[PATCH for-next V1 19/29] IB/mlx4: Added Multicast Groups (MCG) para-virtualization for SRIOV

[Jack Morgenstein <jackm-LDSdmyG8hGV8YrgS2mwiifqBs+8SCbDb@public.gmane.org>]

From: Oren Duer <oren-VPRAkNaXOzVS1MOuV/RT9w@public.gmane.org>

MCG para-virtualization support includes:
- Creating multicast groups by VFs, and keeping accounting of them
- Leaving multicast groups by VFs
- SM will only be updated with real changes in the overall picture of MCGs status
- Creation of MGID=0 groups (let SM choose MGID)

Note that the MCG module maintains its own internal MCG object reference
counts. The reason for this is that the IB core is used to track only
the multicast groups joins generated by the PF it runs over.
The PF ib-core layer is unaware of slaves, so it cannot be used to keep
track of MCG joins they generate.

Signed-off-by: Oren Duer <oren-VPRAkNaXOzVS1MOuV/RT9w@public.gmane.org>
Signed-off-by: Eli Cohen <eli-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
Signed-off-by: Jack Morgenstein <jackm-LDSdmyG8hGV8YrgS2mwiifqBs+8SCbDb@public.gmane.org>
---
 drivers/infiniband/hw/mlx4/Makefile  |    2 +-
 drivers/infiniband/hw/mlx4/mad.c     |   60 ++-
 drivers/infiniband/hw/mlx4/main.c    |   18 +-
 drivers/infiniband/hw/mlx4/mcg.c     | 1187 ++++++++++++++++++++++++++++++++++
 drivers/infiniband/hw/mlx4/mlx4_ib.h |   30 +
 5 files changed, 1285 insertions(+), 12 deletions(-)
 create mode 100644 drivers/infiniband/hw/mlx4/mcg.c

diff --git a/drivers/infiniband/hw/mlx4/Makefile b/drivers/infiniband/hw/mlx4/Makefile
index 70f09c7..20d627d 100644
--- a/drivers/infiniband/hw/mlx4/Makefile
+++ b/drivers/infiniband/hw/mlx4/Makefile
@@ -1,3 +1,3 @@
 obj-$(CONFIG_MLX4_INFINIBAND)	+= mlx4_ib.o
 
-mlx4_ib-y :=	ah.o cq.o doorbell.o mad.o main.o mr.o qp.o srq.o
+mlx4_ib-y :=	ah.o cq.o doorbell.o mad.o main.o mr.o qp.o srq.o mcg.o
diff --git a/drivers/infiniband/hw/mlx4/mad.c b/drivers/infiniband/hw/mlx4/mad.c
index 538003e..513bb0d 100644
--- a/drivers/infiniband/hw/mlx4/mad.c
+++ b/drivers/infiniband/hw/mlx4/mad.c
@@ -75,6 +75,14 @@ struct mlx4_rcv_tunnel_mad {
 	struct ib_mad mad;
 } __packed;
 
+static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num);
+
+__be64 mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx *ctx)
+{
+	return cpu_to_be64(atomic_inc_return(&ctx->tid)) |
+		cpu_to_be64(0xff00000000000000LL);
+}
+
 int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int mad_ifc_flags,
 		 int port, struct ib_wc *in_wc, struct ib_grh *in_grh,
 		 void *in_mad, void *response_mad)
@@ -209,8 +217,7 @@ static void smp_snoop(struct ib_device *ibdev, u8 port_num, struct ib_mad *mad,
 				     pinfo->neighbormtu_mastersmsl & 0xf);
 
 			if (pinfo->clientrereg_resv_subnetto & 0x80)
-				mlx4_ib_dispatch_event(dev, port_num,
-						       IB_EVENT_CLIENT_REREGISTER);
+				handle_client_rereg_event(dev, port_num);
 
 			if (prev_lid != lid)
 				mlx4_ib_dispatch_event(dev, port_num,
@@ -305,7 +312,17 @@ static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, struct ib_mad *ma
 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
 							     struct ib_sa_mad *sa_mad)
 {
-	return 0;
+	int ret = 0;
+
+	/* dispatch to different sa handlers */
+	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
+	case IB_SA_ATTR_MC_MEMBER_REC:
+		ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
+		break;
+	default:
+		break;
+	}
+	return ret;
 }
 
 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
@@ -765,6 +782,16 @@ void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
 	}
 }
 
+static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
+{
+	/* re-configure the mcg's */
+	if (mlx4_is_master(dev->dev)) {
+		if (!dev->sriov.is_going_down)
+			mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
+	}
+	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
+}
+
 void handle_port_mgmt_change_event(struct work_struct *work)
 {
 	struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
@@ -794,8 +821,7 @@ void handle_port_mgmt_change_event(struct work_struct *work)
 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
 
 		if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
-			mlx4_ib_dispatch_event(dev, port,
-					       IB_EVENT_CLIENT_REREGISTER);
+			handle_client_rereg_event(dev, port);
 		break;
 
 	case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
@@ -865,7 +891,17 @@ static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
 		int slave, struct ib_sa_mad *sa_mad)
 {
-	return 0;
+	int ret = 0;
+
+	/* dispatch to different sa handlers */
+	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
+	case IB_SA_ATTR_MC_MEMBER_REC:
+		ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
+		break;
+	default:
+		break;
+	}
+	return ret;
 }
 
 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
@@ -1587,6 +1623,7 @@ static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
 	int ret = 0;
 
 	if (!do_init) {
+		clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
 		/* for master, destroy real sqp resources */
 		if (slave == mlx4_master_func_num(dev->dev))
 			destroy_pv_resources(dev, slave, port,
@@ -1640,10 +1677,16 @@ static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
 		ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
 		if (ret) {
 			ret = -ENOMEM;
-			goto err_wq;
+			goto err_mcg;
 		}
 	}
 
+	ret = mlx4_ib_mcg_port_init(ctx);
+	if (ret) {
+		pr_err("Failed initializing mcg para-virt (%d)\n", ret);
+		goto err_mcg;
+	}
+
 	snprintf(name, sizeof name, "mlx4_ibt%d", port);
 	ctx->wq = create_singlethread_workqueue(name);
 	if (!ctx->wq) {
@@ -1667,6 +1710,8 @@ err_udwq:
 	ctx->wq = NULL;
 
 err_wq:
+	mlx4_ib_mcg_port_cleanup(ctx, 1);
+err_mcg:
 	for (i = 0; i < dev->dev->caps.sqp_demux; i++)
 		free_pv_object(dev, i, port);
 	kfree(ctx->tun);
@@ -1702,6 +1747,7 @@ static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
 	int i;
 	if (ctx) {
 		struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
+		mlx4_ib_mcg_port_cleanup(ctx, 1);
 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
 			if (!ctx->tun[i])
 				continue;
diff --git a/drivers/infiniband/hw/mlx4/main.c b/drivers/infiniband/hw/mlx4/main.c
index 62f70ad..615ce5d 100644
--- a/drivers/infiniband/hw/mlx4/main.c
+++ b/drivers/infiniband/hw/mlx4/main.c
@@ -1578,18 +1578,28 @@ static int __init mlx4_ib_init(void)
 	if (!wq)
 		return -ENOMEM;
 
+	err = mlx4_ib_mcg_init();
+	if (err)
+		goto clean_wq;
+
 	err = mlx4_register_interface(&mlx4_ib_interface);
-	if (err) {
-		destroy_workqueue(wq);
-		return err;
-	}
+	if (err)
+		goto clean_mcg;
 
 	return 0;
+
+clean_mcg:
+	mlx4_ib_mcg_destroy();
+
+clean_wq:
+	destroy_workqueue(wq);
+	return err;
 }
 
 static void __exit mlx4_ib_cleanup(void)
 {
 	mlx4_unregister_interface(&mlx4_ib_interface);
+	mlx4_ib_mcg_destroy();
 	destroy_workqueue(wq);
 }
 
diff --git a/drivers/infiniband/hw/mlx4/mcg.c b/drivers/infiniband/hw/mlx4/mcg.c
new file mode 100644
index 0000000..1ee2e3a
--- /dev/null
+++ b/drivers/infiniband/hw/mlx4/mcg.c
@@ -0,0 +1,1187 @@
+/*
+ * Copyright (c) 2012 Mellanox Technologies. 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 <rdma/ib_mad.h>
+#include <rdma/ib_smi.h>
+#include <rdma/ib_cache.h>
+#include <rdma/ib_sa.h>
+
+#include <linux/mlx4/cmd.h>
+#include <linux/rbtree.h>
+#include <linux/delay.h>
+
+#include "mlx4_ib.h"
+
+#define MAX_VFS		80
+#define MAX_PEND_REQS_PER_FUNC 4
+#define MAD_TIMEOUT_MS	2000
+
+#define mcg_warn(fmt, arg...)	pr_warn("MCG WARNING: " fmt, ##arg)
+#define mcg_error(fmt, arg...)	pr_err(fmt, ##arg)
+#define mcg_warn_group(group, format, arg...) \
+	pr_warn("%s-%d: %16s (port %d): WARNING: " format, __func__, __LINE__,\
+	(group)->name, group->demux->port, ## arg)
+
+#define mcg_error_group(group, format, arg...) \
+	pr_err("  %16s: " format, (group)->name, ## arg)
+
+
+static union ib_gid mgid0;
+
+static struct workqueue_struct *clean_wq;
+
+enum mcast_state {
+	MCAST_NOT_MEMBER = 0,
+	MCAST_MEMBER,
+};
+
+enum mcast_group_state {
+	MCAST_IDLE,
+	MCAST_JOIN_SENT,
+	MCAST_LEAVE_SENT,
+	MCAST_RESP_READY
+};
+
+struct mcast_member {
+	enum mcast_state state;
+	uint8_t			join_state;
+	int			num_pend_reqs;
+	struct list_head	pending;
+};
+
+struct ib_sa_mcmember_data {
+	union ib_gid	mgid;
+	union ib_gid	port_gid;
+	__be32		qkey;
+	__be16		mlid;
+	u8		mtusel_mtu;
+	u8		tclass;
+	__be16		pkey;
+	u8		ratesel_rate;
+	u8		lifetmsel_lifetm;
+	__be32		sl_flowlabel_hoplimit;
+	u8		scope_join_state;
+	u8		proxy_join;
+	u8		reserved[2];
+};
+
+struct mcast_group {
+	struct ib_sa_mcmember_data rec;
+	struct rb_node		node;
+	struct list_head	mgid0_list;
+	struct mlx4_ib_demux_ctx *demux;
+	struct mcast_member	func[MAX_VFS];
+	struct mutex		lock;
+	struct work_struct	work;
+	struct list_head	pending_list;
+	int			members[3];
+	enum mcast_group_state	state;
+	enum mcast_group_state	prev_state;
+	struct ib_sa_mad	response_sa_mad;
+	__be64			last_req_tid;
+
+	char			name[33]; /* MGID string */
+
+	/* refcount is the reference count for the following:
+	   1. Each queued request
+	   2. Each invocation of the worker thread
+	   3. Membership of the port at the SA
+	*/
+	atomic_t		refcount;
+
+	/* delayed work to clean pending SM request */
+	struct delayed_work	timeout_work;
+	struct list_head	cleanup_list;
+};
+
+struct mcast_req {
+	int			func;
+	struct ib_sa_mad	sa_mad;
+	struct list_head	group_list;
+	struct list_head	func_list;
+	struct mcast_group	*group;
+	int			clean;
+};
+
+
+#define safe_atomic_dec(ref) \
+	do {\
+		if (atomic_dec_and_test(ref)) \
+			mcg_warn_group(group, "did not expect to reach zero\n"); \
+	} while (0)
+
+static const char *get_state_string(enum mcast_group_state state)
+{
+	switch (state) {
+	case MCAST_IDLE:
+		return "MCAST_IDLE";
+	case MCAST_JOIN_SENT:
+		return "MCAST_JOIN_SENT";
+	case MCAST_LEAVE_SENT:
+		return "MCAST_LEAVE_SENT";
+	case MCAST_RESP_READY:
+		return "MCAST_RESP_READY";
+	}
+	return "Invalid State";
+}
+
+static struct mcast_group *mcast_find(struct mlx4_ib_demux_ctx *ctx,
+				      union ib_gid *mgid)
+{
+	struct rb_node *node = ctx->mcg_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 mlx4_ib_demux_ctx *ctx,
+					struct mcast_group *group)
+{
+	struct rb_node **link = &ctx->mcg_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
+			return cur_group;
+	}
+	rb_link_node(&group->node, parent, link);
+	rb_insert_color(&group->node, &ctx->mcg_table);
+	return NULL;
+}
+
+static int send_mad_to_wire(struct mlx4_ib_demux_ctx *ctx, struct ib_mad *mad)
+{
+	struct mlx4_ib_dev *dev = ctx->dev;
+	struct ib_ah_attr	ah_attr;
+
+	spin_lock(&dev->sm_lock);
+	if (!dev->sm_ah[ctx->port - 1]) {
+		/* port is not yet Active, sm_ah not ready */
+		spin_unlock(&dev->sm_lock);
+		return -EAGAIN;
+	}
+	mlx4_ib_query_ah(dev->sm_ah[ctx->port - 1], &ah_attr);
+	spin_unlock(&dev->sm_lock);
+	return mlx4_ib_send_to_wire(dev, mlx4_master_func_num(dev->dev), ctx->port,
+				    IB_QPT_GSI, 0, 1, IB_QP1_QKEY, &ah_attr, mad);
+}
+
+static int send_mad_to_slave(int slave, struct mlx4_ib_demux_ctx *ctx,
+			     struct ib_mad *mad)
+{
+	struct mlx4_ib_dev *dev = ctx->dev;
+	struct ib_mad_agent *agent = dev->send_agent[ctx->port - 1][1];
+	struct ib_wc wc;
+	struct ib_ah_attr ah_attr;
+
+	/* Our agent might not yet be registered when mads start to arrive */
+	if (!agent)
+		return -EAGAIN;
+
+	ib_query_ah(dev->sm_ah[ctx->port - 1], &ah_attr);
+
+	wc.pkey_index = 0;
+	wc.sl = 0;
+	wc.dlid_path_bits = 0;
+	wc.port_num = ctx->port;
+	wc.slid = ah_attr.dlid;  /* opensm lid */
+	wc.src_qp = 1;
+	return mlx4_ib_send_to_slave(dev, slave, ctx->port, IB_QPT_GSI, &wc, NULL, mad);
+}
+
+static int send_join_to_wire(struct mcast_group *group, struct ib_sa_mad *sa_mad)
+{
+	struct ib_sa_mad mad;
+	struct ib_sa_mcmember_data *sa_mad_data = (struct ib_sa_mcmember_data *)&mad.data;
+	int ret;
+
+	/* we rely on a mad request as arrived from a VF */
+	memcpy(&mad, sa_mad, sizeof mad);
+
+	/* fix port GID to be the real one (slave 0) */
+	sa_mad_data->port_gid.global.interface_id = group->demux->guid_cache[0];
+
+	/* assign our own TID */
+	mad.mad_hdr.tid = mlx4_ib_get_new_demux_tid(group->demux);
+	group->last_req_tid = mad.mad_hdr.tid; /* keep it for later validation */
+
+	ret = send_mad_to_wire(group->demux, (struct ib_mad *)&mad);
+	/* set timeout handler */
+	if (!ret) {
+		/* calls mlx4_ib_mcg_timeout_handler */
+		queue_delayed_work(group->demux->mcg_wq, &group->timeout_work,
+				msecs_to_jiffies(MAD_TIMEOUT_MS));
+	}
+
+	return ret;
+}
+
+static int send_leave_to_wire(struct mcast_group *group, u8 join_state)
+{
+	struct ib_sa_mad mad;
+	struct ib_sa_mcmember_data *sa_data = (struct ib_sa_mcmember_data *)&mad.data;
+	int ret;
+
+	memset(&mad, 0, sizeof mad);
+	mad.mad_hdr.base_version = 1;
+	mad.mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
+	mad.mad_hdr.class_version = 2;
+	mad.mad_hdr.method = IB_SA_METHOD_DELETE;
+	mad.mad_hdr.status = cpu_to_be16(0);
+	mad.mad_hdr.class_specific = cpu_to_be16(0);
+	mad.mad_hdr.tid = mlx4_ib_get_new_demux_tid(group->demux);
+	group->last_req_tid = mad.mad_hdr.tid; /* keep it for later validation */
+	mad.mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_MC_MEMBER_REC);
+	mad.mad_hdr.attr_mod = cpu_to_be32(0);
+	mad.sa_hdr.sm_key = 0x0;
+	mad.sa_hdr.attr_offset = cpu_to_be16(7);
+	mad.sa_hdr.comp_mask = IB_SA_MCMEMBER_REC_MGID |
+		IB_SA_MCMEMBER_REC_PORT_GID | IB_SA_MCMEMBER_REC_JOIN_STATE;
+
+	*sa_data = group->rec;
+	sa_data->scope_join_state = join_state;
+
+	ret = send_mad_to_wire(group->demux, (struct ib_mad *)&mad);
+	if (ret)
+		group->state = MCAST_IDLE;
+
+	/* set timeout handler */
+	if (!ret) {
+		/* calls mlx4_ib_mcg_timeout_handler */
+		queue_delayed_work(group->demux->mcg_wq, &group->timeout_work,
+				msecs_to_jiffies(MAD_TIMEOUT_MS));
+	}
+
+	return ret;
+}
+
+static int send_reply_to_slave(int slave, struct mcast_group *group,
+		struct ib_sa_mad *req_sa_mad, u16 status)
+{
+	struct ib_sa_mad mad;
+	struct ib_sa_mcmember_data *sa_data = (struct ib_sa_mcmember_data *)&mad.data;
+	struct ib_sa_mcmember_data *req_sa_data = (struct ib_sa_mcmember_data *)&req_sa_mad->data;
+	int ret;
+
+	memset(&mad, 0, sizeof mad);
+	mad.mad_hdr.base_version = 1;
+	mad.mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
+	mad.mad_hdr.class_version = 2;
+	mad.mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
+	mad.mad_hdr.status = cpu_to_be16(status);
+	mad.mad_hdr.class_specific = cpu_to_be16(0);
+	mad.mad_hdr.tid = req_sa_mad->mad_hdr.tid;
+	*(u8 *)&mad.mad_hdr.tid = 0; /* resetting tid to 0 */
+	mad.mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_MC_MEMBER_REC);
+	mad.mad_hdr.attr_mod = cpu_to_be32(0);
+	mad.sa_hdr.sm_key = req_sa_mad->sa_hdr.sm_key;
+	mad.sa_hdr.attr_offset = cpu_to_be16(7);
+	mad.sa_hdr.comp_mask = 0; /* ignored on responses, see IBTA spec */
+
+	*sa_data = group->rec;
+
+	/* reconstruct VF's requested join_state and port_gid */
+	sa_data->scope_join_state &= 0xf0;
+	sa_data->scope_join_state |= (group->func[slave].join_state & 0x0f);
+	memcpy(&sa_data->port_gid, &req_sa_data->port_gid, sizeof req_sa_data->port_gid);
+
+	ret = send_mad_to_slave(slave, group->demux, (struct ib_mad *)&mad);
+	return ret;
+}
+
+static int check_selector(ib_sa_comp_mask comp_mask,
+			  ib_sa_comp_mask selector_mask,
+			  ib_sa_comp_mask value_mask,
+			  u8 src_value, u8 dst_value)
+{
+	int err;
+	u8 selector = dst_value >> 6;
+	dst_value &= 0x3f;
+	src_value &= 0x3f;
+
+	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 u16 cmp_rec(struct ib_sa_mcmember_data *src,
+		   struct ib_sa_mcmember_data *dst, ib_sa_comp_mask comp_mask)
+{
+	/* src is group record, dst is request record */
+	/* MGID must already match */
+	/* Port_GID we always replace to our Port_GID, so it is a match */
+
+#define MAD_STATUS_REQ_INVALID 0x0200
+	if (comp_mask & IB_SA_MCMEMBER_REC_QKEY && src->qkey != dst->qkey)
+		return MAD_STATUS_REQ_INVALID;
+	if (comp_mask & IB_SA_MCMEMBER_REC_MLID && src->mlid != dst->mlid)
+		return MAD_STATUS_REQ_INVALID;
+	if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_MTU_SELECTOR,
+				 IB_SA_MCMEMBER_REC_MTU,
+				 src->mtusel_mtu, dst->mtusel_mtu))
+		return MAD_STATUS_REQ_INVALID;
+	if (comp_mask & IB_SA_MCMEMBER_REC_TRAFFIC_CLASS &&
+	    src->tclass != dst->tclass)
+		return MAD_STATUS_REQ_INVALID;
+	if (comp_mask & IB_SA_MCMEMBER_REC_PKEY && src->pkey != dst->pkey)
+		return MAD_STATUS_REQ_INVALID;
+	if (check_selector(comp_mask, IB_SA_MCMEMBER_REC_RATE_SELECTOR,
+				 IB_SA_MCMEMBER_REC_RATE,
+				 src->ratesel_rate, dst->ratesel_rate))
+		return MAD_STATUS_REQ_INVALID;
+	if (check_selector(comp_mask,
+				 IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME_SELECTOR,
+				 IB_SA_MCMEMBER_REC_PACKET_LIFE_TIME,
+				 src->lifetmsel_lifetm, dst->lifetmsel_lifetm))
+		return MAD_STATUS_REQ_INVALID;
+	if (comp_mask & IB_SA_MCMEMBER_REC_SL &&
+			(be32_to_cpu(src->sl_flowlabel_hoplimit) & 0xf0000000) !=
+			(be32_to_cpu(dst->sl_flowlabel_hoplimit) & 0xf0000000))
+		return MAD_STATUS_REQ_INVALID;
+	if (comp_mask & IB_SA_MCMEMBER_REC_FLOW_LABEL &&
+			(be32_to_cpu(src->sl_flowlabel_hoplimit) & 0x0fffff00) !=
+			(be32_to_cpu(dst->sl_flowlabel_hoplimit) & 0x0fffff00))
+		return MAD_STATUS_REQ_INVALID;
+	if (comp_mask & IB_SA_MCMEMBER_REC_HOP_LIMIT &&
+			(be32_to_cpu(src->sl_flowlabel_hoplimit) & 0x000000ff) !=
+			(be32_to_cpu(dst->sl_flowlabel_hoplimit) & 0x000000ff))
+		return MAD_STATUS_REQ_INVALID;
+	if (comp_mask & IB_SA_MCMEMBER_REC_SCOPE &&
+			(src->scope_join_state & 0xf0) !=
+			(dst->scope_join_state & 0xf0))
+		return MAD_STATUS_REQ_INVALID;
+
+	/* join_state checked separately, proxy_join ignored */
+
+	return 0;
+}
+
+/* release group, return 1 if this was last release and group is destroyed
+ * timout work is canceled sync */
+static int release_group(struct mcast_group *group, int from_timeout_handler)
+{
+	struct mlx4_ib_demux_ctx *ctx = group->demux;
+	int nzgroup;
+
+	mutex_lock(&ctx->mcg_table_lock);
+	mutex_lock(&group->lock);
+	if (atomic_dec_and_test(&group->refcount)) {
+		if (!from_timeout_handler) {
+			if (group->state != MCAST_IDLE &&
+			    !cancel_delayed_work(&group->timeout_work)) {
+				atomic_inc(&group->refcount);
+				mutex_unlock(&group->lock);
+				mutex_unlock(&ctx->mcg_table_lock);
+				return 0;
+			}
+		}
+
+		nzgroup = memcmp(&group->rec.mgid, &mgid0, sizeof mgid0);
+		if (!list_empty(&group->pending_list))
+			mcg_warn_group(group, "releasing a group with non empty pending list\n");
+		if (nzgroup)
+			rb_erase(&group->node, &ctx->mcg_table);
+		list_del_init(&group->mgid0_list);
+		mutex_unlock(&group->lock);
+		mutex_unlock(&ctx->mcg_table_lock);
+		kfree(group);
+		return 1;
+	} else {
+		mutex_unlock(&group->lock);
+		mutex_unlock(&ctx->mcg_table_lock);
+	}
+	return 0;
+}
+
+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;
+}
+
+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 |= (1 << i);
+
+	return leave_state & (group->rec.scope_join_state & 7);
+}
+
+static int join_group(struct mcast_group *group, int slave, u8 join_mask)
+{
+	int ret = 0;
+	u8 join_state;
+
+	/* remove bits that slave is already member of, and adjust */
+	join_state = join_mask & (~group->func[slave].join_state);
+	adjust_membership(group, join_state, 1);
+	group->func[slave].join_state |= join_state;
+	if (group->func[slave].state != MCAST_MEMBER && join_state) {
+		group->func[slave].state = MCAST_MEMBER;
+		ret = 1;
+	}
+	return ret;
+}
+
+static int leave_group(struct mcast_group *group, int slave, u8 leave_state)
+{
+	int ret = 0;
+
+	adjust_membership(group, leave_state, -1);
+	group->func[slave].join_state &= ~leave_state;
+	if (!group->func[slave].join_state) {
+		group->func[slave].state = MCAST_NOT_MEMBER;
+		ret = 1;
+	}
+	return ret;
+}
+
+static int check_leave(struct mcast_group *group, int slave, u8 leave_mask)
+{
+	if (group->func[slave].state != MCAST_MEMBER)
+		return MAD_STATUS_REQ_INVALID;
+
+	/* make sure we're not deleting unset bits */
+	if (~group->func[slave].join_state & leave_mask)
+		return MAD_STATUS_REQ_INVALID;
+
+	if (!leave_mask)
+		return MAD_STATUS_REQ_INVALID;
+
+	return 0;
+}
+
+static void mlx4_ib_mcg_timeout_handler(struct work_struct *work)
+{
+	struct delayed_work *delay = to_delayed_work(work);
+	struct mcast_group *group;
+	struct mcast_req *req = NULL;
+
+	group = container_of(delay, typeof(*group), timeout_work);
+
+	mutex_lock(&group->lock);
+	if (group->state == MCAST_JOIN_SENT) {
+		if (!list_empty(&group->pending_list)) {
+			req = list_first_entry(&group->pending_list, struct mcast_req, group_list);
+			list_del(&req->group_list);
+			list_del(&req->func_list);
+			--group->func[req->func].num_pend_reqs;
+			mutex_unlock(&group->lock);
+			kfree(req);
+			if (memcmp(&group->rec.mgid, &mgid0, sizeof mgid0)) {
+				if (release_group(group, 1))
+					return;
+			} else {
+				kfree(group);
+				return;
+			}
+			mutex_lock(&group->lock);
+		} else
+			mcg_warn_group(group, "DRIVER BUG\n");
+	} else if (group->state == MCAST_LEAVE_SENT) {
+		if (group->rec.scope_join_state & 7)
+			group->rec.scope_join_state &= 0xf8;
+		group->state = MCAST_IDLE;
+		mutex_unlock(&group->lock);
+		if (release_group(group, 1))
+			return;
+		mutex_lock(&group->lock);
+	} else
+		mcg_warn_group(group, "invalid state %s\n", get_state_string(group->state));
+	group->state = MCAST_IDLE;
+	atomic_inc(&group->refcount);
+	if (!queue_work(group->demux->mcg_wq, &group->work))
+		safe_atomic_dec(&group->refcount);
+
+	mutex_unlock(&group->lock);
+}
+
+static int handle_leave_req(struct mcast_group *group, u8 leave_mask,
+			    struct mcast_req *req)
+{
+	u16 status;
+
+	if (req->clean)
+		leave_mask = group->func[req->func].join_state;
+
+	status = check_leave(group, req->func, leave_mask);
+	if (!status)
+		leave_group(group, req->func, leave_mask);
+
+	if (!req->clean)
+		send_reply_to_slave(req->func, group, &req->sa_mad, status);
+	--group->func[req->func].num_pend_reqs;
+	list_del(&req->group_list);
+	list_del(&req->func_list);
+	kfree(req);
+	return 1;
+}
+
+static int handle_join_req(struct mcast_group *group, u8 join_mask,
+			   struct mcast_req *req)
+{
+	u8 group_join_state = group->rec.scope_join_state & 7;
+	int ref = 0;
+	u16 status;
+	struct ib_sa_mcmember_data *sa_data = (struct ib_sa_mcmember_data *)req->sa_mad.data;
+
+	if (join_mask == (group_join_state & join_mask)) {
+		/* port's membership need not change */
+		status = cmp_rec(&group->rec, sa_data, req->sa_mad.sa_hdr.comp_mask);
+		if (!status)
+			join_group(group, req->func, join_mask);
+
+		--group->func[req->func].num_pend_reqs;
+		send_reply_to_slave(req->func, group, &req->sa_mad, status);
+		list_del(&req->group_list);
+		list_del(&req->func_list);
+		kfree(req);
+		++ref;
+	} else {
+		/* port's membership needs to be updated */
+		group->prev_state = group->state;
+		if (send_join_to_wire(group, &req->sa_mad)) {
+			--group->func[req->func].num_pend_reqs;
+			list_del(&req->group_list);
+			list_del(&req->func_list);
+			kfree(req);
+			ref = 1;
+			group->state = group->prev_state;
+		} else
+			group->state = MCAST_JOIN_SENT;
+	}
+
+	return ref;
+}
+
+static void mlx4_ib_mcg_work_handler(struct work_struct *work)
+{
+	struct mcast_group *group;
+	struct mcast_req *req = NULL;
+	struct ib_sa_mcmember_data *sa_data;
+	u8 req_join_state;
+	int rc = 1; /* release_count - this is for the scheduled work */
+	u16 status;
+	u8 method;
+
+	group = container_of(work, typeof(*group), work);
+
+	mutex_lock(&group->lock);
+
+	/* First, let's see if a response from SM is waiting regarding this group.
+	 * If so, we need to update the group's REC. If this is a bad response, we
+	 * may need to send a bad response to a VF waiting for it. If VF is waiting
+	 * and this is a good response, the VF will be answered later in this func. */
+	if (group->state == MCAST_RESP_READY) {
+		/* cancels mlx4_ib_mcg_timeout_handler */
+		cancel_delayed_work(&group->timeout_work);
+		status = be16_to_cpu(group->response_sa_mad.mad_hdr.status);
+		method = group->response_sa_mad.mad_hdr.method;
+		if (group->last_req_tid != group->response_sa_mad.mad_hdr.tid) {
+			mcg_warn_group(group, "Got MAD response to existing MGID but wrong TID, dropping. Resp TID=%llx, group TID=%llx\n",
+				be64_to_cpu(group->response_sa_mad.mad_hdr.tid),
+				be64_to_cpu(group->last_req_tid));
+			group->state = group->prev_state;
+			goto process_requests;
+		}
+		if (status) {
+			if (!list_empty(&group->pending_list))
+				req = list_first_entry(&group->pending_list,
+						struct mcast_req, group_list);
+			if ((method == IB_MGMT_METHOD_GET_RESP)) {
+					if (req) {
+						send_reply_to_slave(req->func, group, &req->sa_mad, status);
+						--group->func[req->func].num_pend_reqs;
+						list_del(&req->group_list);
+						list_del(&req->func_list);
+						kfree(req);
+						++rc;
+					} else
+						mcg_warn_group(group, "no request for failed join\n");
+			} else if (method == IB_SA_METHOD_DELETE_RESP && group->demux->flushing)
+				++rc;
+		} else {
+			u8 resp_join_state;
+			u8 cur_join_state;
+
+			resp_join_state = ((struct ib_sa_mcmember_data *)
+						group->response_sa_mad.data)->scope_join_state & 7;
+			cur_join_state = group->rec.scope_join_state & 7;
+
+			if (method == IB_MGMT_METHOD_GET_RESP) {
+				/* successfull join */
+				if (!cur_join_state && resp_join_state)
+					--rc;
+			} else if (!resp_join_state)
+					++rc;
+			memcpy(&group->rec, group->response_sa_mad.data, sizeof group->rec);
+		}
+		group->state = MCAST_IDLE;
+	}
+
+process_requests:
+	/* We should now go over pending join/leave requests, as long as we are idle. */
+	while (!list_empty(&group->pending_list) && group->state == MCAST_IDLE) {
+		req = list_first_entry(&group->pending_list, struct mcast_req,
+				       group_list);
+		sa_data = (struct ib_sa_mcmember_data *)req->sa_mad.data;
+		req_join_state = sa_data->scope_join_state & 0x7;
+
+		/* For a leave request, we will immediately answer the VF, and
+		 * update our internal counters. The actual leave will be sent
+		 * to SM later, if at all needed. We dequeue the request now. */
+		if (req->sa_mad.mad_hdr.method == IB_SA_METHOD_DELETE)
+			rc += handle_leave_req(group, req_join_state, req);
+		else
+			rc += handle_join_req(group, req_join_state, req);
+	}
+
+	/* Handle leaves */
+	if (group->state == MCAST_IDLE) {
+		req_join_state = get_leave_state(group);
+		if (req_join_state) {
+			group->rec.scope_join_state &= ~req_join_state;
+			group->prev_state = group->state;
+			if (send_leave_to_wire(group, req_join_state)) {
+				group->state = group->prev_state;
+				++rc;
+			} else
+				group->state = MCAST_LEAVE_SENT;
+		}
+	}
+
+	if (!list_empty(&group->pending_list) && group->state == MCAST_IDLE)
+		goto process_requests;
+	mutex_unlock(&group->lock);
+
+	while (rc--)
+		release_group(group, 0);
+}
+
+static struct mcast_group *search_relocate_mgid0_group(struct mlx4_ib_demux_ctx *ctx,
+						       __be64 tid,
+						       union ib_gid *new_mgid)
+{
+	struct mcast_group *group = NULL, *cur_group;
+	struct mcast_req *req;
+	struct list_head *pos;
+	struct list_head *n;
+
+	mutex_lock(&ctx->mcg_table_lock);
+	list_for_each_safe(pos, n, &ctx->mcg_mgid0_list) {
+		group = list_entry(pos, struct mcast_group, mgid0_list);
+		mutex_lock(&group->lock);
+		if (group->last_req_tid == tid) {
+			if (memcmp(new_mgid, &mgid0, sizeof mgid0)) {
+				group->rec.mgid = *new_mgid;
+				sprintf(group->name, "%016llx%016llx",
+						be64_to_cpu(group->rec.mgid.global.subnet_prefix),
+						be64_to_cpu(group->rec.mgid.global.interface_id));
+				list_del_init(&group->mgid0_list);
+				cur_group = mcast_insert(ctx, group);
+				if (cur_group) {
+					/* A race between our code and SM. Silently cleaning the new one */
+					req = list_first_entry(&group->pending_list,
+							       struct mcast_req, group_list);
+					--group->func[req->func].num_pend_reqs;
+					list_del(&req->group_list);
+					list_del(&req->func_list);
+					kfree(req);
+					mutex_unlock(&group->lock);
+					mutex_unlock(&ctx->mcg_table_lock);
+					release_group(group, 0);
+					return NULL;
+				}
+
+				atomic_inc(&group->refcount);
+				mutex_unlock(&group->lock);
+				mutex_unlock(&ctx->mcg_table_lock);
+				return group;
+			} else {
+				struct mcast_req *tmp1, *tmp2;
+
+				list_del(&group->mgid0_list);
+				if (!list_empty(&group->pending_list) && group->state != MCAST_IDLE)
+					cancel_delayed_work_sync(&group->timeout_work);
+
+				list_for_each_entry_safe(tmp1, tmp2, &group->pending_list, group_list) {
+					list_del(&tmp1->group_list);
+					kfree(tmp1);
+				}
+				mutex_unlock(&group->lock);
+				mutex_unlock(&ctx->mcg_table_lock);
+				kfree(group);
+				return NULL;
+			}
+		}
+		mutex_unlock(&group->lock);
+	}
+	mutex_unlock(&ctx->mcg_table_lock);
+
+	return NULL;
+}
+
+static struct mcast_group *acquire_group(struct mlx4_ib_demux_ctx *ctx,
+					 union ib_gid *mgid, int create,
+					 gfp_t gfp_mask)
+{
+	struct mcast_group *group, *cur_group;
+	int is_mgid0;
+	int i;
+
+	is_mgid0 = !memcmp(&mgid0, mgid, sizeof mgid0);
+	if (!is_mgid0) {
+		group = mcast_find(ctx, mgid);
+		if (group)
+			goto found;
+	}
+
+	if (!create)
+		return ERR_PTR(-ENOENT);
+
+	group = kzalloc(sizeof *group, gfp_mask);
+	if (!group)
+		return ERR_PTR(-ENOMEM);
+
+	group->demux = ctx;
+	group->rec.mgid = *mgid;
+	INIT_LIST_HEAD(&group->pending_list);
+	INIT_LIST_HEAD(&group->mgid0_list);
+	for (i = 0; i < MAX_VFS; ++i)
+		INIT_LIST_HEAD(&group->func[i].pending);
+	INIT_WORK(&group->work, mlx4_ib_mcg_work_handler);
+	INIT_DELAYED_WORK(&group->timeout_work, mlx4_ib_mcg_timeout_handler);
+	mutex_init(&group->lock);
+	sprintf(group->name, "%016llx%016llx",
+			be64_to_cpu(group->rec.mgid.global.subnet_prefix),
+			be64_to_cpu(group->rec.mgid.global.interface_id));
+	group->state = MCAST_IDLE;
+
+	if (is_mgid0) {
+		list_add(&group->mgid0_list, &ctx->mcg_mgid0_list);
+		goto found;
+	}
+
+	cur_group = mcast_insert(ctx, group);
+	if (cur_group) {
+		mcg_warn("group just showed up %s - confused\n", cur_group->name);
+		kfree(group);
+		return ERR_PTR(-EINVAL);
+	}
+
+found:
+	atomic_inc(&group->refcount);
+	return group;
+}
+
+static void queue_req(struct mcast_req *req)
+{
+	struct mcast_group *group = req->group;
+
+	atomic_inc(&group->refcount); /* for the request */
+	atomic_inc(&group->refcount); /* for scheduling the work */
+	list_add_tail(&req->group_list, &group->pending_list);
+	list_add_tail(&req->func_list, &group->func[req->func].pending);
+	/* calls mlx4_ib_mcg_work_handler */
+	if (!queue_work(group->demux->mcg_wq, &group->work))
+		safe_atomic_dec(&group->refcount);
+}
+
+int mlx4_ib_mcg_demux_handler(struct ib_device *ibdev, int port, int slave,
+			      struct ib_sa_mad *mad)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	struct ib_sa_mcmember_data *rec = (struct ib_sa_mcmember_data *)mad->data;
+	struct mlx4_ib_demux_ctx *ctx = &dev->sriov.demux[port - 1];
+	struct mcast_group *group;
+
+	switch (mad->mad_hdr.method) {
+	case IB_MGMT_METHOD_GET_RESP:
+	case IB_SA_METHOD_DELETE_RESP:
+		mutex_lock(&ctx->mcg_table_lock);
+		group = acquire_group(ctx, &rec->mgid, 0, GFP_KERNEL);
+		mutex_unlock(&ctx->mcg_table_lock);
+		if (IS_ERR(group)) {
+			if (mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP) {
+				__be64 tid = mad->mad_hdr.tid;
+				*(u8 *)(&tid) = (u8)slave; /* in group we kept the modified TID */
+				group = search_relocate_mgid0_group(ctx, tid, &rec->mgid);
+			} else
+				group = NULL;
+		}
+
+		if (!group)
+			return 1;
+
+		mutex_lock(&group->lock);
+		group->response_sa_mad = *mad;
+		group->prev_state = group->state;
+		group->state = MCAST_RESP_READY;
+		/* calls mlx4_ib_mcg_work_handler */
+		atomic_inc(&group->refcount);
+		if (!queue_work(ctx->mcg_wq, &group->work))
+			safe_atomic_dec(&group->refcount);
+		mutex_unlock(&group->lock);
+		release_group(group, 0);
+		return 1; /* consumed */
+	case IB_MGMT_METHOD_SET:
+	case IB_SA_METHOD_GET_TABLE:
+	case IB_SA_METHOD_GET_TABLE_RESP:
+	case IB_SA_METHOD_DELETE:
+		return 0; /* not consumed, pass-through to guest over tunnel */
+	default:
+		mcg_warn("In demux, port %d: unexpected MCMember method: 0x%x, dropping\n",
+			port, mad->mad_hdr.method);
+		return 1; /* consumed */
+	}
+}
+
+int mlx4_ib_mcg_multiplex_handler(struct ib_device *ibdev, int port,
+				  int slave, struct ib_sa_mad *sa_mad)
+{
+	struct mlx4_ib_dev *dev = to_mdev(ibdev);
+	struct ib_sa_mcmember_data *rec = (struct ib_sa_mcmember_data *)sa_mad->data;
+	struct mlx4_ib_demux_ctx *ctx = &dev->sriov.demux[port - 1];
+	struct mcast_group *group;
+	struct mcast_req *req;
+	int may_create = 0;
+
+	if (ctx->flushing)
+		return -EAGAIN;
+
+	switch (sa_mad->mad_hdr.method) {
+	case IB_MGMT_METHOD_SET:
+		may_create = 1;
+	case IB_SA_METHOD_DELETE:
+		req = kzalloc(sizeof *req, GFP_KERNEL);
+		if (!req)
+			return -ENOMEM;
+
+		req->func = slave;
+		req->sa_mad = *sa_mad;
+
+		mutex_lock(&ctx->mcg_table_lock);
+		group = acquire_group(ctx, &rec->mgid, may_create, GFP_KERNEL);
+		mutex_unlock(&ctx->mcg_table_lock);
+		if (IS_ERR(group)) {
+			kfree(req);
+			return PTR_ERR(group);
+		}
+		mutex_lock(&group->lock);
+		if (group->func[slave].num_pend_reqs > MAX_PEND_REQS_PER_FUNC) {
+			mutex_unlock(&group->lock);
+			mcg_warn_group(group, "Port %d, Func %d has too many pending requests (%d), dropping\n",
+				       port, slave, MAX_PEND_REQS_PER_FUNC);
+			release_group(group, 0);
+			kfree(req);
+			return -ENOMEM;
+		}
+		++group->func[slave].num_pend_reqs;
+		req->group = group;
+		queue_req(req);
+		mutex_unlock(&group->lock);
+		release_group(group, 0);
+		return 1; /* consumed */
+	case IB_SA_METHOD_GET_TABLE:
+	case IB_MGMT_METHOD_GET_RESP:
+	case IB_SA_METHOD_GET_TABLE_RESP:
+	case IB_SA_METHOD_DELETE_RESP:
+		return 0; /* not consumed, pass-through */
+	default:
+		mcg_warn("In multiplex, port %d, func %d: unexpected MCMember method: 0x%x, dropping\n",
+			port, slave, sa_mad->mad_hdr.method);
+		return 1; /* consumed */
+	}
+}
+
+int mlx4_ib_mcg_port_init(struct mlx4_ib_demux_ctx *ctx)
+{
+	char name[20];
+
+	atomic_set(&ctx->tid, 0);
+	sprintf(name, "mlx4_ib_mcg%d", ctx->port);
+	ctx->mcg_wq = create_singlethread_workqueue(name);
+	if (!ctx->mcg_wq)
+		return -ENOMEM;
+
+	mutex_init(&ctx->mcg_table_lock);
+	ctx->mcg_table = RB_ROOT;
+	INIT_LIST_HEAD(&ctx->mcg_mgid0_list);
+	ctx->flushing = 0;
+
+	return 0;
+}
+
+static void force_clean_group(struct mcast_group *group)
+{
+	struct mcast_req *req, *tmp
+		;
+	list_for_each_entry_safe(req, tmp, &group->pending_list, group_list) {
+		list_del(&req->group_list);
+		kfree(req);
+	}
+	rb_erase(&group->node, &group->demux->mcg_table);
+	kfree(group);
+}
+
+static void _mlx4_ib_mcg_port_cleanup(struct mlx4_ib_demux_ctx *ctx, int destroy_wq)
+{
+	int i;
+	struct rb_node *p;
+	struct mcast_group *group;
+	unsigned long end;
+	int count;
+
+	if (ctx->flushing)
+		return;
+
+	ctx->flushing = 1;
+	for (i = 0; i < MAX_VFS; ++i)
+		clean_vf_mcast(ctx, i);
+
+	end = jiffies + msecs_to_jiffies(MAD_TIMEOUT_MS + 3000);
+	do {
+		count = 0;
+		mutex_lock(&ctx->mcg_table_lock);
+		for (p = rb_first(&ctx->mcg_table); p; p = rb_next(p))
+			++count;
+		mutex_unlock(&ctx->mcg_table_lock);
+		if (!count)
+			break;
+
+		msleep(1);
+	} while (time_after(end, jiffies));
+
+	flush_workqueue(ctx->mcg_wq);
+	if (destroy_wq)
+		destroy_workqueue(ctx->mcg_wq);
+
+	mutex_lock(&ctx->mcg_table_lock);
+	while ((p = rb_first(&ctx->mcg_table)) != NULL) {
+		group = rb_entry(p, struct mcast_group, node);
+		if (atomic_read(&group->refcount))
+			mcg_warn_group(group, "group refcount %d!!! (pointer %p)\n", atomic_read(&group->refcount), group);
+
+		force_clean_group(group);
+	}
+	mutex_unlock(&ctx->mcg_table_lock);
+
+	if (!destroy_wq)
+		ctx->flushing = 0;
+}
+
+struct clean_work {
+	struct work_struct work;
+	struct mlx4_ib_demux_ctx *ctx;
+	int destroy_wq;
+};
+
+static void mcg_clean_task(struct work_struct *work)
+{
+	struct clean_work *cw = container_of(work, struct clean_work, work);
+
+	_mlx4_ib_mcg_port_cleanup(cw->ctx, cw->destroy_wq);
+	kfree(cw);
+}
+
+void mlx4_ib_mcg_port_cleanup(struct mlx4_ib_demux_ctx *ctx, int destroy_wq)
+{
+	struct clean_work *work;
+
+	if (destroy_wq) {
+		_mlx4_ib_mcg_port_cleanup(ctx, destroy_wq);
+		return;
+	}
+
+	work = kmalloc(sizeof *work, GFP_KERNEL);
+	if (!work) {
+		mcg_warn("failed allocating work for cleanup\n");
+		return;
+	}
+
+	work->ctx = ctx;
+	work->destroy_wq = destroy_wq;
+	INIT_WORK(&work->work, mcg_clean_task);
+	queue_work(clean_wq, &work->work);
+}
+
+static void build_leave_mad(struct mcast_req *req)
+{
+	struct ib_sa_mad *mad = &req->sa_mad;
+
+	mad->mad_hdr.method = IB_SA_METHOD_DELETE;
+}
+
+
+static void clear_pending_reqs(struct mcast_group *group, int vf)
+{
+	struct mcast_req *req, *tmp, *group_first = NULL;
+	int clear;
+	int pend = 0;
+
+	if (!list_empty(&group->pending_list))
+		group_first = list_first_entry(&group->pending_list, struct mcast_req, group_list);
+
+	list_for_each_entry_safe(req, tmp, &group->func[vf].pending, func_list) {
+		clear = 1;
+		if (group_first == req &&
+		    (group->state == MCAST_JOIN_SENT ||
+		     group->state == MCAST_LEAVE_SENT)) {
+			clear = cancel_delayed_work(&group->timeout_work);
+			pend = !clear;
+			group->state = MCAST_IDLE;
+		}
+		if (clear) {
+			--group->func[vf].num_pend_reqs;
+			list_del(&req->group_list);
+			list_del(&req->func_list);
+			kfree(req);
+			atomic_dec(&group->refcount);
+		}
+	}
+
+	if (!pend && (!list_empty(&group->func[vf].pending) || group->func[vf].num_pend_reqs)) {
+		mcg_warn_group(group, "DRIVER BUG: list_empty %d, num_pend_reqs %d\n",
+			       list_empty(&group->func[vf].pending), group->func[vf].num_pend_reqs);
+	}
+}
+
+static int push_deleteing_req(struct mcast_group *group, int slave)
+{
+	struct mcast_req *req;
+	struct mcast_req *pend_req;
+
+	if (!group->func[slave].join_state)
+		return 0;
+
+	req = kzalloc(sizeof *req, GFP_KERNEL);
+	if (!req) {
+		mcg_warn_group(group, "failed allocation - may leave stall groups\n");
+		return -ENOMEM;
+	}
+
+	if (!list_empty(&group->func[slave].pending)) {
+		pend_req = list_entry(group->func[slave].pending.prev, struct mcast_req, group_list);
+		if (pend_req->clean) {
+			kfree(req);
+			return 0;
+		}
+	}
+
+	req->clean = 1;
+	req->func = slave;
+	req->group = group;
+	++group->func[slave].num_pend_reqs;
+	build_leave_mad(req);
+	queue_req(req);
+	return 0;
+}
+
+void clean_vf_mcast(struct mlx4_ib_demux_ctx *ctx, int slave)
+{
+	struct mcast_group *group;
+	struct rb_node *p;
+
+	mutex_lock(&ctx->mcg_table_lock);
+	for (p = rb_first(&ctx->mcg_table); p; p = rb_next(p)) {
+		group = rb_entry(p, struct mcast_group, node);
+		mutex_lock(&group->lock);
+		if (atomic_read(&group->refcount)) {
+			/* clear pending requests of this VF */
+			clear_pending_reqs(group, slave);
+			push_deleteing_req(group, slave);
+		}
+		mutex_unlock(&group->lock);
+	}
+	mutex_unlock(&ctx->mcg_table_lock);
+}
+
+
+int mlx4_ib_mcg_init(void)
+{
+	clean_wq = create_singlethread_workqueue("mlx4_ib_mcg");
+	if (!clean_wq)
+		return -ENOMEM;
+
+	return 0;
+}
+
+void mlx4_ib_mcg_destroy(void)
+{
+	destroy_workqueue(clean_wq);
+}
diff --git a/drivers/infiniband/hw/mlx4/mlx4_ib.h b/drivers/infiniband/hw/mlx4/mlx4_ib.h
index 348e323..ab1d7f9 100644
--- a/drivers/infiniband/hw/mlx4/mlx4_ib.h
+++ b/drivers/infiniband/hw/mlx4/mlx4_ib.h
@@ -37,9 +37,11 @@
 #include <linux/compiler.h>
 #include <linux/list.h>
 #include <linux/mutex.h>
+#include <linux/idr.h>
 
 #include <rdma/ib_verbs.h>
 #include <rdma/ib_umem.h>
+#include <rdma/ib_mad.h>
 
 #include <linux/mlx4/device.h>
 #include <linux/mlx4/doorbell.h>
@@ -328,7 +330,14 @@ struct mlx4_ib_demux_ctx {
 	__be64 subnet_prefix;
 	__be64 guid_cache[128];
 	struct mlx4_ib_dev *dev;
+	/* the following lock protects both mcg_table and mcg_mgid0_list */
+	struct mutex		mcg_table_lock;
+	struct rb_root		mcg_table;
+	struct list_head	mcg_mgid0_list;
+	struct workqueue_struct	*mcg_wq;
 	struct mlx4_ib_demux_pv_ctx **tun;
+	atomic_t tid;
+	int    flushing; /* flushing the work queue */
 };
 
 struct mlx4_ib_sriov {
@@ -552,6 +561,19 @@ static inline int mlx4_ib_ah_grh_present(struct mlx4_ib_ah *ah)
 	return !!(ah->av.ib.g_slid & 0x80);
 }
 
+int mlx4_ib_mcg_port_init(struct mlx4_ib_demux_ctx *ctx);
+void mlx4_ib_mcg_port_cleanup(struct mlx4_ib_demux_ctx *ctx, int destroy_wq);
+void clean_vf_mcast(struct mlx4_ib_demux_ctx *ctx, int slave);
+int mlx4_ib_mcg_init(void);
+void mlx4_ib_mcg_destroy(void);
+
+int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid);
+
+int mlx4_ib_mcg_multiplex_handler(struct ib_device *ibdev, int port, int slave,
+				  struct ib_sa_mad *sa_mad);
+int mlx4_ib_mcg_demux_handler(struct ib_device *ibdev, int port, int slave,
+			      struct ib_sa_mad *mad);
+
 int mlx4_ib_add_mc(struct mlx4_ib_dev *mdev, struct mlx4_ib_qp *mqp,
 		   union ib_gid *gid);
 
@@ -560,4 +582,12 @@ void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
 
 void mlx4_ib_tunnels_update_work(struct work_struct *work);
 
+int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
+			  enum ib_qp_type qpt, struct ib_wc *wc,
+			  struct ib_grh *grh, struct ib_mad *mad);
+int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
+			 enum ib_qp_type dest_qpt, u16 pkey_index, u32 remote_qpn,
+			 u32 qkey, struct ib_ah_attr *attr, struct ib_mad *mad);
+__be64 mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx *ctx);
+
 #endif /* MLX4_IB_H */
-- 
1.7.1

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html