa_free_multicast(struct ib_sa_multicast *multicast)
+{
+	struct mcast_member *member;
+	struct mcast_group *group;
+
+	member = container_of(multicast, struct mcast_member, multicast);
+	group = member->group;
+
+	spin_lock_irq(&group->lock);
+	if (member->state == MCAST_MEMBER)
+		adjust_membership(group, multicast->rec.join_state, -1);
+
+	list_del_init(&member->list);
+
+	if (group->state == MCAST_IDLE) {
+		group->state = MCAST_BUSY;
+		spin_unlock_irq(&group->lock);
+		/* Continue to hold reference on group until callback */
+		queue_work(mcast_wq, &group->work);
+	} else {
+		spin_unlock_irq(&group->lock);
+		release_group(group);
+	}
+
+	deref_member(member);
+	wait_for_completion(&member->comp);
+	ib_sa_client_put(member->client);
+	kfree(member);
+}
+EXPORT_SYMBOL(ib_sa_free_multicast);
+
+int ib_sa_get_mcmember_rec(struct ib_device *device, u8 port_num,
+			   union ib_gid *mgid, struct ib_sa_mcmember_rec *rec)
+{
+	struct mcast_device *dev;
+	struct mcast_port *port;
+	struct mcast_group *group;
+	unsigned long flags;
+	int ret = 0;
+
+	dev = ib_get_client_data(device, &mcast_client);
+	if (!dev)
+		return -ENODEV;
+
+	port = &dev->port[port_num - dev->start_port];
+	spin_lock_irqsave(&port->lock, flags);
+	group = mcast_find(port, mgid);
+	if (group)
+		*rec = group->rec;
+	else
+		ret = -EADDRNOTAVAIL;
+	spin_unlock_irqrestore(&port->lock, flags);
+
+	return ret;
+}
+EXPORT_SYMBOL(ib_sa_get_mcmember_rec);
+
+int ib_init_ah_from_mcmember(struct ib_device *device, u8 port_num,
+			     struct ib_sa_mcmember_rec *rec,
+			     struct ib_ah_attr *ah_attr)
+{
+	int ret;
+	u16 gid_index;
+	u8 p;
+
+	ret = ib_find_cached_gid(device, &rec->port_gid, &p, &gid_index);
+	if (ret)
+		return ret;
+
+	memset(ah_attr, 0, sizeof *ah_attr);
+	ah_attr->dlid = be16_to_cpu(rec->mlid);
+	ah_attr->sl = rec->sl;
+	ah_attr->port_num = port_num;
+	ah_attr->static_rate = rec->rate;
+
+	ah_attr->ah_flags = IB_AH_GRH;
+	ah_attr->grh.dgid = rec->mgid;
+
+	ah_attr->grh.sgid_index = (u8) gid_index;
+	ah_attr->grh.flow_label = be32_to_cpu(rec->flow_label);
+	ah_attr->grh.hop_limit = rec->hop_limit;
+	ah_attr->grh.traffic_class = rec->traffic_class;
+
+	return 0;
+}
+EXPORT_SYMBOL(ib_init_ah_from_mcmember);
+
+static void mcast_groups_lost(struct mcast_port *port)
+{
+	struct mcast_group *group;
+	struct rb_node *node;
+	unsigned long flags;
+
+	spin_lock_irqsave(&port->lock, flags);
+	for (node = rb_first(&port->table); node; node = rb_next(node)) {
+		group = rb_entry(node, struct mcast_group, node);
+		spin_lock(&group->lock);
+		if (group->state == MCAST_IDLE) {
+			atomic_inc(&group->refcount);
+			queue_work(mcast_wq, &group->work);
+		}
+		group->state = MCAST_ERROR;
+		spin_unlock(&group->lock);
+	}
+	spin_unlock_irqrestore(&port->lock, flags);
+}
+
+static void mcast_event_handler(struct ib_event_handler *handler,
+				struct ib_event *event)
+{
+	struct mcast_device *dev;
+
+	dev = container_of(handler, struct mcast_device, event_handler);
+
+	switch (event->event) {
+	case IB_EVENT_PORT_ERR:
+	case IB_EVENT_LID_CHANGE:
+	case IB_EVENT_SM_CHANGE:
+	case IB_EVENT_CLIENT_REREGISTER:
+		mcast_groups_lost(&dev->port[event->element.port_num -
+					     dev->start_port]);
+		break;
+	default:
+		break;
+	}
+}
+
+static void mcast_add_one(struct ib_device *device)
+{
+	struct mcast_device *dev;
+	struct mcast_port *port;
+	int i;
+
+	if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
+		return;
+
+	dev = kmalloc(sizeof *dev + device->phys_port_cnt * sizeof *port,
+		      GFP_KERNEL);
+	if (!dev)
+		return;
+
+	if (device->node_type == RDMA_NODE_IB_SWITCH)
+		dev->start_port = dev->end_port = 0;
+	else {
+		dev->start_port = 1;
+		dev->end_port = device->phys_port_cnt;
+	}
+
+	for (i = 0; i <= dev->end_port - dev->start_port; i++) {
+		port = &dev->port[i];
+		port->dev = dev;
+		port->port_num = dev->start_port + i;
+		spin_lock_init(&port->lock);
+		port->table = RB_ROOT;
+		init_completion(&port->comp);
+		atomic_set(&port->refcount, 1);
+	}
+
+	dev->device = device;
+	ib_set_client_data(device, &mcast_client, dev);
+
+	INIT_IB_EVENT_HANDLER(&dev->event_handler, device, mcast_event_handler);
+	ib_register_event_handler(&dev->event_handler);
+}
+
+static void mcast_remove_one(struct ib_device *device)
+{
+	struct mcast_device *dev;
+	struct mcast_port *port;
+	int i;
+
+	dev = ib_get_client_data(device, &mcast_client);
+	if (!dev)
+		return;
+
+	ib_unregister_event_handler(&dev->event_handler);
+	flush_workqueue(mcast_wq);
+
+	for (i = 0; i <= dev->end_port - dev->start_port; i++) {
+		port = &dev->port[i];
+		deref_port(port);
+		wait_for_completion(&port->comp);
+	}
+
+	kfree(dev);
+}
+
+int mcast_init(void)
+{
+	int ret;
+
+	mcast_wq = create_singlethread_workqueue("ib_mcast");
+	if (!mcast_wq)
+		return -ENOMEM;
+
+	ib_sa_register_client(&sa_client);
+
+	ret = ib_register_client(&mcast_client);
+	if (ret)
+		goto err;
+	return 0;
+
+err:
+	ib_sa_unregister_client(&sa_client);
+	destroy_workqueue(mcast_wq);
+	return ret;
+}
+
+void mcast_cleanup(void)
+{
+	ib_unregister_client(&mcast_client);
+	ib_sa_unregister_client(&sa_client);
+	destroy_workqueue(mcast_wq);
+}
diff --git a/drivers/infiniband/core/sa.h b/drivers/infiniband/core/sa.h
new file mode 100644
index 00000000000..24c93fd320f
--- /dev/null
+++ b/drivers/infiniband/core/sa.h
@@ -0,0 +1,66 @@
+/*
+ * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Voltaire, Inc.  All rights reserved.
+ * Copyright (c) 2006 Intel Corporation.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef SA_H
+#define SA_H
+
+#include <rdma/ib_sa.h>
+
+static inline void ib_sa_client_get(struct ib_sa_client *client)
+{
+	atomic_inc(&client->users);
+}
+
+static inline void ib_sa_client_put(struct ib_sa_client *client)
+{
+	if (atomic_dec_and_test(&client->users))
+		complete(&client->comp);
+}
+
+int ib_sa_mcmember_rec_query(struct ib_sa_client *client,
+			     struct ib_device *device, u8 port_num,
+			     u8 method,
+			     struct ib_sa_mcmember_rec *rec,
+			     ib_sa_comp_mask comp_mask,
+			     int timeout_ms, gfp_t gfp_mask,
+			     void (*callback)(int status,
+					      struct ib_sa_mcmember_rec *resp,
+					      void *context),
+			     void *context,
+			     struct ib_sa_query **sa_query);
+
+int mcast_init(void);
+void mcast_cleanup(void);
+
+#endif /* SA_H */
diff --git a/drivers/infiniband/core/sa_query.c b/drivers/infiniband/core/sa_query.c
index e45afba7534..d7d4a5309ba 100644
--- a/drivers/infiniband/core/sa_query.c
+++ b/drivers/infiniband/core/sa_query.c
@@ -47,8 +47,8 @@
 #include <linux/workqueue.h>
 
 #include <rdma/ib_pack.h>
-#include <rdma/ib_sa.h>
 #include <rdma/ib_cache.h>
+#include "sa.h"
 
 MODULE_AUTHOR("Roland Dreier");
 MODULE_DESCRIPTION("InfiniBand subnet administration query support");
@@ -425,17 +425,6 @@ void ib_sa_register_client(struct ib_sa_client *client)
 }
 EXPORT_SYMBOL(ib_sa_register_client);
 
-static inline void ib_sa_client_get(struct ib_sa_client *client)
-{
-	atomic_inc(&client->users);
-}
-
-static inline void ib_sa_client_put(struct ib_sa_client *client)
-{
-	if (atomic_dec_and_test(&client->users))
-		complete(&client->comp);
-}
-
 void ib_sa_unregister_client(struct ib_sa_client *client)
 {
 	ib_sa_client_put(client);
@@ -901,7 +890,6 @@ err1:
 	kfree(query);
 	return ret;
 }
-EXPORT_SYMBOL(ib_sa_mcmember_rec_query);
 
 static void send_handler(struct ib_mad_agent *agent,
 			 struct ib_mad_send_wc *mad_send_wc)
@@ -1053,14 +1041,27 @@ static int __init ib_sa_init(void)
 	get_random_bytes(&tid, sizeof tid);
 
 	ret = ib_register_client(&sa_client);
-	if (ret)
+	if (ret) {
 		printk(KERN_ERR "Couldn't register ib_sa client\n");
+		goto err1;
+	}
+
+	ret = mcast_init();
+	if (ret) {
+		printk(KERN_ERR "Couldn't initialize multicast handling\n");
+		goto err2;
+	}
 
+	return 0;
+err2:
+	ib_unregister_client(&sa_client);
+err1:
 	return ret;
 }
 
 static void __exit ib_sa_cleanup(void)
 {
+	mcast_cleanup();
 	ib_unregister_client(&sa_client);
 	idr_destroy(&query_idr);
 }