Merge /spare/repo/linux-2.6/

48 files changed, 12521 insertions, 1333 deletions

diff --git a/drivers/infiniband/Kconfig b/drivers/infiniband/Kconfig
index 3cc3ff0cccb1..79c8e2dd9c33 100644
--- a/drivers/infiniband/Kconfig
+++ b/drivers/infiniband/Kconfig
@@ -7,6 +7,16 @@ config INFINIBAND
 	  any protocols you wish to use as well as drivers for your
 	  InfiniBand hardware.
 
+config INFINIBAND_USER_VERBS
+	tristate "InfiniBand userspace verbs support"
+	depends on INFINIBAND
+	---help---
+	  Userspace InfiniBand verbs support.  This is the kernel side
+	  of userspace verbs, which allows userspace processes to
+	  directly access InfiniBand hardware for fast-path
+	  operations.  You will also need libibverbs and a hardware
+	  driver library from <http://www.openib.org>.
+
 source "drivers/infiniband/hw/mthca/Kconfig"
 
 source "drivers/infiniband/ulp/ipoib/Kconfig"
diff --git a/drivers/infiniband/core/Makefile b/drivers/infiniband/core/Makefile
index d2dbfb52c0a3..10be36731ed7 100644
--- a/drivers/infiniband/core/Makefile
+++ b/drivers/infiniband/core/Makefile
@@ -1,12 +1,20 @@
 EXTRA_CFLAGS += -Idrivers/infiniband/include
 
-obj-$(CONFIG_INFINIBAND) +=	ib_core.o ib_mad.o ib_sa.o ib_umad.o
+obj-$(CONFIG_INFINIBAND) +=		ib_core.o ib_mad.o ib_sa.o \
+					ib_cm.o ib_umad.o ib_ucm.o
+obj-$(CONFIG_INFINIBAND_USER_VERBS) +=	ib_uverbs.o
 
 ib_core-y :=			packer.o ud_header.o verbs.o sysfs.o \
 				device.o fmr_pool.o cache.o
 
-ib_mad-y :=			mad.o smi.o agent.o
+ib_mad-y :=			mad.o smi.o agent.o mad_rmpp.o
 
 ib_sa-y :=			sa_query.o
 
+ib_cm-y :=			cm.o
+
 ib_umad-y :=			user_mad.o
+
+ib_ucm-y :=			ucm.o
+
+ib_uverbs-y :=			uverbs_main.o uverbs_cmd.o uverbs_mem.o
diff --git a/drivers/infiniband/core/agent.c b/drivers/infiniband/core/agent.c
index 23d1957c4b29..729f0b0d983a 100644
--- a/drivers/infiniband/core/agent.c
+++ b/drivers/infiniband/core/agent.c
@@ -134,7 +134,7 @@ static int agent_mad_send(struct ib_mad_agent *mad_agent,
 					  sizeof(mad_priv->mad),
 					  DMA_TO_DEVICE);
 	gather_list.length = sizeof(mad_priv->mad);
-	gather_list.lkey = (*port_priv->mr).lkey;
+	gather_list.lkey = mad_agent->mr->lkey;
 
 	send_wr.next = NULL;
 	send_wr.opcode = IB_WR_SEND;
@@ -156,10 +156,10 @@ static int agent_mad_send(struct ib_mad_agent *mad_agent,
 			/* Should sgid be looked up ? */
 			ah_attr.grh.sgid_index = 0;
 			ah_attr.grh.hop_limit = grh->hop_limit;
-			ah_attr.grh.flow_label = be32_to_cpup(
-				&grh->version_tclass_flow)  & 0xfffff;
-			ah_attr.grh.traffic_class = (be32_to_cpup(
-				&grh->version_tclass_flow) >> 20) & 0xff;
+			ah_attr.grh.flow_label = be32_to_cpu(
+				grh->version_tclass_flow)  & 0xfffff;
+			ah_attr.grh.traffic_class = (be32_to_cpu(
+				grh->version_tclass_flow) >> 20) & 0xff;
 			memcpy(ah_attr.grh.dgid.raw,
 			       grh->sgid.raw,
 			       sizeof(ah_attr.grh.dgid));
@@ -322,22 +322,12 @@ int ib_agent_port_open(struct ib_device *device, int port_num)
 		goto error3;
 	}
 
-	port_priv->mr = ib_get_dma_mr(port_priv->smp_agent->qp->pd,
-				      IB_ACCESS_LOCAL_WRITE);
-	if (IS_ERR(port_priv->mr)) {
-		printk(KERN_ERR SPFX "Couldn't get DMA MR\n");
-		ret = PTR_ERR(port_priv->mr);
-		goto error4;
-	}
-
 	spin_lock_irqsave(&ib_agent_port_list_lock, flags);
 	list_add_tail(&port_priv->port_list, &ib_agent_port_list);
 	spin_unlock_irqrestore(&ib_agent_port_list_lock, flags);
 
 	return 0;
 
-error4:
-	ib_unregister_mad_agent(port_priv->perf_mgmt_agent);
 error3:
 	ib_unregister_mad_agent(port_priv->smp_agent);
 error2:
@@ -361,8 +351,6 @@ int ib_agent_port_close(struct ib_device *device, int port_num)
 	list_del(&port_priv->port_list);
 	spin_unlock_irqrestore(&ib_agent_port_list_lock, flags);
 
-	ib_dereg_mr(port_priv->mr);
-
 	ib_unregister_mad_agent(port_priv->perf_mgmt_agent);
 	ib_unregister_mad_agent(port_priv->smp_agent);
 	kfree(port_priv);
diff --git a/drivers/infiniband/core/agent_priv.h b/drivers/infiniband/core/agent_priv.h
index 17a0cce5813c..17435af1e914 100644
--- a/drivers/infiniband/core/agent_priv.h
+++ b/drivers/infiniband/core/agent_priv.h
@@ -33,7 +33,7 @@
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  *
- * $Id: agent_priv.h 1389 2004-12-27 22:56:47Z roland $
+ * $Id: agent_priv.h 1640 2005-01-24 22:39:02Z halr $
  */
 
 #ifndef __IB_AGENT_PRIV_H__
@@ -57,7 +57,6 @@ struct ib_agent_port_private {
 	int port_num;
 	struct ib_mad_agent *smp_agent;	      /* SM class */
 	struct ib_mad_agent *perf_mgmt_agent; /* PerfMgmt class */
-	struct ib_mr *mr;
 };
 
 #endif	/* __IB_AGENT_PRIV_H__ */
diff --git a/drivers/infiniband/core/cm.c b/drivers/infiniband/core/cm.c
new file mode 100644
index 000000000000..403ed125d8f4
--- /dev/null
+++ b/drivers/infiniband/core/cm.c
@@ -0,0 +1,3324 @@
+/*
+ * Copyright (c) 2004, 2005 Intel Corporation.  All rights reserved.
+ * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
+ * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. 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.
+ *
+ * $Id: cm.c 2821 2005-07-08 17:07:28Z sean.hefty $
+ */
+#include <linux/dma-mapping.h>
+#include <linux/err.h>
+#include <linux/idr.h>
+#include <linux/interrupt.h>
+#include <linux/pci.h>
+#include <linux/rbtree.h>
+#include <linux/spinlock.h>
+#include <linux/workqueue.h>
+
+#include <ib_cache.h>
+#include <ib_cm.h>
+#include "cm_msgs.h"
+
+MODULE_AUTHOR("Sean Hefty");
+MODULE_DESCRIPTION("InfiniBand CM");
+MODULE_LICENSE("Dual BSD/GPL");
+
+static void cm_add_one(struct ib_device *device);
+static void cm_remove_one(struct ib_device *device);
+
+static struct ib_client cm_client = {
+	.name   = "cm",
+	.add    = cm_add_one,
+	.remove = cm_remove_one
+};
+
+static struct ib_cm {
+	spinlock_t lock;
+	struct list_head device_list;
+	rwlock_t device_lock;
+	struct rb_root listen_service_table;
+	u64 listen_service_id;
+	/* struct rb_root peer_service_table; todo: fix peer to peer */
+	struct rb_root remote_qp_table;
+	struct rb_root remote_id_table;
+	struct rb_root remote_sidr_table;
+	struct idr local_id_table;
+	struct workqueue_struct *wq;
+} cm;
+
+struct cm_port {
+	struct cm_device *cm_dev;
+	struct ib_mad_agent *mad_agent;
+	u8 port_num;
+};
+
+struct cm_device {
+	struct list_head list;
+	struct ib_device *device;
+	u64 ca_guid;
+	struct cm_port port[0];
+};
+
+struct cm_av {
+	struct cm_port *port;
+	union ib_gid dgid;
+	struct ib_ah_attr ah_attr;
+	u16 pkey_index;
+	u8 packet_life_time;
+};
+
+struct cm_work {
+	struct work_struct work;
+	struct list_head list;
+	struct cm_port *port;
+	struct ib_mad_recv_wc *mad_recv_wc;	/* Received MADs */
+	u32 local_id;				/* Established / timewait */
+	u32 remote_id;
+	struct ib_cm_event cm_event;
+	struct ib_sa_path_rec path[0];
+};
+
+struct cm_timewait_info {
+	struct cm_work work;			/* Must be first. */
+	struct rb_node remote_qp_node;
+	struct rb_node remote_id_node;
+	u64 remote_ca_guid;
+	u32 remote_qpn;
+	u8 inserted_remote_qp;
+	u8 inserted_remote_id;
+};
+
+struct cm_id_private {
+	struct ib_cm_id	id;
+
+	struct rb_node service_node;
+	struct rb_node sidr_id_node;
+	spinlock_t lock;
+	wait_queue_head_t wait;
+	atomic_t refcount;
+
+	struct ib_mad_send_buf *msg;
+	struct cm_timewait_info *timewait_info;
+	/* todo: use alternate port on send failure */
+	struct cm_av av;
+	struct cm_av alt_av;
+
+	void *private_data;
+	u64 tid;
+	u32 local_qpn;
+	u32 remote_qpn;
+	u32 sq_psn;
+	u32 rq_psn;
+	int timeout_ms;
+	enum ib_mtu path_mtu;
+	u8 private_data_len;
+	u8 max_cm_retries;
+	u8 peer_to_peer;
+	u8 responder_resources;
+	u8 initiator_depth;
+	u8 local_ack_timeout;
+	u8 retry_count;
+	u8 rnr_retry_count;
+	u8 service_timeout;
+
+	struct list_head work_list;
+	atomic_t work_count;
+};
+
+static void cm_work_handler(void *data);
+
+static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
+{
+	if (atomic_dec_and_test(&cm_id_priv->refcount))
+		wake_up(&cm_id_priv->wait);
+}
+
+static int cm_alloc_msg(struct cm_id_private *cm_id_priv,
+			struct ib_mad_send_buf **msg)
+{
+	struct ib_mad_agent *mad_agent;
+	struct ib_mad_send_buf *m;
+	struct ib_ah *ah;
+
+	mad_agent = cm_id_priv->av.port->mad_agent;
+	ah = ib_create_ah(mad_agent->qp->pd, &cm_id_priv->av.ah_attr);
+	if (IS_ERR(ah))
+		return PTR_ERR(ah);
+
+	m = ib_create_send_mad(mad_agent, 1, cm_id_priv->av.pkey_index,
+			       ah, 0, sizeof(struct ib_mad_hdr),
+			       sizeof(struct ib_mad)-sizeof(struct ib_mad_hdr),
+			       GFP_ATOMIC);
+	if (IS_ERR(m)) {
+		ib_destroy_ah(ah);
+		return PTR_ERR(m);
+	}
+
+	/* Timeout set by caller if response is expected. */
+	m->send_wr.wr.ud.retries = cm_id_priv->max_cm_retries;
+
+	atomic_inc(&cm_id_priv->refcount);
+	m->context[0] = cm_id_priv;
+	*msg = m;
+	return 0;
+}
+
+static int cm_alloc_response_msg(struct cm_port *port,
+				 struct ib_mad_recv_wc *mad_recv_wc,
+				 struct ib_mad_send_buf **msg)
+{
+	struct ib_mad_send_buf *m;
+	struct ib_ah *ah;
+
+	ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
+				  mad_recv_wc->recv_buf.grh, port->port_num);
+	if (IS_ERR(ah))
+		return PTR_ERR(ah);
+
+	m = ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
+			       ah, 0, sizeof(struct ib_mad_hdr),
+			       sizeof(struct ib_mad)-sizeof(struct ib_mad_hdr),
+			       GFP_ATOMIC);
+	if (IS_ERR(m)) {
+		ib_destroy_ah(ah);
+		return PTR_ERR(m);
+	}
+	*msg = m;
+	return 0;
+}
+
+static void cm_free_msg(struct ib_mad_send_buf *msg)
+{
+	ib_destroy_ah(msg->send_wr.wr.ud.ah);
+	if (msg->context[0])
+		cm_deref_id(msg->context[0]);
+	ib_free_send_mad(msg);
+}
+
+static void * cm_copy_private_data(const void *private_data,
+				   u8 private_data_len)
+{
+	void *data;
+
+	if (!private_data || !private_data_len)
+		return NULL;
+
+	data = kmalloc(private_data_len, GFP_KERNEL);
+	if (!data)
+		return ERR_PTR(-ENOMEM);
+
+	memcpy(data, private_data, private_data_len);
+	return data;
+}
+
+static void cm_set_private_data(struct cm_id_private *cm_id_priv,
+				 void *private_data, u8 private_data_len)
+{
+	if (cm_id_priv->private_data && cm_id_priv->private_data_len)
+		kfree(cm_id_priv->private_data);
+
+	cm_id_priv->private_data = private_data;
+	cm_id_priv->private_data_len = private_data_len;
+}
+
+static void cm_set_ah_attr(struct ib_ah_attr *ah_attr, u8 port_num,
+			   u16 dlid, u8 sl, u16 src_path_bits)
+{
+	memset(ah_attr, 0, sizeof ah_attr);
+	ah_attr->dlid = be16_to_cpu(dlid);
+	ah_attr->sl = sl;
+	ah_attr->src_path_bits = src_path_bits;
+	ah_attr->port_num = port_num;
+}
+
+static void cm_init_av_for_response(struct cm_port *port,
+				    struct ib_wc *wc, struct cm_av *av)
+{
+	av->port = port;
+	av->pkey_index = wc->pkey_index;
+	cm_set_ah_attr(&av->ah_attr, port->port_num, cpu_to_be16(wc->slid),
+		       wc->sl, wc->dlid_path_bits);
+}
+
+static int cm_init_av_by_path(struct ib_sa_path_rec *path, struct cm_av *av)
+{
+	struct cm_device *cm_dev;
+	struct cm_port *port = NULL;
+	unsigned long flags;
+	int ret;
+	u8 p;
+
+	read_lock_irqsave(&cm.device_lock, flags);
+	list_for_each_entry(cm_dev, &cm.device_list, list) {
+		if (!ib_find_cached_gid(cm_dev->device, &path->sgid,
+					&p, NULL)) {
+			port = &cm_dev->port[p-1];
+			break;
+		}
+	}
+	read_unlock_irqrestore(&cm.device_lock, flags);
+
+	if (!port)
+		return -EINVAL;
+
+	ret = ib_find_cached_pkey(cm_dev->device, port->port_num,
+				  be16_to_cpu(path->pkey), &av->pkey_index);
+	if (ret)
+		return ret;
+
+	av->port = port;
+	cm_set_ah_attr(&av->ah_attr, av->port->port_num, path->dlid,
+		       path->sl, path->slid & 0x7F);
+	av->packet_life_time = path->packet_life_time;
+	return 0;
+}
+
+static int cm_alloc_id(struct cm_id_private *cm_id_priv)
+{
+	unsigned long flags;
+	int ret;
+
+	do {
+		spin_lock_irqsave(&cm.lock, flags);
+		ret = idr_get_new_above(&cm.local_id_table, cm_id_priv, 1,
+					(int *) &cm_id_priv->id.local_id);
+		spin_unlock_irqrestore(&cm.lock, flags);
+	} while( (ret == -EAGAIN) && idr_pre_get(&cm.local_id_table, GFP_KERNEL) );
+	return ret;
+}
+
+static void cm_free_id(u32 local_id)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&cm.lock, flags);
+	idr_remove(&cm.local_id_table, (int) local_id);
+	spin_unlock_irqrestore(&cm.lock, flags);
+}
+
+static struct cm_id_private * cm_get_id(u32 local_id, u32 remote_id)
+{
+	struct cm_id_private *cm_id_priv;
+
+	cm_id_priv = idr_find(&cm.local_id_table, (int) local_id);
+	if (cm_id_priv) {
+		if (cm_id_priv->id.remote_id == remote_id)
+			atomic_inc(&cm_id_priv->refcount);
+		else
+			cm_id_priv = NULL;
+	}
+
+	return cm_id_priv;
+}
+
+static struct cm_id_private * cm_acquire_id(u32 local_id, u32 remote_id)
+{
+	struct cm_id_private *cm_id_priv;
+	unsigned long flags;
+
+	spin_lock_irqsave(&cm.lock, flags);
+	cm_id_priv = cm_get_id(local_id, remote_id);
+	spin_unlock_irqrestore(&cm.lock, flags);
+
+	return cm_id_priv;
+}
+
+static struct cm_id_private * cm_insert_listen(struct cm_id_private *cm_id_priv)
+{
+	struct rb_node **link = &cm.listen_service_table.rb_node;
+	struct rb_node *parent = NULL;
+	struct cm_id_private *cur_cm_id_priv;
+	u64 service_id = cm_id_priv->id.service_id;
+	u64 service_mask = cm_id_priv->id.service_mask;
+
+	while (*link) {
+		parent = *link;
+		cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
+					  service_node);
+		if ((cur_cm_id_priv->id.service_mask & service_id) ==
+		    (service_mask & cur_cm_id_priv->id.service_id))
+			return cm_id_priv;
+		if (service_id < cur_cm_id_priv->id.service_id)
+			link = &(*link)->rb_left;
+		else
+			link = &(*link)->rb_right;
+	}
+	rb_link_node(&cm_id_priv->service_node, parent, link);
+	rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
+	return NULL;
+}
+
+static struct cm_id_private * cm_find_listen(u64 service_id)
+{
+	struct rb_node *node = cm.listen_service_table.rb_node;
+	struct cm_id_private *cm_id_priv;
+
+	while (node) {
+		cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
+		if ((cm_id_priv->id.service_mask & service_id) ==
+		    (cm_id_priv->id.service_mask & cm_id_priv->id.service_id))
+			return cm_id_priv;
+		if (service_id < cm_id_priv->id.service_id)
+			node = node->rb_left;
+		else
+			node = node->rb_right;
+	}
+	return NULL;
+}
+
+static struct cm_timewait_info * cm_insert_remote_id(struct cm_timewait_info
+						     *timewait_info)
+{
+	struct rb_node **link = &cm.remote_id_table.rb_node;
+	struct rb_node *parent = NULL;
+	struct cm_timewait_info *cur_timewait_info;
+	u64 remote_ca_guid = timewait_info->remote_ca_guid;
+	u32 remote_id = timewait_info->work.remote_id;
+
+	while (*link) {
+		parent = *link;
+		cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
+					     remote_id_node);
+		if (remote_id < cur_timewait_info->work.remote_id)
+			link = &(*link)->rb_left;
+		else if (remote_id > cur_timewait_info->work.remote_id)
+			link = &(*link)->rb_right;
+		else if (remote_ca_guid < cur_timewait_info->remote_ca_guid)
+			link = &(*link)->rb_left;
+		else if (remote_ca_guid > cur_timewait_info->remote_ca_guid)
+			link = &(*link)->rb_right;
+		else
+			return cur_timewait_info;
+	}
+	timewait_info->inserted_remote_id = 1;
+	rb_link_node(&timewait_info->remote_id_node, parent, link);
+	rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
+	return NULL;
+}
+
+static struct cm_timewait_info * cm_find_remote_id(u64 remote_ca_guid,
+						   u32 remote_id)
+{
+	struct rb_node *node = cm.remote_id_table.rb_node;
+	struct cm_timewait_info *timewait_info;
+
+	while (node) {
+		timewait_info = rb_entry(node, struct cm_timewait_info,
+					 remote_id_node);
+		if (remote_id < timewait_info->work.remote_id)
+			node = node->rb_left;
+		else if (remote_id > timewait_info->work.remote_id)
+			node = node->rb_right;
+		else if (remote_ca_guid < timewait_info->remote_ca_guid)
+			node = node->rb_left;
+		else if (remote_ca_guid > timewait_info->remote_ca_guid)
+			node = node->rb_right;
+		else
+			return timewait_info;
+	}
+	return NULL;
+}
+
+static struct cm_timewait_info * cm_insert_remote_qpn(struct cm_timewait_info
+						      *timewait_info)
+{
+	struct rb_node **link = &cm.remote_qp_table.rb_node;
+	struct rb_node *parent = NULL;
+	struct cm_timewait_info *cur_timewait_info;
+	u64 remote_ca_guid = timewait_info->remote_ca_guid;
+	u32 remote_qpn = timewait_info->remote_qpn;
+
+	while (*link) {
+		parent = *link;
+		cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
+					     remote_qp_node);
+		if (remote_qpn < cur_timewait_info->remote_qpn)
+			link = &(*link)->rb_left;
+		else if (remote_qpn > cur_timewait_info->remote_qpn)
+			link = &(*link)->rb_right;
+		else if (remote_ca_guid < cur_timewait_info->remote_ca_guid)
+			link = &(*link)->rb_left;
+		else if (remote_ca_guid > cur_timewait_info->remote_ca_guid)
+			link = &(*link)->rb_right;
+		else
+			return cur_timewait_info;
+	}
+	timewait_info->inserted_remote_qp = 1;
+	rb_link_node(&timewait_info->remote_qp_node, parent, link);
+	rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
+	return NULL;
+}
+
+static struct cm_id_private * cm_insert_remote_sidr(struct cm_id_private
+						    *cm_id_priv)
+{
+	struct rb_node **link = &cm.remote_sidr_table.rb_node;
+	struct rb_node *parent = NULL;
+	struct cm_id_private *cur_cm_id_priv;
+	union ib_gid *port_gid = &cm_id_priv->av.dgid;
+	u32 remote_id = cm_id_priv->id.remote_id;
+
+	while (*link) {
+		parent = *link;
+		cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
+					  sidr_id_node);
+		if (remote_id < cur_cm_id_priv->id.remote_id)
+			link = &(*link)->rb_left;
+		else if (remote_id > cur_cm_id_priv->id.remote_id)
+			link = &(*link)->rb_right;
+		else {
+			int cmp;
+			cmp = memcmp(port_gid, &cur_cm_id_priv->av.dgid,
+				     sizeof *port_gid);
+			if (cmp < 0)
+				link = &(*link)->rb_left;
+			else if (cmp > 0)
+				link = &(*link)->rb_right;
+			else
+				return cur_cm_id_priv;
+		}
+	}
+	rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
+	rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
+	return NULL;
+}
+
+static void cm_reject_sidr_req(struct cm_id_private *cm_id_priv,
+			       enum ib_cm_sidr_status status)
+{
+	struct ib_cm_sidr_rep_param param;
+
+	memset(&param, 0, sizeof param);
+	param.status = status;
+	ib_send_cm_sidr_rep(&cm_id_priv->id, &param);
+}
+
+struct ib_cm_id *ib_create_cm_id(ib_cm_handler cm_handler,
+				 void *context)
+{
+	struct cm_id_private *cm_id_priv;
+	int ret;
+
+	cm_id_priv = kmalloc(sizeof *cm_id_priv, GFP_KERNEL);
+	if (!cm_id_priv)
+		return ERR_PTR(-ENOMEM);
+
+	memset(cm_id_priv, 0, sizeof *cm_id_priv);
+	cm_id_priv->id.state = IB_CM_IDLE;
+	cm_id_priv->id.cm_handler = cm_handler;
+	cm_id_priv->id.context = context;
+	ret = cm_alloc_id(cm_id_priv);
+	if (ret)
+		goto error;
+
+	spin_lock_init(&cm_id_priv->lock);
+	init_waitqueue_head(&cm_id_priv->wait);
+	INIT_LIST_HEAD(&cm_id_priv->work_list);
+	atomic_set(&cm_id_priv->work_count, -1);
+	atomic_set(&cm_id_priv->refcount, 1);
+	return &cm_id_priv->id;
+
+error:
+	kfree(cm_id_priv);
+	return ERR_PTR(-ENOMEM);
+}
+EXPORT_SYMBOL(ib_create_cm_id);
+
+static struct cm_work * cm_dequeue_work(struct cm_id_private *cm_id_priv)
+{
+	struct cm_work *work;
+
+	if (list_empty(&cm_id_priv->work_list))
+		return NULL;
+
+	work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
+	list_del(&work->list);
+	return work;
+}
+
+static void cm_free_work(struct cm_work *work)
+{
+	if (work->mad_recv_wc)
+		ib_free_recv_mad(work->mad_recv_wc);
+	kfree(work);
+}
+
+static inline int cm_convert_to_ms(int iba_time)
+{
+	/* approximate conversion to ms from 4.096us x 2^iba_time */
+	return 1 << max(iba_time - 8, 0);
+}
+
+static void cm_cleanup_timewait(struct cm_timewait_info *timewait_info)
+{
+	unsigned long flags;
+
+	if (!timewait_info->inserted_remote_id &&
+	    !timewait_info->inserted_remote_qp)
+	    return;
+
+	spin_lock_irqsave(&cm.lock, flags);
+	if (timewait_info->inserted_remote_id) {
+		rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
+		timewait_info->inserted_remote_id = 0;
+	}
+
+	if (timewait_info->inserted_remote_qp) {
+		rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
+		timewait_info->inserted_remote_qp = 0;
+	}
+	spin_unlock_irqrestore(&cm.lock, flags);
+}
+
+static struct cm_timewait_info * cm_create_timewait_info(u32 local_id)
+{
+	struct cm_timewait_info *timewait_info;
+
+	timewait_info = kmalloc(sizeof *timewait_info, GFP_KERNEL);
+	if (!timewait_info)
+		return ERR_PTR(-ENOMEM);
+	memset(timewait_info, 0, sizeof *timewait_info);
+
+	timewait_info->work.local_id = local_id;
+	INIT_WORK(&timewait_info->work.work, cm_work_handler,
+		  &timewait_info->work);
+	timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
+	return timewait_info;
+}
+
+static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
+{
+	int wait_time;
+
+	/*
+	 * The cm_id could be destroyed by the user before we exit timewait.
+	 * To protect against this, we search for the cm_id after exiting
+	 * timewait before notifying the user that we've exited timewait.
+	 */
+	cm_id_priv->id.state = IB_CM_TIMEWAIT;
+	wait_time = cm_convert_to_ms(cm_id_priv->local_ack_timeout);
+	queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
+			   msecs_to_jiffies(wait_time));
+	cm_id_priv->timewait_info = NULL;
+}
+
+static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
+{
+	cm_id_priv->id.state = IB_CM_IDLE;
+	if (cm_id_priv->timewait_info) {
+		cm_cleanup_timewait(cm_id_priv->timewait_info);
+		kfree(cm_id_priv->timewait_info);
+		cm_id_priv->timewait_info = NULL;
+	}
+}
+
+void ib_destroy_cm_id(struct ib_cm_id *cm_id)
+{
+	struct cm_id_private *cm_id_priv;
+	struct cm_work *work;
+	unsigned long flags;
+
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+retest:
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	switch (cm_id->state) {
+	case IB_CM_LISTEN:
+		cm_id->state = IB_CM_IDLE;
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		spin_lock_irqsave(&cm.lock, flags);
+		rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
+		spin_unlock_irqrestore(&cm.lock, flags);
+		break;
+	case IB_CM_SIDR_REQ_SENT:
+		cm_id->state = IB_CM_IDLE;
+		ib_cancel_mad(cm_id_priv->av.port->mad_agent,
+			      (unsigned long) cm_id_priv->msg);
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		break;
+	case IB_CM_SIDR_REQ_RCVD:
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		cm_reject_sidr_req(cm_id_priv, IB_SIDR_REJECT);
+		break;
+	case IB_CM_REQ_SENT:
+	case IB_CM_MRA_REQ_RCVD:
+	case IB_CM_REP_SENT:
+	case IB_CM_MRA_REP_RCVD:
+		ib_cancel_mad(cm_id_priv->av.port->mad_agent,
+			      (unsigned long) cm_id_priv->msg);
+		/* Fall through */
+	case IB_CM_REQ_RCVD:
+	case IB_CM_MRA_REQ_SENT:
+	case IB_CM_REP_RCVD:
+	case IB_CM_MRA_REP_SENT:
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		ib_send_cm_rej(cm_id, IB_CM_REJ_TIMEOUT,
+			       &cm_id_priv->av.port->cm_dev->ca_guid,
+			       sizeof cm_id_priv->av.port->cm_dev->ca_guid,
+			       NULL, 0);
+		break;
+	case IB_CM_ESTABLISHED:
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		ib_send_cm_dreq(cm_id, NULL, 0);
+		goto retest;
+	case IB_CM_DREQ_SENT:
+		ib_cancel_mad(cm_id_priv->av.port->mad_agent,
+			      (unsigned long) cm_id_priv->msg);
+		cm_enter_timewait(cm_id_priv);
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		break;
+	case IB_CM_DREQ_RCVD:
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		ib_send_cm_drep(cm_id, NULL, 0);
+		break;
+	default:
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		break;
+	}
+
+	cm_free_id(cm_id->local_id);
+	atomic_dec(&cm_id_priv->refcount);
+	wait_event(cm_id_priv->wait, !atomic_read(&cm_id_priv->refcount));
+	while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
+		cm_free_work(work);
+	if (cm_id_priv->private_data && cm_id_priv->private_data_len)
+		kfree(cm_id_priv->private_data);
+	kfree(cm_id_priv);
+}
+EXPORT_SYMBOL(ib_destroy_cm_id);
+
+int ib_cm_listen(struct ib_cm_id *cm_id,
+		 u64 service_id,
+		 u64 service_mask)
+{
+	struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
+	unsigned long flags;
+	int ret = 0;
+
+	service_mask = service_mask ? service_mask : ~0ULL;
+	service_id &= service_mask;
+	if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID &&
+	    (service_id != IB_CM_ASSIGN_SERVICE_ID))
+		return -EINVAL;
+
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+	BUG_ON(cm_id->state != IB_CM_IDLE);
+
+	cm_id->state = IB_CM_LISTEN;
+
+	spin_lock_irqsave(&cm.lock, flags);
+	if (service_id == IB_CM_ASSIGN_SERVICE_ID) {
+		cm_id->service_id = __cpu_to_be64(cm.listen_service_id++);
+		cm_id->service_mask = ~0ULL;
+	} else {
+		cm_id->service_id = service_id;
+		cm_id->service_mask = service_mask;
+	}
+	cur_cm_id_priv = cm_insert_listen(cm_id_priv);
+	spin_unlock_irqrestore(&cm.lock, flags);
+
+	if (cur_cm_id_priv) {
+		cm_id->state = IB_CM_IDLE;
+		ret = -EBUSY;
+	}
+	return ret;
+}
+EXPORT_SYMBOL(ib_cm_listen);
+
+static u64 cm_form_tid(struct cm_id_private *cm_id_priv,
+		       enum cm_msg_sequence msg_seq)
+{
+	u64 hi_tid, low_tid;
+
+	hi_tid   = ((u64) cm_id_priv->av.port->mad_agent->hi_tid) << 32;
+	low_tid  = (u64) (cm_id_priv->id.local_id | (msg_seq << 30));
+	return cpu_to_be64(hi_tid | low_tid);
+}
+
+static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
+			      enum cm_msg_attr_id attr_id, u64 tid)
+{
+	hdr->base_version  = IB_MGMT_BASE_VERSION;
+	hdr->mgmt_class	   = IB_MGMT_CLASS_CM;
+	hdr->class_version = IB_CM_CLASS_VERSION;
+	hdr->method	   = IB_MGMT_METHOD_SEND;
+	hdr->attr_id	   = attr_id;
+	hdr->tid	   = tid;
+}
+
+static void cm_format_req(struct cm_req_msg *req_msg,
+			  struct cm_id_private *cm_id_priv,
+			  struct ib_cm_req_param *param)
+{
+	cm_format_mad_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
+			  cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_REQ));
+
+	req_msg->local_comm_id = cm_id_priv->id.local_id;
+	req_msg->service_id = param->service_id;
+	req_msg->local_ca_guid = cm_id_priv->av.port->cm_dev->ca_guid;
+	cm_req_set_local_qpn(req_msg, cpu_to_be32(param->qp_num));
+	cm_req_set_resp_res(req_msg, param->responder_resources);
+	cm_req_set_init_depth(req_msg, param->initiator_depth);
+	cm_req_set_remote_resp_timeout(req_msg,
+				       param->remote_cm_response_timeout);
+	cm_req_set_qp_type(req_msg, param->qp_type);
+	cm_req_set_flow_ctrl(req_msg, param->flow_control);
+	cm_req_set_starting_psn(req_msg, cpu_to_be32(param->starting_psn));
+	cm_req_set_local_resp_timeout(req_msg,
+				      param->local_cm_response_timeout);
+	cm_req_set_retry_count(req_msg, param->retry_count);
+	req_msg->pkey = param->primary_path->pkey;
+	cm_req_set_path_mtu(req_msg, param->primary_path->mtu);
+	cm_req_set_rnr_retry_count(req_msg, param->rnr_retry_count);
+	cm_req_set_max_cm_retries(req_msg, param->max_cm_retries);
+	cm_req_set_srq(req_msg, param->srq);
+
+	req_msg->primary_local_lid = param->primary_path->slid;
+	req_msg->primary_remote_lid = param->primary_path->dlid;
+	req_msg->primary_local_gid = param->primary_path->sgid;
+	req_msg->primary_remote_gid = param->primary_path->dgid;
+	cm_req_set_primary_flow_label(req_msg, param->primary_path->flow_label);
+	cm_req_set_primary_packet_rate(req_msg, param->primary_path->rate);
+	req_msg->primary_traffic_class = param->primary_path->traffic_class;
+	req_msg->primary_hop_limit = param->primary_path->hop_limit;
+	cm_req_set_primary_sl(req_msg, param->primary_path->sl);
+	cm_req_set_primary_subnet_local(req_msg, 1); /* local only... */
+	cm_req_set_primary_local_ack_timeout(req_msg,
+		min(31, param->primary_path->packet_life_time + 1));
+
+	if (param->alternate_path) {
+		req_msg->alt_local_lid = param->alternate_path->slid;
+		req_msg->alt_remote_lid = param->alternate_path->dlid;
+		req_msg->alt_local_gid = param->alternate_path->sgid;
+		req_msg->alt_remote_gid = param->alternate_path->dgid;
+		cm_req_set_alt_flow_label(req_msg,
+					  param->alternate_path->flow_label);
+		cm_req_set_alt_packet_rate(req_msg, param->alternate_path->rate);
+		req_msg->alt_traffic_class = param->alternate_path->traffic_class;
+		req_msg->alt_hop_limit = param->alternate_path->hop_limit;
+		cm_req_set_alt_sl(req_msg, param->alternate_path->sl);
+		cm_req_set_alt_subnet_local(req_msg, 1); /* local only... */
+		cm_req_set_alt_local_ack_timeout(req_msg,
+			min(31, param->alternate_path->packet_life_time + 1));
+	}
+
+	if (param->private_data && param->private_data_len)
+		memcpy(req_msg->private_data, param->private_data,
+		       param->private_data_len);
+}
+
+static inline int cm_validate_req_param(struct ib_cm_req_param *param)
+{
+	/* peer-to-peer not supported */
+	if (param->peer_to_peer)
+		return -EINVAL;
+
+	if (!param->primary_path)
+		return -EINVAL;
+
+	if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC)
+		return -EINVAL;
+
+	if (param->private_data &&
+	    param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
+		return -EINVAL;
+
+	if (param->alternate_path &&
+	    (param->alternate_path->pkey != param->primary_path->pkey ||
+	     param->alternate_path->mtu != param->primary_path->mtu))
+		return -EINVAL;
+
+	return 0;
+}
+
+int ib_send_cm_req(struct ib_cm_id *cm_id,
+		   struct ib_cm_req_param *param)
+{
+	struct cm_id_private *cm_id_priv;
+	struct ib_send_wr *bad_send_wr;
+	struct cm_req_msg *req_msg;
+	unsigned long flags;
+	int ret;
+
+	ret = cm_validate_req_param(param);
+	if (ret)
+		return ret;
+
+	/* Verify that we're not in timewait. */
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id->state != IB_CM_IDLE) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		ret = -EINVAL;
+		goto out;
+	}
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
+							    id.local_id);
+	if (IS_ERR(cm_id_priv->timewait_info))
+		goto out;
+
+	ret = cm_init_av_by_path(param->primary_path, &cm_id_priv->av);
+	if (ret)
+		goto error1;
+	if (param->alternate_path) {
+		ret = cm_init_av_by_path(param->alternate_path,
+					 &cm_id_priv->alt_av);
+		if (ret)
+			goto error1;
+	}
+	cm_id->service_id = param->service_id;
+	cm_id->service_mask = ~0ULL;
+	cm_id_priv->timeout_ms = cm_convert_to_ms(
+				    param->primary_path->packet_life_time) * 2 +
+				 cm_convert_to_ms(
+				    param->remote_cm_response_timeout);
+	cm_id_priv->max_cm_retries = param->max_cm_retries;
+	cm_id_priv->initiator_depth = param->initiator_depth;
+	cm_id_priv->responder_resources = param->responder_resources;
+	cm_id_priv->retry_count = param->retry_count;
+	cm_id_priv->path_mtu = param->primary_path->mtu;
+
+	ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
+	if (ret)
+		goto error1;
+
+	req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
+	cm_format_req(req_msg, cm_id_priv, param);
+	cm_id_priv->tid = req_msg->hdr.tid;
+	cm_id_priv->msg->send_wr.wr.ud.timeout_ms = cm_id_priv->timeout_ms;
+	cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
+
+	cm_id_priv->local_qpn = cm_req_get_local_qpn(req_msg);
+	cm_id_priv->rq_psn = cm_req_get_starting_psn(req_msg);
+	cm_id_priv->local_ack_timeout =
+				cm_req_get_primary_local_ack_timeout(req_msg);
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
+				&cm_id_priv->msg->send_wr, &bad_send_wr);
+	if (ret) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		goto error2;
+	}
+	BUG_ON(cm_id->state != IB_CM_IDLE);
+	cm_id->state = IB_CM_REQ_SENT;
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	return 0;
+
+error2:	cm_free_msg(cm_id_priv->msg);
+error1:	kfree(cm_id_priv->timewait_info);
+out:	return ret;
+}
+EXPORT_SYMBOL(ib_send_cm_req);
+
+static int cm_issue_rej(struct cm_port *port,
+			struct ib_mad_recv_wc *mad_recv_wc,
+			enum ib_cm_rej_reason reason,
+			enum cm_msg_response msg_rejected,
+			void *ari, u8 ari_length)
+{
+	struct ib_mad_send_buf *msg = NULL;
+	struct ib_send_wr *bad_send_wr;
+	struct cm_rej_msg *rej_msg, *rcv_msg;
+	int ret;
+
+	ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
+	if (ret)
+		return ret;
+
+	/* We just need common CM header information.  Cast to any message. */
+	rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
+	rej_msg = (struct cm_rej_msg *) msg->mad;
+
+	cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
+	rej_msg->remote_comm_id = rcv_msg->local_comm_id;
+	rej_msg->local_comm_id = rcv_msg->remote_comm_id;
+	cm_rej_set_msg_rejected(rej_msg, msg_rejected);
+	rej_msg->reason = reason;
+
+	if (ari && ari_length) {
+		cm_rej_set_reject_info_len(rej_msg, ari_length);
+		memcpy(rej_msg->ari, ari, ari_length);
+	}
+
+	ret = ib_post_send_mad(port->mad_agent, &msg->send_wr, &bad_send_wr);
+	if (ret)
+		cm_free_msg(msg);
+
+	return ret;
+}
+
+static inline int cm_is_active_peer(u64 local_ca_guid, u64 remote_ca_guid,
+				    u32 local_qpn, u32 remote_qpn)
+{
+	return (be64_to_cpu(local_ca_guid) > be64_to_cpu(remote_ca_guid) ||
+		((local_ca_guid == remote_ca_guid) &&
+		 (be32_to_cpu(local_qpn) > be32_to_cpu(remote_qpn))));
+}
+
+static inline void cm_format_paths_from_req(struct cm_req_msg *req_msg,
+					    struct ib_sa_path_rec *primary_path,
+					    struct ib_sa_path_rec *alt_path)
+{
+	memset(primary_path, 0, sizeof *primary_path);
+	primary_path->dgid = req_msg->primary_local_gid;
+	primary_path->sgid = req_msg->primary_remote_gid;
+	primary_path->dlid = req_msg->primary_local_lid;
+	primary_path->slid = req_msg->primary_remote_lid;
+	primary_path->flow_label = cm_req_get_primary_flow_label(req_msg);
+	primary_path->hop_limit = req_msg->primary_hop_limit;
+	primary_path->traffic_class = req_msg->primary_traffic_class;
+	primary_path->reversible = 1;
+	primary_path->pkey = req_msg->pkey;
+	primary_path->sl = cm_req_get_primary_sl(req_msg);
+	primary_path->mtu_selector = IB_SA_EQ;
+	primary_path->mtu = cm_req_get_path_mtu(req_msg);
+	primary_path->rate_selector = IB_SA_EQ;
+	primary_path->rate = cm_req_get_primary_packet_rate(req_msg);
+	primary_path->packet_life_time_selector = IB_SA_EQ;
+	primary_path->packet_life_time =
+		cm_req_get_primary_local_ack_timeout(req_msg);
+	primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
+
+	if (req_msg->alt_local_lid) {
+		memset(alt_path, 0, sizeof *alt_path);
+		alt_path->dgid = req_msg->alt_local_gid;
+		alt_path->sgid = req_msg->alt_remote_gid;
+		alt_path->dlid = req_msg->alt_local_lid;
+		alt_path->slid = req_msg->alt_remote_lid;
+		alt_path->flow_label = cm_req_get_alt_flow_label(req_msg);
+		alt_path->hop_limit = req_msg->alt_hop_limit;
+		alt_path->traffic_class = req_msg->alt_traffic_class;
+		alt_path->reversible = 1;
+		alt_path->pkey = req_msg->pkey;
+		alt_path->sl = cm_req_get_alt_sl(req_msg);
+		alt_path->mtu_selector = IB_SA_EQ;
+		alt_path->mtu = cm_req_get_path_mtu(req_msg);
+		alt_path->rate_selector = IB_SA_EQ;
+		alt_path->rate = cm_req_get_alt_packet_rate(req_msg);
+		alt_path->packet_life_time_selector = IB_SA_EQ;
+		alt_path->packet_life_time =
+			cm_req_get_alt_local_ack_timeout(req_msg);
+		alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
+	}
+}
+
+static void cm_format_req_event(struct cm_work *work,
+				struct cm_id_private *cm_id_priv,
+				struct ib_cm_id *listen_id)
+{
+	struct cm_req_msg *req_msg;
+	struct ib_cm_req_event_param *param;
+
+	req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
+	param = &work->cm_event.param.req_rcvd;
+	param->listen_id = listen_id;
+	param->device = cm_id_priv->av.port->mad_agent->device;
+	param->port = cm_id_priv->av.port->port_num;
+	param->primary_path = &work->path[0];
+	if (req_msg->alt_local_lid)
+		param->alternate_path = &work->path[1];
+	else
+		param->alternate_path = NULL;
+	param->remote_ca_guid = req_msg->local_ca_guid;
+	param->remote_qkey = be32_to_cpu(req_msg->local_qkey);
+	param->remote_qpn = be32_to_cpu(cm_req_get_local_qpn(req_msg));
+	param->qp_type = cm_req_get_qp_type(req_msg);
+	param->starting_psn = be32_to_cpu(cm_req_get_starting_psn(req_msg));
+	param->responder_resources = cm_req_get_init_depth(req_msg);
+	param->initiator_depth = cm_req_get_resp_res(req_msg);
+	param->local_cm_response_timeout =
+					cm_req_get_remote_resp_timeout(req_msg);
+	param->flow_control = cm_req_get_flow_ctrl(req_msg);
+	param->remote_cm_response_timeout =
+					cm_req_get_local_resp_timeout(req_msg);
+	param->retry_count = cm_req_get_retry_count(req_msg);
+	param->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
+	param->srq = cm_req_get_srq(req_msg);
+	work->cm_event.private_data = &req_msg->private_data;
+}
+
+static void cm_process_work(struct cm_id_private *cm_id_priv,
+			    struct cm_work *work)
+{
+	unsigned long flags;
+	int ret;
+
+	/* We will typically only have the current event to report. */
+	ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
+	cm_free_work(work);
+
+	while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
+		spin_lock_irqsave(&cm_id_priv->lock, flags);
+		work = cm_dequeue_work(cm_id_priv);
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		BUG_ON(!work);
+		ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
+						&work->cm_event);
+		cm_free_work(work);
+	}
+	cm_deref_id(cm_id_priv);
+	if (ret)
+		ib_destroy_cm_id(&cm_id_priv->id);
+}
+
+static void cm_format_mra(struct cm_mra_msg *mra_msg,
+			  struct cm_id_private *cm_id_priv,
+			  enum cm_msg_response msg_mraed, u8 service_timeout,
+			  const void *private_data, u8 private_data_len)
+{
+	cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
+	cm_mra_set_msg_mraed(mra_msg, msg_mraed);
+	mra_msg->local_comm_id = cm_id_priv->id.local_id;
+	mra_msg->remote_comm_id = cm_id_priv->id.remote_id;
+	cm_mra_set_service_timeout(mra_msg, service_timeout);
+
+	if (private_data && private_data_len)
+		memcpy(mra_msg->private_data, private_data, private_data_len);
+}
+
+static void cm_format_rej(struct cm_rej_msg *rej_msg,
+			  struct cm_id_private *cm_id_priv,
+			  enum ib_cm_rej_reason reason,
+			  void *ari,
+			  u8 ari_length,
+			  const void *private_data,
+			  u8 private_data_len)
+{
+	cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
+	rej_msg->remote_comm_id = cm_id_priv->id.remote_id;
+
+	switch(cm_id_priv->id.state) {
+	case IB_CM_REQ_RCVD:
+		rej_msg->local_comm_id = 0;
+		cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
+		break;
+	case IB_CM_MRA_REQ_SENT:
+		rej_msg->local_comm_id = cm_id_priv->id.local_id;
+		cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
+		break;
+	case IB_CM_REP_RCVD:
+	case IB_CM_MRA_REP_SENT:
+		rej_msg->local_comm_id = cm_id_priv->id.local_id;
+		cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REP);
+		break;
+	default:
+		rej_msg->local_comm_id = cm_id_priv->id.local_id;
+		cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_OTHER);
+		break;
+	}
+
+	rej_msg->reason = reason;
+	if (ari && ari_length) {
+		cm_rej_set_reject_info_len(rej_msg, ari_length);
+		memcpy(rej_msg->ari, ari, ari_length);
+	}
+
+	if (private_data && private_data_len)
+		memcpy(rej_msg->private_data, private_data, private_data_len);
+}
+
+static void cm_dup_req_handler(struct cm_work *work,
+			       struct cm_id_private *cm_id_priv)
+{
+	struct ib_mad_send_buf *msg = NULL;
+	struct ib_send_wr *bad_send_wr;
+	unsigned long flags;
+	int ret;
+
+	/* Quick state check to discard duplicate REQs. */
+	if (cm_id_priv->id.state == IB_CM_REQ_RCVD)
+		return;
+
+	ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
+	if (ret)
+		return;
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	switch (cm_id_priv->id.state) {
+	case IB_CM_MRA_REQ_SENT:
+		cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
+			      CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
+			      cm_id_priv->private_data,
+			      cm_id_priv->private_data_len);
+		break;
+	case IB_CM_TIMEWAIT:
+		cm_format_rej((struct cm_rej_msg *) msg->mad, cm_id_priv,
+			      IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0);
+		break;
+	default:
+		goto unlock;
+	}
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent, &msg->send_wr,
+			       &bad_send_wr);
+	if (ret)
+		goto free;
+	return;
+
+unlock:	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+free:	cm_free_msg(msg);
+}
+
+static struct cm_id_private * cm_match_req(struct cm_work *work,
+					   struct cm_id_private *cm_id_priv)
+{
+	struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
+	struct cm_timewait_info *timewait_info;
+	struct cm_req_msg *req_msg;
+	unsigned long flags;
+
+	req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
+
+	/* Check for duplicate REQ and stale connections. */
+	spin_lock_irqsave(&cm.lock, flags);
+	timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
+	if (!timewait_info)
+		timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
+
+	if (timewait_info) {
+		cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
+					   timewait_info->work.remote_id);
+		spin_unlock_irqrestore(&cm.lock, flags);
+		if (cur_cm_id_priv) {
+			cm_dup_req_handler(work, cur_cm_id_priv);
+			cm_deref_id(cur_cm_id_priv);
+		} else
+			cm_issue_rej(work->port, work->mad_recv_wc,
+				     IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
+				     NULL, 0);
+		goto error;
+	}
+
+	/* Find matching listen request. */
+	listen_cm_id_priv = cm_find_listen(req_msg->service_id);
+	if (!listen_cm_id_priv) {
+		spin_unlock_irqrestore(&cm.lock, flags);
+		cm_issue_rej(work->port, work->mad_recv_wc,
+			     IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
+			     NULL, 0);
+		goto error;
+	}
+	atomic_inc(&listen_cm_id_priv->refcount);
+	atomic_inc(&cm_id_priv->refcount);
+	cm_id_priv->id.state = IB_CM_REQ_RCVD;
+	atomic_inc(&cm_id_priv->work_count);
+	spin_unlock_irqrestore(&cm.lock, flags);
+	return listen_cm_id_priv;
+
+error:	cm_cleanup_timewait(cm_id_priv->timewait_info);
+	return NULL;
+}
+
+static int cm_req_handler(struct cm_work *work)
+{
+	struct ib_cm_id *cm_id;
+	struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
+	struct cm_req_msg *req_msg;
+	int ret;
+
+	req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
+
+	cm_id = ib_create_cm_id(NULL, NULL);
+	if (IS_ERR(cm_id))
+		return PTR_ERR(cm_id);
+
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+	cm_id_priv->id.remote_id = req_msg->local_comm_id;
+	cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
+				&cm_id_priv->av);
+	cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
+							    id.local_id);
+	if (IS_ERR(cm_id_priv->timewait_info)) {
+		ret = PTR_ERR(cm_id_priv->timewait_info);
+		goto error1;
+	}
+	cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
+	cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
+	cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
+
+	listen_cm_id_priv = cm_match_req(work, cm_id_priv);
+	if (!listen_cm_id_priv) {
+		ret = -EINVAL;
+		goto error2;
+	}
+
+	cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
+	cm_id_priv->id.context = listen_cm_id_priv->id.context;
+	cm_id_priv->id.service_id = req_msg->service_id;
+	cm_id_priv->id.service_mask = ~0ULL;
+
+	cm_format_paths_from_req(req_msg, &work->path[0], &work->path[1]);
+	ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av);
+	if (ret)
+		goto error3;
+	if (req_msg->alt_local_lid) {
+		ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av);
+		if (ret)
+			goto error3;
+	}
+	cm_id_priv->tid = req_msg->hdr.tid;
+	cm_id_priv->timeout_ms = cm_convert_to_ms(
+					cm_req_get_local_resp_timeout(req_msg));
+	cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
+	cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
+	cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
+	cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
+	cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
+	cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
+	cm_id_priv->local_ack_timeout =
+				cm_req_get_primary_local_ack_timeout(req_msg);
+	cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
+	cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
+
+	cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
+	cm_process_work(cm_id_priv, work);
+	cm_deref_id(listen_cm_id_priv);
+	return 0;
+
+error3:	atomic_dec(&cm_id_priv->refcount);
+	cm_deref_id(listen_cm_id_priv);
+	cm_cleanup_timewait(cm_id_priv->timewait_info);
+error2:	kfree(cm_id_priv->timewait_info);
+error1:	ib_destroy_cm_id(&cm_id_priv->id);
+	return ret;
+}
+
+static void cm_format_rep(struct cm_rep_msg *rep_msg,
+			  struct cm_id_private *cm_id_priv,
+			  struct ib_cm_rep_param *param)
+{
+	cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
+	rep_msg->local_comm_id = cm_id_priv->id.local_id;
+	rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
+	cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
+	cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
+	rep_msg->resp_resources = param->responder_resources;
+	rep_msg->initiator_depth = param->initiator_depth;
+	cm_rep_set_target_ack_delay(rep_msg, param->target_ack_delay);
+	cm_rep_set_failover(rep_msg, param->failover_accepted);
+	cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
+	cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
+	cm_rep_set_srq(rep_msg, param->srq);
+	rep_msg->local_ca_guid = cm_id_priv->av.port->cm_dev->ca_guid;
+
+	if (param->private_data && param->private_data_len)
+		memcpy(rep_msg->private_data, param->private_data,
+		       param->private_data_len);
+}
+
+int ib_send_cm_rep(struct ib_cm_id *cm_id,
+		   struct ib_cm_rep_param *param)
+{
+	struct cm_id_private *cm_id_priv;
+	struct ib_mad_send_buf *msg;
+	struct cm_rep_msg *rep_msg;
+	struct ib_send_wr *bad_send_wr;
+	unsigned long flags;
+	int ret;
+
+	if (param->private_data &&
+	    param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
+		return -EINVAL;
+
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id->state != IB_CM_REQ_RCVD &&
+	    cm_id->state != IB_CM_MRA_REQ_SENT) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	ret = cm_alloc_msg(cm_id_priv, &msg);
+	if (ret)
+		goto out;
+
+	rep_msg = (struct cm_rep_msg *) msg->mad;
+	cm_format_rep(rep_msg, cm_id_priv, param);
+	msg->send_wr.wr.ud.timeout_ms = cm_id_priv->timeout_ms;
+	msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
+
+	ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
+			       &msg->send_wr, &bad_send_wr);
+	if (ret) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		cm_free_msg(msg);
+		return ret;
+	}
+
+	cm_id->state = IB_CM_REP_SENT;
+	cm_id_priv->msg = msg;
+	cm_id_priv->initiator_depth = param->initiator_depth;
+	cm_id_priv->responder_resources = param->responder_resources;
+	cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
+	cm_id_priv->local_qpn = cm_rep_get_local_qpn(rep_msg);
+
+out:	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	return ret;
+}
+EXPORT_SYMBOL(ib_send_cm_rep);
+
+static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
+			  struct cm_id_private *cm_id_priv,
+			  const void *private_data,
+			  u8 private_data_len)
+{
+	cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
+	rtu_msg->local_comm_id = cm_id_priv->id.local_id;
+	rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
+
+	if (private_data && private_data_len)
+		memcpy(rtu_msg->private_data, private_data, private_data_len);
+}
+
+int ib_send_cm_rtu(struct ib_cm_id *cm_id,
+		   const void *private_data,
+		   u8 private_data_len)
+{
+	struct cm_id_private *cm_id_priv;
+	struct ib_mad_send_buf *msg;
+	struct ib_send_wr *bad_send_wr;
+	unsigned long flags;
+	void *data;
+	int ret;
+
+	if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
+		return -EINVAL;
+
+	data = cm_copy_private_data(private_data, private_data_len);
+	if (IS_ERR(data))
+		return PTR_ERR(data);
+
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id->state != IB_CM_REP_RCVD &&
+	    cm_id->state != IB_CM_MRA_REP_SENT) {
+		ret = -EINVAL;
+		goto error;
+	}
+
+	ret = cm_alloc_msg(cm_id_priv, &msg);
+	if (ret)
+		goto error;
+
+	cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
+		      private_data, private_data_len);
+
+	ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
+			       &msg->send_wr, &bad_send_wr);
+	if (ret) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		cm_free_msg(msg);
+		kfree(data);
+		return ret;
+	}
+
+	cm_id->state = IB_CM_ESTABLISHED;
+	cm_set_private_data(cm_id_priv, data, private_data_len);
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	return 0;
+
+error:	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	kfree(data);
+	return ret;
+}
+EXPORT_SYMBOL(ib_send_cm_rtu);
+
+static void cm_format_rep_event(struct cm_work *work)
+{
+	struct cm_rep_msg *rep_msg;
+	struct ib_cm_rep_event_param *param;
+
+	rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
+	param = &work->cm_event.param.rep_rcvd;
+	param->remote_ca_guid = rep_msg->local_ca_guid;
+	param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
+	param->remote_qpn = be32_to_cpu(cm_rep_get_local_qpn(rep_msg));
+	param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
+	param->responder_resources = rep_msg->initiator_depth;
+	param->initiator_depth = rep_msg->resp_resources;
+	param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
+	param->failover_accepted = cm_rep_get_failover(rep_msg);
+	param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
+	param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
+	param->srq = cm_rep_get_srq(rep_msg);
+	work->cm_event.private_data = &rep_msg->private_data;
+}
+
+static void cm_dup_rep_handler(struct cm_work *work)
+{
+	struct cm_id_private *cm_id_priv;
+	struct cm_rep_msg *rep_msg;
+	struct ib_mad_send_buf *msg = NULL;
+	struct ib_send_wr *bad_send_wr;
+	unsigned long flags;
+	int ret;
+
+	rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
+	cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
+				   rep_msg->local_comm_id);
+	if (!cm_id_priv)
+		return;
+
+	ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
+	if (ret)
+		goto deref;
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
+		cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
+			      cm_id_priv->private_data,
+			      cm_id_priv->private_data_len);
+	else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
+		cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
+			      CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
+			      cm_id_priv->private_data,
+			      cm_id_priv->private_data_len);
+	else
+		goto unlock;
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent, &msg->send_wr,
+			       &bad_send_wr);
+	if (ret)
+		goto free;
+	goto deref;
+
+unlock:	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+free:	cm_free_msg(msg);
+deref:	cm_deref_id(cm_id_priv);
+}
+
+static int cm_rep_handler(struct cm_work *work)
+{
+	struct cm_id_private *cm_id_priv;
+	struct cm_rep_msg *rep_msg;
+	unsigned long flags;
+	int ret;
+
+	rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
+	cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
+	if (!cm_id_priv) {
+		cm_dup_rep_handler(work);
+		return -EINVAL;
+	}
+
+	cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
+	cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
+	cm_id_priv->timewait_info->remote_qpn = cm_rep_get_local_qpn(rep_msg);
+
+	spin_lock_irqsave(&cm.lock, flags);
+	/* Check for duplicate REP. */
+	if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
+		spin_unlock_irqrestore(&cm.lock, flags);
+		ret = -EINVAL;
+		goto error;
+	}
+	/* Check for a stale connection. */
+	if (cm_insert_remote_qpn(cm_id_priv->timewait_info)) {
+		spin_unlock_irqrestore(&cm.lock, flags);
+		cm_issue_rej(work->port, work->mad_recv_wc,
+			     IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
+			     NULL, 0);
+		ret = -EINVAL;
+		goto error;
+	}
+	spin_unlock_irqrestore(&cm.lock, flags);
+
+	cm_format_rep_event(work);
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	switch (cm_id_priv->id.state) {
+	case IB_CM_REQ_SENT:
+	case IB_CM_MRA_REQ_RCVD:
+		break;
+	default:
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		ret = -EINVAL;
+		goto error;
+	}
+	cm_id_priv->id.state = IB_CM_REP_RCVD;
+	cm_id_priv->id.remote_id = rep_msg->local_comm_id;
+	cm_id_priv->remote_qpn = cm_rep_get_local_qpn(rep_msg);
+	cm_id_priv->initiator_depth = rep_msg->resp_resources;
+	cm_id_priv->responder_resources = rep_msg->initiator_depth;
+	cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
+	cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
+
+	/* todo: handle peer_to_peer */
+
+	ib_cancel_mad(cm_id_priv->av.port->mad_agent,
+		      (unsigned long) cm_id_priv->msg);
+	ret = atomic_inc_and_test(&cm_id_priv->work_count);
+	if (!ret)
+		list_add_tail(&work->list, &cm_id_priv->work_list);
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	if (ret)
+		cm_process_work(cm_id_priv, work);
+	else
+		cm_deref_id(cm_id_priv);
+	return 0;
+
+error:	cm_cleanup_timewait(cm_id_priv->timewait_info);
+	cm_deref_id(cm_id_priv);
+	return ret;
+}
+
+static int cm_establish_handler(struct cm_work *work)
+{
+	struct cm_id_private *cm_id_priv;
+	unsigned long flags;
+	int ret;
+
+	/* See comment in ib_cm_establish about lookup. */
+	cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
+	if (!cm_id_priv)
+		return -EINVAL;
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		goto out;
+	}
+
+	ib_cancel_mad(cm_id_priv->av.port->mad_agent,
+		      (unsigned long) cm_id_priv->msg);
+	ret = atomic_inc_and_test(&cm_id_priv->work_count);
+	if (!ret)
+		list_add_tail(&work->list, &cm_id_priv->work_list);
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	if (ret)
+		cm_process_work(cm_id_priv, work);
+	else
+		cm_deref_id(cm_id_priv);
+	return 0;
+out:
+	cm_deref_id(cm_id_priv);
+	return -EINVAL;
+}
+
+static int cm_rtu_handler(struct cm_work *work)
+{
+	struct cm_id_private *cm_id_priv;
+	struct cm_rtu_msg *rtu_msg;
+	unsigned long flags;
+	int ret;
+
+	rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
+	cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
+				   rtu_msg->local_comm_id);
+	if (!cm_id_priv)
+		return -EINVAL;
+
+	work->cm_event.private_data = &rtu_msg->private_data;
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id_priv->id.state != IB_CM_REP_SENT &&
+	    cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		goto out;
+	}
+	cm_id_priv->id.state = IB_CM_ESTABLISHED;
+
+	ib_cancel_mad(cm_id_priv->av.port->mad_agent,
+		      (unsigned long) cm_id_priv->msg);
+	ret = atomic_inc_and_test(&cm_id_priv->work_count);
+	if (!ret)
+		list_add_tail(&work->list, &cm_id_priv->work_list);
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	if (ret)
+		cm_process_work(cm_id_priv, work);
+	else
+		cm_deref_id(cm_id_priv);
+	return 0;
+out:
+	cm_deref_id(cm_id_priv);
+	return -EINVAL;
+}
+
+static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
+			  struct cm_id_private *cm_id_priv,
+			  const void *private_data,
+			  u8 private_data_len)
+{
+	cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
+			  cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
+	dreq_msg->local_comm_id = cm_id_priv->id.local_id;
+	dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
+	cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
+
+	if (private_data && private_data_len)
+		memcpy(dreq_msg->private_data, private_data, private_data_len);
+}
+
+int ib_send_cm_dreq(struct ib_cm_id *cm_id,
+		    const void *private_data,
+		    u8 private_data_len)
+{
+	struct cm_id_private *cm_id_priv;
+	struct ib_mad_send_buf *msg;
+	struct ib_send_wr *bad_send_wr;
+	unsigned long flags;
+	int ret;
+
+	if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
+		return -EINVAL;
+
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id->state != IB_CM_ESTABLISHED) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	ret = cm_alloc_msg(cm_id_priv, &msg);
+	if (ret) {
+		cm_enter_timewait(cm_id_priv);
+		goto out;
+	}
+
+	cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
+		       private_data, private_data_len);
+	msg->send_wr.wr.ud.timeout_ms = cm_id_priv->timeout_ms;
+	msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
+
+	ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
+			       &msg->send_wr, &bad_send_wr);
+	if (ret) {
+		cm_enter_timewait(cm_id_priv);
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		cm_free_msg(msg);
+		return ret;
+	}
+
+	cm_id->state = IB_CM_DREQ_SENT;
+	cm_id_priv->msg = msg;
+out:	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	return ret;
+}
+EXPORT_SYMBOL(ib_send_cm_dreq);
+
+static void cm_format_drep(struct cm_drep_msg *drep_msg,
+			  struct cm_id_private *cm_id_priv,
+			  const void *private_data,
+			  u8 private_data_len)
+{
+	cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
+	drep_msg->local_comm_id = cm_id_priv->id.local_id;
+	drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
+
+	if (private_data && private_data_len)
+		memcpy(drep_msg->private_data, private_data, private_data_len);
+}
+
+int ib_send_cm_drep(struct ib_cm_id *cm_id,
+		    const void *private_data,
+		    u8 private_data_len)
+{
+	struct cm_id_private *cm_id_priv;
+	struct ib_mad_send_buf *msg;
+	struct ib_send_wr *bad_send_wr;
+	unsigned long flags;
+	void *data;
+	int ret;
+
+	if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
+		return -EINVAL;
+
+	data = cm_copy_private_data(private_data, private_data_len);
+	if (IS_ERR(data))
+		return PTR_ERR(data);
+
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id->state != IB_CM_DREQ_RCVD) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		kfree(data);
+		return -EINVAL;
+	}
+
+	cm_set_private_data(cm_id_priv, data, private_data_len);
+	cm_enter_timewait(cm_id_priv);
+
+	ret = cm_alloc_msg(cm_id_priv, &msg);
+	if (ret)
+		goto out;
+
+	cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
+		       private_data, private_data_len);
+
+	ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent, &msg->send_wr,
+			       &bad_send_wr);
+	if (ret) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		cm_free_msg(msg);
+		return ret;
+	}
+
+out:	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	return ret;
+}
+EXPORT_SYMBOL(ib_send_cm_drep);
+
+static int cm_dreq_handler(struct cm_work *work)
+{
+	struct cm_id_private *cm_id_priv;
+	struct cm_dreq_msg *dreq_msg;
+	struct ib_mad_send_buf *msg = NULL;
+	struct ib_send_wr *bad_send_wr;
+	unsigned long flags;
+	int ret;
+
+	dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
+	cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
+				   dreq_msg->local_comm_id);
+	if (!cm_id_priv)
+		return -EINVAL;
+
+	work->cm_event.private_data = &dreq_msg->private_data;
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
+		goto unlock;
+
+	switch (cm_id_priv->id.state) {
+	case IB_CM_REP_SENT:
+	case IB_CM_DREQ_SENT:
+		ib_cancel_mad(cm_id_priv->av.port->mad_agent,
+			      (unsigned long) cm_id_priv->msg);
+		break;
+	case IB_CM_ESTABLISHED:
+	case IB_CM_MRA_REP_RCVD:
+		break;
+	case IB_CM_TIMEWAIT:
+		if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
+			goto unlock;
+
+		cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
+			       cm_id_priv->private_data,
+			       cm_id_priv->private_data_len);
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+		if (ib_post_send_mad(cm_id_priv->av.port->mad_agent,
+				     &msg->send_wr, &bad_send_wr))
+			cm_free_msg(msg);
+		goto deref;
+	default:
+		goto unlock;
+	}
+	cm_id_priv->id.state = IB_CM_DREQ_RCVD;
+	cm_id_priv->tid = dreq_msg->hdr.tid;
+	ret = atomic_inc_and_test(&cm_id_priv->work_count);
+	if (!ret)
+		list_add_tail(&work->list, &cm_id_priv->work_list);
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	if (ret)
+		cm_process_work(cm_id_priv, work);
+	else
+		cm_deref_id(cm_id_priv);
+	return 0;
+
+unlock:	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+deref:	cm_deref_id(cm_id_priv);
+	return -EINVAL;
+}
+
+static int cm_drep_handler(struct cm_work *work)
+{
+	struct cm_id_private *cm_id_priv;
+	struct cm_drep_msg *drep_msg;
+	unsigned long flags;
+	int ret;
+
+	drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
+	cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
+				   drep_msg->local_comm_id);
+	if (!cm_id_priv)
+		return -EINVAL;
+
+	work->cm_event.private_data = &drep_msg->private_data;
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
+	    cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		goto out;
+	}
+	cm_enter_timewait(cm_id_priv);
+
+	ib_cancel_mad(cm_id_priv->av.port->mad_agent,
+		      (unsigned long) cm_id_priv->msg);
+	ret = atomic_inc_and_test(&cm_id_priv->work_count);
+	if (!ret)
+		list_add_tail(&work->list, &cm_id_priv->work_list);
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	if (ret)
+		cm_process_work(cm_id_priv, work);
+	else
+		cm_deref_id(cm_id_priv);
+	return 0;
+out:
+	cm_deref_id(cm_id_priv);
+	return -EINVAL;
+}
+
+int ib_send_cm_rej(struct ib_cm_id *cm_id,
+		   enum ib_cm_rej_reason reason,
+		   void *ari,
+		   u8 ari_length,
+		   const void *private_data,
+		   u8 private_data_len)
+{
+	struct cm_id_private *cm_id_priv;
+	struct ib_mad_send_buf *msg;
+	struct ib_send_wr *bad_send_wr;
+	unsigned long flags;
+	int ret;
+
+	if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
+	    (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
+		return -EINVAL;
+
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	switch (cm_id->state) {
+	case IB_CM_REQ_SENT:
+	case IB_CM_MRA_REQ_RCVD:
+	case IB_CM_REQ_RCVD:
+	case IB_CM_MRA_REQ_SENT:
+	case IB_CM_REP_RCVD:
+	case IB_CM_MRA_REP_SENT:
+		ret = cm_alloc_msg(cm_id_priv, &msg);
+		if (!ret)
+			cm_format_rej((struct cm_rej_msg *) msg->mad,
+				      cm_id_priv, reason, ari, ari_length,
+				      private_data, private_data_len);
+
+		cm_reset_to_idle(cm_id_priv);
+		break;
+	case IB_CM_REP_SENT:
+	case IB_CM_MRA_REP_RCVD:
+		ret = cm_alloc_msg(cm_id_priv, &msg);
+		if (!ret)
+			cm_format_rej((struct cm_rej_msg *) msg->mad,
+				      cm_id_priv, reason, ari, ari_length,
+				      private_data, private_data_len);
+
+		cm_enter_timewait(cm_id_priv);
+		break;
+	default:
+		ret = -EINVAL;
+		goto out;
+	}
+
+	if (ret)
+		goto out;
+
+	ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
+			       &msg->send_wr, &bad_send_wr);
+	if (ret)
+		cm_free_msg(msg);
+
+out:	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	return ret;
+}
+EXPORT_SYMBOL(ib_send_cm_rej);
+
+static void cm_format_rej_event(struct cm_work *work)
+{
+	struct cm_rej_msg *rej_msg;
+	struct ib_cm_rej_event_param *param;
+
+	rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
+	param = &work->cm_event.param.rej_rcvd;
+	param->ari = rej_msg->ari;
+	param->ari_length = cm_rej_get_reject_info_len(rej_msg);
+	param->reason = rej_msg->reason;
+	work->cm_event.private_data = &rej_msg->private_data;
+}
+
+static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
+{
+	struct cm_timewait_info *timewait_info;
+	struct cm_id_private *cm_id_priv;
+	unsigned long flags;
+	u32 remote_id;
+
+	remote_id = rej_msg->local_comm_id;
+
+	if (rej_msg->reason == IB_CM_REJ_TIMEOUT) {
+		spin_lock_irqsave(&cm.lock, flags);
+		timewait_info = cm_find_remote_id( *((u64 *) rej_msg->ari),
+						  remote_id);
+		if (!timewait_info) {
+			spin_unlock_irqrestore(&cm.lock, flags);
+			return NULL;
+		}
+		cm_id_priv = idr_find(&cm.local_id_table,
+				      (int) timewait_info->work.local_id);
+		if (cm_id_priv) {
+			if (cm_id_priv->id.remote_id == remote_id)
+				atomic_inc(&cm_id_priv->refcount);
+			else
+				cm_id_priv = NULL;
+		}
+		spin_unlock_irqrestore(&cm.lock, flags);
+	} else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
+		cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
+	else
+		cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
+
+	return cm_id_priv;
+}
+
+static int cm_rej_handler(struct cm_work *work)
+{
+	struct cm_id_private *cm_id_priv;
+	struct cm_rej_msg *rej_msg;
+	unsigned long flags;
+	int ret;
+
+	rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
+	cm_id_priv = cm_acquire_rejected_id(rej_msg);
+	if (!cm_id_priv)
+		return -EINVAL;
+
+	cm_format_rej_event(work);
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	switch (cm_id_priv->id.state) {
+	case IB_CM_REQ_SENT:
+	case IB_CM_MRA_REQ_RCVD:
+	case IB_CM_REP_SENT:
+	case IB_CM_MRA_REP_RCVD:
+		ib_cancel_mad(cm_id_priv->av.port->mad_agent,
+			      (unsigned long) cm_id_priv->msg);
+		/* fall through */
+	case IB_CM_REQ_RCVD:
+	case IB_CM_MRA_REQ_SENT:
+		if (rej_msg->reason == IB_CM_REJ_STALE_CONN)
+			cm_enter_timewait(cm_id_priv);
+		else
+			cm_reset_to_idle(cm_id_priv);
+		break;
+	case IB_CM_DREQ_SENT:
+		ib_cancel_mad(cm_id_priv->av.port->mad_agent,
+			      (unsigned long) cm_id_priv->msg);
+		/* fall through */
+	case IB_CM_REP_RCVD:
+	case IB_CM_MRA_REP_SENT:
+	case IB_CM_ESTABLISHED:
+		cm_enter_timewait(cm_id_priv);
+		break;
+	default:
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		ret = -EINVAL;
+		goto out;
+	}
+
+	ret = atomic_inc_and_test(&cm_id_priv->work_count);
+	if (!ret)
+		list_add_tail(&work->list, &cm_id_priv->work_list);
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	if (ret)
+		cm_process_work(cm_id_priv, work);
+	else
+		cm_deref_id(cm_id_priv);
+	return 0;
+out:
+	cm_deref_id(cm_id_priv);
+	return -EINVAL;
+}
+
+int ib_send_cm_mra(struct ib_cm_id *cm_id,
+		   u8 service_timeout,
+		   const void *private_data,
+		   u8 private_data_len)
+{
+	struct cm_id_private *cm_id_priv;
+	struct ib_mad_send_buf *msg;
+	struct ib_send_wr *bad_send_wr;
+	void *data;
+	unsigned long flags;
+	int ret;
+
+	if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
+		return -EINVAL;
+
+	data = cm_copy_private_data(private_data, private_data_len);
+	if (IS_ERR(data))
+		return PTR_ERR(data);
+
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	switch(cm_id_priv->id.state) {
+	case IB_CM_REQ_RCVD:
+		ret = cm_alloc_msg(cm_id_priv, &msg);
+		if (ret)
+			goto error1;
+
+		cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
+			      CM_MSG_RESPONSE_REQ, service_timeout,
+			      private_data, private_data_len);
+		ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
+				       &msg->send_wr, &bad_send_wr);
+		if (ret)
+			goto error2;
+		cm_id->state = IB_CM_MRA_REQ_SENT;
+		break;
+	case IB_CM_REP_RCVD:
+		ret = cm_alloc_msg(cm_id_priv, &msg);
+		if (ret)
+			goto error1;
+
+		cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
+			      CM_MSG_RESPONSE_REP, service_timeout,
+			      private_data, private_data_len);
+		ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
+				       &msg->send_wr, &bad_send_wr);
+		if (ret)
+			goto error2;
+		cm_id->state = IB_CM_MRA_REP_SENT;
+		break;
+	case IB_CM_ESTABLISHED:
+		ret = cm_alloc_msg(cm_id_priv, &msg);
+		if (ret)
+			goto error1;
+
+		cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
+			      CM_MSG_RESPONSE_OTHER, service_timeout,
+			      private_data, private_data_len);
+		ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
+				       &msg->send_wr, &bad_send_wr);
+		if (ret)
+			goto error2;
+		cm_id->lap_state = IB_CM_MRA_LAP_SENT;
+		break;
+	default:
+		ret = -EINVAL;
+		goto error1;
+	}
+	cm_id_priv->service_timeout = service_timeout;
+	cm_set_private_data(cm_id_priv, data, private_data_len);
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	return 0;
+
+error1:	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	kfree(data);
+	return ret;
+
+error2:	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	kfree(data);
+	cm_free_msg(msg);
+	return ret;
+}
+EXPORT_SYMBOL(ib_send_cm_mra);
+
+static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
+{
+	switch (cm_mra_get_msg_mraed(mra_msg)) {
+	case CM_MSG_RESPONSE_REQ:
+		return cm_acquire_id(mra_msg->remote_comm_id, 0);
+	case CM_MSG_RESPONSE_REP:
+	case CM_MSG_RESPONSE_OTHER:
+		return cm_acquire_id(mra_msg->remote_comm_id,
+				     mra_msg->local_comm_id);
+	default:
+		return NULL;
+	}
+}
+
+static int cm_mra_handler(struct cm_work *work)
+{
+	struct cm_id_private *cm_id_priv;
+	struct cm_mra_msg *mra_msg;
+	unsigned long flags;
+	int timeout, ret;
+
+	mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
+	cm_id_priv = cm_acquire_mraed_id(mra_msg);
+	if (!cm_id_priv)
+		return -EINVAL;
+
+	work->cm_event.private_data = &mra_msg->private_data;
+	work->cm_event.param.mra_rcvd.service_timeout =
+					cm_mra_get_service_timeout(mra_msg);
+	timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
+		  cm_convert_to_ms(cm_id_priv->av.packet_life_time);
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	switch (cm_id_priv->id.state) {
+	case IB_CM_REQ_SENT:
+		if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
+		    ib_modify_mad(cm_id_priv->av.port->mad_agent,
+				  (unsigned long) cm_id_priv->msg, timeout))
+			goto out;
+		cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
+		break;
+	case IB_CM_REP_SENT:
+		if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
+		    ib_modify_mad(cm_id_priv->av.port->mad_agent,
+				  (unsigned long) cm_id_priv->msg, timeout))
+			goto out;
+		cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
+		break;
+	case IB_CM_ESTABLISHED:
+		if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
+		    cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
+		    ib_modify_mad(cm_id_priv->av.port->mad_agent,
+				  (unsigned long) cm_id_priv->msg, timeout))
+			goto out;
+		cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
+		break;
+	default:
+		goto out;
+	}
+
+	cm_id_priv->msg->context[1] = (void *) (unsigned long)
+				      cm_id_priv->id.state;
+	ret = atomic_inc_and_test(&cm_id_priv->work_count);
+	if (!ret)
+		list_add_tail(&work->list, &cm_id_priv->work_list);
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	if (ret)
+		cm_process_work(cm_id_priv, work);
+	else
+		cm_deref_id(cm_id_priv);
+	return 0;
+out:
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	cm_deref_id(cm_id_priv);
+	return -EINVAL;
+}
+
+static void cm_format_lap(struct cm_lap_msg *lap_msg,
+			  struct cm_id_private *cm_id_priv,
+			  struct ib_sa_path_rec *alternate_path,
+			  const void *private_data,
+			  u8 private_data_len)
+{
+	cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
+			  cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
+	lap_msg->local_comm_id = cm_id_priv->id.local_id;
+	lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
+	cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
+	/* todo: need remote CM response timeout */
+	cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
+	lap_msg->alt_local_lid = alternate_path->slid;
+	lap_msg->alt_remote_lid = alternate_path->dlid;
+	lap_msg->alt_local_gid = alternate_path->sgid;
+	lap_msg->alt_remote_gid = alternate_path->dgid;
+	cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
+	cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
+	lap_msg->alt_hop_limit = alternate_path->hop_limit;
+	cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
+	cm_lap_set_sl(lap_msg, alternate_path->sl);
+	cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
+	cm_lap_set_local_ack_timeout(lap_msg,
+		min(31, alternate_path->packet_life_time + 1));
+
+	if (private_data && private_data_len)
+		memcpy(lap_msg->private_data, private_data, private_data_len);
+}
+
+int ib_send_cm_lap(struct ib_cm_id *cm_id,
+		   struct ib_sa_path_rec *alternate_path,
+		   const void *private_data,
+		   u8 private_data_len)
+{
+	struct cm_id_private *cm_id_priv;
+	struct ib_mad_send_buf *msg;
+	struct ib_send_wr *bad_send_wr;
+	unsigned long flags;
+	int ret;
+
+	if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
+		return -EINVAL;
+
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id->state != IB_CM_ESTABLISHED ||
+	    cm_id->lap_state != IB_CM_LAP_IDLE) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	ret = cm_alloc_msg(cm_id_priv, &msg);
+	if (ret)
+		goto out;
+
+	cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
+		      alternate_path, private_data, private_data_len);
+	msg->send_wr.wr.ud.timeout_ms = cm_id_priv->timeout_ms;
+	msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
+
+	ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
+			       &msg->send_wr, &bad_send_wr);
+	if (ret) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		cm_free_msg(msg);
+		return ret;
+	}
+
+	cm_id->lap_state = IB_CM_LAP_SENT;
+	cm_id_priv->msg = msg;
+
+out:	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	return ret;
+}
+EXPORT_SYMBOL(ib_send_cm_lap);
+
+static void cm_format_path_from_lap(struct ib_sa_path_rec *path,
+				    struct cm_lap_msg *lap_msg)
+{
+	memset(path, 0, sizeof *path);
+	path->dgid = lap_msg->alt_local_gid;
+	path->sgid = lap_msg->alt_remote_gid;
+	path->dlid = lap_msg->alt_local_lid;
+	path->slid = lap_msg->alt_remote_lid;
+	path->flow_label = cm_lap_get_flow_label(lap_msg);
+	path->hop_limit = lap_msg->alt_hop_limit;
+	path->traffic_class = cm_lap_get_traffic_class(lap_msg);
+	path->reversible = 1;
+	/* pkey is same as in REQ */
+	path->sl = cm_lap_get_sl(lap_msg);
+	path->mtu_selector = IB_SA_EQ;
+	/* mtu is same as in REQ */
+	path->rate_selector = IB_SA_EQ;
+	path->rate = cm_lap_get_packet_rate(lap_msg);
+	path->packet_life_time_selector = IB_SA_EQ;
+	path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
+	path->packet_life_time -= (path->packet_life_time > 0);
+}
+
+static int cm_lap_handler(struct cm_work *work)
+{
+	struct cm_id_private *cm_id_priv;
+	struct cm_lap_msg *lap_msg;
+	struct ib_cm_lap_event_param *param;
+	struct ib_mad_send_buf *msg = NULL;
+	struct ib_send_wr *bad_send_wr;
+	unsigned long flags;
+	int ret;
+
+	/* todo: verify LAP request and send reject APR if invalid. */
+	lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
+	cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
+				   lap_msg->local_comm_id);
+	if (!cm_id_priv)
+		return -EINVAL;
+
+	param = &work->cm_event.param.lap_rcvd;
+	param->alternate_path = &work->path[0];
+	cm_format_path_from_lap(param->alternate_path, lap_msg);
+	work->cm_event.private_data = &lap_msg->private_data;
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
+		goto unlock;
+
+	switch (cm_id_priv->id.lap_state) {
+	case IB_CM_LAP_IDLE:
+		break;
+	case IB_CM_MRA_LAP_SENT:
+		if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
+			goto unlock;
+
+		cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
+			      CM_MSG_RESPONSE_OTHER,
+			      cm_id_priv->service_timeout,
+			      cm_id_priv->private_data,
+			      cm_id_priv->private_data_len);
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+		if (ib_post_send_mad(cm_id_priv->av.port->mad_agent,
+				     &msg->send_wr, &bad_send_wr))
+			cm_free_msg(msg);
+		goto deref;
+	default:
+		goto unlock;
+	}
+
+	cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
+	cm_id_priv->tid = lap_msg->hdr.tid;
+	ret = atomic_inc_and_test(&cm_id_priv->work_count);
+	if (!ret)
+		list_add_tail(&work->list, &cm_id_priv->work_list);
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	if (ret)
+		cm_process_work(cm_id_priv, work);
+	else
+		cm_deref_id(cm_id_priv);
+	return 0;
+
+unlock:	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+deref:	cm_deref_id(cm_id_priv);
+	return -EINVAL;
+}
+
+static void cm_format_apr(struct cm_apr_msg *apr_msg,
+			  struct cm_id_private *cm_id_priv,
+			  enum ib_cm_apr_status status,
+			  void *info,
+			  u8 info_length,
+			  const void *private_data,
+			  u8 private_data_len)
+{
+	cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
+	apr_msg->local_comm_id = cm_id_priv->id.local_id;
+	apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
+	apr_msg->ap_status = (u8) status;
+
+	if (info && info_length) {
+		apr_msg->info_length = info_length;
+		memcpy(apr_msg->info, info, info_length);
+	}
+
+	if (private_data && private_data_len)
+		memcpy(apr_msg->private_data, private_data, private_data_len);
+}
+
+int ib_send_cm_apr(struct ib_cm_id *cm_id,
+		   enum ib_cm_apr_status status,
+		   void *info,
+		   u8 info_length,
+		   const void *private_data,
+		   u8 private_data_len)
+{
+	struct cm_id_private *cm_id_priv;
+	struct ib_mad_send_buf *msg;
+	struct ib_send_wr *bad_send_wr;
+	unsigned long flags;
+	int ret;
+
+	if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
+	    (info && info_length > IB_CM_APR_INFO_LENGTH))
+		return -EINVAL;
+
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id->state != IB_CM_ESTABLISHED ||
+	    (cm_id->lap_state != IB_CM_LAP_RCVD &&
+	     cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	ret = cm_alloc_msg(cm_id_priv, &msg);
+	if (ret)
+		goto out;
+
+	cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
+		      info, info_length, private_data, private_data_len);
+	ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
+			       &msg->send_wr, &bad_send_wr);
+	if (ret) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		cm_free_msg(msg);
+		return ret;
+	}
+
+	cm_id->lap_state = IB_CM_LAP_IDLE;
+out:	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	return ret;
+}
+EXPORT_SYMBOL(ib_send_cm_apr);
+
+static int cm_apr_handler(struct cm_work *work)
+{
+	struct cm_id_private *cm_id_priv;
+	struct cm_apr_msg *apr_msg;
+	unsigned long flags;
+	int ret;
+
+	apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
+	cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
+				   apr_msg->local_comm_id);
+	if (!cm_id_priv)
+		return -EINVAL; /* Unmatched reply. */
+
+	work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
+	work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
+	work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
+	work->cm_event.private_data = &apr_msg->private_data;
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
+	    (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
+	     cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		goto out;
+	}
+	cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
+	ib_cancel_mad(cm_id_priv->av.port->mad_agent,
+		      (unsigned long) cm_id_priv->msg);
+	cm_id_priv->msg = NULL;
+
+	ret = atomic_inc_and_test(&cm_id_priv->work_count);
+	if (!ret)
+		list_add_tail(&work->list, &cm_id_priv->work_list);
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	if (ret)
+		cm_process_work(cm_id_priv, work);
+	else
+		cm_deref_id(cm_id_priv);
+	return 0;
+out:
+	cm_deref_id(cm_id_priv);
+	return -EINVAL;
+}
+
+static int cm_timewait_handler(struct cm_work *work)
+{
+	struct cm_timewait_info *timewait_info;
+	struct cm_id_private *cm_id_priv;
+	unsigned long flags;
+	int ret;
+
+	timewait_info = (struct cm_timewait_info *)work;
+	cm_cleanup_timewait(timewait_info);
+
+	cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
+				   timewait_info->work.remote_id);
+	if (!cm_id_priv)
+		return -EINVAL;
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
+	    cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		goto out;
+	}
+	cm_id_priv->id.state = IB_CM_IDLE;
+	ret = atomic_inc_and_test(&cm_id_priv->work_count);
+	if (!ret)
+		list_add_tail(&work->list, &cm_id_priv->work_list);
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	if (ret)
+		cm_process_work(cm_id_priv, work);
+	else
+		cm_deref_id(cm_id_priv);
+	return 0;
+out:
+	cm_deref_id(cm_id_priv);
+	return -EINVAL;
+}
+
+static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
+			       struct cm_id_private *cm_id_priv,
+			       struct ib_cm_sidr_req_param *param)
+{
+	cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
+			  cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
+	sidr_req_msg->request_id = cm_id_priv->id.local_id;
+	sidr_req_msg->pkey = param->pkey;
+	sidr_req_msg->service_id = param->service_id;
+
+	if (param->private_data && param->private_data_len)
+		memcpy(sidr_req_msg->private_data, param->private_data,
+		       param->private_data_len);
+}
+
+int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
+			struct ib_cm_sidr_req_param *param)
+{
+	struct cm_id_private *cm_id_priv;
+	struct ib_mad_send_buf *msg;
+	struct ib_send_wr *bad_send_wr;
+	unsigned long flags;
+	int ret;
+
+	if (!param->path || (param->private_data &&
+	     param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
+		return -EINVAL;
+
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+	ret = cm_init_av_by_path(param->path, &cm_id_priv->av);
+	if (ret)
+		goto out;
+
+	cm_id->service_id = param->service_id;
+	cm_id->service_mask = ~0ULL;
+	cm_id_priv->timeout_ms = param->timeout_ms;
+	cm_id_priv->max_cm_retries = param->max_cm_retries;
+	ret = cm_alloc_msg(cm_id_priv, &msg);
+	if (ret)
+		goto out;
+
+	cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
+			   param);
+	msg->send_wr.wr.ud.timeout_ms = cm_id_priv->timeout_ms;
+	msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id->state == IB_CM_IDLE)
+		ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
+				       &msg->send_wr, &bad_send_wr);
+	else
+		ret = -EINVAL;
+
+	if (ret) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		cm_free_msg(msg);
+		goto out;
+	}
+	cm_id->state = IB_CM_SIDR_REQ_SENT;
+	cm_id_priv->msg = msg;
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+out:
+	return ret;
+}
+EXPORT_SYMBOL(ib_send_cm_sidr_req);
+
+static void cm_format_sidr_req_event(struct cm_work *work,
+				     struct ib_cm_id *listen_id)
+{
+	struct cm_sidr_req_msg *sidr_req_msg;
+	struct ib_cm_sidr_req_event_param *param;
+
+	sidr_req_msg = (struct cm_sidr_req_msg *)
+				work->mad_recv_wc->recv_buf.mad;
+	param = &work->cm_event.param.sidr_req_rcvd;
+	param->pkey = sidr_req_msg->pkey;
+	param->listen_id = listen_id;
+	param->device = work->port->mad_agent->device;
+	param->port = work->port->port_num;
+	work->cm_event.private_data = &sidr_req_msg->private_data;
+}
+
+static int cm_sidr_req_handler(struct cm_work *work)
+{
+	struct ib_cm_id *cm_id;
+	struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
+	struct cm_sidr_req_msg *sidr_req_msg;
+	struct ib_wc *wc;
+	unsigned long flags;
+
+	cm_id = ib_create_cm_id(NULL, NULL);
+	if (IS_ERR(cm_id))
+		return PTR_ERR(cm_id);
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+
+	/* Record SGID/SLID and request ID for lookup. */
+	sidr_req_msg = (struct cm_sidr_req_msg *)
+				work->mad_recv_wc->recv_buf.mad;
+	wc = work->mad_recv_wc->wc;
+	cm_id_priv->av.dgid.global.subnet_prefix = wc->slid;
+	cm_id_priv->av.dgid.global.interface_id = 0;
+	cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
+				&cm_id_priv->av);
+	cm_id_priv->id.remote_id = sidr_req_msg->request_id;
+	cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
+	cm_id_priv->tid = sidr_req_msg->hdr.tid;
+	atomic_inc(&cm_id_priv->work_count);
+
+	spin_lock_irqsave(&cm.lock, flags);
+	cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
+	if (cur_cm_id_priv) {
+		spin_unlock_irqrestore(&cm.lock, flags);
+		goto out; /* Duplicate message. */
+	}
+	cur_cm_id_priv = cm_find_listen(sidr_req_msg->service_id);
+	if (!cur_cm_id_priv) {
+		rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
+		spin_unlock_irqrestore(&cm.lock, flags);
+		/* todo: reply with no match */
+		goto out; /* No match. */
+	}
+	atomic_inc(&cur_cm_id_priv->refcount);
+	spin_unlock_irqrestore(&cm.lock, flags);
+
+	cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
+	cm_id_priv->id.context = cur_cm_id_priv->id.context;
+	cm_id_priv->id.service_id = sidr_req_msg->service_id;
+	cm_id_priv->id.service_mask = ~0ULL;
+
+	cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
+	cm_process_work(cm_id_priv, work);
+	cm_deref_id(cur_cm_id_priv);
+	return 0;
+out:
+	ib_destroy_cm_id(&cm_id_priv->id);
+	return -EINVAL;
+}
+
+static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
+			       struct cm_id_private *cm_id_priv,
+			       struct ib_cm_sidr_rep_param *param)
+{
+	cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
+			  cm_id_priv->tid);
+	sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
+	sidr_rep_msg->status = param->status;
+	cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
+	sidr_rep_msg->service_id = cm_id_priv->id.service_id;
+	sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
+
+	if (param->info && param->info_length)
+		memcpy(sidr_rep_msg->info, param->info, param->info_length);
+
+	if (param->private_data && param->private_data_len)
+		memcpy(sidr_rep_msg->private_data, param->private_data,
+		       param->private_data_len);
+}
+
+int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
+			struct ib_cm_sidr_rep_param *param)
+{
+	struct cm_id_private *cm_id_priv;
+	struct ib_mad_send_buf *msg;
+	struct ib_send_wr *bad_send_wr;
+	unsigned long flags;
+	int ret;
+
+	if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
+	    (param->private_data &&
+	     param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
+		return -EINVAL;
+
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
+		ret = -EINVAL;
+		goto error;
+	}
+
+	ret = cm_alloc_msg(cm_id_priv, &msg);
+	if (ret)
+		goto error;
+
+	cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
+			   param);
+	ret = ib_post_send_mad(cm_id_priv->av.port->mad_agent,
+			       &msg->send_wr, &bad_send_wr);
+	if (ret) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		cm_free_msg(msg);
+		return ret;
+	}
+	cm_id->state = IB_CM_IDLE;
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	spin_lock_irqsave(&cm.lock, flags);
+	rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
+	spin_unlock_irqrestore(&cm.lock, flags);
+	return 0;
+
+error:	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	return ret;
+}
+EXPORT_SYMBOL(ib_send_cm_sidr_rep);
+
+static void cm_format_sidr_rep_event(struct cm_work *work)
+{
+	struct cm_sidr_rep_msg *sidr_rep_msg;
+	struct ib_cm_sidr_rep_event_param *param;
+
+	sidr_rep_msg = (struct cm_sidr_rep_msg *)
+				work->mad_recv_wc->recv_buf.mad;
+	param = &work->cm_event.param.sidr_rep_rcvd;
+	param->status = sidr_rep_msg->status;
+	param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
+	param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
+	param->info = &sidr_rep_msg->info;
+	param->info_len = sidr_rep_msg->info_length;
+	work->cm_event.private_data = &sidr_rep_msg->private_data;
+}
+
+static int cm_sidr_rep_handler(struct cm_work *work)
+{
+	struct cm_sidr_rep_msg *sidr_rep_msg;
+	struct cm_id_private *cm_id_priv;
+	unsigned long flags;
+
+	sidr_rep_msg = (struct cm_sidr_rep_msg *)
+				work->mad_recv_wc->recv_buf.mad;
+	cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
+	if (!cm_id_priv)
+		return -EINVAL; /* Unmatched reply. */
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
+		spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+		goto out;
+	}
+	cm_id_priv->id.state = IB_CM_IDLE;
+	ib_cancel_mad(cm_id_priv->av.port->mad_agent,
+		      (unsigned long) cm_id_priv->msg);
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	cm_format_sidr_rep_event(work);
+	cm_process_work(cm_id_priv, work);
+	return 0;
+out:
+	cm_deref_id(cm_id_priv);
+	return -EINVAL;
+}
+
+static void cm_process_send_error(struct ib_mad_send_buf *msg,
+				  enum ib_wc_status wc_status)
+{
+	struct cm_id_private *cm_id_priv;
+	struct ib_cm_event cm_event;
+	enum ib_cm_state state;
+	unsigned long flags;
+	int ret;
+
+	memset(&cm_event, 0, sizeof cm_event);
+	cm_id_priv = msg->context[0];
+
+	/* Discard old sends or ones without a response. */
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	state = (enum ib_cm_state) (unsigned long) msg->context[1];
+	if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
+		goto discard;
+
+	switch (state) {
+	case IB_CM_REQ_SENT:
+	case IB_CM_MRA_REQ_RCVD:
+		cm_reset_to_idle(cm_id_priv);
+		cm_event.event = IB_CM_REQ_ERROR;
+		break;
+	case IB_CM_REP_SENT:
+	case IB_CM_MRA_REP_RCVD:
+		cm_reset_to_idle(cm_id_priv);
+		cm_event.event = IB_CM_REP_ERROR;
+		break;
+	case IB_CM_DREQ_SENT:
+		cm_enter_timewait(cm_id_priv);
+		cm_event.event = IB_CM_DREQ_ERROR;
+		break;
+	case IB_CM_SIDR_REQ_SENT:
+		cm_id_priv->id.state = IB_CM_IDLE;
+		cm_event.event = IB_CM_SIDR_REQ_ERROR;
+		break;
+	default:
+		goto discard;
+	}
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	cm_event.param.send_status = wc_status;
+
+	/* No other events can occur on the cm_id at this point. */
+	ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
+	cm_free_msg(msg);
+	if (ret)
+		ib_destroy_cm_id(&cm_id_priv->id);
+	return;
+discard:
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	cm_free_msg(msg);
+}
+
+static void cm_send_handler(struct ib_mad_agent *mad_agent,
+			    struct ib_mad_send_wc *mad_send_wc)
+{
+	struct ib_mad_send_buf *msg;
+
+	msg = (struct ib_mad_send_buf *)(unsigned long)mad_send_wc->wr_id;
+
+	switch (mad_send_wc->status) {
+	case IB_WC_SUCCESS:
+	case IB_WC_WR_FLUSH_ERR:
+		cm_free_msg(msg);
+		break;
+	default:
+		if (msg->context[0] && msg->context[1])
+			cm_process_send_error(msg, mad_send_wc->status);
+		else
+			cm_free_msg(msg);
+		break;
+	}
+}
+
+static void cm_work_handler(void *data)
+{
+	struct cm_work *work = data;
+	int ret;
+
+	switch (work->cm_event.event) {
+	case IB_CM_REQ_RECEIVED:
+		ret = cm_req_handler(work);
+		break;
+	case IB_CM_MRA_RECEIVED:
+		ret = cm_mra_handler(work);
+		break;
+	case IB_CM_REJ_RECEIVED:
+		ret = cm_rej_handler(work);
+		break;
+	case IB_CM_REP_RECEIVED:
+		ret = cm_rep_handler(work);
+		break;
+	case IB_CM_RTU_RECEIVED:
+		ret = cm_rtu_handler(work);
+		break;
+	case IB_CM_USER_ESTABLISHED:
+		ret = cm_establish_handler(work);
+		break;
+	case IB_CM_DREQ_RECEIVED:
+		ret = cm_dreq_handler(work);
+		break;
+	case IB_CM_DREP_RECEIVED:
+		ret = cm_drep_handler(work);
+		break;
+	case IB_CM_SIDR_REQ_RECEIVED:
+		ret = cm_sidr_req_handler(work);
+		break;
+	case IB_CM_SIDR_REP_RECEIVED:
+		ret = cm_sidr_rep_handler(work);
+		break;
+	case IB_CM_LAP_RECEIVED:
+		ret = cm_lap_handler(work);
+		break;
+	case IB_CM_APR_RECEIVED:
+		ret = cm_apr_handler(work);
+		break;
+	case IB_CM_TIMEWAIT_EXIT:
+		ret = cm_timewait_handler(work);
+		break;
+	default:
+		ret = -EINVAL;
+		break;
+	}
+	if (ret)
+		cm_free_work(work);
+}
+
+int ib_cm_establish(struct ib_cm_id *cm_id)
+{
+	struct cm_id_private *cm_id_priv;
+	struct cm_work *work;
+	unsigned long flags;
+	int ret = 0;
+
+	work = kmalloc(sizeof *work, GFP_ATOMIC);
+	if (!work)
+		return -ENOMEM;
+
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	switch (cm_id->state)
+	{
+	case IB_CM_REP_SENT:
+	case IB_CM_MRA_REP_RCVD:
+		cm_id->state = IB_CM_ESTABLISHED;
+		break;
+	case IB_CM_ESTABLISHED:
+		ret = -EISCONN;
+		break;
+	default:
+		ret = -EINVAL;
+		break;
+	}
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+
+	if (ret) {
+		kfree(work);
+		goto out;
+	}
+
+	/*
+	 * The CM worker thread may try to destroy the cm_id before it
+	 * can execute this work item.  To prevent potential deadlock,
+	 * we need to find the cm_id once we're in the context of the
+	 * worker thread, rather than holding a reference on it.
+	 */
+	INIT_WORK(&work->work, cm_work_handler, work);
+	work->local_id = cm_id->local_id;
+	work->remote_id = cm_id->remote_id;
+	work->mad_recv_wc = NULL;
+	work->cm_event.event = IB_CM_USER_ESTABLISHED;
+	queue_work(cm.wq, &work->work);
+out:
+	return ret;
+}
+EXPORT_SYMBOL(ib_cm_establish);
+
+static void cm_recv_handler(struct ib_mad_agent *mad_agent,
+			    struct ib_mad_recv_wc *mad_recv_wc)
+{
+	struct cm_work *work;
+	enum ib_cm_event_type event;
+	int paths = 0;
+
+	switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
+	case CM_REQ_ATTR_ID:
+		paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
+						    alt_local_lid != 0);
+		event = IB_CM_REQ_RECEIVED;
+		break;
+	case CM_MRA_ATTR_ID:
+		event = IB_CM_MRA_RECEIVED;
+		break;
+	case CM_REJ_ATTR_ID:
+		event = IB_CM_REJ_RECEIVED;
+		break;
+	case CM_REP_ATTR_ID:
+		event = IB_CM_REP_RECEIVED;
+		break;
+	case CM_RTU_ATTR_ID:
+		event = IB_CM_RTU_RECEIVED;
+		break;
+	case CM_DREQ_ATTR_ID:
+		event = IB_CM_DREQ_RECEIVED;
+		break;
+	case CM_DREP_ATTR_ID:
+		event = IB_CM_DREP_RECEIVED;
+		break;
+	case CM_SIDR_REQ_ATTR_ID:
+		event = IB_CM_SIDR_REQ_RECEIVED;
+		break;
+	case CM_SIDR_REP_ATTR_ID:
+		event = IB_CM_SIDR_REP_RECEIVED;
+		break;
+	case CM_LAP_ATTR_ID:
+		paths = 1;
+		event = IB_CM_LAP_RECEIVED;
+		break;
+	case CM_APR_ATTR_ID:
+		event = IB_CM_APR_RECEIVED;
+		break;
+	default:
+		ib_free_recv_mad(mad_recv_wc);
+		return;
+	}
+
+	work = kmalloc(sizeof *work + sizeof(struct ib_sa_path_rec) * paths,
+		       GFP_KERNEL);
+	if (!work) {
+		ib_free_recv_mad(mad_recv_wc);
+		return;
+	}
+
+	INIT_WORK(&work->work, cm_work_handler, work);
+	work->cm_event.event = event;
+	work->mad_recv_wc = mad_recv_wc;
+	work->port = (struct cm_port *)mad_agent->context;
+	queue_work(cm.wq, &work->work);
+}
+
+static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
+				struct ib_qp_attr *qp_attr,
+				int *qp_attr_mask)
+{
+	unsigned long flags;
+	int ret;
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	switch (cm_id_priv->id.state) {
+	case IB_CM_REQ_SENT:
+	case IB_CM_MRA_REQ_RCVD:
+	case IB_CM_REQ_RCVD:
+	case IB_CM_MRA_REQ_SENT:
+	case IB_CM_REP_RCVD:
+	case IB_CM_MRA_REP_SENT:
+	case IB_CM_REP_SENT:
+	case IB_CM_MRA_REP_RCVD:
+	case IB_CM_ESTABLISHED:
+		*qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
+				IB_QP_PKEY_INDEX | IB_QP_PORT;
+		qp_attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE;
+		if (cm_id_priv->responder_resources)
+			qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_WRITE |
+						    IB_ACCESS_REMOTE_READ;
+		qp_attr->pkey_index = cm_id_priv->av.pkey_index;
+		qp_attr->port_num = cm_id_priv->av.port->port_num;
+		ret = 0;
+		break;
+	default:
+		ret = -EINVAL;
+		break;
+	}
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	return ret;
+}
+
+static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
+			       struct ib_qp_attr *qp_attr,
+			       int *qp_attr_mask)
+{
+	unsigned long flags;
+	int ret;
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	switch (cm_id_priv->id.state) {
+	case IB_CM_REQ_RCVD:
+	case IB_CM_MRA_REQ_SENT:
+	case IB_CM_REP_RCVD:
+	case IB_CM_MRA_REP_SENT:
+	case IB_CM_REP_SENT:
+	case IB_CM_MRA_REP_RCVD:
+	case IB_CM_ESTABLISHED:
+		*qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
+				IB_QP_DEST_QPN | IB_QP_RQ_PSN |
+				IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
+		qp_attr->ah_attr = cm_id_priv->av.ah_attr;
+		qp_attr->path_mtu = cm_id_priv->path_mtu;
+		qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
+		qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
+		qp_attr->max_dest_rd_atomic = cm_id_priv->responder_resources;
+		qp_attr->min_rnr_timer = 0;
+		if (cm_id_priv->alt_av.ah_attr.dlid) {
+			*qp_attr_mask |= IB_QP_ALT_PATH;
+			qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
+		}
+		ret = 0;
+		break;
+	default:
+		ret = -EINVAL;
+		break;
+	}
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	return ret;
+}
+
+static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
+			       struct ib_qp_attr *qp_attr,
+			       int *qp_attr_mask)
+{
+	unsigned long flags;
+	int ret;
+
+	spin_lock_irqsave(&cm_id_priv->lock, flags);
+	switch (cm_id_priv->id.state) {
+	case IB_CM_REP_RCVD:
+	case IB_CM_MRA_REP_SENT:
+	case IB_CM_REP_SENT:
+	case IB_CM_MRA_REP_RCVD:
+	case IB_CM_ESTABLISHED:
+		*qp_attr_mask = IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
+				IB_QP_RNR_RETRY | IB_QP_SQ_PSN |
+				IB_QP_MAX_QP_RD_ATOMIC;
+		qp_attr->timeout = cm_id_priv->local_ack_timeout;
+		qp_attr->retry_cnt = cm_id_priv->retry_count;
+		qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
+		qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
+		qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
+		if (cm_id_priv->alt_av.ah_attr.dlid) {
+			*qp_attr_mask |= IB_QP_PATH_MIG_STATE;
+			qp_attr->path_mig_state = IB_MIG_REARM;
+		}
+		ret = 0;
+		break;
+	default:
+		ret = -EINVAL;
+		break;
+	}
+	spin_unlock_irqrestore(&cm_id_priv->lock, flags);
+	return ret;
+}
+
+int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
+		       struct ib_qp_attr *qp_attr,
+		       int *qp_attr_mask)
+{
+	struct cm_id_private *cm_id_priv;
+	int ret;
+
+	cm_id_priv = container_of(cm_id, struct cm_id_private, id);
+	switch (qp_attr->qp_state) {
+	case IB_QPS_INIT:
+		ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
+		break;
+	case IB_QPS_RTR:
+		ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
+		break;
+	case IB_QPS_RTS:
+		ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
+		break;
+	default:
+		ret = -EINVAL;
+		break;
+	}
+	return ret;
+}
+EXPORT_SYMBOL(ib_cm_init_qp_attr);
+
+static u64 cm_get_ca_guid(struct ib_device *device)
+{
+	struct ib_device_attr *device_attr;
+	u64 guid;
+	int ret;
+
+	device_attr = kmalloc(sizeof *device_attr, GFP_KERNEL);
+	if (!device_attr)
+		return 0;
+
+	ret = ib_query_device(device, device_attr);
+	guid = ret ? 0 : device_attr->node_guid;
+	kfree(device_attr);
+	return guid;
+}
+
+static void cm_add_one(struct ib_device *device)
+{
+	struct cm_device *cm_dev;
+	struct cm_port *port;
+	struct ib_mad_reg_req reg_req = {
+		.mgmt_class = IB_MGMT_CLASS_CM,
+		.mgmt_class_version = IB_CM_CLASS_VERSION
+	};
+	struct ib_port_modify port_modify = {
+		.set_port_cap_mask = IB_PORT_CM_SUP
+	};
+	unsigned long flags;
+	int ret;
+	u8 i;
+
+	cm_dev = kmalloc(sizeof(*cm_dev) + sizeof(*port) *
+			 device->phys_port_cnt, GFP_KERNEL);
+	if (!cm_dev)
+		return;
+
+	cm_dev->device = device;
+	cm_dev->ca_guid = cm_get_ca_guid(device);
+	if (!cm_dev->ca_guid)
+		goto error1;
+
+	set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
+	for (i = 1; i <= device->phys_port_cnt; i++) {
+		port = &cm_dev->port[i-1];
+		port->cm_dev = cm_dev;
+		port->port_num = i;
+		port->mad_agent = ib_register_mad_agent(device, i,
+							IB_QPT_GSI,
+							&reg_req,
+							0,
+							cm_send_handler,
+							cm_recv_handler,
+							port);
+		if (IS_ERR(port->mad_agent))
+			goto error2;
+
+		ret = ib_modify_port(device, i, 0, &port_modify);
+		if (ret)
+			goto error3;
+	}
+	ib_set_client_data(device, &cm_client, cm_dev);
+
+	write_lock_irqsave(&cm.device_lock, flags);
+	list_add_tail(&cm_dev->list, &cm.device_list);
+	write_unlock_irqrestore(&cm.device_lock, flags);
+	return;
+
+error3:
+	ib_unregister_mad_agent(port->mad_agent);
+error2:
+	port_modify.set_port_cap_mask = 0;
+	port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
+	while (--i) {
+		port = &cm_dev->port[i-1];
+		ib_modify_port(device, port->port_num, 0, &port_modify);
+		ib_unregister_mad_agent(port->mad_agent);
+	}
+error1:
+	kfree(cm_dev);
+}
+
+static void cm_remove_one(struct ib_device *device)
+{
+	struct cm_device *cm_dev;
+	struct cm_port *port;
+	struct ib_port_modify port_modify = {
+		.clr_port_cap_mask = IB_PORT_CM_SUP
+	};
+	unsigned long flags;
+	int i;
+
+	cm_dev = ib_get_client_data(device, &cm_client);
+	if (!cm_dev)
+		return;
+
+	write_lock_irqsave(&cm.device_lock, flags);
+	list_del(&cm_dev->list);
+	write_unlock_irqrestore(&cm.device_lock, flags);
+
+	for (i = 1; i <= device->phys_port_cnt; i++) {
+		port = &cm_dev->port[i-1];
+		ib_modify_port(device, port->port_num, 0, &port_modify);
+		ib_unregister_mad_agent(port->mad_agent);
+	}
+	kfree(cm_dev);
+}
+
+static int __init ib_cm_init(void)
+{
+	int ret;
+
+	memset(&cm, 0, sizeof cm);
+	INIT_LIST_HEAD(&cm.device_list);
+	rwlock_init(&cm.device_lock);
+	spin_lock_init(&cm.lock);
+	cm.listen_service_table = RB_ROOT;
+	cm.listen_service_id = __constant_be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
+	cm.remote_id_table = RB_ROOT;
+	cm.remote_qp_table = RB_ROOT;
+	cm.remote_sidr_table = RB_ROOT;
+	idr_init(&cm.local_id_table);
+	idr_pre_get(&cm.local_id_table, GFP_KERNEL);
+
+	cm.wq = create_workqueue("ib_cm");
+	if (!cm.wq)
+		return -ENOMEM;
+
+	ret = ib_register_client(&cm_client);
+	if (ret)
+		goto error;
+
+	return 0;
+error:
+	destroy_workqueue(cm.wq);
+	return ret;
+}
+
+static void __exit ib_cm_cleanup(void)
+{
+	flush_workqueue(cm.wq);
+	destroy_workqueue(cm.wq);
+	ib_unregister_client(&cm_client);
+}
+
+module_init(ib_cm_init);
+module_exit(ib_cm_cleanup);
+
diff --git a/drivers/infiniband/core/cm_msgs.h b/drivers/infiniband/core/cm_msgs.h
new file mode 100644
index 000000000000..15a309a77b2b
--- /dev/null
+++ b/drivers/infiniband/core/cm_msgs.h
@@ -0,0 +1,819 @@
+/*
+ * Copyright (c) 2004 Intel Corporation.  All rights reserved.
+ * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
+ * Copyright (c) 2004 Voltaire 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 the madirectory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use source and binary forms, with or
+ *     withmodification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retathe above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHWARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS THE
+ * SOFTWARE.
+ */
+#if !defined(CM_MSGS_H)
+#define CM_MSGS_H
+
+#include <ib_mad.h>
+
+/*
+ * Parameters to routines below should be in network-byte order, and values
+ * are returned in network-byte order.
+ */
+
+#define IB_CM_CLASS_VERSION	2 /* IB specification 1.2 */
+
+enum cm_msg_attr_id {
+	CM_REQ_ATTR_ID	    = __constant_htons(0x0010),
+	CM_MRA_ATTR_ID	    = __constant_htons(0x0011),
+	CM_REJ_ATTR_ID	    = __constant_htons(0x0012),
+	CM_REP_ATTR_ID	    = __constant_htons(0x0013),
+	CM_RTU_ATTR_ID	    = __constant_htons(0x0014),
+	CM_DREQ_ATTR_ID	    = __constant_htons(0x0015),
+	CM_DREP_ATTR_ID	    = __constant_htons(0x0016),
+	CM_SIDR_REQ_ATTR_ID = __constant_htons(0x0017),
+	CM_SIDR_REP_ATTR_ID = __constant_htons(0x0018),
+	CM_LAP_ATTR_ID      = __constant_htons(0x0019),
+	CM_APR_ATTR_ID      = __constant_htons(0x001A)
+};
+
+enum cm_msg_sequence {
+	CM_MSG_SEQUENCE_REQ,
+	CM_MSG_SEQUENCE_LAP,
+	CM_MSG_SEQUENCE_DREQ,
+	CM_MSG_SEQUENCE_SIDR
+};
+
+struct cm_req_msg {
+	struct ib_mad_hdr hdr;
+
+	u32 local_comm_id;
+	u32 rsvd4;
+	u64 service_id;
+	u64 local_ca_guid;
+	u32 rsvd24;
+	u32 local_qkey;
+	/* local QPN:24, responder resources:8 */
+	u32 offset32;
+	/* local EECN:24, initiator depth:8 */
+	u32 offset36;
+	/*
+	 * remote EECN:24, remote CM response timeout:5,
+	 * transport service type:2, end-to-end flow control:1
+	 */
+	u32 offset40;
+	/* starting PSN:24, local CM response timeout:5, retry count:3 */
+	u32 offset44;
+	u16 pkey;
+	/* path MTU:4, RDC exists:1, RNR retry count:3. */
+	u8 offset50;
+	/* max CM Retries:4, SRQ:1, rsvd:3 */
+	u8 offset51;
+
+	u16 primary_local_lid;
+	u16 primary_remote_lid;
+	union ib_gid primary_local_gid;
+	union ib_gid primary_remote_gid;
+	/* flow label:20, rsvd:6, packet rate:6 */
+	u32 primary_offset88;
+	u8 primary_traffic_class;
+	u8 primary_hop_limit;
+	/* SL:4, subnet local:1, rsvd:3 */
+	u8 primary_offset94;
+	/* local ACK timeout:5, rsvd:3 */
+	u8 primary_offset95;
+
+	u16 alt_local_lid;
+	u16 alt_remote_lid;
+	union ib_gid alt_local_gid;
+	union ib_gid alt_remote_gid;
+	/* flow label:20, rsvd:6, packet rate:6 */
+	u32 alt_offset132;
+	u8 alt_traffic_class;
+	u8 alt_hop_limit;
+	/* SL:4, subnet local:1, rsvd:3 */
+	u8 alt_offset138;
+	/* local ACK timeout:5, rsvd:3 */
+	u8 alt_offset139;
+
+	u8 private_data[IB_CM_REQ_PRIVATE_DATA_SIZE];
+
+} __attribute__ ((packed));
+
+static inline u32 cm_req_get_local_qpn(struct cm_req_msg *req_msg)
+{
+	return cpu_to_be32(be32_to_cpu(req_msg->offset32) >> 8);
+}
+
+static inline void cm_req_set_local_qpn(struct cm_req_msg *req_msg, u32 qpn)
+{
+	req_msg->offset32 = cpu_to_be32((be32_to_cpu(qpn) << 8) |
+					 (be32_to_cpu(req_msg->offset32) &
+					  0x000000FF));
+}
+
+static inline u8 cm_req_get_resp_res(struct cm_req_msg *req_msg)
+{
+	return (u8) be32_to_cpu(req_msg->offset32);
+}
+
+static inline void cm_req_set_resp_res(struct cm_req_msg *req_msg, u8 resp_res)
+{
+	req_msg->offset32 = cpu_to_be32(resp_res |
+					(be32_to_cpu(req_msg->offset32) &
+					 0xFFFFFF00));
+}
+
+static inline u8 cm_req_get_init_depth(struct cm_req_msg *req_msg)
+{
+	return (u8) be32_to_cpu(req_msg->offset36);
+}
+
+static inline void cm_req_set_init_depth(struct cm_req_msg *req_msg,
+					 u8 init_depth)
+{
+	req_msg->offset36 = cpu_to_be32(init_depth |
+					(be32_to_cpu(req_msg->offset36) &
+					 0xFFFFFF00));
+}
+
+static inline u8 cm_req_get_remote_resp_timeout(struct cm_req_msg *req_msg)
+{
+	return (u8) ((be32_to_cpu(req_msg->offset40) & 0xF8) >> 3);
+}
+
+static inline void cm_req_set_remote_resp_timeout(struct cm_req_msg *req_msg,
+						  u8 resp_timeout)
+{
+	req_msg->offset40 = cpu_to_be32((resp_timeout << 3) |
+					 (be32_to_cpu(req_msg->offset40) &
+					  0xFFFFFF07));
+}
+
+static inline enum ib_qp_type cm_req_get_qp_type(struct cm_req_msg *req_msg)
+{
+	u8 transport_type = (u8) (be32_to_cpu(req_msg->offset40) & 0x06) >> 1;
+	switch(transport_type) {
+	case 0: return IB_QPT_RC;
+	case 1: return IB_QPT_UC;
+	default: return 0;
+	}
+}
+
+static inline void cm_req_set_qp_type(struct cm_req_msg *req_msg,
+				      enum ib_qp_type qp_type)
+{
+	switch(qp_type) {
+	case IB_QPT_UC:
+		req_msg->offset40 = cpu_to_be32((be32_to_cpu(
+						  req_msg->offset40) &
+						   0xFFFFFFF9) | 0x2);
+	default:
+		req_msg->offset40 = cpu_to_be32(be32_to_cpu(
+						 req_msg->offset40) &
+						  0xFFFFFFF9);
+	}
+}
+
+static inline u8 cm_req_get_flow_ctrl(struct cm_req_msg *req_msg)
+{
+	return be32_to_cpu(req_msg->offset40) & 0x1;
+}
+
+static inline void cm_req_set_flow_ctrl(struct cm_req_msg *req_msg,
+					u8 flow_ctrl)
+{
+	req_msg->offset40 = cpu_to_be32((flow_ctrl & 0x1) |
+					 (be32_to_cpu(req_msg->offset40) &
+					  0xFFFFFFFE));
+}
+
+static inline u32 cm_req_get_starting_psn(struct cm_req_msg *req_msg)
+{
+	return cpu_to_be32(be32_to_cpu(req_msg->offset44) >> 8);
+}
+
+static inline void cm_req_set_starting_psn(struct cm_req_msg *req_msg,
+					   u32 starting_psn)
+{
+	req_msg->offset44 = cpu_to_be32((be32_to_cpu(starting_psn) << 8) |
+			    (be32_to_cpu(req_msg->offset44) & 0x000000FF));
+}
+
+static inline u8 cm_req_get_local_resp_timeout(struct cm_req_msg *req_msg)
+{
+	return (u8) ((be32_to_cpu(req_msg->offset44) & 0xF8) >> 3);
+}
+
+static inline void cm_req_set_local_resp_timeout(struct cm_req_msg *req_msg,
+						 u8 resp_timeout)
+{
+	req_msg->offset44 = cpu_to_be32((resp_timeout << 3) |
+			    (be32_to_cpu(req_msg->offset44) & 0xFFFFFF07));
+}
+
+static inline u8 cm_req_get_retry_count(struct cm_req_msg *req_msg)
+{
+	return (u8) (be32_to_cpu(req_msg->offset44) & 0x7);
+}
+
+static inline void cm_req_set_retry_count(struct cm_req_msg *req_msg,
+					  u8 retry_count)
+{
+	req_msg->offset44 = cpu_to_be32((retry_count & 0x7) |
+			    (be32_to_cpu(req_msg->offset44) & 0xFFFFFFF8));
+}
+
+static inline u8 cm_req_get_path_mtu(struct cm_req_msg *req_msg)
+{
+	return req_msg->offset50 >> 4;
+}
+
+static inline void cm_req_set_path_mtu(struct cm_req_msg *req_msg, u8 path_mtu)
+{
+	req_msg->offset50 = (u8) ((req_msg->offset50 & 0xF) | (path_mtu << 4));
+}
+
+static inline u8 cm_req_get_rnr_retry_count(struct cm_req_msg *req_msg)
+{
+	return req_msg->offset50 & 0x7;
+}
+
+static inline void cm_req_set_rnr_retry_count(struct cm_req_msg *req_msg,
+					      u8 rnr_retry_count)
+{
+	req_msg->offset50 = (u8) ((req_msg->offset50 & 0xF8) |
+				  (rnr_retry_count & 0x7));
+}
+
+static inline u8 cm_req_get_max_cm_retries(struct cm_req_msg *req_msg)
+{
+	return req_msg->offset51 >> 4;
+}
+
+static inline void cm_req_set_max_cm_retries(struct cm_req_msg *req_msg,
+					     u8 retries)
+{
+	req_msg->offset51 = (u8) ((req_msg->offset51 & 0xF) | (retries << 4));
+}
+
+static inline u8 cm_req_get_srq(struct cm_req_msg *req_msg)
+{
+	return (req_msg->offset51 & 0x8) >> 3;
+}
+
+static inline void cm_req_set_srq(struct cm_req_msg *req_msg, u8 srq)
+{
+	req_msg->offset51 = (u8) ((req_msg->offset51 & 0xF7) |
+				  ((srq & 0x1) << 3));
+}
+
+static inline u32 cm_req_get_primary_flow_label(struct cm_req_msg *req_msg)
+{
+	return cpu_to_be32((be32_to_cpu(req_msg->primary_offset88) >> 12));
+}
+
+static inline void cm_req_set_primary_flow_label(struct cm_req_msg *req_msg,
+						 u32 flow_label)
+{
+	req_msg->primary_offset88 = cpu_to_be32(
+				    (be32_to_cpu(req_msg->primary_offset88) &
+				     0x00000FFF) |
+				     (be32_to_cpu(flow_label) << 12));
+}
+
+static inline u8 cm_req_get_primary_packet_rate(struct cm_req_msg *req_msg)
+{
+	return (u8) (be32_to_cpu(req_msg->primary_offset88) & 0x3F);
+}
+
+static inline void cm_req_set_primary_packet_rate(struct cm_req_msg *req_msg,
+						  u8 rate)
+{
+	req_msg->primary_offset88 = cpu_to_be32(
+				    (be32_to_cpu(req_msg->primary_offset88) &
+				     0xFFFFFFC0) | (rate & 0x3F));
+}
+
+static inline u8 cm_req_get_primary_sl(struct cm_req_msg *req_msg)
+{
+	return (u8) (req_msg->primary_offset94 >> 4);
+}
+
+static inline void cm_req_set_primary_sl(struct cm_req_msg *req_msg, u8 sl)
+{
+	req_msg->primary_offset94 = (u8) ((req_msg->primary_offset94 & 0x0F) |
+					  (sl << 4));
+}
+
+static inline u8 cm_req_get_primary_subnet_local(struct cm_req_msg *req_msg)
+{
+	return (u8) ((req_msg->primary_offset94 & 0x08) >> 3);
+}
+
+static inline void cm_req_set_primary_subnet_local(struct cm_req_msg *req_msg,
+						   u8 subnet_local)
+{
+	req_msg->primary_offset94 = (u8) ((req_msg->primary_offset94 & 0xF7) |
+					  ((subnet_local & 0x1) << 3));
+}
+
+static inline u8 cm_req_get_primary_local_ack_timeout(struct cm_req_msg *req_msg)
+{
+	return (u8) (req_msg->primary_offset95 >> 3);
+}
+
+static inline void cm_req_set_primary_local_ack_timeout(struct cm_req_msg *req_msg,
+							u8 local_ack_timeout)
+{
+	req_msg->primary_offset95 = (u8) ((req_msg->primary_offset95 & 0x07) |
+					  (local_ack_timeout << 3));
+}
+
+static inline u32 cm_req_get_alt_flow_label(struct cm_req_msg *req_msg)
+{
+	return cpu_to_be32((be32_to_cpu(req_msg->alt_offset132) >> 12));
+}
+
+static inline void cm_req_set_alt_flow_label(struct cm_req_msg *req_msg,
+					     u32 flow_label)
+{
+	req_msg->alt_offset132 = cpu_to_be32(
+				 (be32_to_cpu(req_msg->alt_offset132) &
+				  0x00000FFF) |
+				  (be32_to_cpu(flow_label) << 12));
+}
+
+static inline u8 cm_req_get_alt_packet_rate(struct cm_req_msg *req_msg)
+{
+	return (u8) (be32_to_cpu(req_msg->alt_offset132) & 0x3F);
+}
+
+static inline void cm_req_set_alt_packet_rate(struct cm_req_msg *req_msg,
+					      u8 rate)
+{
+	req_msg->alt_offset132 = cpu_to_be32(
+				 (be32_to_cpu(req_msg->alt_offset132) &
+				  0xFFFFFFC0) | (rate & 0x3F));
+}
+
+static inline u8 cm_req_get_alt_sl(struct cm_req_msg *req_msg)
+{
+	return (u8) (req_msg->alt_offset138 >> 4);
+}
+
+static inline void cm_req_set_alt_sl(struct cm_req_msg *req_msg, u8 sl)
+{
+	req_msg->alt_offset138 = (u8) ((req_msg->alt_offset138 & 0x0F) |
+				       (sl << 4));
+}
+
+static inline u8 cm_req_get_alt_subnet_local(struct cm_req_msg *req_msg)
+{
+	return (u8) ((req_msg->alt_offset138 & 0x08) >> 3);
+}
+
+static inline void cm_req_set_alt_subnet_local(struct cm_req_msg *req_msg,
+					       u8 subnet_local)
+{
+	req_msg->alt_offset138 = (u8) ((req_msg->alt_offset138 & 0xF7) |
+				       ((subnet_local & 0x1) << 3));
+}
+
+static inline u8 cm_req_get_alt_local_ack_timeout(struct cm_req_msg *req_msg)
+{
+	return (u8) (req_msg->alt_offset139 >> 3);
+}
+
+static inline void cm_req_set_alt_local_ack_timeout(struct cm_req_msg *req_msg,
+						    u8 local_ack_timeout)
+{
+	req_msg->alt_offset139 = (u8) ((req_msg->alt_offset139 & 0x07) |
+				       (local_ack_timeout << 3));
+}
+
+/* Message REJected or MRAed */
+enum cm_msg_response {
+	CM_MSG_RESPONSE_REQ = 0x0,
+	CM_MSG_RESPONSE_REP = 0x1,
+	CM_MSG_RESPONSE_OTHER = 0x2
+};
+
+ struct cm_mra_msg {
+	struct ib_mad_hdr hdr;
+
+	u32 local_comm_id;
+	u32 remote_comm_id;
+	/* message MRAed:2, rsvd:6 */
+	u8 offset8;
+	/* service timeout:5, rsvd:3 */
+	u8 offset9;
+
+	u8 private_data[IB_CM_MRA_PRIVATE_DATA_SIZE];
+
+} __attribute__ ((packed));
+
+static inline u8 cm_mra_get_msg_mraed(struct cm_mra_msg *mra_msg)
+{
+	return (u8) (mra_msg->offset8 >> 6);
+}
+
+static inline void cm_mra_set_msg_mraed(struct cm_mra_msg *mra_msg, u8 msg)
+{
+	mra_msg->offset8 = (u8) ((mra_msg->offset8 & 0x3F) | (msg << 6));
+}
+
+static inline u8 cm_mra_get_service_timeout(struct cm_mra_msg *mra_msg)
+{
+	return (u8) (mra_msg->offset9 >> 3);
+}
+
+static inline void cm_mra_set_service_timeout(struct cm_mra_msg *mra_msg,
+					      u8 service_timeout)
+{
+	mra_msg->offset9 = (u8) ((mra_msg->offset9 & 0x07) |
+				 (service_timeout << 3));
+}
+
+struct cm_rej_msg {
+	struct ib_mad_hdr hdr;
+
+	u32 local_comm_id;
+	u32 remote_comm_id;
+	/* message REJected:2, rsvd:6 */
+	u8 offset8;
+	/* reject info length:7, rsvd:1. */
+	u8 offset9;
+	u16 reason;
+	u8 ari[IB_CM_REJ_ARI_LENGTH];
+
+	u8 private_data[IB_CM_REJ_PRIVATE_DATA_SIZE];
+
+} __attribute__ ((packed));
+
+static inline u8 cm_rej_get_msg_rejected(struct cm_rej_msg *rej_msg)
+{
+	return (u8) (rej_msg->offset8 >> 6);
+}
+
+static inline void cm_rej_set_msg_rejected(struct cm_rej_msg *rej_msg, u8 msg)
+{
+	rej_msg->offset8 = (u8) ((rej_msg->offset8 & 0x3F) | (msg << 6));
+}
+
+static inline u8 cm_rej_get_reject_info_len(struct cm_rej_msg *rej_msg)
+{
+	return (u8) (rej_msg->offset9 >> 1);
+}
+
+static inline void cm_rej_set_reject_info_len(struct cm_rej_msg *rej_msg,
+					      u8 len)
+{
+	rej_msg->offset9 = (u8) ((rej_msg->offset9 & 0x1) | (len << 1));
+}
+
+struct cm_rep_msg {
+	struct ib_mad_hdr hdr;
+
+	u32 local_comm_id;
+	u32 remote_comm_id;
+	u32 local_qkey;
+	/* local QPN:24, rsvd:8 */
+	u32 offset12;
+	/* local EECN:24, rsvd:8 */
+	u32 offset16;
+	/* starting PSN:24 rsvd:8 */
+	u32 offset20;
+	u8 resp_resources;
+	u8 initiator_depth;
+	/* target ACK delay:5, failover accepted:2, end-to-end flow control:1 */
+	u8 offset26;
+	/* RNR retry count:3, SRQ:1, rsvd:5 */
+	u8 offset27;
+	u64 local_ca_guid;
+
+	u8 private_data[IB_CM_REP_PRIVATE_DATA_SIZE];
+
+} __attribute__ ((packed));
+
+static inline u32 cm_rep_get_local_qpn(struct cm_rep_msg *rep_msg)
+{
+	return cpu_to_be32(be32_to_cpu(rep_msg->offset12) >> 8);
+}
+
+static inline void cm_rep_set_local_qpn(struct cm_rep_msg *rep_msg, u32 qpn)
+{
+	rep_msg->offset12 = cpu_to_be32((be32_to_cpu(qpn) << 8) |
+			    (be32_to_cpu(rep_msg->offset12) & 0x000000FF));
+}
+
+static inline u32 cm_rep_get_starting_psn(struct cm_rep_msg *rep_msg)
+{
+	return cpu_to_be32(be32_to_cpu(rep_msg->offset20) >> 8);
+}
+
+static inline void cm_rep_set_starting_psn(struct cm_rep_msg *rep_msg,
+					   u32 starting_psn)
+{
+	rep_msg->offset20 = cpu_to_be32((be32_to_cpu(starting_psn) << 8) |
+			    (be32_to_cpu(rep_msg->offset20) & 0x000000FF));
+}
+
+static inline u8 cm_rep_get_target_ack_delay(struct cm_rep_msg *rep_msg)
+{
+	return (u8) (rep_msg->offset26 >> 3);
+}
+
+static inline void cm_rep_set_target_ack_delay(struct cm_rep_msg *rep_msg,
+					       u8 target_ack_delay)
+{
+	rep_msg->offset26 = (u8) ((rep_msg->offset26 & 0x07) |
+				  (target_ack_delay << 3));
+}
+
+static inline u8 cm_rep_get_failover(struct cm_rep_msg *rep_msg)
+{
+	return (u8) ((rep_msg->offset26 & 0x06) >> 1);
+}
+
+static inline void cm_rep_set_failover(struct cm_rep_msg *rep_msg, u8 failover)
+{
+	rep_msg->offset26 = (u8) ((rep_msg->offset26 & 0xF9) |
+				  ((failover & 0x3) << 1));
+}
+
+static inline u8 cm_rep_get_flow_ctrl(struct cm_rep_msg *rep_msg)
+{
+	return (u8) (rep_msg->offset26 & 0x01);
+}
+
+static inline void cm_rep_set_flow_ctrl(struct cm_rep_msg *rep_msg,
+					    u8 flow_ctrl)
+{
+	rep_msg->offset26 = (u8) ((rep_msg->offset26 & 0xFE) |
+				  (flow_ctrl & 0x1));
+}
+
+static inline u8 cm_rep_get_rnr_retry_count(struct cm_rep_msg *rep_msg)
+{
+	return (u8) (rep_msg->offset27 >> 5);
+}
+
+static inline void cm_rep_set_rnr_retry_count(struct cm_rep_msg *rep_msg,
+					      u8 rnr_retry_count)
+{
+	rep_msg->offset27 = (u8) ((rep_msg->offset27 & 0x1F) |
+				  (rnr_retry_count << 5));
+}
+
+static inline u8 cm_rep_get_srq(struct cm_rep_msg *rep_msg)
+{
+	return (u8) ((rep_msg->offset27 >> 4) & 0x1);
+}
+
+static inline void cm_rep_set_srq(struct cm_rep_msg *rep_msg, u8 srq)
+{
+	rep_msg->offset27 = (u8) ((rep_msg->offset27 & 0xEF) |
+				  ((srq & 0x1) << 4));
+}
+
+struct cm_rtu_msg {
+	struct ib_mad_hdr hdr;
+
+	u32 local_comm_id;
+	u32 remote_comm_id;
+
+	u8 private_data[IB_CM_RTU_PRIVATE_DATA_SIZE];
+
+} __attribute__ ((packed));
+
+struct cm_dreq_msg {
+	struct ib_mad_hdr hdr;
+
+	u32 local_comm_id;
+	u32 remote_comm_id;
+	/* remote QPN/EECN:24, rsvd:8 */
+	u32 offset8;
+
+	u8 private_data[IB_CM_DREQ_PRIVATE_DATA_SIZE];
+
+} __attribute__ ((packed));
+
+static inline u32 cm_dreq_get_remote_qpn(struct cm_dreq_msg *dreq_msg)
+{
+	return cpu_to_be32(be32_to_cpu(dreq_msg->offset8) >> 8);
+}
+
+static inline void cm_dreq_set_remote_qpn(struct cm_dreq_msg *dreq_msg, u32 qpn)
+{
+	dreq_msg->offset8 = cpu_to_be32((be32_to_cpu(qpn) << 8) |
+			    (be32_to_cpu(dreq_msg->offset8) & 0x000000FF));
+}
+
+struct cm_drep_msg {
+	struct ib_mad_hdr hdr;
+
+	u32 local_comm_id;
+	u32 remote_comm_id;
+
+	u8 private_data[IB_CM_DREP_PRIVATE_DATA_SIZE];
+
+} __attribute__ ((packed));
+
+struct cm_lap_msg {
+	struct ib_mad_hdr hdr;
+
+	u32 local_comm_id;
+	u32 remote_comm_id;
+
+	u32 rsvd8;
+	/* remote QPN/EECN:24, remote CM response timeout:5, rsvd:3 */
+	u32 offset12;
+	u32 rsvd16;
+
+	u16 alt_local_lid;
+	u16 alt_remote_lid;
+	union ib_gid alt_local_gid;
+	union ib_gid alt_remote_gid;
+	/* flow label:20, rsvd:4, traffic class:8 */
+	u32 offset56;
+	u8 alt_hop_limit;
+	/* rsvd:2, packet rate:6 */
+	uint8_t offset61;
+	/* SL:4, subnet local:1, rsvd:3 */
+	uint8_t offset62;
+	/* local ACK timeout:5, rsvd:3 */
+	uint8_t offset63;
+
+	u8 private_data[IB_CM_LAP_PRIVATE_DATA_SIZE];
+} __attribute__  ((packed));
+
+static inline u32 cm_lap_get_remote_qpn(struct cm_lap_msg *lap_msg)
+{
+	return cpu_to_be32(be32_to_cpu(lap_msg->offset12) >> 8);
+}
+
+static inline void cm_lap_set_remote_qpn(struct cm_lap_msg *lap_msg, u32 qpn)
+{
+	lap_msg->offset12 = cpu_to_be32((be32_to_cpu(qpn) << 8) |
+					 (be32_to_cpu(lap_msg->offset12) &
+					  0x000000FF));
+}
+
+static inline u8 cm_lap_get_remote_resp_timeout(struct cm_lap_msg *lap_msg)
+{
+	return (u8) ((be32_to_cpu(lap_msg->offset12) & 0xF8) >> 3);
+}
+
+static inline void cm_lap_set_remote_resp_timeout(struct cm_lap_msg *lap_msg,
+						  u8 resp_timeout)
+{
+	lap_msg->offset12 = cpu_to_be32((resp_timeout << 3) |
+					 (be32_to_cpu(lap_msg->offset12) &
+					  0xFFFFFF07));
+}
+
+static inline u32 cm_lap_get_flow_label(struct cm_lap_msg *lap_msg)
+{
+	return be32_to_cpu(lap_msg->offset56) >> 12;
+}
+
+static inline void cm_lap_set_flow_label(struct cm_lap_msg *lap_msg,
+					 u32 flow_label)
+{
+	lap_msg->offset56 = cpu_to_be32((flow_label << 12) |
+					 (be32_to_cpu(lap_msg->offset56) &
+					  0x00000FFF));
+}
+
+static inline u8 cm_lap_get_traffic_class(struct cm_lap_msg *lap_msg)
+{
+	return (u8) be32_to_cpu(lap_msg->offset56);
+}
+
+static inline void cm_lap_set_traffic_class(struct cm_lap_msg *lap_msg,
+					    u8 traffic_class)
+{
+	lap_msg->offset56 = cpu_to_be32(traffic_class |
+					 (be32_to_cpu(lap_msg->offset56) &
+					  0xFFFFFF00));
+}
+
+static inline u8 cm_lap_get_packet_rate(struct cm_lap_msg *lap_msg)
+{
+	return lap_msg->offset61 & 0x3F;
+}
+
+static inline void cm_lap_set_packet_rate(struct cm_lap_msg *lap_msg,
+					  u8 packet_rate)
+{
+	lap_msg->offset61 = (packet_rate & 0x3F) | (lap_msg->offset61 & 0xC0);
+}
+
+static inline u8 cm_lap_get_sl(struct cm_lap_msg *lap_msg)
+{
+	return lap_msg->offset62 >> 4;
+}
+
+static inline void cm_lap_set_sl(struct cm_lap_msg *lap_msg, u8 sl)
+{
+	lap_msg->offset62 = (sl << 4) | (lap_msg->offset62 & 0x0F);
+}
+
+static inline u8 cm_lap_get_subnet_local(struct cm_lap_msg *lap_msg)
+{
+	return (lap_msg->offset62 >> 3) & 0x1;
+}
+
+static inline void cm_lap_set_subnet_local(struct cm_lap_msg *lap_msg,
+					   u8 subnet_local)
+{
+	lap_msg->offset62 = ((subnet_local & 0x1) << 3) |
+			     (lap_msg->offset61 & 0xF7);
+}
+static inline u8 cm_lap_get_local_ack_timeout(struct cm_lap_msg *lap_msg)
+{
+	return lap_msg->offset63 >> 3;
+}
+
+static inline void cm_lap_set_local_ack_timeout(struct cm_lap_msg *lap_msg,
+						u8 local_ack_timeout)
+{
+	lap_msg->offset63 = (local_ack_timeout << 3) |
+			    (lap_msg->offset63 & 0x07);
+}
+
+struct cm_apr_msg {
+	struct ib_mad_hdr hdr;
+
+	u32 local_comm_id;
+	u32 remote_comm_id;
+
+	u8 info_length;
+	u8 ap_status;
+	u8 info[IB_CM_APR_INFO_LENGTH];
+
+	u8 private_data[IB_CM_APR_PRIVATE_DATA_SIZE];
+} __attribute__ ((packed));
+
+struct cm_sidr_req_msg {
+	struct ib_mad_hdr hdr;
+
+	u32 request_id;
+	u16 pkey;
+	u16 rsvd;
+	u64 service_id;
+
+	u8 private_data[IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE];
+} __attribute__ ((packed));
+
+struct cm_sidr_rep_msg {
+	struct ib_mad_hdr hdr;
+
+	u32 request_id;
+	u8 status;
+	u8 info_length;
+	u16 rsvd;
+	/* QPN:24, rsvd:8 */
+	u32 offset8;
+	u64 service_id;
+	u32 qkey;
+	u8 info[IB_CM_SIDR_REP_INFO_LENGTH];
+
+	u8 private_data[IB_CM_SIDR_REP_PRIVATE_DATA_SIZE];
+} __attribute__ ((packed));
+
+static inline u32 cm_sidr_rep_get_qpn(struct cm_sidr_rep_msg *sidr_rep_msg)
+{
+	return cpu_to_be32(be32_to_cpu(sidr_rep_msg->offset8) >> 8);
+}
+
+static inline void cm_sidr_rep_set_qpn(struct cm_sidr_rep_msg *sidr_rep_msg,
+				       u32 qpn)
+{
+	sidr_rep_msg->offset8 = cpu_to_be32((be32_to_cpu(qpn) << 8) |
+					(be32_to_cpu(sidr_rep_msg->offset8) &
+					 0x000000FF));
+}
+
+#endif /* CM_MSGS_H */
diff --git a/drivers/infiniband/core/fmr_pool.c b/drivers/infiniband/core/fmr_pool.c
index 328feae2a5be..7763b31abba7 100644
--- a/drivers/infiniband/core/fmr_pool.c
+++ b/drivers/infiniband/core/fmr_pool.c
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. 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
@@ -29,7 +30,7 @@
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  *
- * $Id: fmr_pool.c 1349 2004-12-16 21:09:43Z roland $
+ * $Id: fmr_pool.c 2730 2005-06-28 16:43:03Z sean.hefty $
  */
 
 #include <linux/errno.h>
@@ -329,7 +330,7 @@ EXPORT_SYMBOL(ib_create_fmr_pool);
  *
  * Destroy an FMR pool and free all associated resources.
  */
-int ib_destroy_fmr_pool(struct ib_fmr_pool *pool)
+void ib_destroy_fmr_pool(struct ib_fmr_pool *pool)
 {
 	struct ib_pool_fmr *fmr;
 	struct ib_pool_fmr *tmp;
@@ -352,8 +353,6 @@ int ib_destroy_fmr_pool(struct ib_fmr_pool *pool)
 
 	kfree(pool->cache_bucket);
 	kfree(pool);
-
-	return 0;
 }
 EXPORT_SYMBOL(ib_destroy_fmr_pool);
 
diff --git a/drivers/infiniband/core/mad.c b/drivers/infiniband/core/mad.c
index 23628c622a50..b97e210ce9c8 100644
--- a/drivers/infiniband/core/mad.c
+++ b/drivers/infiniband/core/mad.c
@@ -1,5 +1,7 @@
 /*
  * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
+ * Copyright (c) 2005 Intel Corporation.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies Ltd.  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
@@ -29,12 +31,12 @@
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  *
- * $Id: mad.c 1389 2004-12-27 22:56:47Z roland $
+ * $Id: mad.c 2817 2005-07-07 11:29:26Z halr $
  */
-
 #include <linux/dma-mapping.h>
 
 #include "mad_priv.h"
+#include "mad_rmpp.h"
 #include "smi.h"
 #include "agent.h"
 
@@ -45,6 +47,7 @@ MODULE_AUTHOR("Sean Hefty");
 
 
 kmem_cache_t *ib_mad_cache;
+
 static struct list_head ib_mad_port_list;
 static u32 ib_mad_client_id = 0;
 
@@ -58,16 +61,12 @@ static int method_in_use(struct ib_mad_mgmt_method_table **method,
 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
 static struct ib_mad_agent_private *find_mad_agent(
 					struct ib_mad_port_private *port_priv,
-					struct ib_mad *mad, int solicited);
+					struct ib_mad *mad);
 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
 				    struct ib_mad_private *mad);
 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
-static void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
-				    struct ib_mad_send_wc *mad_send_wc);
 static void timeout_sends(void *data);
-static void cancel_sends(void *data);
 static void local_completions(void *data);
-static int solicited_mad(struct ib_mad *mad);
 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
 			      struct ib_mad_agent_private *agent_priv,
 			      u8 mgmt_class);
@@ -197,8 +196,8 @@ struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
 	if (qpn == -1)
 		goto error1;
 
-	if (rmpp_version)
-		goto error1;	/* XXX: until RMPP implemented */
+	if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
+		goto error1;
 
 	/* Validate MAD registration request if supplied */
 	if (mad_reg_req) {
@@ -261,22 +260,29 @@ struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
 		ret = ERR_PTR(-ENOMEM);
 		goto error1;
 	}
+	memset(mad_agent_priv, 0, sizeof *mad_agent_priv);
+
+	mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
+						 IB_ACCESS_LOCAL_WRITE);
+	if (IS_ERR(mad_agent_priv->agent.mr)) {
+		ret = ERR_PTR(-ENOMEM);
+		goto error2;
+	}
 
 	if (mad_reg_req) {
 		reg_req = kmalloc(sizeof *reg_req, GFP_KERNEL);
 		if (!reg_req) {
 			ret = ERR_PTR(-ENOMEM);
-			goto error2;
+			goto error3;
 		}
 		/* Make a copy of the MAD registration request */
 		memcpy(reg_req, mad_reg_req, sizeof *reg_req);
 	}
 
 	/* Now, fill in the various structures */
-	memset(mad_agent_priv, 0, sizeof *mad_agent_priv);
 	mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
 	mad_agent_priv->reg_req = reg_req;
-	mad_agent_priv->rmpp_version = rmpp_version;
+	mad_agent_priv->agent.rmpp_version = rmpp_version;
 	mad_agent_priv->agent.device = device;
 	mad_agent_priv->agent.recv_handler = recv_handler;
 	mad_agent_priv->agent.send_handler = send_handler;
@@ -301,7 +307,7 @@ struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
 				if (method) {
 					if (method_in_use(&method,
 							   mad_reg_req))
-						goto error3;
+						goto error4;
 				}
 			}
 			ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
@@ -317,14 +323,14 @@ struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
 					if (is_vendor_method_in_use(
 							vendor_class,
 							mad_reg_req))
-						goto error3;
+						goto error4;
 				}
 			}
 			ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
 		}
 		if (ret2) {
 			ret = ERR_PTR(ret2);
-			goto error3;
+			goto error4;
 		}
 	}
 
@@ -335,22 +341,24 @@ struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
 	spin_lock_init(&mad_agent_priv->lock);
 	INIT_LIST_HEAD(&mad_agent_priv->send_list);
 	INIT_LIST_HEAD(&mad_agent_priv->wait_list);
+	INIT_LIST_HEAD(&mad_agent_priv->done_list);
+	INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
 	INIT_WORK(&mad_agent_priv->timed_work, timeout_sends, mad_agent_priv);
 	INIT_LIST_HEAD(&mad_agent_priv->local_list);
 	INIT_WORK(&mad_agent_priv->local_work, local_completions,
 		   mad_agent_priv);
-	INIT_LIST_HEAD(&mad_agent_priv->canceled_list);
-	INIT_WORK(&mad_agent_priv->canceled_work, cancel_sends, mad_agent_priv);
 	atomic_set(&mad_agent_priv->refcount, 1);
 	init_waitqueue_head(&mad_agent_priv->wait);
 
 	return &mad_agent_priv->agent;
 
-error3:
+error4:
 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 	kfree(reg_req);
-error2:
+error3:
 	kfree(mad_agent_priv);
+error2:
+	ib_dereg_mr(mad_agent_priv->agent.mr);
 error1:
 	return ret;
 }
@@ -487,18 +495,16 @@ static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 	 * MADs, preventing us from queuing additional work
 	 */
 	cancel_mads(mad_agent_priv);
-
 	port_priv = mad_agent_priv->qp_info->port_priv;
-
 	cancel_delayed_work(&mad_agent_priv->timed_work);
-	flush_workqueue(port_priv->wq);
 
 	spin_lock_irqsave(&port_priv->reg_lock, flags);
 	remove_mad_reg_req(mad_agent_priv);
 	list_del(&mad_agent_priv->agent_list);
 	spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 
-	/* XXX: Cleanup pending RMPP receives for this agent */
+	flush_workqueue(port_priv->wq);
+	ib_cancel_rmpp_recvs(mad_agent_priv);
 
 	atomic_dec(&mad_agent_priv->refcount);
 	wait_event(mad_agent_priv->wait,
@@ -506,6 +512,7 @@ static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 
 	if (mad_agent_priv->reg_req)
 		kfree(mad_agent_priv->reg_req);
+	ib_dereg_mr(mad_agent_priv->agent.mr);
 	kfree(mad_agent_priv);
 }
 
@@ -551,6 +558,13 @@ int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
 }
 EXPORT_SYMBOL(ib_unregister_mad_agent);
 
+static inline int response_mad(struct ib_mad *mad)
+{
+	/* Trap represses are responses although response bit is reset */
+	return ((mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
+		(mad->mad_hdr.method & IB_MGMT_METHOD_RESP));
+}
+
 static void dequeue_mad(struct ib_mad_list_head *mad_list)
 {
 	struct ib_mad_queue *mad_queue;
@@ -643,7 +657,7 @@ static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
 				  struct ib_smp *smp,
 				  struct ib_send_wr *send_wr)
 {
-	int ret, solicited;
+	int ret;
 	unsigned long flags;
 	struct ib_mad_local_private *local;
 	struct ib_mad_private *mad_priv;
@@ -689,11 +703,7 @@ static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
 	switch (ret)
 	{
 	case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
-		/*
-		 * See if response is solicited and
-		 * there is a recv handler
-		 */
-		if (solicited_mad(&mad_priv->mad.mad) &&
+		if (response_mad(&mad_priv->mad.mad) &&
 		    mad_agent_priv->agent.recv_handler) {
 			local->mad_priv = mad_priv;
 			local->recv_mad_agent = mad_agent_priv;
@@ -710,15 +720,13 @@ static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
 		break;
 	case IB_MAD_RESULT_SUCCESS:
 		/* Treat like an incoming receive MAD */
-		solicited = solicited_mad(&mad_priv->mad.mad);
 		port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
 					    mad_agent_priv->agent.port_num);
 		if (port_priv) {
 			mad_priv->mad.mad.mad_hdr.tid =
 				((struct ib_mad *)smp)->mad_hdr.tid;
 			recv_mad_agent = find_mad_agent(port_priv,
-						       &mad_priv->mad.mad,
-							solicited);
+						        &mad_priv->mad.mad);
 		}
 		if (!port_priv || !recv_mad_agent) {
 			kmem_cache_free(ib_mad_cache, mad_priv);
@@ -750,43 +758,133 @@ static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
 	list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 	queue_work(mad_agent_priv->qp_info->port_priv->wq,
-		  &mad_agent_priv->local_work);
+		   &mad_agent_priv->local_work);
 	ret = 1;
 out:
 	return ret;
 }
 
-static int ib_send_mad(struct ib_mad_agent_private *mad_agent_priv,
-		       struct ib_mad_send_wr_private *mad_send_wr)
+static int get_buf_length(int hdr_len, int data_len)
+{
+	int seg_size, pad;
+
+	seg_size = sizeof(struct ib_mad) - hdr_len;
+	if (data_len && seg_size) {
+		pad = seg_size - data_len % seg_size;
+		if (pad == seg_size)
+			pad = 0;
+	} else
+		pad = seg_size;
+	return hdr_len + data_len + pad;
+}
+
+struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
+					    u32 remote_qpn, u16 pkey_index,
+					    struct ib_ah *ah, int rmpp_active,
+					    int hdr_len, int data_len,
+					    unsigned int __nocast gfp_mask)
+{
+	struct ib_mad_agent_private *mad_agent_priv;
+	struct ib_mad_send_buf *send_buf;
+	int buf_size;
+	void *buf;
+
+	mad_agent_priv = container_of(mad_agent,
+				      struct ib_mad_agent_private, agent);
+	buf_size = get_buf_length(hdr_len, data_len);
+
+	if ((!mad_agent->rmpp_version &&
+	     (rmpp_active || buf_size > sizeof(struct ib_mad))) ||
+	    (!rmpp_active && buf_size > sizeof(struct ib_mad)))
+		return ERR_PTR(-EINVAL);
+
+	buf = kmalloc(sizeof *send_buf + buf_size, gfp_mask);
+	if (!buf)
+		return ERR_PTR(-ENOMEM);
+	memset(buf, 0, sizeof *send_buf + buf_size);
+
+	send_buf = buf + buf_size;
+	send_buf->mad = buf;
+
+	send_buf->sge.addr = dma_map_single(mad_agent->device->dma_device,
+					    buf, buf_size, DMA_TO_DEVICE);
+	pci_unmap_addr_set(send_buf, mapping, send_buf->sge.addr);
+	send_buf->sge.length = buf_size;
+	send_buf->sge.lkey = mad_agent->mr->lkey;
+
+	send_buf->send_wr.wr_id = (unsigned long) send_buf;
+	send_buf->send_wr.sg_list = &send_buf->sge;
+	send_buf->send_wr.num_sge = 1;
+	send_buf->send_wr.opcode = IB_WR_SEND;
+	send_buf->send_wr.send_flags = IB_SEND_SIGNALED;
+	send_buf->send_wr.wr.ud.ah = ah;
+	send_buf->send_wr.wr.ud.mad_hdr = &send_buf->mad->mad_hdr;
+	send_buf->send_wr.wr.ud.remote_qpn = remote_qpn;
+	send_buf->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
+	send_buf->send_wr.wr.ud.pkey_index = pkey_index;
+
+	if (rmpp_active) {
+		struct ib_rmpp_mad *rmpp_mad;
+		rmpp_mad = (struct ib_rmpp_mad *)send_buf->mad;
+		rmpp_mad->rmpp_hdr.paylen_newwin = cpu_to_be32(hdr_len -
+			offsetof(struct ib_rmpp_mad, data) + data_len);
+		rmpp_mad->rmpp_hdr.rmpp_version = mad_agent->rmpp_version;
+		rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
+		ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr,
+				  IB_MGMT_RMPP_FLAG_ACTIVE);
+	}
+
+	send_buf->mad_agent = mad_agent;
+	atomic_inc(&mad_agent_priv->refcount);
+	return send_buf;
+}
+EXPORT_SYMBOL(ib_create_send_mad);
+
+void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
+{
+	struct ib_mad_agent_private *mad_agent_priv;
+
+	mad_agent_priv = container_of(send_buf->mad_agent,
+				      struct ib_mad_agent_private, agent);
+
+	dma_unmap_single(send_buf->mad_agent->device->dma_device,
+			 pci_unmap_addr(send_buf, mapping),
+			 send_buf->sge.length, DMA_TO_DEVICE);
+	kfree(send_buf->mad);
+
+	if (atomic_dec_and_test(&mad_agent_priv->refcount))
+		wake_up(&mad_agent_priv->wait);
+}
+EXPORT_SYMBOL(ib_free_send_mad);
+
+int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
 {
 	struct ib_mad_qp_info *qp_info;
 	struct ib_send_wr *bad_send_wr;
+	struct list_head *list;
 	unsigned long flags;
 	int ret;
 
-	/* Replace user's WR ID with our own to find WR upon completion */
-	qp_info = mad_agent_priv->qp_info;
-	mad_send_wr->wr_id = mad_send_wr->send_wr.wr_id;
+	/* Set WR ID to find mad_send_wr upon completion */
+	qp_info = mad_send_wr->mad_agent_priv->qp_info;
 	mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
 	mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
 
 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
-	if (qp_info->send_queue.count++ < qp_info->send_queue.max_active) {
-		list_add_tail(&mad_send_wr->mad_list.list,
-			      &qp_info->send_queue.list);
-		spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
-		ret = ib_post_send(mad_agent_priv->agent.qp,
+	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
+		ret = ib_post_send(mad_send_wr->mad_agent_priv->agent.qp,
 				   &mad_send_wr->send_wr, &bad_send_wr);
-		if (ret) {
-			printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
-			dequeue_mad(&mad_send_wr->mad_list);
-		}
+		list = &qp_info->send_queue.list;
 	} else {
-		list_add_tail(&mad_send_wr->mad_list.list,
-			      &qp_info->overflow_list);
-		spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
 		ret = 0;
+		list = &qp_info->overflow_list;
 	}
+
+	if (!ret) {
+		qp_info->send_queue.count++;
+		list_add_tail(&mad_send_wr->mad_list.list, list);
+	}
+	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
 	return ret;
 }
 
@@ -860,18 +958,19 @@ int ib_post_send_mad(struct ib_mad_agent *mad_agent,
 			ret = -ENOMEM;
 			goto error2;
 		}
+		memset(mad_send_wr, 0, sizeof *mad_send_wr);
 
 		mad_send_wr->send_wr = *send_wr;
 		mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
 		memcpy(mad_send_wr->sg_list, send_wr->sg_list,
 		       sizeof *send_wr->sg_list * send_wr->num_sge);
-		mad_send_wr->send_wr.next = NULL;
+		mad_send_wr->wr_id = send_wr->wr_id;
 		mad_send_wr->tid = send_wr->wr.ud.mad_hdr->tid;
-		mad_send_wr->agent = mad_agent;
+		mad_send_wr->mad_agent_priv = mad_agent_priv;
 		/* Timeout will be updated after send completes */
 		mad_send_wr->timeout = msecs_to_jiffies(send_wr->wr.
 							ud.timeout_ms);
-		mad_send_wr->retry = 0;
+		mad_send_wr->retries = mad_send_wr->send_wr.wr.ud.retries;
 		/* One reference for each work request to QP + response */
 		mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
 		mad_send_wr->status = IB_WC_SUCCESS;
@@ -883,8 +982,13 @@ int ib_post_send_mad(struct ib_mad_agent *mad_agent,
 			      &mad_agent_priv->send_list);
 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 
-		ret = ib_send_mad(mad_agent_priv, mad_send_wr);
-		if (ret) {
+		if (mad_agent_priv->agent.rmpp_version) {
+			ret = ib_send_rmpp_mad(mad_send_wr);
+			if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
+				ret = ib_send_mad(mad_send_wr);
+		} else
+			ret = ib_send_mad(mad_send_wr);
+		if (ret < 0) {
 			/* Fail send request */
 			spin_lock_irqsave(&mad_agent_priv->lock, flags);
 			list_del(&mad_send_wr->agent_list);
@@ -910,41 +1014,28 @@ EXPORT_SYMBOL(ib_post_send_mad);
  */
 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
 {
-	struct ib_mad_recv_buf *entry;
+	struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
 	struct ib_mad_private_header *mad_priv_hdr;
 	struct ib_mad_private *priv;
+	struct list_head free_list;
 
-	mad_priv_hdr = container_of(mad_recv_wc,
-				    struct ib_mad_private_header,
-				    recv_wc);
-	priv = container_of(mad_priv_hdr, struct ib_mad_private, header);
+	INIT_LIST_HEAD(&free_list);
+	list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
 
-	/*
-	 * Walk receive buffer list associated with this WC
-	 * No need to remove them from list of receive buffers
-	 */
-	list_for_each_entry(entry, &mad_recv_wc->recv_buf.list, list) {
-		/* Free previous receive buffer */
-		kmem_cache_free(ib_mad_cache, priv);
+	list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
+					&free_list, list) {
+		mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
+					   recv_buf);
 		mad_priv_hdr = container_of(mad_recv_wc,
 					    struct ib_mad_private_header,
 					    recv_wc);
 		priv = container_of(mad_priv_hdr, struct ib_mad_private,
 				    header);
+		kmem_cache_free(ib_mad_cache, priv);
 	}
-
-	/* Free last buffer */
-	kmem_cache_free(ib_mad_cache, priv);
 }
 EXPORT_SYMBOL(ib_free_recv_mad);
 
-void ib_coalesce_recv_mad(struct ib_mad_recv_wc *mad_recv_wc,
-			  void *buf)
-{
-	printk(KERN_ERR PFX "ib_coalesce_recv_mad() not implemented yet\n");
-}
-EXPORT_SYMBOL(ib_coalesce_recv_mad);
-
 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
 					u8 rmpp_version,
 					ib_mad_send_handler send_handler,
@@ -1338,42 +1429,15 @@ out:
 	return;
 }
 
-static int response_mad(struct ib_mad *mad)
-{
-	/* Trap represses are responses although response bit is reset */
-	return ((mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
-		(mad->mad_hdr.method & IB_MGMT_METHOD_RESP));
-}
-
-static int solicited_mad(struct ib_mad *mad)
-{
-	/* CM MADs are never solicited */
-	if (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CM) {
-		return 0;
-	}
-
-	/* XXX: Determine whether MAD is using RMPP */
-
-	/* Not using RMPP */
-	/* Is this MAD a response to a previous MAD ? */
-	return response_mad(mad);
-}
-
 static struct ib_mad_agent_private *
 find_mad_agent(struct ib_mad_port_private *port_priv,
-	       struct ib_mad *mad,
-	       int solicited)
+	       struct ib_mad *mad)
 {
 	struct ib_mad_agent_private *mad_agent = NULL;
 	unsigned long flags;
 
 	spin_lock_irqsave(&port_priv->reg_lock, flags);
-
-	/*
-	 * Whether MAD was solicited determines type of routing to
-	 * MAD client.
-	 */
-	if (solicited) {
+	if (response_mad(mad)) {
 		u32 hi_tid;
 		struct ib_mad_agent_private *entry;
 
@@ -1477,21 +1541,20 @@ out:
 	return valid;
 }
 
-/*
- * Return start of fully reassembled MAD, or NULL, if MAD isn't assembled yet
- */
-static struct ib_mad_private *
-reassemble_recv(struct ib_mad_agent_private *mad_agent_priv,
-		struct ib_mad_private *recv)
+static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
+		       struct ib_mad_hdr *mad_hdr)
 {
-	/* Until we have RMPP, all receives are reassembled!... */
-	INIT_LIST_HEAD(&recv->header.recv_wc.recv_buf.list);
-	return recv;
+	struct ib_rmpp_mad *rmpp_mad;
+
+	rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
+	return !mad_agent_priv->agent.rmpp_version ||
+		!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
+				    IB_MGMT_RMPP_FLAG_ACTIVE) ||
+		(rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
 }
 
-static struct ib_mad_send_wr_private*
-find_send_req(struct ib_mad_agent_private *mad_agent_priv,
-	      u64 tid)
+struct ib_mad_send_wr_private*
+ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv, u64 tid)
 {
 	struct ib_mad_send_wr_private *mad_send_wr;
 
@@ -1507,7 +1570,9 @@ find_send_req(struct ib_mad_agent_private *mad_agent_priv,
 	 */
 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
 			    agent_list) {
-		if (mad_send_wr->tid == tid && mad_send_wr->timeout) {
+		if (is_data_mad(mad_agent_priv,
+				mad_send_wr->send_wr.wr.ud.mad_hdr) &&
+		    mad_send_wr->tid == tid && mad_send_wr->timeout) {
 			/* Verify request has not been canceled */
 			return (mad_send_wr->status == IB_WC_SUCCESS) ?
 				mad_send_wr : NULL;
@@ -1516,43 +1581,55 @@ find_send_req(struct ib_mad_agent_private *mad_agent_priv,
 	return NULL;
 }
 
+void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
+{
+	mad_send_wr->timeout = 0;
+	if (mad_send_wr->refcount == 1) {
+		list_del(&mad_send_wr->agent_list);
+		list_add_tail(&mad_send_wr->agent_list,
+			      &mad_send_wr->mad_agent_priv->done_list);
+	}
+}
+
 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
-				 struct ib_mad_private *recv,
-				 int solicited)
+				 struct ib_mad_recv_wc *mad_recv_wc)
 {
 	struct ib_mad_send_wr_private *mad_send_wr;
 	struct ib_mad_send_wc mad_send_wc;
 	unsigned long flags;
-
-	/* Fully reassemble receive before processing */
-	recv = reassemble_recv(mad_agent_priv, recv);
-	if (!recv) {
-		if (atomic_dec_and_test(&mad_agent_priv->refcount))
-			wake_up(&mad_agent_priv->wait);
-		return;
+	u64 tid;
+
+	INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
+	list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
+	if (mad_agent_priv->agent.rmpp_version) {
+		mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
+						      mad_recv_wc);
+		if (!mad_recv_wc) {
+			if (atomic_dec_and_test(&mad_agent_priv->refcount))
+				wake_up(&mad_agent_priv->wait);
+			return;
+		}
 	}
 
 	/* Complete corresponding request */
-	if (solicited) {
+	if (response_mad(mad_recv_wc->recv_buf.mad)) {
+		tid = mad_recv_wc->recv_buf.mad->mad_hdr.tid;
 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
-		mad_send_wr = find_send_req(mad_agent_priv,
-					    recv->mad.mad.mad_hdr.tid);
+		mad_send_wr = ib_find_send_mad(mad_agent_priv, tid);
 		if (!mad_send_wr) {
 			spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
-			ib_free_recv_mad(&recv->header.recv_wc);
+			ib_free_recv_mad(mad_recv_wc);
 			if (atomic_dec_and_test(&mad_agent_priv->refcount))
 				wake_up(&mad_agent_priv->wait);
 			return;
 		}
-		/* Timeout = 0 means that we won't wait for a response */
-		mad_send_wr->timeout = 0;
+		ib_mark_mad_done(mad_send_wr);
 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 
 		/* Defined behavior is to complete response before request */
-		recv->header.recv_wc.wc->wr_id = mad_send_wr->wr_id;
-		mad_agent_priv->agent.recv_handler(
-						&mad_agent_priv->agent,
-						&recv->header.recv_wc);
+		mad_recv_wc->wc->wr_id = mad_send_wr->wr_id;
+		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
+						   mad_recv_wc);
 		atomic_dec(&mad_agent_priv->refcount);
 
 		mad_send_wc.status = IB_WC_SUCCESS;
@@ -1560,9 +1637,8 @@ static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
 		mad_send_wc.wr_id = mad_send_wr->wr_id;
 		ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
 	} else {
-		mad_agent_priv->agent.recv_handler(
-						&mad_agent_priv->agent,
-						&recv->header.recv_wc);
+		mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
+						   mad_recv_wc);
 		if (atomic_dec_and_test(&mad_agent_priv->refcount))
 			wake_up(&mad_agent_priv->wait);
 	}
@@ -1576,7 +1652,6 @@ static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
 	struct ib_mad_private *recv, *response;
 	struct ib_mad_list_head *mad_list;
 	struct ib_mad_agent_private *mad_agent;
-	int solicited;
 
 	response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
 	if (!response)
@@ -1662,11 +1737,9 @@ local:
 		}
 	}
 
-	/* Determine corresponding MAD agent for incoming receive MAD */
-	solicited = solicited_mad(&recv->mad.mad);
-	mad_agent = find_mad_agent(port_priv, &recv->mad.mad, solicited);
+	mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
 	if (mad_agent) {
-		ib_mad_complete_recv(mad_agent, recv, solicited);
+		ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
 		/*
 		 * recv is freed up in error cases in ib_mad_complete_recv
 		 * or via recv_handler in ib_mad_complete_recv()
@@ -1710,26 +1783,31 @@ static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
 	}
 }
 
-static void wait_for_response(struct ib_mad_agent_private *mad_agent_priv,
-			      struct ib_mad_send_wr_private *mad_send_wr )
+static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
 {
+	struct ib_mad_agent_private *mad_agent_priv;
 	struct ib_mad_send_wr_private *temp_mad_send_wr;
 	struct list_head *list_item;
 	unsigned long delay;
 
+	mad_agent_priv = mad_send_wr->mad_agent_priv;
 	list_del(&mad_send_wr->agent_list);
 
 	delay = mad_send_wr->timeout;
 	mad_send_wr->timeout += jiffies;
 
-	list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
-		temp_mad_send_wr = list_entry(list_item,
-					      struct ib_mad_send_wr_private,
-					      agent_list);
-		if (time_after(mad_send_wr->timeout,
-			       temp_mad_send_wr->timeout))
-			break;
+	if (delay) {
+		list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
+			temp_mad_send_wr = list_entry(list_item,
+						struct ib_mad_send_wr_private,
+						agent_list);
+			if (time_after(mad_send_wr->timeout,
+				       temp_mad_send_wr->timeout))
+				break;
+		}
 	}
+	else
+		list_item = &mad_agent_priv->wait_list;
 	list_add(&mad_send_wr->agent_list, list_item);
 
 	/* Reschedule a work item if we have a shorter timeout */
@@ -1740,19 +1818,32 @@ static void wait_for_response(struct ib_mad_agent_private *mad_agent_priv,
 	}
 }
 
+void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
+			  int timeout_ms)
+{
+	mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
+	wait_for_response(mad_send_wr);
+}
+
 /*
  * Process a send work completion
  */
-static void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
-				    struct ib_mad_send_wc *mad_send_wc)
+void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
+			     struct ib_mad_send_wc *mad_send_wc)
 {
 	struct ib_mad_agent_private	*mad_agent_priv;
 	unsigned long			flags;
+	int				ret;
 
-	mad_agent_priv = container_of(mad_send_wr->agent,
-				      struct ib_mad_agent_private, agent);
-
+	mad_agent_priv = mad_send_wr->mad_agent_priv;
 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
+	if (mad_agent_priv->agent.rmpp_version) {
+		ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
+		if (ret == IB_RMPP_RESULT_CONSUMED)
+			goto done;
+	} else
+		ret = IB_RMPP_RESULT_UNHANDLED;
+
 	if (mad_send_wc->status != IB_WC_SUCCESS &&
 	    mad_send_wr->status == IB_WC_SUCCESS) {
 		mad_send_wr->status = mad_send_wc->status;
@@ -1762,10 +1853,9 @@ static void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
 	if (--mad_send_wr->refcount > 0) {
 		if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
 		    mad_send_wr->status == IB_WC_SUCCESS) {
-			wait_for_response(mad_agent_priv, mad_send_wr);
+			wait_for_response(mad_send_wr);
 		}
-		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
-		return;
+		goto done;
 	}
 
 	/* Remove send from MAD agent and notify client of completion */
@@ -1775,14 +1865,18 @@ static void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
 
 	if (mad_send_wr->status != IB_WC_SUCCESS )
 		mad_send_wc->status = mad_send_wr->status;
-	mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
-					    mad_send_wc);
+	if (ret != IB_RMPP_RESULT_INTERNAL)
+		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
+						   mad_send_wc);
 
 	/* Release reference on agent taken when sending */
 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
 		wake_up(&mad_agent_priv->wait);
 
 	kfree(mad_send_wr);
+	return;
+done:
+	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 }
 
 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
@@ -1961,6 +2055,8 @@ static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
 
 	/* Empty wait list to prevent receives from finding a request */
 	list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
+	/* Empty local completion list as well */
+	list_splice_init(&mad_agent_priv->local_list, &cancel_list);
 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 
 	/* Report all cancelled requests */
@@ -1980,8 +2076,7 @@ static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
 }
 
 static struct ib_mad_send_wr_private*
-find_send_by_wr_id(struct ib_mad_agent_private *mad_agent_priv,
-		   u64 wr_id)
+find_send_by_wr_id(struct ib_mad_agent_private *mad_agent_priv, u64 wr_id)
 {
 	struct ib_mad_send_wr_private *mad_send_wr;
 
@@ -1993,79 +2088,50 @@ find_send_by_wr_id(struct ib_mad_agent_private *mad_agent_priv,
 
 	list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
 			    agent_list) {
-		if (mad_send_wr->wr_id == wr_id)
+		if (is_data_mad(mad_agent_priv,
+				mad_send_wr->send_wr.wr.ud.mad_hdr) &&
+		    mad_send_wr->wr_id == wr_id)
 			return mad_send_wr;
 	}
 	return NULL;
 }
 
-void cancel_sends(void *data)
-{
-	struct ib_mad_agent_private *mad_agent_priv;
-	struct ib_mad_send_wr_private *mad_send_wr;
-	struct ib_mad_send_wc mad_send_wc;
-	unsigned long flags;
-
-	mad_agent_priv = data;
-
-	mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
-	mad_send_wc.vendor_err = 0;
-
-	spin_lock_irqsave(&mad_agent_priv->lock, flags);
-	while (!list_empty(&mad_agent_priv->canceled_list)) {
-		mad_send_wr = list_entry(mad_agent_priv->canceled_list.next,
-					 struct ib_mad_send_wr_private,
-					 agent_list);
-
-		list_del(&mad_send_wr->agent_list);
-		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
-
-		mad_send_wc.wr_id = mad_send_wr->wr_id;
-		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
-						   &mad_send_wc);
-
-		kfree(mad_send_wr);
-		if (atomic_dec_and_test(&mad_agent_priv->refcount))
-			wake_up(&mad_agent_priv->wait);
-		spin_lock_irqsave(&mad_agent_priv->lock, flags);
-	}
-	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
-}
-
-void ib_cancel_mad(struct ib_mad_agent *mad_agent,
-		  u64 wr_id)
+int ib_modify_mad(struct ib_mad_agent *mad_agent, u64 wr_id, u32 timeout_ms)
 {
 	struct ib_mad_agent_private *mad_agent_priv;
 	struct ib_mad_send_wr_private *mad_send_wr;
 	unsigned long flags;
+	int active;
 
 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
 				      agent);
 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
 	mad_send_wr = find_send_by_wr_id(mad_agent_priv, wr_id);
-	if (!mad_send_wr) {
+	if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
-		goto out;
+		return -EINVAL;
 	}
 
-	if (mad_send_wr->status == IB_WC_SUCCESS)
-		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
-
-	if (mad_send_wr->refcount != 0) {
+	active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
+	if (!timeout_ms) {
 		mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
-		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
-		goto out;
+		mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
 	}
 
-	list_del(&mad_send_wr->agent_list);
-	list_add_tail(&mad_send_wr->agent_list, &mad_agent_priv->canceled_list);
-	adjust_timeout(mad_agent_priv);
+	mad_send_wr->send_wr.wr.ud.timeout_ms = timeout_ms;
+	if (active)
+		mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
+	else
+		ib_reset_mad_timeout(mad_send_wr, timeout_ms);
+
 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
+	return 0;
+}
+EXPORT_SYMBOL(ib_modify_mad);
 
-	queue_work(mad_agent_priv->qp_info->port_priv->wq,
-		   &mad_agent_priv->canceled_work);
-out:
-	return;
+void ib_cancel_mad(struct ib_mad_agent *mad_agent, u64 wr_id)
+{
+	ib_modify_mad(mad_agent, wr_id, 0);
 }
 EXPORT_SYMBOL(ib_cancel_mad);
 
@@ -2075,6 +2141,7 @@ static void local_completions(void *data)
 	struct ib_mad_local_private *local;
 	struct ib_mad_agent_private *recv_mad_agent;
 	unsigned long flags;
+	int recv = 0;
 	struct ib_wc wc;
 	struct ib_mad_send_wc mad_send_wc;
 
@@ -2090,10 +2157,10 @@ static void local_completions(void *data)
 			recv_mad_agent = local->recv_mad_agent;
 			if (!recv_mad_agent) {
 				printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
-				kmem_cache_free(ib_mad_cache, local->mad_priv);
 				goto local_send_completion;
 			}
 
+			recv = 1;
 			/*
 			 * Defined behavior is to complete response
 			 * before request
@@ -2105,7 +2172,9 @@ static void local_completions(void *data)
 			local->mad_priv->header.recv_wc.wc = &wc;
 			local->mad_priv->header.recv_wc.mad_len =
 						sizeof(struct ib_mad);
-			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.recv_buf.list);
+			INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
+			list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
+				 &local->mad_priv->header.recv_wc.rmpp_list);
 			local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
 			local->mad_priv->header.recv_wc.recv_buf.mad =
 						&local->mad_priv->mad.mad;
@@ -2136,11 +2205,47 @@ local_send_completion:
 		spin_lock_irqsave(&mad_agent_priv->lock, flags);
 		list_del(&local->completion_list);
 		atomic_dec(&mad_agent_priv->refcount);
+		if (!recv)
+			kmem_cache_free(ib_mad_cache, local->mad_priv);
 		kfree(local);
 	}
 	spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 }
 
+static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
+{
+	int ret;
+
+	if (!mad_send_wr->retries--)
+		return -ETIMEDOUT;
+
+	mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_wr.
+						wr.ud.timeout_ms);
+
+	if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
+		ret = ib_retry_rmpp(mad_send_wr);
+		switch (ret) {
+		case IB_RMPP_RESULT_UNHANDLED:
+			ret = ib_send_mad(mad_send_wr);
+			break;
+		case IB_RMPP_RESULT_CONSUMED:
+			ret = 0;
+			break;
+		default:
+			ret = -ECOMM;
+			break;
+		}
+	} else
+		ret = ib_send_mad(mad_send_wr);
+
+	if (!ret) {
+		mad_send_wr->refcount++;
+		list_add_tail(&mad_send_wr->agent_list,
+			      &mad_send_wr->mad_agent_priv->send_list);
+	}
+	return ret;
+}
+
 static void timeout_sends(void *data)
 {
 	struct ib_mad_agent_private *mad_agent_priv;
@@ -2149,8 +2254,6 @@ static void timeout_sends(void *data)
 	unsigned long flags, delay;
 
 	mad_agent_priv = (struct ib_mad_agent_private *)data;
-
-	mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
 	mad_send_wc.vendor_err = 0;
 
 	spin_lock_irqsave(&mad_agent_priv->lock, flags);
@@ -2170,8 +2273,16 @@ static void timeout_sends(void *data)
 		}
 
 		list_del(&mad_send_wr->agent_list);
+		if (mad_send_wr->status == IB_WC_SUCCESS &&
+		    !retry_send(mad_send_wr))
+			continue;
+
 		spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 
+		if (mad_send_wr->status == IB_WC_SUCCESS)
+			mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
+		else
+			mad_send_wc.status = mad_send_wr->status;
 		mad_send_wc.wr_id = mad_send_wr->wr_id;
 		mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
 						   &mad_send_wc);
@@ -2447,14 +2558,6 @@ static int ib_mad_port_open(struct ib_device *device,
 	unsigned long flags;
 	char name[sizeof "ib_mad123"];
 
-	/* First, check if port already open at MAD layer */
-	port_priv = ib_get_mad_port(device, port_num);
-	if (port_priv) {
-		printk(KERN_DEBUG PFX "%s port %d already open\n",
-		       device->name, port_num);
-		return 0;
-	}
-
 	/* Create new device info */
 	port_priv = kmalloc(sizeof *port_priv, GFP_KERNEL);
 	if (!port_priv) {
@@ -2579,7 +2682,7 @@ static int ib_mad_port_close(struct ib_device *device, int port_num)
 
 static void ib_mad_init_device(struct ib_device *device)
 {
-	int ret, num_ports, cur_port, i, ret2;
+	int num_ports, cur_port, i;
 
 	if (device->node_type == IB_NODE_SWITCH) {
 		num_ports = 1;
@@ -2589,47 +2692,37 @@ static void ib_mad_init_device(struct ib_device *device)
 		cur_port = 1;
 	}
 	for (i = 0; i < num_ports; i++, cur_port++) {
-		ret = ib_mad_port_open(device, cur_port);
-		if (ret) {
+		if (ib_mad_port_open(device, cur_port)) {
 			printk(KERN_ERR PFX "Couldn't open %s port %d\n",
 			       device->name, cur_port);
 			goto error_device_open;
 		}
-		ret = ib_agent_port_open(device, cur_port);
-		if (ret) {
+		if (ib_agent_port_open(device, cur_port)) {
 			printk(KERN_ERR PFX "Couldn't open %s port %d "
 			       "for agents\n",
 			       device->name, cur_port);
 			goto error_device_open;
 		}
 	}
-
-	goto error_device_query;
+	return;
 
 error_device_open:
 	while (i > 0) {
 		cur_port--;
-		ret2 = ib_agent_port_close(device, cur_port);
-		if (ret2) {
+		if (ib_agent_port_close(device, cur_port))
 			printk(KERN_ERR PFX "Couldn't close %s port %d "
 			       "for agents\n",
 			       device->name, cur_port);
-		}
-		ret2 = ib_mad_port_close(device, cur_port);
-		if (ret2) {
+		if (ib_mad_port_close(device, cur_port))
 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
 			       device->name, cur_port);
-		}
 		i--;
 	}
-
-error_device_query:
-	return;
 }
 
 static void ib_mad_remove_device(struct ib_device *device)
 {
-	int ret = 0, i, num_ports, cur_port, ret2;
+	int i, num_ports, cur_port;
 
 	if (device->node_type == IB_NODE_SWITCH) {
 		num_ports = 1;
@@ -2639,21 +2732,13 @@ static void ib_mad_remove_device(struct ib_device *device)
 		cur_port = 1;
 	}
 	for (i = 0; i < num_ports; i++, cur_port++) {
-		ret2 = ib_agent_port_close(device, cur_port);
-		if (ret2) {
+		if (ib_agent_port_close(device, cur_port))
 			printk(KERN_ERR PFX "Couldn't close %s port %d "
 			       "for agents\n",
 			       device->name, cur_port);
-			if (!ret)
-				ret = ret2;
-		}
-		ret2 = ib_mad_port_close(device, cur_port);
-		if (ret2) {
+		if (ib_mad_port_close(device, cur_port))
 			printk(KERN_ERR PFX "Couldn't close %s port %d\n",
 			       device->name, cur_port);
-			if (!ret)
-				ret = ret2;
-		}
 	}
 }
 
@@ -2709,3 +2794,4 @@ static void __exit ib_mad_cleanup_module(void)
 
 module_init(ib_mad_init_module);
 module_exit(ib_mad_cleanup_module);
+
diff --git a/drivers/infiniband/core/mad_priv.h b/drivers/infiniband/core/mad_priv.h
index 008cbcb94b15..568da10b05ab 100644
--- a/drivers/infiniband/core/mad_priv.h
+++ b/drivers/infiniband/core/mad_priv.h
@@ -1,5 +1,7 @@
 /*
  * Copyright (c) 2004, 2005, Voltaire, Inc. All rights reserved.
+ * Copyright (c) 2005 Intel Corporation. All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. 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
@@ -29,7 +31,7 @@
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  *
- * $Id: mad_priv.h 1389 2004-12-27 22:56:47Z roland $
+ * $Id: mad_priv.h 2730 2005-06-28 16:43:03Z sean.hefty $
  */
 
 #ifndef __IB_MAD_PRIV_H__
@@ -92,16 +94,15 @@ struct ib_mad_agent_private {
 	spinlock_t lock;
 	struct list_head send_list;
 	struct list_head wait_list;
+	struct list_head done_list;
 	struct work_struct timed_work;
 	unsigned long timeout;
 	struct list_head local_list;
 	struct work_struct local_work;
-	struct list_head canceled_list;
-	struct work_struct canceled_work;
+	struct list_head rmpp_list;
 
 	atomic_t refcount;
 	wait_queue_head_t wait;
-	u8 rmpp_version;
 };
 
 struct ib_mad_snoop_private {
@@ -116,15 +117,24 @@ struct ib_mad_snoop_private {
 struct ib_mad_send_wr_private {
 	struct ib_mad_list_head mad_list;
 	struct list_head agent_list;
-	struct ib_mad_agent *agent;
+	struct ib_mad_agent_private *mad_agent_priv;
 	struct ib_send_wr send_wr;
 	struct ib_sge sg_list[IB_MAD_SEND_REQ_MAX_SG];
 	u64 wr_id;			/* client WR ID */
 	u64 tid;
 	unsigned long timeout;
+	int retries;
 	int retry;
 	int refcount;
 	enum ib_wc_status status;
+
+	/* RMPP control */
+	int last_ack;
+	int seg_num;
+	int newwin;
+	int total_seg;
+	int data_offset;
+	int pad;
 };
 
 struct ib_mad_local_private {
@@ -197,4 +207,17 @@ struct ib_mad_port_private {
 
 extern kmem_cache_t *ib_mad_cache;
 
+int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr);
+
+struct ib_mad_send_wr_private *
+ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv, u64 tid);
+
+void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
+			     struct ib_mad_send_wc *mad_send_wc);
+
+void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr);
+
+void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
+			  int timeout_ms);
+
 #endif	/* __IB_MAD_PRIV_H__ */
diff --git a/drivers/infiniband/core/mad_rmpp.c b/drivers/infiniband/core/mad_rmpp.c
new file mode 100644
index 000000000000..8f1eb80e421f
--- /dev/null
+++ b/drivers/infiniband/core/mad_rmpp.c
@@ -0,0 +1,765 @@
+/*
+ * Copyright (c) 2005 Intel Inc. All rights reserved.
+ * Copyright (c) 2005 Voltaire, Inc. 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.
+ *
+ * $Id: mad_rmpp.c 1921 2005-03-02 22:58:44Z sean.hefty $
+ */
+
+#include <linux/dma-mapping.h>
+
+#include "mad_priv.h"
+#include "mad_rmpp.h"
+
+enum rmpp_state {
+	RMPP_STATE_ACTIVE,
+	RMPP_STATE_TIMEOUT,
+	RMPP_STATE_COMPLETE
+};
+
+struct mad_rmpp_recv {
+	struct ib_mad_agent_private *agent;
+	struct list_head list;
+	struct work_struct timeout_work;
+	struct work_struct cleanup_work;
+	wait_queue_head_t wait;
+	enum rmpp_state state;
+	spinlock_t lock;
+	atomic_t refcount;
+
+	struct ib_ah *ah;
+	struct ib_mad_recv_wc *rmpp_wc;
+	struct ib_mad_recv_buf *cur_seg_buf;
+	int last_ack;
+	int seg_num;
+	int newwin;
+
+	u64 tid;
+	u32 src_qp;
+	u16 slid;
+	u8 mgmt_class;
+	u8 class_version;
+	u8 method;
+};
+
+static void destroy_rmpp_recv(struct mad_rmpp_recv *rmpp_recv)
+{
+	atomic_dec(&rmpp_recv->refcount);
+	wait_event(rmpp_recv->wait, !atomic_read(&rmpp_recv->refcount));
+	ib_destroy_ah(rmpp_recv->ah);
+	kfree(rmpp_recv);
+}
+
+void ib_cancel_rmpp_recvs(struct ib_mad_agent_private *agent)
+{
+	struct mad_rmpp_recv *rmpp_recv, *temp_rmpp_recv;
+	unsigned long flags;
+
+	spin_lock_irqsave(&agent->lock, flags);
+	list_for_each_entry(rmpp_recv, &agent->rmpp_list, list) {
+		cancel_delayed_work(&rmpp_recv->timeout_work);
+		cancel_delayed_work(&rmpp_recv->cleanup_work);
+	}
+	spin_unlock_irqrestore(&agent->lock, flags);
+
+	flush_workqueue(agent->qp_info->port_priv->wq);
+
+	list_for_each_entry_safe(rmpp_recv, temp_rmpp_recv,
+				 &agent->rmpp_list, list) {
+		list_del(&rmpp_recv->list);
+		if (rmpp_recv->state != RMPP_STATE_COMPLETE)
+			ib_free_recv_mad(rmpp_recv->rmpp_wc);
+		destroy_rmpp_recv(rmpp_recv);
+	}
+}
+
+static void recv_timeout_handler(void *data)
+{
+	struct mad_rmpp_recv *rmpp_recv = data;
+	struct ib_mad_recv_wc *rmpp_wc;
+	unsigned long flags;
+
+	spin_lock_irqsave(&rmpp_recv->agent->lock, flags);
+	if (rmpp_recv->state != RMPP_STATE_ACTIVE) {
+		spin_unlock_irqrestore(&rmpp_recv->agent->lock, flags);
+		return;
+	}
+	rmpp_recv->state = RMPP_STATE_TIMEOUT;
+	list_del(&rmpp_recv->list);
+	spin_unlock_irqrestore(&rmpp_recv->agent->lock, flags);
+
+	/* TODO: send abort. */
+	rmpp_wc = rmpp_recv->rmpp_wc;
+	destroy_rmpp_recv(rmpp_recv);
+	ib_free_recv_mad(rmpp_wc);
+}
+
+static void recv_cleanup_handler(void *data)
+{
+	struct mad_rmpp_recv *rmpp_recv = data;
+	unsigned long flags;
+
+	spin_lock_irqsave(&rmpp_recv->agent->lock, flags);
+	list_del(&rmpp_recv->list);
+	spin_unlock_irqrestore(&rmpp_recv->agent->lock, flags);
+	destroy_rmpp_recv(rmpp_recv);
+}
+
+static struct mad_rmpp_recv *
+create_rmpp_recv(struct ib_mad_agent_private *agent,
+		 struct ib_mad_recv_wc *mad_recv_wc)
+{
+	struct mad_rmpp_recv *rmpp_recv;
+	struct ib_mad_hdr *mad_hdr;
+
+	rmpp_recv = kmalloc(sizeof *rmpp_recv, GFP_KERNEL);
+	if (!rmpp_recv)
+		return NULL;
+
+	rmpp_recv->ah = ib_create_ah_from_wc(agent->agent.qp->pd,
+					     mad_recv_wc->wc,
+					     mad_recv_wc->recv_buf.grh,
+					     agent->agent.port_num);
+	if (IS_ERR(rmpp_recv->ah))
+		goto error;
+
+	rmpp_recv->agent = agent;
+	init_waitqueue_head(&rmpp_recv->wait);
+	INIT_WORK(&rmpp_recv->timeout_work, recv_timeout_handler, rmpp_recv);
+	INIT_WORK(&rmpp_recv->cleanup_work, recv_cleanup_handler, rmpp_recv);
+	spin_lock_init(&rmpp_recv->lock);
+	rmpp_recv->state = RMPP_STATE_ACTIVE;
+	atomic_set(&rmpp_recv->refcount, 1);
+
+	rmpp_recv->rmpp_wc = mad_recv_wc;
+	rmpp_recv->cur_seg_buf = &mad_recv_wc->recv_buf;
+	rmpp_recv->newwin = 1;
+	rmpp_recv->seg_num = 1;
+	rmpp_recv->last_ack = 0;
+
+	mad_hdr = &mad_recv_wc->recv_buf.mad->mad_hdr;
+	rmpp_recv->tid = mad_hdr->tid;
+	rmpp_recv->src_qp = mad_recv_wc->wc->src_qp;
+	rmpp_recv->slid = mad_recv_wc->wc->slid;
+	rmpp_recv->mgmt_class = mad_hdr->mgmt_class;
+	rmpp_recv->class_version = mad_hdr->class_version;
+	rmpp_recv->method  = mad_hdr->method;
+	return rmpp_recv;
+
+error:	kfree(rmpp_recv);
+	return NULL;
+}
+
+static inline void deref_rmpp_recv(struct mad_rmpp_recv *rmpp_recv)
+{
+	if (atomic_dec_and_test(&rmpp_recv->refcount))
+		wake_up(&rmpp_recv->wait);
+}
+
+static struct mad_rmpp_recv *
+find_rmpp_recv(struct ib_mad_agent_private *agent,
+	       struct ib_mad_recv_wc *mad_recv_wc)
+{
+	struct mad_rmpp_recv *rmpp_recv;
+	struct ib_mad_hdr *mad_hdr = &mad_recv_wc->recv_buf.mad->mad_hdr;
+
+	list_for_each_entry(rmpp_recv, &agent->rmpp_list, list) {
+		if (rmpp_recv->tid == mad_hdr->tid &&
+		    rmpp_recv->src_qp == mad_recv_wc->wc->src_qp &&
+		    rmpp_recv->slid == mad_recv_wc->wc->slid &&
+		    rmpp_recv->mgmt_class == mad_hdr->mgmt_class &&
+		    rmpp_recv->class_version == mad_hdr->class_version &&
+		    rmpp_recv->method == mad_hdr->method)
+			return rmpp_recv;
+	}
+	return NULL;
+}
+
+static struct mad_rmpp_recv *
+acquire_rmpp_recv(struct ib_mad_agent_private *agent,
+		  struct ib_mad_recv_wc *mad_recv_wc)
+{
+	struct mad_rmpp_recv *rmpp_recv;
+	unsigned long flags;
+
+	spin_lock_irqsave(&agent->lock, flags);
+	rmpp_recv = find_rmpp_recv(agent, mad_recv_wc);
+	if (rmpp_recv)
+		atomic_inc(&rmpp_recv->refcount);
+	spin_unlock_irqrestore(&agent->lock, flags);
+	return rmpp_recv;
+}
+
+static struct mad_rmpp_recv *
+insert_rmpp_recv(struct ib_mad_agent_private *agent,
+		 struct mad_rmpp_recv *rmpp_recv)
+{
+	struct mad_rmpp_recv *cur_rmpp_recv;
+
+	cur_rmpp_recv = find_rmpp_recv(agent, rmpp_recv->rmpp_wc);
+	if (!cur_rmpp_recv)
+		list_add_tail(&rmpp_recv->list, &agent->rmpp_list);
+
+	return cur_rmpp_recv;
+}
+
+static int data_offset(u8 mgmt_class)
+{
+	if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
+		return offsetof(struct ib_sa_mad, data);
+	else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
+		 (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
+		return offsetof(struct ib_vendor_mad, data);
+	else
+		return offsetof(struct ib_rmpp_mad, data);
+}
+
+static void format_ack(struct ib_rmpp_mad *ack,
+		       struct ib_rmpp_mad *data,
+		       struct mad_rmpp_recv *rmpp_recv)
+{
+	unsigned long flags;
+
+	memcpy(&ack->mad_hdr, &data->mad_hdr,
+	       data_offset(data->mad_hdr.mgmt_class));
+
+	ack->mad_hdr.method ^= IB_MGMT_METHOD_RESP;
+	ack->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_ACK;
+	ib_set_rmpp_flags(&ack->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
+
+	spin_lock_irqsave(&rmpp_recv->lock, flags);
+	rmpp_recv->last_ack = rmpp_recv->seg_num;
+	ack->rmpp_hdr.seg_num = cpu_to_be32(rmpp_recv->seg_num);
+	ack->rmpp_hdr.paylen_newwin = cpu_to_be32(rmpp_recv->newwin);
+	spin_unlock_irqrestore(&rmpp_recv->lock, flags);
+}
+
+static void ack_recv(struct mad_rmpp_recv *rmpp_recv,
+		     struct ib_mad_recv_wc *recv_wc)
+{
+	struct ib_mad_send_buf *msg;
+	struct ib_send_wr *bad_send_wr;
+	int hdr_len, ret;
+
+	hdr_len = sizeof(struct ib_mad_hdr) + sizeof(struct ib_rmpp_hdr);
+	msg = ib_create_send_mad(&rmpp_recv->agent->agent, recv_wc->wc->src_qp,
+				 recv_wc->wc->pkey_index, rmpp_recv->ah, 1,
+				 hdr_len, sizeof(struct ib_rmpp_mad) - hdr_len,
+				 GFP_KERNEL);
+	if (!msg)
+		return;
+
+	format_ack((struct ib_rmpp_mad *) msg->mad,
+		   (struct ib_rmpp_mad *) recv_wc->recv_buf.mad, rmpp_recv);
+	ret = ib_post_send_mad(&rmpp_recv->agent->agent, &msg->send_wr,
+			       &bad_send_wr);
+	if (ret)
+		ib_free_send_mad(msg);
+}
+
+static inline int get_last_flag(struct ib_mad_recv_buf *seg)
+{
+	struct ib_rmpp_mad *rmpp_mad;
+
+	rmpp_mad = (struct ib_rmpp_mad *) seg->mad;
+	return ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) & IB_MGMT_RMPP_FLAG_LAST;
+}
+
+static inline int get_seg_num(struct ib_mad_recv_buf *seg)
+{
+	struct ib_rmpp_mad *rmpp_mad;
+
+	rmpp_mad = (struct ib_rmpp_mad *) seg->mad;
+	return be32_to_cpu(rmpp_mad->rmpp_hdr.seg_num);
+}
+
+static inline struct ib_mad_recv_buf * get_next_seg(struct list_head *rmpp_list,
+						    struct ib_mad_recv_buf *seg)
+{
+	if (seg->list.next == rmpp_list)
+		return NULL;
+
+	return container_of(seg->list.next, struct ib_mad_recv_buf, list);
+}
+
+static inline int window_size(struct ib_mad_agent_private *agent)
+{
+	return max(agent->qp_info->recv_queue.max_active >> 3, 1);
+}
+
+static struct ib_mad_recv_buf * find_seg_location(struct list_head *rmpp_list,
+						  int seg_num)
+{
+        struct ib_mad_recv_buf *seg_buf;
+	int cur_seg_num;
+
+	list_for_each_entry_reverse(seg_buf, rmpp_list, list) {
+		cur_seg_num = get_seg_num(seg_buf);
+		if (seg_num > cur_seg_num)
+			return seg_buf;
+		if (seg_num == cur_seg_num)
+			break;
+	}
+	return NULL;
+}
+
+static void update_seg_num(struct mad_rmpp_recv *rmpp_recv,
+			   struct ib_mad_recv_buf *new_buf)
+{
+	struct list_head *rmpp_list = &rmpp_recv->rmpp_wc->rmpp_list;
+
+	while (new_buf && (get_seg_num(new_buf) == rmpp_recv->seg_num + 1)) {
+		rmpp_recv->cur_seg_buf = new_buf;
+		rmpp_recv->seg_num++;
+		new_buf = get_next_seg(rmpp_list, new_buf);
+	}
+}
+
+static inline int get_mad_len(struct mad_rmpp_recv *rmpp_recv)
+{
+	struct ib_rmpp_mad *rmpp_mad;
+	int hdr_size, data_size, pad;
+
+	rmpp_mad = (struct ib_rmpp_mad *)rmpp_recv->cur_seg_buf->mad;
+
+	hdr_size = data_offset(rmpp_mad->mad_hdr.mgmt_class);
+	data_size = sizeof(struct ib_rmpp_mad) - hdr_size;
+	pad = data_size - be32_to_cpu(rmpp_mad->rmpp_hdr.paylen_newwin);
+	if (pad > data_size || pad < 0)
+		pad = 0;
+
+	return hdr_size + rmpp_recv->seg_num * data_size - pad;
+}
+
+static struct ib_mad_recv_wc * complete_rmpp(struct mad_rmpp_recv *rmpp_recv)
+{
+	struct ib_mad_recv_wc *rmpp_wc;
+
+	ack_recv(rmpp_recv, rmpp_recv->rmpp_wc);
+	if (rmpp_recv->seg_num > 1)
+		cancel_delayed_work(&rmpp_recv->timeout_work);
+
+	rmpp_wc = rmpp_recv->rmpp_wc;
+	rmpp_wc->mad_len = get_mad_len(rmpp_recv);
+	/* 10 seconds until we can find the packet lifetime */
+	queue_delayed_work(rmpp_recv->agent->qp_info->port_priv->wq,
+			   &rmpp_recv->cleanup_work, msecs_to_jiffies(10000));
+	return rmpp_wc;
+}
+
+void ib_coalesce_recv_mad(struct ib_mad_recv_wc *mad_recv_wc, void *buf)
+{
+	struct ib_mad_recv_buf *seg_buf;
+	struct ib_rmpp_mad *rmpp_mad;
+	void *data;
+	int size, len, offset;
+	u8 flags;
+
+	len = mad_recv_wc->mad_len;
+	if (len <= sizeof(struct ib_mad)) {
+		memcpy(buf, mad_recv_wc->recv_buf.mad, len);
+		return;
+	}
+
+	offset = data_offset(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class);
+
+	list_for_each_entry(seg_buf, &mad_recv_wc->rmpp_list, list) {
+		rmpp_mad = (struct ib_rmpp_mad *)seg_buf->mad;
+		flags = ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr);
+
+		if (flags & IB_MGMT_RMPP_FLAG_FIRST) {
+			data = rmpp_mad;
+			size = sizeof(*rmpp_mad);
+		} else {
+			data = (void *) rmpp_mad + offset;
+			if (flags & IB_MGMT_RMPP_FLAG_LAST)
+				size = len;
+			else
+				size = sizeof(*rmpp_mad) - offset;
+		}
+
+		memcpy(buf, data, size);
+		len -= size;
+		buf += size;
+	}
+}
+EXPORT_SYMBOL(ib_coalesce_recv_mad);
+
+static struct ib_mad_recv_wc *
+continue_rmpp(struct ib_mad_agent_private *agent,
+	      struct ib_mad_recv_wc *mad_recv_wc)
+{
+	struct mad_rmpp_recv *rmpp_recv;
+	struct ib_mad_recv_buf *prev_buf;
+	struct ib_mad_recv_wc *done_wc;
+	int seg_num;
+	unsigned long flags;
+
+	rmpp_recv = acquire_rmpp_recv(agent, mad_recv_wc);
+	if (!rmpp_recv)
+		goto drop1;
+
+	seg_num = get_seg_num(&mad_recv_wc->recv_buf);
+
+	spin_lock_irqsave(&rmpp_recv->lock, flags);
+	if ((rmpp_recv->state == RMPP_STATE_TIMEOUT) ||
+	    (seg_num > rmpp_recv->newwin))
+		goto drop3;
+
+	if ((seg_num <= rmpp_recv->last_ack) ||
+	    (rmpp_recv->state == RMPP_STATE_COMPLETE)) {
+		spin_unlock_irqrestore(&rmpp_recv->lock, flags);
+		ack_recv(rmpp_recv, mad_recv_wc);
+		goto drop2;
+	}
+
+	prev_buf = find_seg_location(&rmpp_recv->rmpp_wc->rmpp_list, seg_num);
+	if (!prev_buf)
+		goto drop3;
+
+	done_wc = NULL;
+	list_add(&mad_recv_wc->recv_buf.list, &prev_buf->list);
+	if (rmpp_recv->cur_seg_buf == prev_buf) {
+		update_seg_num(rmpp_recv, &mad_recv_wc->recv_buf);
+		if (get_last_flag(rmpp_recv->cur_seg_buf)) {
+			rmpp_recv->state = RMPP_STATE_COMPLETE;
+			spin_unlock_irqrestore(&rmpp_recv->lock, flags);
+			done_wc = complete_rmpp(rmpp_recv);
+			goto out;
+		} else if (rmpp_recv->seg_num == rmpp_recv->newwin) {
+			rmpp_recv->newwin += window_size(agent);
+			spin_unlock_irqrestore(&rmpp_recv->lock, flags);
+			ack_recv(rmpp_recv, mad_recv_wc);
+			goto out;
+		}
+	}
+	spin_unlock_irqrestore(&rmpp_recv->lock, flags);
+out:
+	deref_rmpp_recv(rmpp_recv);
+	return done_wc;
+
+drop3:	spin_unlock_irqrestore(&rmpp_recv->lock, flags);
+drop2:	deref_rmpp_recv(rmpp_recv);
+drop1:	ib_free_recv_mad(mad_recv_wc);
+	return NULL;
+}
+
+static struct ib_mad_recv_wc *
+start_rmpp(struct ib_mad_agent_private *agent,
+	   struct ib_mad_recv_wc *mad_recv_wc)
+{
+	struct mad_rmpp_recv *rmpp_recv;
+	unsigned long flags;
+
+	rmpp_recv = create_rmpp_recv(agent, mad_recv_wc);
+	if (!rmpp_recv) {
+		ib_free_recv_mad(mad_recv_wc);
+		return NULL;
+	}
+
+	spin_lock_irqsave(&agent->lock, flags);
+	if (insert_rmpp_recv(agent, rmpp_recv)) {
+		spin_unlock_irqrestore(&agent->lock, flags);
+		/* duplicate first MAD */
+		destroy_rmpp_recv(rmpp_recv);
+		return continue_rmpp(agent, mad_recv_wc);
+	}
+	atomic_inc(&rmpp_recv->refcount);
+
+	if (get_last_flag(&mad_recv_wc->recv_buf)) {
+		rmpp_recv->state = RMPP_STATE_COMPLETE;
+		spin_unlock_irqrestore(&agent->lock, flags);
+		complete_rmpp(rmpp_recv);
+	} else {
+		spin_unlock_irqrestore(&agent->lock, flags);
+		/* 40 seconds until we can find the packet lifetimes */
+		queue_delayed_work(agent->qp_info->port_priv->wq,
+				   &rmpp_recv->timeout_work,
+				   msecs_to_jiffies(40000));
+		rmpp_recv->newwin += window_size(agent);
+		ack_recv(rmpp_recv, mad_recv_wc);
+		mad_recv_wc = NULL;
+	}
+	deref_rmpp_recv(rmpp_recv);
+	return mad_recv_wc;
+}
+
+static inline u64 get_seg_addr(struct ib_mad_send_wr_private *mad_send_wr)
+{
+	return mad_send_wr->sg_list[0].addr + mad_send_wr->data_offset +
+	       (sizeof(struct ib_rmpp_mad) - mad_send_wr->data_offset) *
+	       (mad_send_wr->seg_num - 1);
+}
+
+static int send_next_seg(struct ib_mad_send_wr_private *mad_send_wr)
+{
+	struct ib_rmpp_mad *rmpp_mad;
+	int timeout;
+
+	rmpp_mad = (struct ib_rmpp_mad *)mad_send_wr->send_wr.wr.ud.mad_hdr;
+	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
+	rmpp_mad->rmpp_hdr.seg_num = cpu_to_be32(mad_send_wr->seg_num);
+
+	if (mad_send_wr->seg_num == 1) {
+		rmpp_mad->rmpp_hdr.rmpp_rtime_flags |= IB_MGMT_RMPP_FLAG_FIRST;
+		rmpp_mad->rmpp_hdr.paylen_newwin =
+			cpu_to_be32(mad_send_wr->total_seg *
+				    (sizeof(struct ib_rmpp_mad) -
+				       offsetof(struct ib_rmpp_mad, data)));
+		mad_send_wr->sg_list[0].length = sizeof(struct ib_rmpp_mad);
+	} else {
+		mad_send_wr->send_wr.num_sge = 2;
+		mad_send_wr->sg_list[0].length = mad_send_wr->data_offset;
+		mad_send_wr->sg_list[1].addr = get_seg_addr(mad_send_wr);
+		mad_send_wr->sg_list[1].length = sizeof(struct ib_rmpp_mad) -
+						 mad_send_wr->data_offset;
+		mad_send_wr->sg_list[1].lkey = mad_send_wr->sg_list[0].lkey;
+	}
+
+	if (mad_send_wr->seg_num == mad_send_wr->total_seg) {
+		rmpp_mad->rmpp_hdr.rmpp_rtime_flags |= IB_MGMT_RMPP_FLAG_LAST;
+		rmpp_mad->rmpp_hdr.paylen_newwin =
+			cpu_to_be32(sizeof(struct ib_rmpp_mad) -
+				    offsetof(struct ib_rmpp_mad, data) -
+				    mad_send_wr->pad);
+	}
+
+	/* 2 seconds for an ACK until we can find the packet lifetime */
+	timeout = mad_send_wr->send_wr.wr.ud.timeout_ms;
+	if (!timeout || timeout > 2000)
+		mad_send_wr->timeout = msecs_to_jiffies(2000);
+	mad_send_wr->seg_num++;
+	return ib_send_mad(mad_send_wr);
+}
+
+static void process_rmpp_ack(struct ib_mad_agent_private *agent,
+			     struct ib_mad_recv_wc *mad_recv_wc)
+{
+	struct ib_mad_send_wr_private *mad_send_wr;
+	struct ib_rmpp_mad *rmpp_mad;
+	unsigned long flags;
+	int seg_num, newwin, ret;
+
+	rmpp_mad = (struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad;
+	if (rmpp_mad->rmpp_hdr.rmpp_status)
+		return;
+
+	seg_num = be32_to_cpu(rmpp_mad->rmpp_hdr.seg_num);
+	newwin = be32_to_cpu(rmpp_mad->rmpp_hdr.paylen_newwin);
+
+	spin_lock_irqsave(&agent->lock, flags);
+	mad_send_wr = ib_find_send_mad(agent, rmpp_mad->mad_hdr.tid);
+	if (!mad_send_wr)
+		goto out;	/* Unmatched ACK */
+
+	if ((mad_send_wr->last_ack == mad_send_wr->total_seg) ||
+	    (!mad_send_wr->timeout) || (mad_send_wr->status != IB_WC_SUCCESS))
+		goto out;	/* Send is already done */
+
+	if (seg_num > mad_send_wr->total_seg)
+		goto out;	/* Bad ACK */
+
+	if (newwin < mad_send_wr->newwin || seg_num < mad_send_wr->last_ack)
+		goto out;	/* Old ACK */
+
+	if (seg_num > mad_send_wr->last_ack) {
+		mad_send_wr->last_ack = seg_num;
+		mad_send_wr->retries = mad_send_wr->send_wr.wr.ud.retries;
+	}
+	mad_send_wr->newwin = newwin;
+	if (mad_send_wr->last_ack == mad_send_wr->total_seg) {
+		/* If no response is expected, the ACK completes the send */
+		if (!mad_send_wr->send_wr.wr.ud.timeout_ms) {
+			struct ib_mad_send_wc wc;
+
+			ib_mark_mad_done(mad_send_wr);
+			spin_unlock_irqrestore(&agent->lock, flags);
+
+			wc.status = IB_WC_SUCCESS;
+			wc.vendor_err = 0;
+			wc.wr_id = mad_send_wr->wr_id;
+			ib_mad_complete_send_wr(mad_send_wr, &wc);
+			return;
+		}
+		if (mad_send_wr->refcount == 1)
+			ib_reset_mad_timeout(mad_send_wr, mad_send_wr->
+					     send_wr.wr.ud.timeout_ms);
+	} else if (mad_send_wr->refcount == 1 &&
+		   mad_send_wr->seg_num < mad_send_wr->newwin &&
+		   mad_send_wr->seg_num <= mad_send_wr->total_seg) {
+		/* Send failure will just result in a timeout/retry */
+		ret = send_next_seg(mad_send_wr);
+		if (ret)
+			goto out;
+
+		mad_send_wr->refcount++;
+		list_del(&mad_send_wr->agent_list);
+		list_add_tail(&mad_send_wr->agent_list,
+			      &mad_send_wr->mad_agent_priv->send_list);
+	}
+out:
+	spin_unlock_irqrestore(&agent->lock, flags);
+}
+
+struct ib_mad_recv_wc *
+ib_process_rmpp_recv_wc(struct ib_mad_agent_private *agent,
+			struct ib_mad_recv_wc *mad_recv_wc)
+{
+	struct ib_rmpp_mad *rmpp_mad;
+
+	rmpp_mad = (struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad;
+	if (!(rmpp_mad->rmpp_hdr.rmpp_rtime_flags & IB_MGMT_RMPP_FLAG_ACTIVE))
+		return mad_recv_wc;
+
+	if (rmpp_mad->rmpp_hdr.rmpp_version != IB_MGMT_RMPP_VERSION)
+		goto out;
+
+	switch (rmpp_mad->rmpp_hdr.rmpp_type) {
+	case IB_MGMT_RMPP_TYPE_DATA:
+		if (rmpp_mad->rmpp_hdr.seg_num == __constant_htonl(1))
+			return start_rmpp(agent, mad_recv_wc);
+		else
+			return continue_rmpp(agent, mad_recv_wc);
+	case IB_MGMT_RMPP_TYPE_ACK:
+		process_rmpp_ack(agent, mad_recv_wc);
+		break;
+	case IB_MGMT_RMPP_TYPE_STOP:
+	case IB_MGMT_RMPP_TYPE_ABORT:
+		/* TODO: process_rmpp_nack(agent, mad_recv_wc); */
+		break;
+	default:
+		break;
+	}
+out:
+	ib_free_recv_mad(mad_recv_wc);
+	return NULL;
+}
+
+int ib_send_rmpp_mad(struct ib_mad_send_wr_private *mad_send_wr)
+{
+	struct ib_rmpp_mad *rmpp_mad;
+	int i, total_len, ret;
+
+	rmpp_mad = (struct ib_rmpp_mad *)mad_send_wr->send_wr.wr.ud.mad_hdr;
+	if (!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
+	      IB_MGMT_RMPP_FLAG_ACTIVE))
+		return IB_RMPP_RESULT_UNHANDLED;
+
+	if (rmpp_mad->rmpp_hdr.rmpp_type != IB_MGMT_RMPP_TYPE_DATA)
+		return IB_RMPP_RESULT_INTERNAL;
+
+	if (mad_send_wr->send_wr.num_sge > 1)
+		return -EINVAL;		/* TODO: support num_sge > 1 */
+
+	mad_send_wr->seg_num = 1;
+	mad_send_wr->newwin = 1;
+	mad_send_wr->data_offset = data_offset(rmpp_mad->mad_hdr.mgmt_class);
+
+	total_len = 0;
+	for (i = 0; i < mad_send_wr->send_wr.num_sge; i++)
+		total_len += mad_send_wr->send_wr.sg_list[i].length;
+
+        mad_send_wr->total_seg = (total_len - mad_send_wr->data_offset) /
+			(sizeof(struct ib_rmpp_mad) - mad_send_wr->data_offset);
+	mad_send_wr->pad = total_len - offsetof(struct ib_rmpp_mad, data) -
+			   be32_to_cpu(rmpp_mad->rmpp_hdr.paylen_newwin);
+
+	/* We need to wait for the final ACK even if there isn't a response */
+	mad_send_wr->refcount += (mad_send_wr->timeout == 0);
+	ret = send_next_seg(mad_send_wr);
+	if (!ret)
+		return IB_RMPP_RESULT_CONSUMED;
+	return ret;
+}
+
+int ib_process_rmpp_send_wc(struct ib_mad_send_wr_private *mad_send_wr,
+			    struct ib_mad_send_wc *mad_send_wc)
+{
+	struct ib_rmpp_mad *rmpp_mad;
+	struct ib_mad_send_buf *msg;
+	int ret;
+
+	rmpp_mad = (struct ib_rmpp_mad *)mad_send_wr->send_wr.wr.ud.mad_hdr;
+	if (!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
+	      IB_MGMT_RMPP_FLAG_ACTIVE))
+		return IB_RMPP_RESULT_UNHANDLED; /* RMPP not active */
+
+	if (rmpp_mad->rmpp_hdr.rmpp_type != IB_MGMT_RMPP_TYPE_DATA) {
+		msg = (struct ib_mad_send_buf *) (unsigned long)
+		      mad_send_wc->wr_id;
+		ib_free_send_mad(msg);
+		return IB_RMPP_RESULT_INTERNAL;	 /* ACK, STOP, or ABORT */
+	}
+
+	if (mad_send_wc->status != IB_WC_SUCCESS ||
+	    mad_send_wr->status != IB_WC_SUCCESS)
+		return IB_RMPP_RESULT_PROCESSED; /* Canceled or send error */
+
+	if (!mad_send_wr->timeout)
+		return IB_RMPP_RESULT_PROCESSED; /* Response received */
+
+	if (mad_send_wr->last_ack == mad_send_wr->total_seg) {
+		mad_send_wr->timeout =
+			msecs_to_jiffies(mad_send_wr->send_wr.wr.ud.timeout_ms);
+		return IB_RMPP_RESULT_PROCESSED; /* Send done */
+	}
+
+	if (mad_send_wr->seg_num > mad_send_wr->newwin ||
+	    mad_send_wr->seg_num > mad_send_wr->total_seg)
+		return IB_RMPP_RESULT_PROCESSED; /* Wait for ACK */
+
+	ret = send_next_seg(mad_send_wr);
+	if (ret) {
+		mad_send_wc->status = IB_WC_GENERAL_ERR;
+		return IB_RMPP_RESULT_PROCESSED;
+	}
+	return IB_RMPP_RESULT_CONSUMED;
+}
+
+int ib_retry_rmpp(struct ib_mad_send_wr_private *mad_send_wr)
+{
+	struct ib_rmpp_mad *rmpp_mad;
+	int ret;
+
+	rmpp_mad = (struct ib_rmpp_mad *)mad_send_wr->send_wr.wr.ud.mad_hdr;
+	if (!(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
+	      IB_MGMT_RMPP_FLAG_ACTIVE))
+		return IB_RMPP_RESULT_UNHANDLED; /* RMPP not active */
+
+	if (mad_send_wr->last_ack == mad_send_wr->total_seg)
+		return IB_RMPP_RESULT_PROCESSED;
+
+	mad_send_wr->seg_num = mad_send_wr->last_ack + 1;
+	ret = send_next_seg(mad_send_wr);
+	if (ret)
+		return IB_RMPP_RESULT_PROCESSED;
+
+	return IB_RMPP_RESULT_CONSUMED;
+}
diff --git a/drivers/infiniband/core/mad_rmpp.h b/drivers/infiniband/core/mad_rmpp.h
new file mode 100644
index 000000000000..c4924dfb8e75
--- /dev/null
+++ b/drivers/infiniband/core/mad_rmpp.h
@@ -0,0 +1,58 @@
+/*
+ * Copyright (c) 2005 Intel Inc. 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.
+ *
+ * $Id: mad_rmpp.h 1921 2005-02-25 22:58:44Z sean.hefty $
+ */
+
+#ifndef __MAD_RMPP_H__
+#define __MAD_RMPP_H__
+
+enum {
+	IB_RMPP_RESULT_PROCESSED,
+	IB_RMPP_RESULT_CONSUMED,
+	IB_RMPP_RESULT_INTERNAL,
+	IB_RMPP_RESULT_UNHANDLED
+};
+
+int ib_send_rmpp_mad(struct ib_mad_send_wr_private *mad_send_wr);
+
+struct ib_mad_recv_wc *
+ib_process_rmpp_recv_wc(struct ib_mad_agent_private *agent,
+			struct ib_mad_recv_wc *mad_recv_wc);
+
+int ib_process_rmpp_send_wc(struct ib_mad_send_wr_private *mad_send_wr,
+			    struct ib_mad_send_wc *mad_send_wc);
+
+void ib_cancel_rmpp_recvs(struct ib_mad_agent_private *agent);
+
+int ib_retry_rmpp(struct ib_mad_send_wr_private *mad_send_wr);
+
+#endif	/* __MAD_RMPP_H__ */
diff --git a/drivers/infiniband/core/packer.c b/drivers/infiniband/core/packer.c
index 5f15feffeae2..eb5ff54c10d7 100644
--- a/drivers/infiniband/core/packer.c
+++ b/drivers/infiniband/core/packer.c
@@ -96,7 +96,7 @@ void ib_pack(const struct ib_field        *desc,
 			else
 				val = 0;
 
-			mask = cpu_to_be64(((1ull << desc[i].size_bits) - 1) << shift);
+			mask = cpu_to_be64((~0ull >> (64 - desc[i].size_bits)) << shift);
 			addr = (__be64 *) ((__be32 *) buf + desc[i].offset_words);
 			*addr = (*addr & ~mask) | (cpu_to_be64(val) & mask);
 		} else {
@@ -176,7 +176,7 @@ void ib_unpack(const struct ib_field        *desc,
 			__be64 *addr;
 
 			shift = 64 - desc[i].offset_bits - desc[i].size_bits;
-			mask = ((1ull << desc[i].size_bits) - 1) << shift;
+			mask = (~0ull >> (64 - desc[i].size_bits)) << shift;
 			addr = (__be64 *) buf + desc[i].offset_words;
 			val = (be64_to_cpup(addr) & mask) >> shift;
 			value_write(desc[i].struct_offset_bytes,
diff --git a/drivers/infiniband/core/sa_query.c b/drivers/infiniband/core/sa_query.c
index 276e1a53010d..795184931c83 100644
--- a/drivers/infiniband/core/sa_query.c
+++ b/drivers/infiniband/core/sa_query.c
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Voltaire, Inc.  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
@@ -29,7 +30,7 @@
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  *
- * $Id: sa_query.c 1389 2004-12-27 22:56:47Z roland $
+ * $Id: sa_query.c 2811 2005-07-06 18:11:43Z halr $
  */
 
 #include <linux/module.h>
@@ -50,26 +51,6 @@ MODULE_AUTHOR("Roland Dreier");
 MODULE_DESCRIPTION("InfiniBand subnet administration query support");
 MODULE_LICENSE("Dual BSD/GPL");
 
-/*
- * These two structures must be packed because they have 64-bit fields
- * that are only 32-bit aligned.  64-bit architectures will lay them
- * out wrong otherwise.  (And unfortunately they are sent on the wire
- * so we can't change the layout)
- */
-struct ib_sa_hdr {
-	u64			sm_key;
-	u16			attr_offset;
-	u16			reserved;
-	ib_sa_comp_mask		comp_mask;
-} __attribute__ ((packed));
-
-struct ib_sa_mad {
-	struct ib_mad_hdr	mad_hdr;
-	struct ib_rmpp_hdr	rmpp_hdr;
-	struct ib_sa_hdr	sa_hdr;
-	u8			data[200];
-} __attribute__ ((packed));
-
 struct ib_sa_sm_ah {
 	struct ib_ah        *ah;
 	struct kref          ref;
@@ -77,7 +58,6 @@ struct ib_sa_sm_ah {
 
 struct ib_sa_port {
 	struct ib_mad_agent *agent;
-	struct ib_mr        *mr;
 	struct ib_sa_sm_ah  *sm_ah;
 	struct work_struct   update_task;
 	spinlock_t           ah_lock;
@@ -100,6 +80,12 @@ struct ib_sa_query {
 	int                 id;
 };
 
+struct ib_sa_service_query {
+	void (*callback)(int, struct ib_sa_service_rec *, void *);
+	void *context;
+	struct ib_sa_query sa_query;
+};
+
 struct ib_sa_path_query {
 	void (*callback)(int, struct ib_sa_path_rec *, void *);
 	void *context;
@@ -341,6 +327,54 @@ static const struct ib_field mcmember_rec_table[] = {
 	  .size_bits    = 23 },
 };
 
+#define SERVICE_REC_FIELD(field) \
+	.struct_offset_bytes = offsetof(struct ib_sa_service_rec, field),	\
+	.struct_size_bytes   = sizeof ((struct ib_sa_service_rec *) 0)->field,	\
+	.field_name          = "sa_service_rec:" #field
+
+static const struct ib_field service_rec_table[] = {
+	{ SERVICE_REC_FIELD(id),
+	  .offset_words = 0,
+	  .offset_bits  = 0,
+	  .size_bits    = 64 },
+	{ SERVICE_REC_FIELD(gid),
+	  .offset_words = 2,
+	  .offset_bits  = 0,
+	  .size_bits    = 128 },
+	{ SERVICE_REC_FIELD(pkey),
+	  .offset_words = 6,
+	  .offset_bits  = 0,
+	  .size_bits    = 16 },
+	{ SERVICE_REC_FIELD(lease),
+	  .offset_words = 7,
+	  .offset_bits  = 0,
+	  .size_bits    = 32 },
+	{ SERVICE_REC_FIELD(key),
+	  .offset_words = 8,
+	  .offset_bits  = 0,
+	  .size_bits    = 128 },
+	{ SERVICE_REC_FIELD(name),
+	  .offset_words = 12,
+	  .offset_bits  = 0,
+	  .size_bits    = 64*8 },
+	{ SERVICE_REC_FIELD(data8),
+	  .offset_words = 28,
+	  .offset_bits  = 0,
+	  .size_bits    = 16*8 },
+	{ SERVICE_REC_FIELD(data16),
+	  .offset_words = 32,
+	  .offset_bits  = 0,
+	  .size_bits    = 8*16 },
+	{ SERVICE_REC_FIELD(data32),
+	  .offset_words = 36,
+	  .offset_bits  = 0,
+	  .size_bits    = 4*32 },
+	{ SERVICE_REC_FIELD(data64),
+	  .offset_words = 40,
+	  .offset_bits  = 0,
+	  .size_bits    = 2*64 },
+};
+
 static void free_sm_ah(struct kref *kref)
 {
 	struct ib_sa_sm_ah *sm_ah = container_of(kref, struct ib_sa_sm_ah, ref);
@@ -463,7 +497,7 @@ static int send_mad(struct ib_sa_query *query, int timeout_ms)
 				 .mad_hdr     = &query->mad->mad_hdr,
 				 .remote_qpn  = 1,
 				 .remote_qkey = IB_QP1_QKEY,
-				 .timeout_ms  = timeout_ms
+				 .timeout_ms  = timeout_ms,
 			 }
 		 }
 	};
@@ -492,7 +526,7 @@ retry:
 					    sizeof (struct ib_sa_mad),
 					    DMA_TO_DEVICE);
 	gather_list.length = sizeof (struct ib_sa_mad);
-	gather_list.lkey   = port->mr->lkey;
+	gather_list.lkey   = port->agent->mr->lkey;
 	pci_unmap_addr_set(query, mapping, gather_list.addr);
 
 	ret = ib_post_send_mad(port->agent, &wr, &bad_wr);
@@ -507,7 +541,13 @@ retry:
 		spin_unlock_irqrestore(&idr_lock, flags);
 	}
 
-	return ret;
+	/*
+	 * It's not safe to dereference query any more, because the
+	 * send may already have completed and freed the query in
+	 * another context.  So use wr.wr_id, which has a copy of the
+	 * query's id.
+	 */
+	return ret ? ret : wr.wr_id;
 }
 
 static void ib_sa_path_rec_callback(struct ib_sa_query *sa_query,
@@ -560,7 +600,7 @@ static void ib_sa_path_rec_release(struct ib_sa_query *sa_query)
 int ib_sa_path_rec_get(struct ib_device *device, u8 port_num,
 		       struct ib_sa_path_rec *rec,
 		       ib_sa_comp_mask comp_mask,
-		       int timeout_ms, int gfp_mask,
+		       int timeout_ms, unsigned int __nocast gfp_mask,
 		       void (*callback)(int status,
 					struct ib_sa_path_rec *resp,
 					void *context),
@@ -598,17 +638,126 @@ int ib_sa_path_rec_get(struct ib_device *device, u8 port_num,
 		rec, query->sa_query.mad->data);
 
 	*sa_query = &query->sa_query;
+
 	ret = send_mad(&query->sa_query, timeout_ms);
-	if (ret) {
+	if (ret < 0) {
 		*sa_query = NULL;
 		kfree(query->sa_query.mad);
 		kfree(query);
 	}
 
-	return ret ? ret : query->sa_query.id;
+	return ret;
 }
 EXPORT_SYMBOL(ib_sa_path_rec_get);
 
+static void ib_sa_service_rec_callback(struct ib_sa_query *sa_query,
+				    int status,
+				    struct ib_sa_mad *mad)
+{
+	struct ib_sa_service_query *query =
+		container_of(sa_query, struct ib_sa_service_query, sa_query);
+
+	if (mad) {
+		struct ib_sa_service_rec rec;
+
+		ib_unpack(service_rec_table, ARRAY_SIZE(service_rec_table),
+			  mad->data, &rec);
+		query->callback(status, &rec, query->context);
+	} else
+		query->callback(status, NULL, query->context);
+}
+
+static void ib_sa_service_rec_release(struct ib_sa_query *sa_query)
+{
+	kfree(sa_query->mad);
+	kfree(container_of(sa_query, struct ib_sa_service_query, sa_query));
+}
+
+/**
+ * ib_sa_service_rec_query - Start Service Record operation
+ * @device:device to send request on
+ * @port_num: port number to send request on
+ * @method:SA method - should be get, set, or delete
+ * @rec:Service Record to send in request
+ * @comp_mask:component mask to send in request
+ * @timeout_ms:time to wait for response
+ * @gfp_mask:GFP mask to use for internal allocations
+ * @callback:function called when request completes, times out or is
+ * canceled
+ * @context:opaque user context passed to callback
+ * @sa_query:request context, used to cancel request
+ *
+ * Send a Service Record set/get/delete to the SA to register,
+ * unregister or query a service record.
+ * The callback function will be called when the request completes (or
+ * fails); status is 0 for a successful response, -EINTR if the query
+ * is canceled, -ETIMEDOUT is the query timed out, or -EIO if an error
+ * occurred sending the query.  The resp parameter of the callback is
+ * only valid if status is 0.
+ *
+ * If the return value of ib_sa_service_rec_query() is negative, it is an
+ * error code.  Otherwise it is a request ID that can be used to cancel
+ * the query.
+ */
+int ib_sa_service_rec_query(struct ib_device *device, u8 port_num, u8 method,
+			    struct ib_sa_service_rec *rec,
+			    ib_sa_comp_mask comp_mask,
+			    int timeout_ms, unsigned int __nocast gfp_mask,
+			    void (*callback)(int status,
+					     struct ib_sa_service_rec *resp,
+					     void *context),
+			    void *context,
+			    struct ib_sa_query **sa_query)
+{
+	struct ib_sa_service_query *query;
+	struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
+	struct ib_sa_port   *port   = &sa_dev->port[port_num - sa_dev->start_port];
+	struct ib_mad_agent *agent  = port->agent;
+	int ret;
+
+	if (method != IB_MGMT_METHOD_GET &&
+	    method != IB_MGMT_METHOD_SET &&
+	    method != IB_SA_METHOD_DELETE)
+		return -EINVAL;
+
+	query = kmalloc(sizeof *query, gfp_mask);
+	if (!query)
+		return -ENOMEM;
+	query->sa_query.mad = kmalloc(sizeof *query->sa_query.mad, gfp_mask);
+	if (!query->sa_query.mad) {
+		kfree(query);
+		return -ENOMEM;
+	}
+
+	query->callback = callback;
+	query->context  = context;
+
+	init_mad(query->sa_query.mad, agent);
+
+	query->sa_query.callback              = callback ? ib_sa_service_rec_callback : NULL;
+	query->sa_query.release               = ib_sa_service_rec_release;
+	query->sa_query.port                  = port;
+	query->sa_query.mad->mad_hdr.method   = method;
+	query->sa_query.mad->mad_hdr.attr_id  =
+				cpu_to_be16(IB_SA_ATTR_SERVICE_REC);
+	query->sa_query.mad->sa_hdr.comp_mask = comp_mask;
+
+	ib_pack(service_rec_table, ARRAY_SIZE(service_rec_table),
+		rec, query->sa_query.mad->data);
+
+	*sa_query = &query->sa_query;
+
+	ret = send_mad(&query->sa_query, timeout_ms);
+	if (ret < 0) {
+		*sa_query = NULL;
+		kfree(query->sa_query.mad);
+		kfree(query);
+	}
+
+	return ret;
+}
+EXPORT_SYMBOL(ib_sa_service_rec_query);
+
 static void ib_sa_mcmember_rec_callback(struct ib_sa_query *sa_query,
 					int status,
 					struct ib_sa_mad *mad)
@@ -636,7 +785,7 @@ int ib_sa_mcmember_rec_query(struct ib_device *device, u8 port_num,
 			     u8 method,
 			     struct ib_sa_mcmember_rec *rec,
 			     ib_sa_comp_mask comp_mask,
-			     int timeout_ms, int gfp_mask,
+			     int timeout_ms, unsigned int __nocast gfp_mask,
 			     void (*callback)(int status,
 					      struct ib_sa_mcmember_rec *resp,
 					      void *context),
@@ -674,14 +823,15 @@ int ib_sa_mcmember_rec_query(struct ib_device *device, u8 port_num,
 		rec, query->sa_query.mad->data);
 
 	*sa_query = &query->sa_query;
+
 	ret = send_mad(&query->sa_query, timeout_ms);
-	if (ret) {
+	if (ret < 0) {
 		*sa_query = NULL;
 		kfree(query->sa_query.mad);
 		kfree(query);
 	}
 
-	return ret ? ret : query->sa_query.id;
+	return ret;
 }
 EXPORT_SYMBOL(ib_sa_mcmember_rec_query);
 
@@ -772,7 +922,6 @@ static void ib_sa_add_one(struct ib_device *device)
 	sa_dev->end_port   = e;
 
 	for (i = 0; i <= e - s; ++i) {
-		sa_dev->port[i].mr       = NULL;
 		sa_dev->port[i].sm_ah    = NULL;
 		sa_dev->port[i].port_num = i + s;
 		spin_lock_init(&sa_dev->port[i].ah_lock);
@@ -784,13 +933,6 @@ static void ib_sa_add_one(struct ib_device *device)
 		if (IS_ERR(sa_dev->port[i].agent))
 			goto err;
 
-		sa_dev->port[i].mr = ib_get_dma_mr(sa_dev->port[i].agent->qp->pd,
-						   IB_ACCESS_LOCAL_WRITE);
-		if (IS_ERR(sa_dev->port[i].mr)) {
-			ib_unregister_mad_agent(sa_dev->port[i].agent);
-			goto err;
-		}
-
 		INIT_WORK(&sa_dev->port[i].update_task,
 			  update_sm_ah, &sa_dev->port[i]);
 	}
@@ -814,10 +956,8 @@ static void ib_sa_add_one(struct ib_device *device)
 	return;
 
 err:
-	while (--i >= 0) {
-		ib_dereg_mr(sa_dev->port[i].mr);
+	while (--i >= 0)
 		ib_unregister_mad_agent(sa_dev->port[i].agent);
-	}
 
 	kfree(sa_dev);
 
diff --git a/drivers/infiniband/core/ucm.c b/drivers/infiniband/core/ucm.c
new file mode 100644
index 000000000000..61d07c732f49
--- /dev/null
+++ b/drivers/infiniband/core/ucm.c
@@ -0,0 +1,1387 @@
+/*
+ * Copyright (c) 2005 Topspin Communications.  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.
+ *
+ * $Id: ucm.c 2594 2005-06-13 19:46:02Z libor $
+ */
+#include <linux/init.h>
+#include <linux/fs.h>
+#include <linux/module.h>
+#include <linux/device.h>
+#include <linux/err.h>
+#include <linux/poll.h>
+#include <linux/file.h>
+#include <linux/mount.h>
+#include <linux/cdev.h>
+
+#include <asm/uaccess.h>
+
+#include "ucm.h"
+
+MODULE_AUTHOR("Libor Michalek");
+MODULE_DESCRIPTION("InfiniBand userspace Connection Manager access");
+MODULE_LICENSE("Dual BSD/GPL");
+
+static int ucm_debug_level;
+
+module_param_named(debug_level, ucm_debug_level, int, 0644);
+MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
+
+enum {
+	IB_UCM_MAJOR = 231,
+	IB_UCM_MINOR = 255
+};
+
+#define IB_UCM_DEV MKDEV(IB_UCM_MAJOR, IB_UCM_MINOR)
+
+#define PFX "UCM: "
+
+#define ucm_dbg(format, arg...)			\
+	do {					\
+		if (ucm_debug_level > 0)	\
+			printk(KERN_DEBUG PFX format, ## arg); \
+	} while (0)
+
+static struct semaphore ctx_id_mutex;
+static struct idr       ctx_id_table;
+static int              ctx_id_rover = 0;
+
+static struct ib_ucm_context *ib_ucm_ctx_get(int id)
+{
+	struct ib_ucm_context *ctx;
+
+	down(&ctx_id_mutex);
+	ctx = idr_find(&ctx_id_table, id);
+	if (ctx)
+		ctx->ref++;
+	up(&ctx_id_mutex);
+
+	return ctx;
+}
+
+static void ib_ucm_ctx_put(struct ib_ucm_context *ctx)
+{
+	struct ib_ucm_event *uevent;
+
+	down(&ctx_id_mutex);
+
+	ctx->ref--;
+	if (!ctx->ref)
+		idr_remove(&ctx_id_table, ctx->id);
+
+	up(&ctx_id_mutex);
+
+	if (ctx->ref)
+		return;
+
+	down(&ctx->file->mutex);
+
+	list_del(&ctx->file_list);
+	while (!list_empty(&ctx->events)) {
+
+		uevent = list_entry(ctx->events.next,
+				    struct ib_ucm_event, ctx_list);
+		list_del(&uevent->file_list);
+		list_del(&uevent->ctx_list);
+
+		/* clear incoming connections. */
+		if (uevent->cm_id)
+			ib_destroy_cm_id(uevent->cm_id);
+
+		kfree(uevent);
+	}
+
+	up(&ctx->file->mutex);
+
+	ucm_dbg("Destroyed CM ID <%d>\n", ctx->id);
+
+	ib_destroy_cm_id(ctx->cm_id);
+	kfree(ctx);
+}
+
+static struct ib_ucm_context *ib_ucm_ctx_alloc(struct ib_ucm_file *file)
+{
+	struct ib_ucm_context *ctx;
+	int result;
+
+	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
+	if (!ctx)
+		return NULL;
+
+	ctx->ref  = 1; /* user reference */
+	ctx->file = file;
+
+	INIT_LIST_HEAD(&ctx->events);
+	init_MUTEX(&ctx->mutex);
+
+	list_add_tail(&ctx->file_list, &file->ctxs);
+
+	ctx_id_rover = (ctx_id_rover + 1) & INT_MAX;
+retry:
+	result = idr_pre_get(&ctx_id_table, GFP_KERNEL);
+	if (!result)
+		goto error;
+
+	down(&ctx_id_mutex);
+	result = idr_get_new_above(&ctx_id_table, ctx, ctx_id_rover, &ctx->id);
+	up(&ctx_id_mutex);
+
+	if (result == -EAGAIN)
+		goto retry;
+	if (result)
+		goto error;
+
+	ucm_dbg("Allocated CM ID <%d>\n", ctx->id);
+
+	return ctx;
+error:
+	list_del(&ctx->file_list);
+	kfree(ctx);
+
+	return NULL;
+}
+/*
+ * Event portion of the API, handle CM events
+ * and allow event polling.
+ */
+static void ib_ucm_event_path_get(struct ib_ucm_path_rec *upath,
+				  struct ib_sa_path_rec	 *kpath)
+{
+	if (!kpath || !upath)
+		return;
+
+	memcpy(upath->dgid, kpath->dgid.raw, sizeof(union ib_gid));
+	memcpy(upath->sgid, kpath->sgid.raw, sizeof(union ib_gid));
+
+	upath->dlid             = kpath->dlid;
+	upath->slid             = kpath->slid;
+	upath->raw_traffic      = kpath->raw_traffic;
+	upath->flow_label       = kpath->flow_label;
+	upath->hop_limit        = kpath->hop_limit;
+	upath->traffic_class    = kpath->traffic_class;
+	upath->reversible       = kpath->reversible;
+	upath->numb_path        = kpath->numb_path;
+	upath->pkey             = kpath->pkey;
+	upath->sl	        = kpath->sl;
+	upath->mtu_selector     = kpath->mtu_selector;
+	upath->mtu              = kpath->mtu;
+	upath->rate_selector    = kpath->rate_selector;
+	upath->rate             = kpath->rate;
+	upath->packet_life_time = kpath->packet_life_time;
+	upath->preference       = kpath->preference;
+
+	upath->packet_life_time_selector =
+		kpath->packet_life_time_selector;
+}
+
+static void ib_ucm_event_req_get(struct ib_ucm_req_event_resp *ureq,
+				 struct ib_cm_req_event_param *kreq)
+{
+	ureq->listen_id = (long)kreq->listen_id->context;
+
+	ureq->remote_ca_guid             = kreq->remote_ca_guid;
+	ureq->remote_qkey                = kreq->remote_qkey;
+	ureq->remote_qpn                 = kreq->remote_qpn;
+	ureq->qp_type                    = kreq->qp_type;
+	ureq->starting_psn               = kreq->starting_psn;
+	ureq->responder_resources        = kreq->responder_resources;
+	ureq->initiator_depth            = kreq->initiator_depth;
+	ureq->local_cm_response_timeout  = kreq->local_cm_response_timeout;
+	ureq->flow_control               = kreq->flow_control;
+	ureq->remote_cm_response_timeout = kreq->remote_cm_response_timeout;
+	ureq->retry_count                = kreq->retry_count;
+	ureq->rnr_retry_count            = kreq->rnr_retry_count;
+	ureq->srq                        = kreq->srq;
+
+	ib_ucm_event_path_get(&ureq->primary_path, kreq->primary_path);
+	ib_ucm_event_path_get(&ureq->alternate_path, kreq->alternate_path);
+}
+
+static void ib_ucm_event_rep_get(struct ib_ucm_rep_event_resp *urep,
+				 struct ib_cm_rep_event_param *krep)
+{
+	urep->remote_ca_guid      = krep->remote_ca_guid;
+	urep->remote_qkey         = krep->remote_qkey;
+	urep->remote_qpn          = krep->remote_qpn;
+	urep->starting_psn        = krep->starting_psn;
+	urep->responder_resources = krep->responder_resources;
+	urep->initiator_depth     = krep->initiator_depth;
+	urep->target_ack_delay    = krep->target_ack_delay;
+	urep->failover_accepted   = krep->failover_accepted;
+	urep->flow_control        = krep->flow_control;
+	urep->rnr_retry_count     = krep->rnr_retry_count;
+	urep->srq                 = krep->srq;
+}
+
+static void ib_ucm_event_rej_get(struct ib_ucm_rej_event_resp *urej,
+				 struct ib_cm_rej_event_param *krej)
+{
+	urej->reason = krej->reason;
+}
+
+static void ib_ucm_event_mra_get(struct ib_ucm_mra_event_resp *umra,
+				 struct ib_cm_mra_event_param *kmra)
+{
+	umra->timeout = kmra->service_timeout;
+}
+
+static void ib_ucm_event_lap_get(struct ib_ucm_lap_event_resp *ulap,
+				 struct ib_cm_lap_event_param *klap)
+{
+	ib_ucm_event_path_get(&ulap->path, klap->alternate_path);
+}
+
+static void ib_ucm_event_apr_get(struct ib_ucm_apr_event_resp *uapr,
+				 struct ib_cm_apr_event_param *kapr)
+{
+	uapr->status = kapr->ap_status;
+}
+
+static void ib_ucm_event_sidr_req_get(struct ib_ucm_sidr_req_event_resp *ureq,
+				      struct ib_cm_sidr_req_event_param *kreq)
+{
+	ureq->listen_id = (long)kreq->listen_id->context;
+	ureq->pkey      = kreq->pkey;
+}
+
+static void ib_ucm_event_sidr_rep_get(struct ib_ucm_sidr_rep_event_resp *urep,
+				      struct ib_cm_sidr_rep_event_param *krep)
+{
+	urep->status = krep->status;
+	urep->qkey   = krep->qkey;
+	urep->qpn    = krep->qpn;
+};
+
+static int ib_ucm_event_process(struct ib_cm_event *evt,
+				struct ib_ucm_event *uvt)
+{
+	void *info = NULL;
+	int result;
+
+	switch (evt->event) {
+	case IB_CM_REQ_RECEIVED:
+		ib_ucm_event_req_get(&uvt->resp.u.req_resp,
+				     &evt->param.req_rcvd);
+		uvt->data_len      = IB_CM_REQ_PRIVATE_DATA_SIZE;
+		uvt->resp.present |= (evt->param.req_rcvd.primary_path ?
+				      IB_UCM_PRES_PRIMARY : 0);
+		uvt->resp.present |= (evt->param.req_rcvd.alternate_path ?
+				      IB_UCM_PRES_ALTERNATE : 0);
+		break;
+	case IB_CM_REP_RECEIVED:
+		ib_ucm_event_rep_get(&uvt->resp.u.rep_resp,
+				     &evt->param.rep_rcvd);
+		uvt->data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
+
+		break;
+	case IB_CM_RTU_RECEIVED:
+		uvt->data_len = IB_CM_RTU_PRIVATE_DATA_SIZE;
+		uvt->resp.u.send_status = evt->param.send_status;
+
+		break;
+	case IB_CM_DREQ_RECEIVED:
+		uvt->data_len = IB_CM_DREQ_PRIVATE_DATA_SIZE;
+		uvt->resp.u.send_status = evt->param.send_status;
+
+		break;
+	case IB_CM_DREP_RECEIVED:
+		uvt->data_len = IB_CM_DREP_PRIVATE_DATA_SIZE;
+		uvt->resp.u.send_status = evt->param.send_status;
+
+		break;
+	case IB_CM_MRA_RECEIVED:
+		ib_ucm_event_mra_get(&uvt->resp.u.mra_resp,
+				     &evt->param.mra_rcvd);
+		uvt->data_len = IB_CM_MRA_PRIVATE_DATA_SIZE;
+
+		break;
+	case IB_CM_REJ_RECEIVED:
+		ib_ucm_event_rej_get(&uvt->resp.u.rej_resp,
+				     &evt->param.rej_rcvd);
+		uvt->data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
+		uvt->info_len = evt->param.rej_rcvd.ari_length;
+		info	      = evt->param.rej_rcvd.ari;
+
+		break;
+	case IB_CM_LAP_RECEIVED:
+		ib_ucm_event_lap_get(&uvt->resp.u.lap_resp,
+				     &evt->param.lap_rcvd);
+		uvt->data_len = IB_CM_LAP_PRIVATE_DATA_SIZE;
+		uvt->resp.present |= (evt->param.lap_rcvd.alternate_path ?
+				      IB_UCM_PRES_ALTERNATE : 0);
+		break;
+	case IB_CM_APR_RECEIVED:
+		ib_ucm_event_apr_get(&uvt->resp.u.apr_resp,
+				     &evt->param.apr_rcvd);
+		uvt->data_len = IB_CM_APR_PRIVATE_DATA_SIZE;
+		uvt->info_len = evt->param.apr_rcvd.info_len;
+		info	      = evt->param.apr_rcvd.apr_info;
+
+		break;
+	case IB_CM_SIDR_REQ_RECEIVED:
+		ib_ucm_event_sidr_req_get(&uvt->resp.u.sidr_req_resp,
+					  &evt->param.sidr_req_rcvd);
+		uvt->data_len = IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE;
+
+		break;
+	case IB_CM_SIDR_REP_RECEIVED:
+		ib_ucm_event_sidr_rep_get(&uvt->resp.u.sidr_rep_resp,
+					  &evt->param.sidr_rep_rcvd);
+		uvt->data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
+		uvt->info_len = evt->param.sidr_rep_rcvd.info_len;
+		info	      = evt->param.sidr_rep_rcvd.info;
+
+		break;
+	default:
+		uvt->resp.u.send_status = evt->param.send_status;
+
+		break;
+	}
+
+	if (uvt->data_len && evt->private_data) {
+
+		uvt->data = kmalloc(uvt->data_len, GFP_KERNEL);
+		if (!uvt->data) {
+			result = -ENOMEM;
+			goto error;
+		}
+
+		memcpy(uvt->data, evt->private_data, uvt->data_len);
+		uvt->resp.present |= IB_UCM_PRES_DATA;
+	}
+
+	if (uvt->info_len && info) {
+
+		uvt->info = kmalloc(uvt->info_len, GFP_KERNEL);
+		if (!uvt->info) {
+			result = -ENOMEM;
+			goto error;
+		}
+
+		memcpy(uvt->info, info, uvt->info_len);
+		uvt->resp.present |= IB_UCM_PRES_INFO;
+	}
+
+	return 0;
+error:
+	kfree(uvt->info);
+	kfree(uvt->data);
+	return result;
+}
+
+static int ib_ucm_event_handler(struct ib_cm_id *cm_id,
+				struct ib_cm_event *event)
+{
+	struct ib_ucm_event *uevent;
+	struct ib_ucm_context *ctx;
+	int result = 0;
+	int id;
+	/*
+	 * lookup correct context based on event type.
+	 */
+	switch (event->event) {
+	case IB_CM_REQ_RECEIVED:
+		id = (long)event->param.req_rcvd.listen_id->context;
+		break;
+	case IB_CM_SIDR_REQ_RECEIVED:
+		id = (long)event->param.sidr_req_rcvd.listen_id->context;
+		break;
+	default:
+		id = (long)cm_id->context;
+		break;
+	}
+
+	ucm_dbg("Event. CM ID <%d> event <%d>\n", id, event->event);
+
+	ctx = ib_ucm_ctx_get(id);
+	if (!ctx)
+		return -ENOENT;
+
+	if (event->event == IB_CM_REQ_RECEIVED ||
+	    event->event == IB_CM_SIDR_REQ_RECEIVED)
+		id = IB_UCM_CM_ID_INVALID;
+
+	uevent = kmalloc(sizeof(*uevent), GFP_KERNEL);
+	if (!uevent) {
+		result = -ENOMEM;
+		goto done;
+	}
+
+	memset(uevent, 0, sizeof(*uevent));
+
+	uevent->resp.id    = id;
+	uevent->resp.event = event->event;
+
+	result = ib_ucm_event_process(event, uevent);
+	if (result)
+		goto done;
+
+	uevent->ctx   = ctx;
+	uevent->cm_id = ((event->event == IB_CM_REQ_RECEIVED ||
+			  event->event == IB_CM_SIDR_REQ_RECEIVED ) ?
+			 cm_id : NULL);
+
+	down(&ctx->file->mutex);
+
+	list_add_tail(&uevent->file_list, &ctx->file->events);
+	list_add_tail(&uevent->ctx_list, &ctx->events);
+
+	wake_up_interruptible(&ctx->file->poll_wait);
+
+	up(&ctx->file->mutex);
+done:
+	ctx->error = result;
+	ib_ucm_ctx_put(ctx); /* func reference */
+	return result;
+}
+
+static ssize_t ib_ucm_event(struct ib_ucm_file *file,
+			    const char __user *inbuf,
+			    int in_len, int out_len)
+{
+	struct ib_ucm_context *ctx;
+	struct ib_ucm_event_get cmd;
+	struct ib_ucm_event *uevent = NULL;
+	int result = 0;
+	DEFINE_WAIT(wait);
+
+	if (out_len < sizeof(struct ib_ucm_event_resp))
+		return -ENOSPC;
+
+	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+		return -EFAULT;
+	/*
+	 * wait
+	 */
+	down(&file->mutex);
+
+	while (list_empty(&file->events)) {
+
+		if (file->filp->f_flags & O_NONBLOCK) {
+			result = -EAGAIN;
+			break;
+		}
+
+		if (signal_pending(current)) {
+			result = -ERESTARTSYS;
+			break;
+		}
+
+		prepare_to_wait(&file->poll_wait, &wait, TASK_INTERRUPTIBLE);
+
+		up(&file->mutex);
+		schedule();
+		down(&file->mutex);
+
+		finish_wait(&file->poll_wait, &wait);
+	}
+
+	if (result)
+		goto done;
+
+	uevent = list_entry(file->events.next, struct ib_ucm_event, file_list);
+
+	if (!uevent->cm_id)
+		goto user;
+
+	ctx = ib_ucm_ctx_alloc(file);
+	if (!ctx) {
+		result = -ENOMEM;
+		goto done;
+	}
+
+	ctx->cm_id             = uevent->cm_id;
+	ctx->cm_id->cm_handler = ib_ucm_event_handler;
+	ctx->cm_id->context    = (void *)(unsigned long)ctx->id;
+
+	uevent->resp.id = ctx->id;
+
+user:
+	if (copy_to_user((void __user *)(unsigned long)cmd.response,
+			 &uevent->resp, sizeof(uevent->resp))) {
+		result = -EFAULT;
+		goto done;
+	}
+
+	if (uevent->data) {
+
+		if (cmd.data_len < uevent->data_len) {
+			result = -ENOMEM;
+			goto done;
+		}
+
+		if (copy_to_user((void __user *)(unsigned long)cmd.data,
+				 uevent->data, uevent->data_len)) {
+			result = -EFAULT;
+			goto done;
+		}
+	}
+
+	if (uevent->info) {
+
+		if (cmd.info_len < uevent->info_len) {
+			result = -ENOMEM;
+			goto done;
+		}
+
+		if (copy_to_user((void __user *)(unsigned long)cmd.info,
+				 uevent->info, uevent->info_len)) {
+			result = -EFAULT;
+			goto done;
+		}
+	}
+
+	list_del(&uevent->file_list);
+	list_del(&uevent->ctx_list);
+
+	kfree(uevent->data);
+	kfree(uevent->info);
+	kfree(uevent);
+done:
+	up(&file->mutex);
+	return result;
+}
+
+
+static ssize_t ib_ucm_create_id(struct ib_ucm_file *file,
+				const char __user *inbuf,
+				int in_len, int out_len)
+{
+	struct ib_ucm_create_id cmd;
+	struct ib_ucm_create_id_resp resp;
+	struct ib_ucm_context *ctx;
+	int result;
+
+	if (out_len < sizeof(resp))
+		return -ENOSPC;
+
+	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+		return -EFAULT;
+
+	ctx = ib_ucm_ctx_alloc(file);
+	if (!ctx)
+		return -ENOMEM;
+
+	ctx->cm_id = ib_create_cm_id(ib_ucm_event_handler,
+				     (void *)(unsigned long)ctx->id);
+	if (!ctx->cm_id) {
+		result = -ENOMEM;
+		goto err_cm;
+	}
+
+	resp.id = ctx->id;
+	if (copy_to_user((void __user *)(unsigned long)cmd.response,
+			 &resp, sizeof(resp))) {
+		result = -EFAULT;
+		goto err_ret;
+	}
+
+	return 0;
+err_ret:
+	ib_destroy_cm_id(ctx->cm_id);
+err_cm:
+	ib_ucm_ctx_put(ctx); /* user reference */
+
+	return result;
+}
+
+static ssize_t ib_ucm_destroy_id(struct ib_ucm_file *file,
+				 const char __user *inbuf,
+				 int in_len, int out_len)
+{
+	struct ib_ucm_destroy_id cmd;
+	struct ib_ucm_context *ctx;
+
+	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+		return -EFAULT;
+
+	ctx = ib_ucm_ctx_get(cmd.id);
+	if (!ctx)
+		return -ENOENT;
+
+	ib_ucm_ctx_put(ctx); /* user reference */
+	ib_ucm_ctx_put(ctx); /* func reference */
+
+	return 0;
+}
+
+static ssize_t ib_ucm_attr_id(struct ib_ucm_file *file,
+			      const char __user *inbuf,
+			      int in_len, int out_len)
+{
+	struct ib_ucm_attr_id_resp resp;
+	struct ib_ucm_attr_id cmd;
+	struct ib_ucm_context *ctx;
+	int result = 0;
+
+	if (out_len < sizeof(resp))
+		return -ENOSPC;
+
+	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+		return -EFAULT;
+
+	ctx = ib_ucm_ctx_get(cmd.id);
+	if (!ctx)
+		return -ENOENT;
+
+ 	down(&ctx->file->mutex);
+	if (ctx->file != file) {
+		result = -EINVAL;
+		goto done;
+	}
+
+	resp.service_id   = ctx->cm_id->service_id;
+	resp.service_mask = ctx->cm_id->service_mask;
+	resp.local_id     = ctx->cm_id->local_id;
+	resp.remote_id    = ctx->cm_id->remote_id;
+
+	if (copy_to_user((void __user *)(unsigned long)cmd.response,
+			 &resp, sizeof(resp)))
+		result = -EFAULT;
+
+done:
+	up(&ctx->file->mutex);
+	ib_ucm_ctx_put(ctx); /* func reference */
+	return result;
+}
+
+static ssize_t ib_ucm_listen(struct ib_ucm_file *file,
+			     const char __user *inbuf,
+			     int in_len, int out_len)
+{
+	struct ib_ucm_listen cmd;
+	struct ib_ucm_context *ctx;
+	int result;
+
+	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+		return -EFAULT;
+
+	ctx = ib_ucm_ctx_get(cmd.id);
+	if (!ctx)
+		return -ENOENT;
+
+ 	down(&ctx->file->mutex);
+	if (ctx->file != file)
+		result = -EINVAL;
+	else
+		result = ib_cm_listen(ctx->cm_id, cmd.service_id,
+				      cmd.service_mask);
+
+	up(&ctx->file->mutex);
+	ib_ucm_ctx_put(ctx); /* func reference */
+	return result;
+}
+
+static ssize_t ib_ucm_establish(struct ib_ucm_file *file,
+				const char __user *inbuf,
+				int in_len, int out_len)
+{
+	struct ib_ucm_establish cmd;
+	struct ib_ucm_context *ctx;
+	int result;
+
+	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+		return -EFAULT;
+
+	ctx = ib_ucm_ctx_get(cmd.id);
+	if (!ctx)
+		return -ENOENT;
+
+ 	down(&ctx->file->mutex);
+	if (ctx->file != file)
+		result = -EINVAL;
+	else
+		result = ib_cm_establish(ctx->cm_id);
+
+	up(&ctx->file->mutex);
+	ib_ucm_ctx_put(ctx); /* func reference */
+	return result;
+}
+
+static int ib_ucm_alloc_data(const void **dest, u64 src, u32 len)
+{
+	void *data;
+
+	*dest = NULL;
+
+	if (!len)
+		return 0;
+
+	data = kmalloc(len, GFP_KERNEL);
+	if (!data)
+		return -ENOMEM;
+
+	if (copy_from_user(data, (void __user *)(unsigned long)src, len)) {
+		kfree(data);
+		return -EFAULT;
+	}
+
+	*dest = data;
+	return 0;
+}
+
+static int ib_ucm_path_get(struct ib_sa_path_rec **path, u64 src)
+{
+	struct ib_ucm_path_rec ucm_path;
+	struct ib_sa_path_rec  *sa_path;
+
+	*path = NULL;
+
+	if (!src)
+		return 0;
+
+	sa_path = kmalloc(sizeof(*sa_path), GFP_KERNEL);
+	if (!sa_path)
+		return -ENOMEM;
+
+	if (copy_from_user(&ucm_path, (void __user *)(unsigned long)src,
+			   sizeof(ucm_path))) {
+
+		kfree(sa_path);
+		return -EFAULT;
+	}
+
+	memcpy(sa_path->dgid.raw, ucm_path.dgid, sizeof(union ib_gid));
+	memcpy(sa_path->sgid.raw, ucm_path.sgid, sizeof(union ib_gid));
+
+	sa_path->dlid	          = ucm_path.dlid;
+	sa_path->slid	          = ucm_path.slid;
+	sa_path->raw_traffic      = ucm_path.raw_traffic;
+	sa_path->flow_label       = ucm_path.flow_label;
+	sa_path->hop_limit        = ucm_path.hop_limit;
+	sa_path->traffic_class    = ucm_path.traffic_class;
+	sa_path->reversible       = ucm_path.reversible;
+	sa_path->numb_path        = ucm_path.numb_path;
+	sa_path->pkey             = ucm_path.pkey;
+	sa_path->sl               = ucm_path.sl;
+	sa_path->mtu_selector     = ucm_path.mtu_selector;
+	sa_path->mtu              = ucm_path.mtu;
+	sa_path->rate_selector    = ucm_path.rate_selector;
+	sa_path->rate             = ucm_path.rate;
+	sa_path->packet_life_time = ucm_path.packet_life_time;
+	sa_path->preference       = ucm_path.preference;
+
+	sa_path->packet_life_time_selector =
+		ucm_path.packet_life_time_selector;
+
+	*path = sa_path;
+	return 0;
+}
+
+static ssize_t ib_ucm_send_req(struct ib_ucm_file *file,
+			       const char __user *inbuf,
+			       int in_len, int out_len)
+{
+	struct ib_cm_req_param param;
+	struct ib_ucm_context *ctx;
+	struct ib_ucm_req cmd;
+	int result;
+
+	param.private_data   = NULL;
+	param.primary_path   = NULL;
+	param.alternate_path = NULL;
+
+	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+		return -EFAULT;
+
+	result = ib_ucm_alloc_data(&param.private_data, cmd.data, cmd.len);
+	if (result)
+		goto done;
+
+	result = ib_ucm_path_get(&param.primary_path, cmd.primary_path);
+	if (result)
+		goto done;
+
+	result = ib_ucm_path_get(&param.alternate_path, cmd.alternate_path);
+	if (result)
+		goto done;
+
+	param.private_data_len           = cmd.len;
+	param.service_id                 = cmd.sid;
+	param.qp_num                     = cmd.qpn;
+	param.qp_type                    = cmd.qp_type;
+	param.starting_psn               = cmd.psn;
+	param.peer_to_peer               = cmd.peer_to_peer;
+	param.responder_resources        = cmd.responder_resources;
+	param.initiator_depth            = cmd.initiator_depth;
+	param.remote_cm_response_timeout = cmd.remote_cm_response_timeout;
+	param.flow_control               = cmd.flow_control;
+	param.local_cm_response_timeout  = cmd.local_cm_response_timeout;
+	param.retry_count                = cmd.retry_count;
+	param.rnr_retry_count            = cmd.rnr_retry_count;
+	param.max_cm_retries             = cmd.max_cm_retries;
+	param.srq                        = cmd.srq;
+
+	ctx = ib_ucm_ctx_get(cmd.id);
+	if (!ctx) {
+		result = -ENOENT;
+		goto done;
+	}
+
+ 	down(&ctx->file->mutex);
+	if (ctx->file != file)
+		result = -EINVAL;
+	else
+		result = ib_send_cm_req(ctx->cm_id, &param);
+
+	up(&ctx->file->mutex);
+	ib_ucm_ctx_put(ctx); /* func reference */
+done:
+	kfree(param.private_data);
+	kfree(param.primary_path);
+	kfree(param.alternate_path);
+
+	return result;
+}
+
+static ssize_t ib_ucm_send_rep(struct ib_ucm_file *file,
+			       const char __user *inbuf,
+			       int in_len, int out_len)
+{
+	struct ib_cm_rep_param param;
+	struct ib_ucm_context *ctx;
+	struct ib_ucm_rep cmd;
+	int result;
+
+	param.private_data = NULL;
+
+	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+		return -EFAULT;
+
+	result = ib_ucm_alloc_data(&param.private_data, cmd.data, cmd.len);
+	if (result)
+		return result;
+
+	param.qp_num              = cmd.qpn;
+	param.starting_psn        = cmd.psn;
+	param.private_data_len    = cmd.len;
+	param.responder_resources = cmd.responder_resources;
+	param.initiator_depth     = cmd.initiator_depth;
+	param.target_ack_delay    = cmd.target_ack_delay;
+	param.failover_accepted   = cmd.failover_accepted;
+	param.flow_control        = cmd.flow_control;
+	param.rnr_retry_count     = cmd.rnr_retry_count;
+	param.srq                 = cmd.srq;
+
+	ctx = ib_ucm_ctx_get(cmd.id);
+	if (!ctx) {
+		result = -ENOENT;
+		goto done;
+	}
+
+ 	down(&ctx->file->mutex);
+	if (ctx->file != file)
+		result = -EINVAL;
+	else
+		result = ib_send_cm_rep(ctx->cm_id, &param);
+
+	up(&ctx->file->mutex);
+	ib_ucm_ctx_put(ctx); /* func reference */
+done:
+	kfree(param.private_data);
+
+	return result;
+}
+
+static ssize_t ib_ucm_send_private_data(struct ib_ucm_file *file,
+					const char __user *inbuf, int in_len,
+					int (*func)(struct ib_cm_id *cm_id,
+						    const void *private_data,
+						    u8 private_data_len))
+{
+	struct ib_ucm_private_data cmd;
+	struct ib_ucm_context *ctx;
+	const void *private_data = NULL;
+	int result;
+
+	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+		return -EFAULT;
+
+	result = ib_ucm_alloc_data(&private_data, cmd.data, cmd.len);
+	if (result)
+		return result;
+
+	ctx = ib_ucm_ctx_get(cmd.id);
+	if (!ctx) {
+		result = -ENOENT;
+		goto done;
+	}
+
+ 	down(&ctx->file->mutex);
+	if (ctx->file != file)
+		result = -EINVAL;
+	else
+		result = func(ctx->cm_id, private_data, cmd.len);
+
+	up(&ctx->file->mutex);
+	ib_ucm_ctx_put(ctx); /* func reference */
+done:
+	kfree(private_data);
+
+	return result;
+}
+
+static ssize_t ib_ucm_send_rtu(struct ib_ucm_file *file,
+			       const char __user *inbuf,
+			       int in_len, int out_len)
+{
+	return ib_ucm_send_private_data(file, inbuf, in_len, ib_send_cm_rtu);
+}
+
+static ssize_t ib_ucm_send_dreq(struct ib_ucm_file *file,
+				const char __user *inbuf,
+				int in_len, int out_len)
+{
+	return ib_ucm_send_private_data(file, inbuf, in_len, ib_send_cm_dreq);
+}
+
+static ssize_t ib_ucm_send_drep(struct ib_ucm_file *file,
+				const char __user *inbuf,
+				int in_len, int out_len)
+{
+	return ib_ucm_send_private_data(file, inbuf, in_len, ib_send_cm_drep);
+}
+
+static ssize_t ib_ucm_send_info(struct ib_ucm_file *file,
+				const char __user *inbuf, int in_len,
+				int (*func)(struct ib_cm_id *cm_id,
+					    int status,
+					    const void *info,
+					    u8 info_len,
+					    const void *data,
+					    u8 data_len))
+{
+	struct ib_ucm_context *ctx;
+	struct ib_ucm_info cmd;
+	const void *data = NULL;
+	const void *info = NULL;
+	int result;
+
+	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+		return -EFAULT;
+
+	result = ib_ucm_alloc_data(&data, cmd.data, cmd.data_len);
+	if (result)
+		goto done;
+
+	result = ib_ucm_alloc_data(&info, cmd.info, cmd.info_len);
+	if (result)
+		goto done;
+
+	ctx = ib_ucm_ctx_get(cmd.id);
+	if (!ctx) {
+		result = -ENOENT;
+		goto done;
+	}
+
+ 	down(&ctx->file->mutex);
+	if (ctx->file != file)
+		result = -EINVAL;
+	else
+		result = func(ctx->cm_id, cmd.status,
+			      info, cmd.info_len,
+			      data, cmd.data_len);
+
+	up(&ctx->file->mutex);
+	ib_ucm_ctx_put(ctx); /* func reference */
+done:
+	kfree(data);
+	kfree(info);
+
+	return result;
+}
+
+static ssize_t ib_ucm_send_rej(struct ib_ucm_file *file,
+			       const char __user *inbuf,
+			       int in_len, int out_len)
+{
+	return ib_ucm_send_info(file, inbuf, in_len, (void *)ib_send_cm_rej);
+}
+
+static ssize_t ib_ucm_send_apr(struct ib_ucm_file *file,
+			       const char __user *inbuf,
+			       int in_len, int out_len)
+{
+	return ib_ucm_send_info(file, inbuf, in_len, (void *)ib_send_cm_apr);
+}
+
+static ssize_t ib_ucm_send_mra(struct ib_ucm_file *file,
+			       const char __user *inbuf,
+			       int in_len, int out_len)
+{
+	struct ib_ucm_context *ctx;
+	struct ib_ucm_mra cmd;
+	const void *data = NULL;
+	int result;
+
+	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+		return -EFAULT;
+
+	result = ib_ucm_alloc_data(&data, cmd.data, cmd.len);
+	if (result)
+		return result;
+
+	ctx = ib_ucm_ctx_get(cmd.id);
+	if (!ctx) {
+		result = -ENOENT;
+		goto done;
+	}
+
+ 	down(&ctx->file->mutex);
+	if (ctx->file != file)
+		result = -EINVAL;
+	else
+		result = ib_send_cm_mra(ctx->cm_id, cmd.timeout,
+					data, cmd.len);
+
+	up(&ctx->file->mutex);
+	ib_ucm_ctx_put(ctx); /* func reference */
+done:
+	kfree(data);
+
+	return result;
+}
+
+static ssize_t ib_ucm_send_lap(struct ib_ucm_file *file,
+			       const char __user *inbuf,
+			       int in_len, int out_len)
+{
+	struct ib_ucm_context *ctx;
+	struct ib_sa_path_rec *path = NULL;
+	struct ib_ucm_lap cmd;
+	const void *data = NULL;
+	int result;
+
+	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+		return -EFAULT;
+
+	result = ib_ucm_alloc_data(&data, cmd.data, cmd.len);
+	if (result)
+		goto done;
+
+	result = ib_ucm_path_get(&path, cmd.path);
+	if (result)
+		goto done;
+
+	ctx = ib_ucm_ctx_get(cmd.id);
+	if (!ctx) {
+		result = -ENOENT;
+		goto done;
+	}
+
+ 	down(&ctx->file->mutex);
+	if (ctx->file != file)
+		result = -EINVAL;
+	else
+		result = ib_send_cm_lap(ctx->cm_id, path, data, cmd.len);
+
+	up(&ctx->file->mutex);
+	ib_ucm_ctx_put(ctx); /* func reference */
+done:
+	kfree(data);
+	kfree(path);
+
+	return result;
+}
+
+static ssize_t ib_ucm_send_sidr_req(struct ib_ucm_file *file,
+				    const char __user *inbuf,
+				    int in_len, int out_len)
+{
+	struct ib_cm_sidr_req_param param;
+	struct ib_ucm_context *ctx;
+	struct ib_ucm_sidr_req cmd;
+	int result;
+
+	param.private_data = NULL;
+	param.path = NULL;
+
+	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+		return -EFAULT;
+
+	result = ib_ucm_alloc_data(&param.private_data, cmd.data, cmd.len);
+	if (result)
+		goto done;
+
+	result = ib_ucm_path_get(&param.path, cmd.path);
+	if (result)
+		goto done;
+
+	param.private_data_len = cmd.len;
+	param.service_id       = cmd.sid;
+	param.timeout_ms       = cmd.timeout;
+	param.max_cm_retries   = cmd.max_cm_retries;
+	param.pkey             = cmd.pkey;
+
+	ctx = ib_ucm_ctx_get(cmd.id);
+	if (!ctx) {
+		result = -ENOENT;
+		goto done;
+	}
+
+ 	down(&ctx->file->mutex);
+	if (ctx->file != file)
+		result = -EINVAL;
+	else
+		result = ib_send_cm_sidr_req(ctx->cm_id, &param);
+
+	up(&ctx->file->mutex);
+	ib_ucm_ctx_put(ctx); /* func reference */
+done:
+	kfree(param.private_data);
+	kfree(param.path);
+
+	return result;
+}
+
+static ssize_t ib_ucm_send_sidr_rep(struct ib_ucm_file *file,
+				    const char __user *inbuf,
+				    int in_len, int out_len)
+{
+	struct ib_cm_sidr_rep_param param;
+	struct ib_ucm_sidr_rep cmd;
+	struct ib_ucm_context *ctx;
+	int result;
+
+	param.info = NULL;
+
+	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+		return -EFAULT;
+
+	result = ib_ucm_alloc_data(&param.private_data,
+				   cmd.data, cmd.data_len);
+	if (result)
+		goto done;
+
+	result = ib_ucm_alloc_data(&param.info, cmd.info, cmd.info_len);
+	if (result)
+		goto done;
+
+	param.qp_num	   = cmd.qpn;
+	param.qkey	     = cmd.qkey;
+	param.status	   = cmd.status;
+	param.info_length      = cmd.info_len;
+	param.private_data_len = cmd.data_len;
+
+	ctx = ib_ucm_ctx_get(cmd.id);
+	if (!ctx) {
+		result = -ENOENT;
+		goto done;
+	}
+
+ 	down(&ctx->file->mutex);
+	if (ctx->file != file)
+		result = -EINVAL;
+	else
+		result = ib_send_cm_sidr_rep(ctx->cm_id, &param);
+
+	up(&ctx->file->mutex);
+	ib_ucm_ctx_put(ctx); /* func reference */
+done:
+	kfree(param.private_data);
+	kfree(param.info);
+
+	return result;
+}
+
+static ssize_t (*ucm_cmd_table[])(struct ib_ucm_file *file,
+				  const char __user *inbuf,
+				  int in_len, int out_len) = {
+	[IB_USER_CM_CMD_CREATE_ID]     = ib_ucm_create_id,
+	[IB_USER_CM_CMD_DESTROY_ID]    = ib_ucm_destroy_id,
+	[IB_USER_CM_CMD_ATTR_ID]       = ib_ucm_attr_id,
+	[IB_USER_CM_CMD_LISTEN]        = ib_ucm_listen,
+	[IB_USER_CM_CMD_ESTABLISH]     = ib_ucm_establish,
+	[IB_USER_CM_CMD_SEND_REQ]      = ib_ucm_send_req,
+	[IB_USER_CM_CMD_SEND_REP]      = ib_ucm_send_rep,
+	[IB_USER_CM_CMD_SEND_RTU]      = ib_ucm_send_rtu,
+	[IB_USER_CM_CMD_SEND_DREQ]     = ib_ucm_send_dreq,
+	[IB_USER_CM_CMD_SEND_DREP]     = ib_ucm_send_drep,
+	[IB_USER_CM_CMD_SEND_REJ]      = ib_ucm_send_rej,
+	[IB_USER_CM_CMD_SEND_MRA]      = ib_ucm_send_mra,
+	[IB_USER_CM_CMD_SEND_LAP]      = ib_ucm_send_lap,
+	[IB_USER_CM_CMD_SEND_APR]      = ib_ucm_send_apr,
+	[IB_USER_CM_CMD_SEND_SIDR_REQ] = ib_ucm_send_sidr_req,
+	[IB_USER_CM_CMD_SEND_SIDR_REP] = ib_ucm_send_sidr_rep,
+	[IB_USER_CM_CMD_EVENT]	       = ib_ucm_event,
+};
+
+static ssize_t ib_ucm_write(struct file *filp, const char __user *buf,
+			    size_t len, loff_t *pos)
+{
+	struct ib_ucm_file *file = filp->private_data;
+	struct ib_ucm_cmd_hdr hdr;
+	ssize_t result;
+
+	if (len < sizeof(hdr))
+		return -EINVAL;
+
+	if (copy_from_user(&hdr, buf, sizeof(hdr)))
+		return -EFAULT;
+
+	ucm_dbg("Write. cmd <%d> in <%d> out <%d> len <%Zu>\n",
+		hdr.cmd, hdr.in, hdr.out, len);
+
+	if (hdr.cmd < 0 || hdr.cmd >= ARRAY_SIZE(ucm_cmd_table))
+		return -EINVAL;
+
+	if (hdr.in + sizeof(hdr) > len)
+		return -EINVAL;
+
+	result = ucm_cmd_table[hdr.cmd](file, buf + sizeof(hdr),
+					hdr.in, hdr.out);
+	if (!result)
+		result = len;
+
+	return result;
+}
+
+static unsigned int ib_ucm_poll(struct file *filp,
+				struct poll_table_struct *wait)
+{
+	struct ib_ucm_file *file = filp->private_data;
+	unsigned int mask = 0;
+
+	poll_wait(filp, &file->poll_wait, wait);
+
+	if (!list_empty(&file->events))
+		mask = POLLIN | POLLRDNORM;
+
+	return mask;
+}
+
+static int ib_ucm_open(struct inode *inode, struct file *filp)
+{
+	struct ib_ucm_file *file;
+
+	file = kmalloc(sizeof(*file), GFP_KERNEL);
+	if (!file)
+		return -ENOMEM;
+
+	INIT_LIST_HEAD(&file->events);
+	INIT_LIST_HEAD(&file->ctxs);
+	init_waitqueue_head(&file->poll_wait);
+
+	init_MUTEX(&file->mutex);
+
+	filp->private_data = file;
+	file->filp = filp;
+
+	ucm_dbg("Created struct\n");
+
+	return 0;
+}
+
+static int ib_ucm_close(struct inode *inode, struct file *filp)
+{
+	struct ib_ucm_file *file = filp->private_data;
+	struct ib_ucm_context *ctx;
+
+	down(&file->mutex);
+
+	while (!list_empty(&file->ctxs)) {
+
+		ctx = list_entry(file->ctxs.next,
+				 struct ib_ucm_context, file_list);
+
+		up(&ctx->file->mutex);
+		ib_ucm_ctx_put(ctx); /* user reference */
+		down(&file->mutex);
+	}
+
+	up(&file->mutex);
+
+	kfree(file);
+
+	ucm_dbg("Deleted struct\n");
+	return 0;
+}
+
+static struct file_operations ib_ucm_fops = {
+	.owner 	 = THIS_MODULE,
+	.open 	 = ib_ucm_open,
+	.release = ib_ucm_close,
+	.write 	 = ib_ucm_write,
+	.poll    = ib_ucm_poll,
+};
+
+
+static struct class *ib_ucm_class;
+static struct cdev	  ib_ucm_cdev;
+
+static int __init ib_ucm_init(void)
+{
+	int result;
+
+	result = register_chrdev_region(IB_UCM_DEV, 1, "infiniband_cm");
+	if (result) {
+		ucm_dbg("Error <%d> registering dev\n", result);
+		goto err_chr;
+	}
+
+	cdev_init(&ib_ucm_cdev, &ib_ucm_fops);
+
+	result = cdev_add(&ib_ucm_cdev, IB_UCM_DEV, 1);
+	if (result) {
+ 		ucm_dbg("Error <%d> adding cdev\n", result);
+		goto err_cdev;
+	}
+
+	ib_ucm_class = class_create(THIS_MODULE, "infiniband_cm");
+	if (IS_ERR(ib_ucm_class)) {
+		result = PTR_ERR(ib_ucm_class);
+ 		ucm_dbg("Error <%d> creating class\n", result);
+		goto err_class;
+	}
+
+	class_device_create(ib_ucm_class, IB_UCM_DEV, NULL, "ucm");
+
+	idr_init(&ctx_id_table);
+	init_MUTEX(&ctx_id_mutex);
+
+	return 0;
+err_class:
+	cdev_del(&ib_ucm_cdev);
+err_cdev:
+	unregister_chrdev_region(IB_UCM_DEV, 1);
+err_chr:
+	return result;
+}
+
+static void __exit ib_ucm_cleanup(void)
+{
+	class_device_destroy(ib_ucm_class, IB_UCM_DEV);
+	class_destroy(ib_ucm_class);
+	cdev_del(&ib_ucm_cdev);
+	unregister_chrdev_region(IB_UCM_DEV, 1);
+}
+
+module_init(ib_ucm_init);
+module_exit(ib_ucm_cleanup);
diff --git a/drivers/infiniband/core/ucm.h b/drivers/infiniband/core/ucm.h
new file mode 100644
index 000000000000..6d36606151b2
--- /dev/null
+++ b/drivers/infiniband/core/ucm.h
@@ -0,0 +1,89 @@
+/*
+ * Copyright (c) 2005 Topspin Communications.  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.
+ *
+ * $Id: ucm.h 2208 2005-04-22 23:24:31Z libor $
+ */
+
+#ifndef UCM_H
+#define UCM_H
+
+#include <linux/fs.h>
+#include <linux/device.h>
+#include <linux/cdev.h>
+#include <linux/idr.h>
+
+#include <ib_cm.h>
+#include <ib_user_cm.h>
+
+#define IB_UCM_CM_ID_INVALID 0xffffffff
+
+struct ib_ucm_file {
+	struct semaphore mutex;
+	struct file *filp;
+	/*
+	 * list of pending events
+	 */
+	struct list_head  ctxs;   /* list of active connections */
+	struct list_head  events; /* list of pending events */
+	wait_queue_head_t poll_wait;
+};
+
+struct ib_ucm_context {
+	int                 id;
+	int                 ref;
+	int                 error;
+
+	struct ib_ucm_file *file;
+	struct ib_cm_id    *cm_id;
+	struct semaphore    mutex;
+
+	struct list_head    events;    /* list of pending events. */
+	struct list_head    file_list; /* member in file ctx list */
+};
+
+struct ib_ucm_event {
+	struct ib_ucm_context *ctx;
+	struct list_head file_list; /* member in file event list */
+	struct list_head ctx_list;  /* member in ctx event list */
+
+	struct ib_ucm_event_resp resp;
+	void *data;
+	void *info;
+	int data_len;
+	int info_len;
+	/*
+	 * new connection identifiers needs to be saved until
+	 * userspace can get a handle on them.
+	 */
+	struct ib_cm_id *cm_id;
+};
+
+#endif /* UCM_H */
diff --git a/drivers/infiniband/core/user_mad.c b/drivers/infiniband/core/user_mad.c
index 9d912d6877ff..2e38792df533 100644
--- a/drivers/infiniband/core/user_mad.c
+++ b/drivers/infiniband/core/user_mad.c
@@ -1,5 +1,7 @@
 /*
  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Voltaire, Inc. All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. 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
@@ -29,7 +31,7 @@
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  *
- * $Id: user_mad.c 1389 2004-12-27 22:56:47Z roland $
+ * $Id: user_mad.c 2814 2005-07-06 19:14:09Z halr $
  */
 
 #include <linux/module.h>
@@ -94,10 +96,12 @@ struct ib_umad_file {
 };
 
 struct ib_umad_packet {
-	struct ib_user_mad mad;
 	struct ib_ah      *ah;
+	struct ib_mad_send_buf *msg;
 	struct list_head   list;
+	int		   length;
 	DECLARE_PCI_UNMAP_ADDR(mapping)
+	struct ib_user_mad mad;
 };
 
 static const dev_t base_dev = MKDEV(IB_UMAD_MAJOR, IB_UMAD_MINOR_BASE);
@@ -114,10 +118,10 @@ static int queue_packet(struct ib_umad_file *file,
 	int ret = 1;
 
 	down_read(&file->agent_mutex);
-	for (packet->mad.id = 0;
-	     packet->mad.id < IB_UMAD_MAX_AGENTS;
-	     packet->mad.id++)
-		if (agent == file->agent[packet->mad.id]) {
+	for (packet->mad.hdr.id = 0;
+	     packet->mad.hdr.id < IB_UMAD_MAX_AGENTS;
+	     packet->mad.hdr.id++)
+		if (agent == file->agent[packet->mad.hdr.id]) {
 			spin_lock_irq(&file->recv_lock);
 			list_add_tail(&packet->list, &file->recv_list);
 			spin_unlock_irq(&file->recv_lock);
@@ -135,22 +139,30 @@ static void send_handler(struct ib_mad_agent *agent,
 			 struct ib_mad_send_wc *send_wc)
 {
 	struct ib_umad_file *file = agent->context;
-	struct ib_umad_packet *packet =
+	struct ib_umad_packet *timeout, *packet =
 		(void *) (unsigned long) send_wc->wr_id;
 
-	dma_unmap_single(agent->device->dma_device,
-			 pci_unmap_addr(packet, mapping),
-			 sizeof packet->mad.data,
-			 DMA_TO_DEVICE);
-	ib_destroy_ah(packet->ah);
+	ib_destroy_ah(packet->msg->send_wr.wr.ud.ah);
+	ib_free_send_mad(packet->msg);
 
 	if (send_wc->status == IB_WC_RESP_TIMEOUT_ERR) {
-		packet->mad.status = ETIMEDOUT;
+		timeout = kmalloc(sizeof *timeout + sizeof (struct ib_mad_hdr),
+				  GFP_KERNEL);
+		if (!timeout)
+			goto out;
 
-		if (!queue_packet(file, agent, packet))
-			return;
-	}
+		memset(timeout, 0, sizeof *timeout + sizeof (struct ib_mad_hdr));
 
+		timeout->length = sizeof (struct ib_mad_hdr);
+		timeout->mad.hdr.id = packet->mad.hdr.id;
+		timeout->mad.hdr.status = ETIMEDOUT;
+		memcpy(timeout->mad.data, packet->mad.data,
+		       sizeof (struct ib_mad_hdr));
+
+		if (!queue_packet(file, agent, timeout))
+				return;
+	}
+out:
 	kfree(packet);
 }
 
@@ -159,30 +171,35 @@ static void recv_handler(struct ib_mad_agent *agent,
 {
 	struct ib_umad_file *file = agent->context;
 	struct ib_umad_packet *packet;
+	int length;
 
 	if (mad_recv_wc->wc->status != IB_WC_SUCCESS)
 		goto out;
 
-	packet = kmalloc(sizeof *packet, GFP_KERNEL);
+	length = mad_recv_wc->mad_len;
+	packet = kmalloc(sizeof *packet + length, GFP_KERNEL);
 	if (!packet)
 		goto out;
 
-	memset(packet, 0, sizeof *packet);
+	memset(packet, 0, sizeof *packet + length);
+	packet->length = length;
+
+	ib_coalesce_recv_mad(mad_recv_wc, packet->mad.data);
 
-	memcpy(packet->mad.data, mad_recv_wc->recv_buf.mad, sizeof packet->mad.data);
-	packet->mad.status        = 0;
-	packet->mad.qpn 	  = cpu_to_be32(mad_recv_wc->wc->src_qp);
-	packet->mad.lid 	  = cpu_to_be16(mad_recv_wc->wc->slid);
-	packet->mad.sl  	  = mad_recv_wc->wc->sl;
-	packet->mad.path_bits 	  = mad_recv_wc->wc->dlid_path_bits;
-	packet->mad.grh_present   = !!(mad_recv_wc->wc->wc_flags & IB_WC_GRH);
-	if (packet->mad.grh_present) {
+	packet->mad.hdr.status    = 0;
+	packet->mad.hdr.length    = length + sizeof (struct ib_user_mad);
+	packet->mad.hdr.qpn 	  = cpu_to_be32(mad_recv_wc->wc->src_qp);
+	packet->mad.hdr.lid 	  = cpu_to_be16(mad_recv_wc->wc->slid);
+	packet->mad.hdr.sl  	  = mad_recv_wc->wc->sl;
+	packet->mad.hdr.path_bits = mad_recv_wc->wc->dlid_path_bits;
+	packet->mad.hdr.grh_present = !!(mad_recv_wc->wc->wc_flags & IB_WC_GRH);
+	if (packet->mad.hdr.grh_present) {
 		/* XXX parse GRH */
-		packet->mad.gid_index 	  = 0;
-		packet->mad.hop_limit 	  = 0;
-		packet->mad.traffic_class = 0;
-		memset(packet->mad.gid, 0, 16);
-		packet->mad.flow_label 	  = 0;
+		packet->mad.hdr.gid_index 	= 0;
+		packet->mad.hdr.hop_limit 	= 0;
+		packet->mad.hdr.traffic_class	= 0;
+		memset(packet->mad.hdr.gid, 0, 16);
+		packet->mad.hdr.flow_label	= 0;
 	}
 
 	if (queue_packet(file, agent, packet))
@@ -199,7 +216,7 @@ static ssize_t ib_umad_read(struct file *filp, char __user *buf,
 	struct ib_umad_packet *packet;
 	ssize_t ret;
 
-	if (count < sizeof (struct ib_user_mad))
+	if (count < sizeof (struct ib_user_mad) + sizeof (struct ib_mad))
 		return -EINVAL;
 
 	spin_lock_irq(&file->recv_lock);
@@ -222,12 +239,25 @@ static ssize_t ib_umad_read(struct file *filp, char __user *buf,
 
 	spin_unlock_irq(&file->recv_lock);
 
-	if (copy_to_user(buf, &packet->mad, sizeof packet->mad))
+	if (count < packet->length + sizeof (struct ib_user_mad)) {
+		/* Return length needed (and first RMPP segment) if too small */
+		if (copy_to_user(buf, &packet->mad,
+				 sizeof (struct ib_user_mad) + sizeof (struct ib_mad)))
+			ret = -EFAULT;
+		else
+			ret = -ENOSPC;
+	} else if (copy_to_user(buf, &packet->mad,
+			      packet->length + sizeof (struct ib_user_mad)))
 		ret = -EFAULT;
 	else
-		ret = sizeof packet->mad;
-
-	kfree(packet);
+		ret = packet->length + sizeof (struct ib_user_mad);
+	if (ret < 0) {
+		/* Requeue packet */
+		spin_lock_irq(&file->recv_lock);
+		list_add(&packet->list, &file->recv_list);
+		spin_unlock_irq(&file->recv_lock);
+	} else
+		kfree(packet);
 	return ret;
 }
 
@@ -238,69 +268,57 @@ static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
 	struct ib_umad_packet *packet;
 	struct ib_mad_agent *agent;
 	struct ib_ah_attr ah_attr;
-	struct ib_sge      gather_list;
-	struct ib_send_wr *bad_wr, wr = {
-		.opcode      = IB_WR_SEND,
-		.sg_list     = &gather_list,
-		.num_sge     = 1,
-		.send_flags  = IB_SEND_SIGNALED,
-	};
+	struct ib_send_wr *bad_wr;
+	struct ib_rmpp_mad *rmpp_mad;
 	u8 method;
 	u64 *tid;
-	int ret;
+	int ret, length, hdr_len, data_len, rmpp_hdr_size;
+	int rmpp_active = 0;
 
 	if (count < sizeof (struct ib_user_mad))
 		return -EINVAL;
 
-	packet = kmalloc(sizeof *packet, GFP_KERNEL);
+	length = count - sizeof (struct ib_user_mad);
+	packet = kmalloc(sizeof *packet + sizeof(struct ib_mad_hdr) +
+			 sizeof(struct ib_rmpp_hdr), GFP_KERNEL);
 	if (!packet)
 		return -ENOMEM;
 
-	if (copy_from_user(&packet->mad, buf, sizeof packet->mad)) {
-		kfree(packet);
-		return -EFAULT;
+	if (copy_from_user(&packet->mad, buf,
+			    sizeof (struct ib_user_mad) +
+			    sizeof(struct ib_mad_hdr) +
+			    sizeof(struct ib_rmpp_hdr))) {
+		ret = -EFAULT;
+		goto err;
 	}
 
-	if (packet->mad.id < 0 || packet->mad.id >= IB_UMAD_MAX_AGENTS) {
+	if (packet->mad.hdr.id < 0 ||
+	    packet->mad.hdr.id >= IB_UMAD_MAX_AGENTS) {
 		ret = -EINVAL;
 		goto err;
 	}
 
+	packet->length = length;
+
 	down_read(&file->agent_mutex);
 
-	agent = file->agent[packet->mad.id];
+	agent = file->agent[packet->mad.hdr.id];
 	if (!agent) {
 		ret = -EINVAL;
 		goto err_up;
 	}
 
-	/*
-	 * If userspace is generating a request that will generate a
-	 * response, we need to make sure the high-order part of the
-	 * transaction ID matches the agent being used to send the
-	 * MAD.
-	 */
-	method = ((struct ib_mad_hdr *) packet->mad.data)->method;
-
-	if (!(method & IB_MGMT_METHOD_RESP)       &&
-	    method != IB_MGMT_METHOD_TRAP_REPRESS &&
-	    method != IB_MGMT_METHOD_SEND) {
-		tid = &((struct ib_mad_hdr *) packet->mad.data)->tid;
-		*tid = cpu_to_be64(((u64) agent->hi_tid) << 32 |
-				   (be64_to_cpup(tid) & 0xffffffff));
-	}
-
 	memset(&ah_attr, 0, sizeof ah_attr);
-	ah_attr.dlid          = be16_to_cpu(packet->mad.lid);
-	ah_attr.sl            = packet->mad.sl;
-	ah_attr.src_path_bits = packet->mad.path_bits;
+	ah_attr.dlid          = be16_to_cpu(packet->mad.hdr.lid);
+	ah_attr.sl            = packet->mad.hdr.sl;
+	ah_attr.src_path_bits = packet->mad.hdr.path_bits;
 	ah_attr.port_num      = file->port->port_num;
-	if (packet->mad.grh_present) {
+	if (packet->mad.hdr.grh_present) {
 		ah_attr.ah_flags = IB_AH_GRH;
-		memcpy(ah_attr.grh.dgid.raw, packet->mad.gid, 16);
-		ah_attr.grh.flow_label 	   = packet->mad.flow_label;
-		ah_attr.grh.hop_limit  	   = packet->mad.hop_limit;
-		ah_attr.grh.traffic_class  = packet->mad.traffic_class;
+		memcpy(ah_attr.grh.dgid.raw, packet->mad.hdr.gid, 16);
+		ah_attr.grh.flow_label 	   = packet->mad.hdr.flow_label;
+		ah_attr.grh.hop_limit  	   = packet->mad.hdr.hop_limit;
+		ah_attr.grh.traffic_class  = packet->mad.hdr.traffic_class;
 	}
 
 	packet->ah = ib_create_ah(agent->qp->pd, &ah_attr);
@@ -309,34 +327,104 @@ static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
 		goto err_up;
 	}
 
-	gather_list.addr = dma_map_single(agent->device->dma_device,
-					  packet->mad.data,
-					  sizeof packet->mad.data,
-					  DMA_TO_DEVICE);
-	gather_list.length = sizeof packet->mad.data;
-	gather_list.lkey   = file->mr[packet->mad.id]->lkey;
-	pci_unmap_addr_set(packet, mapping, gather_list.addr);
+	rmpp_mad = (struct ib_rmpp_mad *) packet->mad.data;
+	if (ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) & IB_MGMT_RMPP_FLAG_ACTIVE) {
+		/* RMPP active */
+		if (!agent->rmpp_version) {
+			ret = -EINVAL;
+			goto err_ah;
+		}
+		/* Validate that management class can support RMPP */
+		if (rmpp_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_ADM) {
+			hdr_len = offsetof(struct ib_sa_mad, data);
+			data_len = length;
+		} else if ((rmpp_mad->mad_hdr.mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
+			    (rmpp_mad->mad_hdr.mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)) {
+				hdr_len = offsetof(struct ib_vendor_mad, data);
+				data_len = length - hdr_len;
+		} else {
+			ret = -EINVAL;
+			goto err_ah;
+		}
+		rmpp_active = 1;
+	} else {
+		if (length > sizeof(struct ib_mad)) {
+			ret = -EINVAL;
+			goto err_ah;
+		}
+		hdr_len = offsetof(struct ib_mad, data);
+		data_len = length - hdr_len;
+	}
+
+	packet->msg = ib_create_send_mad(agent,
+					 be32_to_cpu(packet->mad.hdr.qpn),
+					 0, packet->ah, rmpp_active,
+					 hdr_len, data_len,
+					 GFP_KERNEL);
+	if (IS_ERR(packet->msg)) {
+		ret = PTR_ERR(packet->msg);
+		goto err_ah;
+	}
 
-	wr.wr.ud.mad_hdr     = (struct ib_mad_hdr *) packet->mad.data;
-	wr.wr.ud.ah          = packet->ah;
-	wr.wr.ud.remote_qpn  = be32_to_cpu(packet->mad.qpn);
-	wr.wr.ud.remote_qkey = be32_to_cpu(packet->mad.qkey);
-	wr.wr.ud.timeout_ms  = packet->mad.timeout_ms;
+	packet->msg->send_wr.wr.ud.timeout_ms  = packet->mad.hdr.timeout_ms;
+	packet->msg->send_wr.wr.ud.retries = packet->mad.hdr.retries;
 
-	wr.wr_id            = (unsigned long) packet;
+	/* Override send WR WRID initialized in ib_create_send_mad */
+	packet->msg->send_wr.wr_id = (unsigned long) packet;
 
-	ret = ib_post_send_mad(agent, &wr, &bad_wr);
-	if (ret) {
-		dma_unmap_single(agent->device->dma_device,
-				 pci_unmap_addr(packet, mapping),
-				 sizeof packet->mad.data,
-				 DMA_TO_DEVICE);
-		goto err_up;
+	if (!rmpp_active) {
+		/* Copy message from user into send buffer */
+		if (copy_from_user(packet->msg->mad,
+				   buf + sizeof(struct ib_user_mad), length)) {
+			ret = -EFAULT;
+			goto err_msg;
+		}
+	} else {
+		rmpp_hdr_size = sizeof(struct ib_mad_hdr) +
+				sizeof(struct ib_rmpp_hdr);
+
+		/* Only copy MAD headers (RMPP header in place) */
+		memcpy(packet->msg->mad, packet->mad.data,
+		       sizeof(struct ib_mad_hdr));
+
+		/* Now, copy rest of message from user into send buffer */
+		if (copy_from_user(((struct ib_rmpp_mad *) packet->msg->mad)->data,
+				   buf + sizeof (struct ib_user_mad) + rmpp_hdr_size,
+				   length - rmpp_hdr_size)) {
+			ret = -EFAULT;
+			goto err_msg;
+		}
+	}
+
+	/*
+	 * If userspace is generating a request that will generate a
+	 * response, we need to make sure the high-order part of the
+	 * transaction ID matches the agent being used to send the
+	 * MAD.
+	 */
+	method = packet->msg->mad->mad_hdr.method;
+
+	if (!(method & IB_MGMT_METHOD_RESP)       &&
+	    method != IB_MGMT_METHOD_TRAP_REPRESS &&
+	    method != IB_MGMT_METHOD_SEND) {
+		tid = &packet->msg->mad->mad_hdr.tid;
+		*tid = cpu_to_be64(((u64) agent->hi_tid) << 32 |
+				   (be64_to_cpup(tid) & 0xffffffff));
 	}
 
+	ret = ib_post_send_mad(agent, &packet->msg->send_wr, &bad_wr);
+	if (ret)
+		goto err_msg;
+
 	up_read(&file->agent_mutex);
 
-	return sizeof packet->mad;
+	return sizeof (struct ib_user_mad_hdr) + packet->length;
+
+err_msg:
+	ib_free_send_mad(packet->msg);
+
+err_ah:
+	ib_destroy_ah(packet->ah);
 
 err_up:
 	up_read(&file->agent_mutex);
@@ -399,7 +487,8 @@ found:
 	agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
 				      ureq.qpn ? IB_QPT_GSI : IB_QPT_SMI,
 				      ureq.mgmt_class ? &req : NULL,
-				      0, send_handler, recv_handler, file);
+				      ureq.rmpp_version,
+				      send_handler, recv_handler, file);
 	if (IS_ERR(agent)) {
 		ret = PTR_ERR(agent);
 		goto out;
@@ -460,8 +549,8 @@ out:
 	return ret;
 }
 
-static long ib_umad_ioctl(struct file *filp,
-			 unsigned int cmd, unsigned long arg)
+static long ib_umad_ioctl(struct file *filp, unsigned int cmd,
+			  unsigned long arg)
 {
 	switch (cmd) {
 	case IB_USER_MAD_REGISTER_AGENT:
@@ -517,14 +606,14 @@ static int ib_umad_close(struct inode *inode, struct file *filp)
 }
 
 static struct file_operations umad_fops = {
-	.owner 	        = THIS_MODULE,
-	.read 	        = ib_umad_read,
-	.write 	        = ib_umad_write,
-	.poll 	        = ib_umad_poll,
+	.owner 	 	= THIS_MODULE,
+	.read 	 	= ib_umad_read,
+	.write 	 	= ib_umad_write,
+	.poll 	 	= ib_umad_poll,
 	.unlocked_ioctl = ib_umad_ioctl,
-	.compat_ioctl   = ib_umad_ioctl,
-	.open 	        = ib_umad_open,
-	.release        = ib_umad_close
+	.compat_ioctl 	= ib_umad_ioctl,
+	.open 	 	= ib_umad_open,
+	.release 	= ib_umad_close
 };
 
 static int ib_umad_sm_open(struct inode *inode, struct file *filp)
diff --git a/drivers/infiniband/core/uverbs.h b/drivers/infiniband/core/uverbs.h
new file mode 100644
index 000000000000..7696022f9a4e
--- /dev/null
+++ b/drivers/infiniband/core/uverbs.h
@@ -0,0 +1,133 @@
+/*
+ * Copyright (c) 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  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.
+ *
+ * $Id: uverbs.h 2559 2005-06-06 19:43:16Z roland $
+ */
+
+#ifndef UVERBS_H
+#define UVERBS_H
+
+/* Include device.h and fs.h until cdev.h is self-sufficient */
+#include <linux/fs.h>
+#include <linux/device.h>
+#include <linux/cdev.h>
+#include <linux/kref.h>
+#include <linux/idr.h>
+
+#include <ib_verbs.h>
+#include <ib_user_verbs.h>
+
+struct ib_uverbs_device {
+	int					devnum;
+	struct cdev				dev;
+	struct class_device			class_dev;
+	struct ib_device		       *ib_dev;
+	int					num_comp;
+};
+
+struct ib_uverbs_event_file {
+	struct kref				ref;
+	struct ib_uverbs_file		       *uverbs_file;
+	spinlock_t				lock;
+	int					fd;
+	int					is_async;
+	wait_queue_head_t			poll_wait;
+	struct fasync_struct		       *async_queue;
+	struct list_head			event_list;
+};
+
+struct ib_uverbs_file {
+	struct kref				ref;
+	struct ib_uverbs_device		       *device;
+	struct ib_ucontext		       *ucontext;
+	struct ib_event_handler			event_handler;
+	struct ib_uverbs_event_file	        async_file;
+	struct ib_uverbs_event_file	        comp_file[1];
+};
+
+struct ib_uverbs_async_event {
+	struct ib_uverbs_async_event_desc	desc;
+	struct list_head			list;
+};
+
+struct ib_uverbs_comp_event {
+	struct ib_uverbs_comp_event_desc	desc;
+	struct list_head			list;
+};
+
+struct ib_uobject_mr {
+	struct ib_uobject			uobj;
+	struct page			       *page_list;
+	struct scatterlist		       *sg_list;
+};
+
+extern struct semaphore ib_uverbs_idr_mutex;
+extern struct idr ib_uverbs_pd_idr;
+extern struct idr ib_uverbs_mr_idr;
+extern struct idr ib_uverbs_mw_idr;
+extern struct idr ib_uverbs_ah_idr;
+extern struct idr ib_uverbs_cq_idr;
+extern struct idr ib_uverbs_qp_idr;
+
+void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context);
+void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr);
+void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr);
+
+int ib_umem_get(struct ib_device *dev, struct ib_umem *mem,
+		void *addr, size_t size, int write);
+void ib_umem_release(struct ib_device *dev, struct ib_umem *umem);
+void ib_umem_release_on_close(struct ib_device *dev, struct ib_umem *umem);
+
+#define IB_UVERBS_DECLARE_CMD(name)					\
+	ssize_t ib_uverbs_##name(struct ib_uverbs_file *file,		\
+				 const char __user *buf, int in_len,	\
+				 int out_len)
+
+IB_UVERBS_DECLARE_CMD(query_params);
+IB_UVERBS_DECLARE_CMD(get_context);
+IB_UVERBS_DECLARE_CMD(query_device);
+IB_UVERBS_DECLARE_CMD(query_port);
+IB_UVERBS_DECLARE_CMD(query_gid);
+IB_UVERBS_DECLARE_CMD(query_pkey);
+IB_UVERBS_DECLARE_CMD(alloc_pd);
+IB_UVERBS_DECLARE_CMD(dealloc_pd);
+IB_UVERBS_DECLARE_CMD(reg_mr);
+IB_UVERBS_DECLARE_CMD(dereg_mr);
+IB_UVERBS_DECLARE_CMD(create_cq);
+IB_UVERBS_DECLARE_CMD(destroy_cq);
+IB_UVERBS_DECLARE_CMD(create_qp);
+IB_UVERBS_DECLARE_CMD(modify_qp);
+IB_UVERBS_DECLARE_CMD(destroy_qp);
+IB_UVERBS_DECLARE_CMD(attach_mcast);
+IB_UVERBS_DECLARE_CMD(detach_mcast);
+
+#endif /* UVERBS_H */
diff --git a/drivers/infiniband/core/uverbs_cmd.c b/drivers/infiniband/core/uverbs_cmd.c
new file mode 100644
index 000000000000..5f2bbcda4c73
--- /dev/null
+++ b/drivers/infiniband/core/uverbs_cmd.c
@@ -0,0 +1,1006 @@
+/*
+ * Copyright (c) 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  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.
+ *
+ * $Id: uverbs_cmd.c 2708 2005-06-24 17:27:21Z roland $
+ */
+
+#include <asm/uaccess.h>
+
+#include "uverbs.h"
+
+#define INIT_UDATA(udata, ibuf, obuf, ilen, olen)			\
+	do {								\
+		(udata)->inbuf  = (void __user *) (ibuf);		\
+		(udata)->outbuf = (void __user *) (obuf);		\
+		(udata)->inlen  = (ilen);				\
+		(udata)->outlen = (olen);				\
+	} while (0)
+
+ssize_t ib_uverbs_query_params(struct ib_uverbs_file *file,
+			       const char __user *buf,
+			       int in_len, int out_len)
+{
+	struct ib_uverbs_query_params      cmd;
+	struct ib_uverbs_query_params_resp resp;
+
+	if (out_len < sizeof resp)
+		return -ENOSPC;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	memset(&resp, 0, sizeof resp);
+
+	resp.num_cq_events = file->device->num_comp;
+
+	if (copy_to_user((void __user *) (unsigned long) cmd.response, &resp, sizeof resp))
+	    return -EFAULT;
+
+	return in_len;
+}
+
+ssize_t ib_uverbs_get_context(struct ib_uverbs_file *file,
+			      const char __user *buf,
+			      int in_len, int out_len)
+{
+	struct ib_uverbs_get_context      cmd;
+	struct ib_uverbs_get_context_resp resp;
+	struct ib_udata                   udata;
+	struct ib_device                 *ibdev = file->device->ib_dev;
+	int i;
+	int ret = in_len;
+
+	if (out_len < sizeof resp)
+		return -ENOSPC;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	INIT_UDATA(&udata, buf + sizeof cmd,
+		   (unsigned long) cmd.response + sizeof resp,
+		   in_len - sizeof cmd, out_len - sizeof resp);
+
+	file->ucontext = ibdev->alloc_ucontext(ibdev, &udata);
+	if (IS_ERR(file->ucontext)) {
+		ret = PTR_ERR(file->ucontext);
+		file->ucontext = NULL;
+		return ret;
+	}
+
+	file->ucontext->device = ibdev;
+	INIT_LIST_HEAD(&file->ucontext->pd_list);
+	INIT_LIST_HEAD(&file->ucontext->mr_list);
+	INIT_LIST_HEAD(&file->ucontext->mw_list);
+	INIT_LIST_HEAD(&file->ucontext->cq_list);
+	INIT_LIST_HEAD(&file->ucontext->qp_list);
+	INIT_LIST_HEAD(&file->ucontext->srq_list);
+	INIT_LIST_HEAD(&file->ucontext->ah_list);
+	spin_lock_init(&file->ucontext->lock);
+
+	resp.async_fd = file->async_file.fd;
+	for (i = 0; i < file->device->num_comp; ++i)
+		if (copy_to_user((void __user *) (unsigned long) cmd.cq_fd_tab +
+				 i * sizeof (__u32),
+				 &file->comp_file[i].fd, sizeof (__u32)))
+			goto err;
+
+	if (copy_to_user((void __user *) (unsigned long) cmd.response,
+			 &resp, sizeof resp))
+		goto err;
+
+	return in_len;
+
+err:
+	ibdev->dealloc_ucontext(file->ucontext);
+	file->ucontext = NULL;
+
+	return -EFAULT;
+}
+
+ssize_t ib_uverbs_query_device(struct ib_uverbs_file *file,
+			       const char __user *buf,
+			       int in_len, int out_len)
+{
+	struct ib_uverbs_query_device      cmd;
+	struct ib_uverbs_query_device_resp resp;
+	struct ib_device_attr              attr;
+	int                                ret;
+
+	if (out_len < sizeof resp)
+		return -ENOSPC;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	ret = ib_query_device(file->device->ib_dev, &attr);
+	if (ret)
+		return ret;
+
+	memset(&resp, 0, sizeof resp);
+
+	resp.fw_ver 		       = attr.fw_ver;
+	resp.node_guid 		       = attr.node_guid;
+	resp.sys_image_guid 	       = attr.sys_image_guid;
+	resp.max_mr_size 	       = attr.max_mr_size;
+	resp.page_size_cap 	       = attr.page_size_cap;
+	resp.vendor_id 		       = attr.vendor_id;
+	resp.vendor_part_id 	       = attr.vendor_part_id;
+	resp.hw_ver 		       = attr.hw_ver;
+	resp.max_qp 		       = attr.max_qp;
+	resp.max_qp_wr 		       = attr.max_qp_wr;
+	resp.device_cap_flags 	       = attr.device_cap_flags;
+	resp.max_sge 		       = attr.max_sge;
+	resp.max_sge_rd 	       = attr.max_sge_rd;
+	resp.max_cq 		       = attr.max_cq;
+	resp.max_cqe 		       = attr.max_cqe;
+	resp.max_mr 		       = attr.max_mr;
+	resp.max_pd 		       = attr.max_pd;
+	resp.max_qp_rd_atom 	       = attr.max_qp_rd_atom;
+	resp.max_ee_rd_atom 	       = attr.max_ee_rd_atom;
+	resp.max_res_rd_atom 	       = attr.max_res_rd_atom;
+	resp.max_qp_init_rd_atom       = attr.max_qp_init_rd_atom;
+	resp.max_ee_init_rd_atom       = attr.max_ee_init_rd_atom;
+	resp.atomic_cap 	       = attr.atomic_cap;
+	resp.max_ee 		       = attr.max_ee;
+	resp.max_rdd 		       = attr.max_rdd;
+	resp.max_mw 		       = attr.max_mw;
+	resp.max_raw_ipv6_qp 	       = attr.max_raw_ipv6_qp;
+	resp.max_raw_ethy_qp 	       = attr.max_raw_ethy_qp;
+	resp.max_mcast_grp 	       = attr.max_mcast_grp;
+	resp.max_mcast_qp_attach       = attr.max_mcast_qp_attach;
+	resp.max_total_mcast_qp_attach = attr.max_total_mcast_qp_attach;
+	resp.max_ah 		       = attr.max_ah;
+	resp.max_fmr 		       = attr.max_fmr;
+	resp.max_map_per_fmr 	       = attr.max_map_per_fmr;
+	resp.max_srq 		       = attr.max_srq;
+	resp.max_srq_wr 	       = attr.max_srq_wr;
+	resp.max_srq_sge 	       = attr.max_srq_sge;
+	resp.max_pkeys 		       = attr.max_pkeys;
+	resp.local_ca_ack_delay        = attr.local_ca_ack_delay;
+	resp.phys_port_cnt	       = file->device->ib_dev->phys_port_cnt;
+
+	if (copy_to_user((void __user *) (unsigned long) cmd.response,
+			 &resp, sizeof resp))
+		return -EFAULT;
+
+	return in_len;
+}
+
+ssize_t ib_uverbs_query_port(struct ib_uverbs_file *file,
+			     const char __user *buf,
+			     int in_len, int out_len)
+{
+	struct ib_uverbs_query_port      cmd;
+	struct ib_uverbs_query_port_resp resp;
+	struct ib_port_attr              attr;
+	int                              ret;
+
+	if (out_len < sizeof resp)
+		return -ENOSPC;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	ret = ib_query_port(file->device->ib_dev, cmd.port_num, &attr);
+	if (ret)
+		return ret;
+
+	memset(&resp, 0, sizeof resp);
+
+	resp.state 	     = attr.state;
+	resp.max_mtu 	     = attr.max_mtu;
+	resp.active_mtu      = attr.active_mtu;
+	resp.gid_tbl_len     = attr.gid_tbl_len;
+	resp.port_cap_flags  = attr.port_cap_flags;
+	resp.max_msg_sz      = attr.max_msg_sz;
+	resp.bad_pkey_cntr   = attr.bad_pkey_cntr;
+	resp.qkey_viol_cntr  = attr.qkey_viol_cntr;
+	resp.pkey_tbl_len    = attr.pkey_tbl_len;
+	resp.lid 	     = attr.lid;
+	resp.sm_lid 	     = attr.sm_lid;
+	resp.lmc 	     = attr.lmc;
+	resp.max_vl_num      = attr.max_vl_num;
+	resp.sm_sl 	     = attr.sm_sl;
+	resp.subnet_timeout  = attr.subnet_timeout;
+	resp.init_type_reply = attr.init_type_reply;
+	resp.active_width    = attr.active_width;
+	resp.active_speed    = attr.active_speed;
+	resp.phys_state      = attr.phys_state;
+
+	if (copy_to_user((void __user *) (unsigned long) cmd.response,
+			 &resp, sizeof resp))
+		return -EFAULT;
+
+	return in_len;
+}
+
+ssize_t ib_uverbs_query_gid(struct ib_uverbs_file *file,
+			    const char __user *buf,
+			    int in_len, int out_len)
+{
+	struct ib_uverbs_query_gid      cmd;
+	struct ib_uverbs_query_gid_resp resp;
+	int                             ret;
+
+	if (out_len < sizeof resp)
+		return -ENOSPC;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	memset(&resp, 0, sizeof resp);
+
+	ret = ib_query_gid(file->device->ib_dev, cmd.port_num, cmd.index,
+			   (union ib_gid *) resp.gid);
+	if (ret)
+		return ret;
+
+	if (copy_to_user((void __user *) (unsigned long) cmd.response,
+			 &resp, sizeof resp))
+		return -EFAULT;
+
+	return in_len;
+}
+
+ssize_t ib_uverbs_query_pkey(struct ib_uverbs_file *file,
+			     const char __user *buf,
+			     int in_len, int out_len)
+{
+	struct ib_uverbs_query_pkey      cmd;
+	struct ib_uverbs_query_pkey_resp resp;
+	int                              ret;
+
+	if (out_len < sizeof resp)
+		return -ENOSPC;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	memset(&resp, 0, sizeof resp);
+
+	ret = ib_query_pkey(file->device->ib_dev, cmd.port_num, cmd.index,
+			    &resp.pkey);
+	if (ret)
+		return ret;
+
+	if (copy_to_user((void __user *) (unsigned long) cmd.response,
+			 &resp, sizeof resp))
+		return -EFAULT;
+
+	return in_len;
+}
+
+ssize_t ib_uverbs_alloc_pd(struct ib_uverbs_file *file,
+			   const char __user *buf,
+			   int in_len, int out_len)
+{
+	struct ib_uverbs_alloc_pd      cmd;
+	struct ib_uverbs_alloc_pd_resp resp;
+	struct ib_udata                udata;
+	struct ib_uobject             *uobj;
+	struct ib_pd                  *pd;
+	int                            ret;
+
+	if (out_len < sizeof resp)
+		return -ENOSPC;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	INIT_UDATA(&udata, buf + sizeof cmd,
+		   (unsigned long) cmd.response + sizeof resp,
+		   in_len - sizeof cmd, out_len - sizeof resp);
+
+	uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
+	if (!uobj)
+		return -ENOMEM;
+
+	uobj->context = file->ucontext;
+
+	pd = file->device->ib_dev->alloc_pd(file->device->ib_dev,
+					    file->ucontext, &udata);
+	if (IS_ERR(pd)) {
+		ret = PTR_ERR(pd);
+		goto err;
+	}
+
+	pd->device  = file->device->ib_dev;
+	pd->uobject = uobj;
+	atomic_set(&pd->usecnt, 0);
+
+retry:
+	if (!idr_pre_get(&ib_uverbs_pd_idr, GFP_KERNEL)) {
+		ret = -ENOMEM;
+		goto err_pd;
+	}
+
+	down(&ib_uverbs_idr_mutex);
+	ret = idr_get_new(&ib_uverbs_pd_idr, pd, &uobj->id);
+	up(&ib_uverbs_idr_mutex);
+
+	if (ret == -EAGAIN)
+		goto retry;
+	if (ret)
+		goto err_pd;
+
+	spin_lock_irq(&file->ucontext->lock);
+	list_add_tail(&uobj->list, &file->ucontext->pd_list);
+	spin_unlock_irq(&file->ucontext->lock);
+
+	memset(&resp, 0, sizeof resp);
+	resp.pd_handle = uobj->id;
+
+	if (copy_to_user((void __user *) (unsigned long) cmd.response,
+			 &resp, sizeof resp)) {
+		ret = -EFAULT;
+		goto err_list;
+	}
+
+	return in_len;
+
+err_list:
+ 	spin_lock_irq(&file->ucontext->lock);
+	list_del(&uobj->list);
+	spin_unlock_irq(&file->ucontext->lock);
+
+	down(&ib_uverbs_idr_mutex);
+	idr_remove(&ib_uverbs_pd_idr, uobj->id);
+	up(&ib_uverbs_idr_mutex);
+
+err_pd:
+	ib_dealloc_pd(pd);
+
+err:
+	kfree(uobj);
+	return ret;
+}
+
+ssize_t ib_uverbs_dealloc_pd(struct ib_uverbs_file *file,
+			     const char __user *buf,
+			     int in_len, int out_len)
+{
+	struct ib_uverbs_dealloc_pd cmd;
+	struct ib_pd               *pd;
+	struct ib_uobject          *uobj;
+	int                         ret = -EINVAL;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	down(&ib_uverbs_idr_mutex);
+
+	pd = idr_find(&ib_uverbs_pd_idr, cmd.pd_handle);
+	if (!pd || pd->uobject->context != file->ucontext)
+		goto out;
+
+	uobj = pd->uobject;
+
+	ret = ib_dealloc_pd(pd);
+	if (ret)
+		goto out;
+
+	idr_remove(&ib_uverbs_pd_idr, cmd.pd_handle);
+
+	spin_lock_irq(&file->ucontext->lock);
+	list_del(&uobj->list);
+	spin_unlock_irq(&file->ucontext->lock);
+
+	kfree(uobj);
+
+out:
+	up(&ib_uverbs_idr_mutex);
+
+	return ret ? ret : in_len;
+}
+
+ssize_t ib_uverbs_reg_mr(struct ib_uverbs_file *file,
+			 const char __user *buf, int in_len,
+			 int out_len)
+{
+	struct ib_uverbs_reg_mr      cmd;
+	struct ib_uverbs_reg_mr_resp resp;
+	struct ib_udata              udata;
+	struct ib_umem_object       *obj;
+	struct ib_pd                *pd;
+	struct ib_mr                *mr;
+	int                          ret;
+
+	if (out_len < sizeof resp)
+		return -ENOSPC;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	INIT_UDATA(&udata, buf + sizeof cmd,
+		   (unsigned long) cmd.response + sizeof resp,
+		   in_len - sizeof cmd, out_len - sizeof resp);
+
+	if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
+		return -EINVAL;
+
+	obj = kmalloc(sizeof *obj, GFP_KERNEL);
+	if (!obj)
+		return -ENOMEM;
+
+	obj->uobject.context = file->ucontext;
+
+	/*
+	 * We ask for writable memory if any access flags other than
+	 * "remote read" are set.  "Local write" and "remote write"
+	 * obviously require write access.  "Remote atomic" can do
+	 * things like fetch and add, which will modify memory, and
+	 * "MW bind" can change permissions by binding a window.
+	 */
+	ret = ib_umem_get(file->device->ib_dev, &obj->umem,
+			  (void *) (unsigned long) cmd.start, cmd.length,
+			  !!(cmd.access_flags & ~IB_ACCESS_REMOTE_READ));
+	if (ret)
+		goto err_free;
+
+	obj->umem.virt_base = cmd.hca_va;
+
+	down(&ib_uverbs_idr_mutex);
+
+	pd = idr_find(&ib_uverbs_pd_idr, cmd.pd_handle);
+	if (!pd || pd->uobject->context != file->ucontext) {
+		ret = -EINVAL;
+		goto err_up;
+	}
+
+	if (!pd->device->reg_user_mr) {
+		ret = -ENOSYS;
+		goto err_up;
+	}
+
+	mr = pd->device->reg_user_mr(pd, &obj->umem, cmd.access_flags, &udata);
+	if (IS_ERR(mr)) {
+		ret = PTR_ERR(mr);
+		goto err_up;
+	}
+
+	mr->device  = pd->device;
+	mr->pd      = pd;
+	mr->uobject = &obj->uobject;
+	atomic_inc(&pd->usecnt);
+	atomic_set(&mr->usecnt, 0);
+
+	memset(&resp, 0, sizeof resp);
+	resp.lkey = mr->lkey;
+	resp.rkey = mr->rkey;
+
+retry:
+	if (!idr_pre_get(&ib_uverbs_mr_idr, GFP_KERNEL)) {
+		ret = -ENOMEM;
+		goto err_unreg;
+	}
+
+	ret = idr_get_new(&ib_uverbs_mr_idr, mr, &obj->uobject.id);
+
+	if (ret == -EAGAIN)
+		goto retry;
+	if (ret)
+		goto err_unreg;
+
+	resp.mr_handle = obj->uobject.id;
+
+	spin_lock_irq(&file->ucontext->lock);
+	list_add_tail(&obj->uobject.list, &file->ucontext->mr_list);
+	spin_unlock_irq(&file->ucontext->lock);
+
+	if (copy_to_user((void __user *) (unsigned long) cmd.response,
+			 &resp, sizeof resp)) {
+		ret = -EFAULT;
+		goto err_list;
+	}
+
+	up(&ib_uverbs_idr_mutex);
+
+	return in_len;
+
+err_list:
+	spin_lock_irq(&file->ucontext->lock);
+	list_del(&obj->uobject.list);
+	spin_unlock_irq(&file->ucontext->lock);
+
+err_unreg:
+	ib_dereg_mr(mr);
+
+err_up:
+	up(&ib_uverbs_idr_mutex);
+
+	ib_umem_release(file->device->ib_dev, &obj->umem);
+
+err_free:
+	kfree(obj);
+	return ret;
+}
+
+ssize_t ib_uverbs_dereg_mr(struct ib_uverbs_file *file,
+			   const char __user *buf, int in_len,
+			   int out_len)
+{
+	struct ib_uverbs_dereg_mr cmd;
+	struct ib_mr             *mr;
+	struct ib_umem_object    *memobj;
+	int                       ret = -EINVAL;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	down(&ib_uverbs_idr_mutex);
+
+	mr = idr_find(&ib_uverbs_mr_idr, cmd.mr_handle);
+	if (!mr || mr->uobject->context != file->ucontext)
+		goto out;
+
+	memobj = container_of(mr->uobject, struct ib_umem_object, uobject);
+
+	ret = ib_dereg_mr(mr);
+	if (ret)
+		goto out;
+
+	idr_remove(&ib_uverbs_mr_idr, cmd.mr_handle);
+
+	spin_lock_irq(&file->ucontext->lock);
+	list_del(&memobj->uobject.list);
+	spin_unlock_irq(&file->ucontext->lock);
+
+	ib_umem_release(file->device->ib_dev, &memobj->umem);
+	kfree(memobj);
+
+out:
+	up(&ib_uverbs_idr_mutex);
+
+	return ret ? ret : in_len;
+}
+
+ssize_t ib_uverbs_create_cq(struct ib_uverbs_file *file,
+			    const char __user *buf, int in_len,
+			    int out_len)
+{
+	struct ib_uverbs_create_cq      cmd;
+	struct ib_uverbs_create_cq_resp resp;
+	struct ib_udata                 udata;
+	struct ib_uobject              *uobj;
+	struct ib_cq                   *cq;
+	int                             ret;
+
+	if (out_len < sizeof resp)
+		return -ENOSPC;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	INIT_UDATA(&udata, buf + sizeof cmd,
+		   (unsigned long) cmd.response + sizeof resp,
+		   in_len - sizeof cmd, out_len - sizeof resp);
+
+	if (cmd.event_handler >= file->device->num_comp)
+		return -EINVAL;
+
+	uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
+	if (!uobj)
+		return -ENOMEM;
+
+	uobj->user_handle = cmd.user_handle;
+	uobj->context     = file->ucontext;
+
+	cq = file->device->ib_dev->create_cq(file->device->ib_dev, cmd.cqe,
+					     file->ucontext, &udata);
+	if (IS_ERR(cq)) {
+		ret = PTR_ERR(cq);
+		goto err;
+	}
+
+	cq->device        = file->device->ib_dev;
+	cq->uobject       = uobj;
+	cq->comp_handler  = ib_uverbs_comp_handler;
+	cq->event_handler = ib_uverbs_cq_event_handler;
+	cq->cq_context    = file;
+	atomic_set(&cq->usecnt, 0);
+
+retry:
+	if (!idr_pre_get(&ib_uverbs_cq_idr, GFP_KERNEL)) {
+		ret = -ENOMEM;
+		goto err_cq;
+	}
+
+	down(&ib_uverbs_idr_mutex);
+	ret = idr_get_new(&ib_uverbs_cq_idr, cq, &uobj->id);
+	up(&ib_uverbs_idr_mutex);
+
+	if (ret == -EAGAIN)
+		goto retry;
+	if (ret)
+		goto err_cq;
+
+	spin_lock_irq(&file->ucontext->lock);
+	list_add_tail(&uobj->list, &file->ucontext->cq_list);
+	spin_unlock_irq(&file->ucontext->lock);
+
+	memset(&resp, 0, sizeof resp);
+	resp.cq_handle = uobj->id;
+	resp.cqe       = cq->cqe;
+
+	if (copy_to_user((void __user *) (unsigned long) cmd.response,
+			 &resp, sizeof resp)) {
+		ret = -EFAULT;
+		goto err_list;
+	}
+
+	return in_len;
+
+err_list:
+ 	spin_lock_irq(&file->ucontext->lock);
+	list_del(&uobj->list);
+	spin_unlock_irq(&file->ucontext->lock);
+
+	down(&ib_uverbs_idr_mutex);
+	idr_remove(&ib_uverbs_cq_idr, uobj->id);
+	up(&ib_uverbs_idr_mutex);
+
+err_cq:
+	ib_destroy_cq(cq);
+
+err:
+	kfree(uobj);
+	return ret;
+}
+
+ssize_t ib_uverbs_destroy_cq(struct ib_uverbs_file *file,
+			     const char __user *buf, int in_len,
+			     int out_len)
+{
+	struct ib_uverbs_destroy_cq cmd;
+	struct ib_cq               *cq;
+	struct ib_uobject          *uobj;
+	int                         ret = -EINVAL;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	down(&ib_uverbs_idr_mutex);
+
+	cq = idr_find(&ib_uverbs_cq_idr, cmd.cq_handle);
+	if (!cq || cq->uobject->context != file->ucontext)
+		goto out;
+
+	uobj = cq->uobject;
+
+	ret = ib_destroy_cq(cq);
+	if (ret)
+		goto out;
+
+	idr_remove(&ib_uverbs_cq_idr, cmd.cq_handle);
+
+	spin_lock_irq(&file->ucontext->lock);
+	list_del(&uobj->list);
+	spin_unlock_irq(&file->ucontext->lock);
+
+	kfree(uobj);
+
+out:
+	up(&ib_uverbs_idr_mutex);
+
+	return ret ? ret : in_len;
+}
+
+ssize_t ib_uverbs_create_qp(struct ib_uverbs_file *file,
+			    const char __user *buf, int in_len,
+			    int out_len)
+{
+	struct ib_uverbs_create_qp      cmd;
+	struct ib_uverbs_create_qp_resp resp;
+	struct ib_udata                 udata;
+	struct ib_uobject              *uobj;
+	struct ib_pd                   *pd;
+	struct ib_cq                   *scq, *rcq;
+	struct ib_qp                   *qp;
+	struct ib_qp_init_attr          attr;
+	int ret;
+
+	if (out_len < sizeof resp)
+		return -ENOSPC;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	INIT_UDATA(&udata, buf + sizeof cmd,
+		   (unsigned long) cmd.response + sizeof resp,
+		   in_len - sizeof cmd, out_len - sizeof resp);
+
+	uobj = kmalloc(sizeof *uobj, GFP_KERNEL);
+	if (!uobj)
+		return -ENOMEM;
+
+	down(&ib_uverbs_idr_mutex);
+
+	pd  = idr_find(&ib_uverbs_pd_idr, cmd.pd_handle);
+	scq = idr_find(&ib_uverbs_cq_idr, cmd.send_cq_handle);
+	rcq = idr_find(&ib_uverbs_cq_idr, cmd.recv_cq_handle);
+
+	if (!pd  || pd->uobject->context  != file->ucontext ||
+	    !scq || scq->uobject->context != file->ucontext ||
+	    !rcq || rcq->uobject->context != file->ucontext) {
+		ret = -EINVAL;
+		goto err_up;
+	}
+
+	attr.event_handler = ib_uverbs_qp_event_handler;
+	attr.qp_context    = file;
+	attr.send_cq       = scq;
+	attr.recv_cq       = rcq;
+	attr.srq           = NULL;
+	attr.sq_sig_type   = cmd.sq_sig_all ? IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
+	attr.qp_type       = cmd.qp_type;
+
+	attr.cap.max_send_wr     = cmd.max_send_wr;
+	attr.cap.max_recv_wr     = cmd.max_recv_wr;
+	attr.cap.max_send_sge    = cmd.max_send_sge;
+	attr.cap.max_recv_sge    = cmd.max_recv_sge;
+	attr.cap.max_inline_data = cmd.max_inline_data;
+
+	uobj->user_handle = cmd.user_handle;
+	uobj->context     = file->ucontext;
+
+	qp = pd->device->create_qp(pd, &attr, &udata);
+	if (IS_ERR(qp)) {
+		ret = PTR_ERR(qp);
+		goto err_up;
+	}
+
+	qp->device     	  = pd->device;
+	qp->pd         	  = pd;
+	qp->send_cq    	  = attr.send_cq;
+	qp->recv_cq    	  = attr.recv_cq;
+	qp->srq	       	  = attr.srq;
+	qp->uobject       = uobj;
+	qp->event_handler = attr.event_handler;
+	qp->qp_context    = attr.qp_context;
+	qp->qp_type	  = attr.qp_type;
+	atomic_inc(&pd->usecnt);
+	atomic_inc(&attr.send_cq->usecnt);
+	atomic_inc(&attr.recv_cq->usecnt);
+	if (attr.srq)
+		atomic_inc(&attr.srq->usecnt);
+
+	memset(&resp, 0, sizeof resp);
+	resp.qpn = qp->qp_num;
+
+retry:
+	if (!idr_pre_get(&ib_uverbs_qp_idr, GFP_KERNEL)) {
+		ret = -ENOMEM;
+		goto err_destroy;
+	}
+
+	ret = idr_get_new(&ib_uverbs_qp_idr, qp, &uobj->id);
+
+	if (ret == -EAGAIN)
+		goto retry;
+	if (ret)
+		goto err_destroy;
+
+	resp.qp_handle = uobj->id;
+
+	spin_lock_irq(&file->ucontext->lock);
+	list_add_tail(&uobj->list, &file->ucontext->qp_list);
+	spin_unlock_irq(&file->ucontext->lock);
+
+	if (copy_to_user((void __user *) (unsigned long) cmd.response,
+			 &resp, sizeof resp)) {
+		ret = -EFAULT;
+		goto err_list;
+	}
+
+	up(&ib_uverbs_idr_mutex);
+
+	return in_len;
+
+err_list:
+	spin_lock_irq(&file->ucontext->lock);
+	list_del(&uobj->list);
+	spin_unlock_irq(&file->ucontext->lock);
+
+err_destroy:
+	ib_destroy_qp(qp);
+
+err_up:
+	up(&ib_uverbs_idr_mutex);
+
+	kfree(uobj);
+	return ret;
+}
+
+ssize_t ib_uverbs_modify_qp(struct ib_uverbs_file *file,
+			    const char __user *buf, int in_len,
+			    int out_len)
+{
+	struct ib_uverbs_modify_qp cmd;
+	struct ib_qp              *qp;
+	struct ib_qp_attr         *attr;
+	int                        ret;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	attr = kmalloc(sizeof *attr, GFP_KERNEL);
+	if (!attr)
+		return -ENOMEM;
+
+	down(&ib_uverbs_idr_mutex);
+
+	qp = idr_find(&ib_uverbs_qp_idr, cmd.qp_handle);
+	if (!qp || qp->uobject->context != file->ucontext) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	attr->qp_state 		  = cmd.qp_state;
+	attr->cur_qp_state 	  = cmd.cur_qp_state;
+	attr->path_mtu 		  = cmd.path_mtu;
+	attr->path_mig_state 	  = cmd.path_mig_state;
+	attr->qkey 		  = cmd.qkey;
+	attr->rq_psn 		  = cmd.rq_psn;
+	attr->sq_psn 		  = cmd.sq_psn;
+	attr->dest_qp_num 	  = cmd.dest_qp_num;
+	attr->qp_access_flags 	  = cmd.qp_access_flags;
+	attr->pkey_index 	  = cmd.pkey_index;
+	attr->alt_pkey_index 	  = cmd.pkey_index;
+	attr->en_sqd_async_notify = cmd.en_sqd_async_notify;
+	attr->max_rd_atomic 	  = cmd.max_rd_atomic;
+	attr->max_dest_rd_atomic  = cmd.max_dest_rd_atomic;
+	attr->min_rnr_timer 	  = cmd.min_rnr_timer;
+	attr->port_num 		  = cmd.port_num;
+	attr->timeout 		  = cmd.timeout;
+	attr->retry_cnt 	  = cmd.retry_cnt;
+	attr->rnr_retry 	  = cmd.rnr_retry;
+	attr->alt_port_num 	  = cmd.alt_port_num;
+	attr->alt_timeout 	  = cmd.alt_timeout;
+
+	memcpy(attr->ah_attr.grh.dgid.raw, cmd.dest.dgid, 16);
+	attr->ah_attr.grh.flow_label        = cmd.dest.flow_label;
+	attr->ah_attr.grh.sgid_index        = cmd.dest.sgid_index;
+	attr->ah_attr.grh.hop_limit         = cmd.dest.hop_limit;
+	attr->ah_attr.grh.traffic_class     = cmd.dest.traffic_class;
+	attr->ah_attr.dlid 	    	    = cmd.dest.dlid;
+	attr->ah_attr.sl   	    	    = cmd.dest.sl;
+	attr->ah_attr.src_path_bits 	    = cmd.dest.src_path_bits;
+	attr->ah_attr.static_rate   	    = cmd.dest.static_rate;
+	attr->ah_attr.ah_flags 	    	    = cmd.dest.is_global ? IB_AH_GRH : 0;
+	attr->ah_attr.port_num 	    	    = cmd.dest.port_num;
+
+	memcpy(attr->alt_ah_attr.grh.dgid.raw, cmd.alt_dest.dgid, 16);
+	attr->alt_ah_attr.grh.flow_label    = cmd.alt_dest.flow_label;
+	attr->alt_ah_attr.grh.sgid_index    = cmd.alt_dest.sgid_index;
+	attr->alt_ah_attr.grh.hop_limit     = cmd.alt_dest.hop_limit;
+	attr->alt_ah_attr.grh.traffic_class = cmd.alt_dest.traffic_class;
+	attr->alt_ah_attr.dlid 	    	    = cmd.alt_dest.dlid;
+	attr->alt_ah_attr.sl   	    	    = cmd.alt_dest.sl;
+	attr->alt_ah_attr.src_path_bits     = cmd.alt_dest.src_path_bits;
+	attr->alt_ah_attr.static_rate       = cmd.alt_dest.static_rate;
+	attr->alt_ah_attr.ah_flags 	    = cmd.alt_dest.is_global ? IB_AH_GRH : 0;
+	attr->alt_ah_attr.port_num 	    = cmd.alt_dest.port_num;
+
+	ret = ib_modify_qp(qp, attr, cmd.attr_mask);
+	if (ret)
+		goto out;
+
+	ret = in_len;
+
+out:
+	up(&ib_uverbs_idr_mutex);
+	kfree(attr);
+
+	return ret;
+}
+
+ssize_t ib_uverbs_destroy_qp(struct ib_uverbs_file *file,
+			     const char __user *buf, int in_len,
+			     int out_len)
+{
+	struct ib_uverbs_destroy_qp cmd;
+	struct ib_qp               *qp;
+	struct ib_uobject          *uobj;
+	int                         ret = -EINVAL;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	down(&ib_uverbs_idr_mutex);
+
+	qp = idr_find(&ib_uverbs_qp_idr, cmd.qp_handle);
+	if (!qp || qp->uobject->context != file->ucontext)
+		goto out;
+
+	uobj = qp->uobject;
+
+	ret = ib_destroy_qp(qp);
+	if (ret)
+		goto out;
+
+	idr_remove(&ib_uverbs_qp_idr, cmd.qp_handle);
+
+	spin_lock_irq(&file->ucontext->lock);
+	list_del(&uobj->list);
+	spin_unlock_irq(&file->ucontext->lock);
+
+	kfree(uobj);
+
+out:
+	up(&ib_uverbs_idr_mutex);
+
+	return ret ? ret : in_len;
+}
+
+ssize_t ib_uverbs_attach_mcast(struct ib_uverbs_file *file,
+			       const char __user *buf, int in_len,
+			       int out_len)
+{
+	struct ib_uverbs_attach_mcast cmd;
+	struct ib_qp                 *qp;
+	int                           ret = -EINVAL;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	down(&ib_uverbs_idr_mutex);
+
+	qp = idr_find(&ib_uverbs_qp_idr, cmd.qp_handle);
+	if (qp && qp->uobject->context == file->ucontext)
+		ret = ib_attach_mcast(qp, (union ib_gid *) cmd.gid, cmd.mlid);
+
+	up(&ib_uverbs_idr_mutex);
+
+	return ret ? ret : in_len;
+}
+
+ssize_t ib_uverbs_detach_mcast(struct ib_uverbs_file *file,
+			       const char __user *buf, int in_len,
+			       int out_len)
+{
+	struct ib_uverbs_detach_mcast cmd;
+	struct ib_qp                 *qp;
+	int                           ret = -EINVAL;
+
+	if (copy_from_user(&cmd, buf, sizeof cmd))
+		return -EFAULT;
+
+	down(&ib_uverbs_idr_mutex);
+
+	qp = idr_find(&ib_uverbs_qp_idr, cmd.qp_handle);
+	if (qp && qp->uobject->context == file->ucontext)
+		ret = ib_detach_mcast(qp, (union ib_gid *) cmd.gid, cmd.mlid);
+
+	up(&ib_uverbs_idr_mutex);
+
+	return ret ? ret : in_len;
+}
diff --git a/drivers/infiniband/core/uverbs_main.c b/drivers/infiniband/core/uverbs_main.c
new file mode 100644
index 000000000000..eb99e693dec2
--- /dev/null
+++ b/drivers/infiniband/core/uverbs_main.c
@@ -0,0 +1,710 @@
+/*
+ * Copyright (c) 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  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.
+ *
+ * $Id: uverbs_main.c 2733 2005-06-28 19:14:34Z roland $
+ */
+
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/device.h>
+#include <linux/err.h>
+#include <linux/fs.h>
+#include <linux/poll.h>
+#include <linux/file.h>
+#include <linux/mount.h>
+
+#include <asm/uaccess.h>
+
+#include "uverbs.h"
+
+MODULE_AUTHOR("Roland Dreier");
+MODULE_DESCRIPTION("InfiniBand userspace verbs access");
+MODULE_LICENSE("Dual BSD/GPL");
+
+#define INFINIBANDEVENTFS_MAGIC	0x49426576	/* "IBev" */
+
+enum {
+	IB_UVERBS_MAJOR       = 231,
+	IB_UVERBS_BASE_MINOR  = 192,
+	IB_UVERBS_MAX_DEVICES = 32
+};
+
+#define IB_UVERBS_BASE_DEV	MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
+
+DECLARE_MUTEX(ib_uverbs_idr_mutex);
+DEFINE_IDR(ib_uverbs_pd_idr);
+DEFINE_IDR(ib_uverbs_mr_idr);
+DEFINE_IDR(ib_uverbs_mw_idr);
+DEFINE_IDR(ib_uverbs_ah_idr);
+DEFINE_IDR(ib_uverbs_cq_idr);
+DEFINE_IDR(ib_uverbs_qp_idr);
+
+static spinlock_t map_lock;
+static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
+
+static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
+				     const char __user *buf, int in_len,
+				     int out_len) = {
+	[IB_USER_VERBS_CMD_QUERY_PARAMS]  = ib_uverbs_query_params,
+	[IB_USER_VERBS_CMD_GET_CONTEXT]   = ib_uverbs_get_context,
+	[IB_USER_VERBS_CMD_QUERY_DEVICE]  = ib_uverbs_query_device,
+	[IB_USER_VERBS_CMD_QUERY_PORT]    = ib_uverbs_query_port,
+	[IB_USER_VERBS_CMD_QUERY_GID]     = ib_uverbs_query_gid,
+	[IB_USER_VERBS_CMD_QUERY_PKEY]    = ib_uverbs_query_pkey,
+	[IB_USER_VERBS_CMD_ALLOC_PD]      = ib_uverbs_alloc_pd,
+	[IB_USER_VERBS_CMD_DEALLOC_PD]    = ib_uverbs_dealloc_pd,
+	[IB_USER_VERBS_CMD_REG_MR]        = ib_uverbs_reg_mr,
+	[IB_USER_VERBS_CMD_DEREG_MR]      = ib_uverbs_dereg_mr,
+	[IB_USER_VERBS_CMD_CREATE_CQ]     = ib_uverbs_create_cq,
+	[IB_USER_VERBS_CMD_DESTROY_CQ]    = ib_uverbs_destroy_cq,
+	[IB_USER_VERBS_CMD_CREATE_QP]     = ib_uverbs_create_qp,
+	[IB_USER_VERBS_CMD_MODIFY_QP]     = ib_uverbs_modify_qp,
+	[IB_USER_VERBS_CMD_DESTROY_QP]    = ib_uverbs_destroy_qp,
+	[IB_USER_VERBS_CMD_ATTACH_MCAST]  = ib_uverbs_attach_mcast,
+	[IB_USER_VERBS_CMD_DETACH_MCAST]  = ib_uverbs_detach_mcast,
+};
+
+static struct vfsmount *uverbs_event_mnt;
+
+static void ib_uverbs_add_one(struct ib_device *device);
+static void ib_uverbs_remove_one(struct ib_device *device);
+
+static int ib_dealloc_ucontext(struct ib_ucontext *context)
+{
+	struct ib_uobject *uobj, *tmp;
+
+	if (!context)
+		return 0;
+
+	down(&ib_uverbs_idr_mutex);
+
+	/* XXX Free AHs */
+
+	list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
+		struct ib_qp *qp = idr_find(&ib_uverbs_qp_idr, uobj->id);
+		idr_remove(&ib_uverbs_qp_idr, uobj->id);
+		ib_destroy_qp(qp);
+		list_del(&uobj->list);
+		kfree(uobj);
+	}
+
+	list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
+		struct ib_cq *cq = idr_find(&ib_uverbs_cq_idr, uobj->id);
+		idr_remove(&ib_uverbs_cq_idr, uobj->id);
+		ib_destroy_cq(cq);
+		list_del(&uobj->list);
+		kfree(uobj);
+	}
+
+	/* XXX Free SRQs */
+	/* XXX Free MWs */
+
+	list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
+		struct ib_mr *mr = idr_find(&ib_uverbs_mr_idr, uobj->id);
+		struct ib_umem_object *memobj;
+
+		idr_remove(&ib_uverbs_mr_idr, uobj->id);
+		ib_dereg_mr(mr);
+
+		memobj = container_of(uobj, struct ib_umem_object, uobject);
+		ib_umem_release_on_close(mr->device, &memobj->umem);
+
+		list_del(&uobj->list);
+		kfree(memobj);
+	}
+
+	list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
+		struct ib_pd *pd = idr_find(&ib_uverbs_pd_idr, uobj->id);
+		idr_remove(&ib_uverbs_pd_idr, uobj->id);
+		ib_dealloc_pd(pd);
+		list_del(&uobj->list);
+		kfree(uobj);
+	}
+
+	up(&ib_uverbs_idr_mutex);
+
+	return context->device->dealloc_ucontext(context);
+}
+
+static void ib_uverbs_release_file(struct kref *ref)
+{
+	struct ib_uverbs_file *file =
+		container_of(ref, struct ib_uverbs_file, ref);
+
+	module_put(file->device->ib_dev->owner);
+	kfree(file);
+}
+
+static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
+				    size_t count, loff_t *pos)
+{
+	struct ib_uverbs_event_file *file = filp->private_data;
+	void *event;
+	int eventsz;
+	int ret = 0;
+
+	spin_lock_irq(&file->lock);
+
+	while (list_empty(&file->event_list) && file->fd >= 0) {
+		spin_unlock_irq(&file->lock);
+
+		if (filp->f_flags & O_NONBLOCK)
+			return -EAGAIN;
+
+		if (wait_event_interruptible(file->poll_wait,
+					     !list_empty(&file->event_list) ||
+					     file->fd < 0))
+			return -ERESTARTSYS;
+
+		spin_lock_irq(&file->lock);
+	}
+
+	if (file->fd < 0) {
+		spin_unlock_irq(&file->lock);
+		return -ENODEV;
+	}
+
+	if (file->is_async) {
+		event   = list_entry(file->event_list.next,
+				     struct ib_uverbs_async_event, list);
+		eventsz = sizeof (struct ib_uverbs_async_event_desc);
+	} else {
+		event   = list_entry(file->event_list.next,
+				     struct ib_uverbs_comp_event, list);
+		eventsz = sizeof (struct ib_uverbs_comp_event_desc);
+	}
+
+	if (eventsz > count) {
+		ret   = -EINVAL;
+		event = NULL;
+	} else
+		list_del(file->event_list.next);
+
+	spin_unlock_irq(&file->lock);
+
+	if (event) {
+		if (copy_to_user(buf, event, eventsz))
+			ret = -EFAULT;
+		else
+			ret = eventsz;
+	}
+
+	kfree(event);
+
+	return ret;
+}
+
+static unsigned int ib_uverbs_event_poll(struct file *filp,
+					 struct poll_table_struct *wait)
+{
+	unsigned int pollflags = 0;
+	struct ib_uverbs_event_file *file = filp->private_data;
+
+	poll_wait(filp, &file->poll_wait, wait);
+
+	spin_lock_irq(&file->lock);
+	if (file->fd < 0)
+		pollflags = POLLERR;
+	else if (!list_empty(&file->event_list))
+		pollflags = POLLIN | POLLRDNORM;
+	spin_unlock_irq(&file->lock);
+
+	return pollflags;
+}
+
+static void ib_uverbs_event_release(struct ib_uverbs_event_file *file)
+{
+	struct list_head *entry, *tmp;
+
+	spin_lock_irq(&file->lock);
+	if (file->fd != -1) {
+		file->fd = -1;
+		list_for_each_safe(entry, tmp, &file->event_list)
+			if (file->is_async)
+				kfree(list_entry(entry, struct ib_uverbs_async_event, list));
+			else
+				kfree(list_entry(entry, struct ib_uverbs_comp_event, list));
+	}
+	spin_unlock_irq(&file->lock);
+}
+
+static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
+{
+	struct ib_uverbs_event_file *file = filp->private_data;
+
+	return fasync_helper(fd, filp, on, &file->async_queue);
+}
+
+static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
+{
+	struct ib_uverbs_event_file *file = filp->private_data;
+
+	ib_uverbs_event_release(file);
+	ib_uverbs_event_fasync(-1, filp, 0);
+	kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
+
+	return 0;
+}
+
+static struct file_operations uverbs_event_fops = {
+	/*
+	 * No .owner field since we artificially create event files,
+	 * so there is no increment to the module reference count in
+	 * the open path.  All event files come from a uverbs command
+	 * file, which already takes a module reference, so this is OK.
+	 */
+	.read 	 = ib_uverbs_event_read,
+	.poll    = ib_uverbs_event_poll,
+	.release = ib_uverbs_event_close,
+	.fasync  = ib_uverbs_event_fasync
+};
+
+void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
+{
+	struct ib_uverbs_file       *file = cq_context;
+	struct ib_uverbs_comp_event *entry;
+	unsigned long                flags;
+
+	entry = kmalloc(sizeof *entry, GFP_ATOMIC);
+	if (!entry)
+		return;
+
+	entry->desc.cq_handle = cq->uobject->user_handle;
+
+	spin_lock_irqsave(&file->comp_file[0].lock, flags);
+	list_add_tail(&entry->list, &file->comp_file[0].event_list);
+	spin_unlock_irqrestore(&file->comp_file[0].lock, flags);
+
+	wake_up_interruptible(&file->comp_file[0].poll_wait);
+	kill_fasync(&file->comp_file[0].async_queue, SIGIO, POLL_IN);
+}
+
+static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
+				    __u64 element, __u64 event)
+{
+	struct ib_uverbs_async_event *entry;
+	unsigned long flags;
+
+	entry = kmalloc(sizeof *entry, GFP_ATOMIC);
+	if (!entry)
+		return;
+
+	entry->desc.element    = element;
+	entry->desc.event_type = event;
+
+	spin_lock_irqsave(&file->async_file.lock, flags);
+	list_add_tail(&entry->list, &file->async_file.event_list);
+	spin_unlock_irqrestore(&file->async_file.lock, flags);
+
+	wake_up_interruptible(&file->async_file.poll_wait);
+	kill_fasync(&file->async_file.async_queue, SIGIO, POLL_IN);
+}
+
+void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
+{
+	ib_uverbs_async_handler(context_ptr,
+				event->element.cq->uobject->user_handle,
+				event->event);
+}
+
+void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
+{
+	ib_uverbs_async_handler(context_ptr,
+				event->element.qp->uobject->user_handle,
+				event->event);
+}
+
+static void ib_uverbs_event_handler(struct ib_event_handler *handler,
+				    struct ib_event *event)
+{
+	struct ib_uverbs_file *file =
+		container_of(handler, struct ib_uverbs_file, event_handler);
+
+	ib_uverbs_async_handler(file, event->element.port_num, event->event);
+}
+
+static int ib_uverbs_event_init(struct ib_uverbs_event_file *file,
+				struct ib_uverbs_file *uverbs_file)
+{
+	struct file *filp;
+
+	spin_lock_init(&file->lock);
+	INIT_LIST_HEAD(&file->event_list);
+	init_waitqueue_head(&file->poll_wait);
+	file->uverbs_file = uverbs_file;
+	file->async_queue = NULL;
+
+	file->fd = get_unused_fd();
+	if (file->fd < 0)
+		return file->fd;
+
+	filp = get_empty_filp();
+	if (!filp) {
+		put_unused_fd(file->fd);
+		return -ENFILE;
+	}
+
+	filp->f_op 	   = &uverbs_event_fops;
+	filp->f_vfsmnt 	   = mntget(uverbs_event_mnt);
+	filp->f_dentry 	   = dget(uverbs_event_mnt->mnt_root);
+	filp->f_mapping    = filp->f_dentry->d_inode->i_mapping;
+	filp->f_flags      = O_RDONLY;
+	filp->f_mode       = FMODE_READ;
+	filp->private_data = file;
+
+	fd_install(file->fd, filp);
+
+	return 0;
+}
+
+static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
+			     size_t count, loff_t *pos)
+{
+	struct ib_uverbs_file *file = filp->private_data;
+	struct ib_uverbs_cmd_hdr hdr;
+
+	if (count < sizeof hdr)
+		return -EINVAL;
+
+	if (copy_from_user(&hdr, buf, sizeof hdr))
+		return -EFAULT;
+
+	if (hdr.in_words * 4 != count)
+		return -EINVAL;
+
+	if (hdr.command < 0 || hdr.command >= ARRAY_SIZE(uverbs_cmd_table))
+		return -EINVAL;
+
+	if (!file->ucontext                               &&
+	    hdr.command != IB_USER_VERBS_CMD_QUERY_PARAMS &&
+	    hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
+		return -EINVAL;
+
+	return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
+					     hdr.in_words * 4, hdr.out_words * 4);
+}
+
+static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
+{
+	struct ib_uverbs_file *file = filp->private_data;
+
+	if (!file->ucontext)
+		return -ENODEV;
+	else
+		return file->device->ib_dev->mmap(file->ucontext, vma);
+}
+
+static int ib_uverbs_open(struct inode *inode, struct file *filp)
+{
+	struct ib_uverbs_device *dev =
+		container_of(inode->i_cdev, struct ib_uverbs_device, dev);
+	struct ib_uverbs_file *file;
+	int i = 0;
+	int ret;
+
+	if (!try_module_get(dev->ib_dev->owner))
+		return -ENODEV;
+
+	file = kmalloc(sizeof *file +
+		       (dev->num_comp - 1) * sizeof (struct ib_uverbs_event_file),
+		       GFP_KERNEL);
+	if (!file)
+		return -ENOMEM;
+
+	file->device = dev;
+	kref_init(&file->ref);
+
+	file->ucontext = NULL;
+
+	ret = ib_uverbs_event_init(&file->async_file, file);
+	if (ret)
+		goto err;
+
+	file->async_file.is_async = 1;
+
+	kref_get(&file->ref);
+
+	for (i = 0; i < dev->num_comp; ++i) {
+		ret = ib_uverbs_event_init(&file->comp_file[i], file);
+		if (ret)
+			goto err_async;
+		kref_get(&file->ref);
+		file->comp_file[i].is_async = 0;
+	}
+
+
+	filp->private_data = file;
+
+	INIT_IB_EVENT_HANDLER(&file->event_handler, dev->ib_dev,
+			      ib_uverbs_event_handler);
+	if (ib_register_event_handler(&file->event_handler))
+		goto err_async;
+
+	return 0;
+
+err_async:
+	while (i--)
+		ib_uverbs_event_release(&file->comp_file[i]);
+
+	ib_uverbs_event_release(&file->async_file);
+
+err:
+	kref_put(&file->ref, ib_uverbs_release_file);
+
+	return ret;
+}
+
+static int ib_uverbs_close(struct inode *inode, struct file *filp)
+{
+	struct ib_uverbs_file *file = filp->private_data;
+	int i;
+
+	ib_unregister_event_handler(&file->event_handler);
+	ib_uverbs_event_release(&file->async_file);
+	ib_dealloc_ucontext(file->ucontext);
+
+	for (i = 0; i < file->device->num_comp; ++i)
+		ib_uverbs_event_release(&file->comp_file[i]);
+
+	kref_put(&file->ref, ib_uverbs_release_file);
+
+	return 0;
+}
+
+static struct file_operations uverbs_fops = {
+	.owner 	 = THIS_MODULE,
+	.write 	 = ib_uverbs_write,
+	.open 	 = ib_uverbs_open,
+	.release = ib_uverbs_close
+};
+
+static struct file_operations uverbs_mmap_fops = {
+	.owner 	 = THIS_MODULE,
+	.write 	 = ib_uverbs_write,
+	.mmap    = ib_uverbs_mmap,
+	.open 	 = ib_uverbs_open,
+	.release = ib_uverbs_close
+};
+
+static struct ib_client uverbs_client = {
+	.name   = "uverbs",
+	.add    = ib_uverbs_add_one,
+	.remove = ib_uverbs_remove_one
+};
+
+static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
+{
+	struct ib_uverbs_device *dev =
+		container_of(class_dev, struct ib_uverbs_device, class_dev);
+
+	return sprintf(buf, "%s\n", dev->ib_dev->name);
+}
+static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
+
+static void ib_uverbs_release_class_dev(struct class_device *class_dev)
+{
+	struct ib_uverbs_device *dev =
+		container_of(class_dev, struct ib_uverbs_device, class_dev);
+
+	cdev_del(&dev->dev);
+	clear_bit(dev->devnum, dev_map);
+	kfree(dev);
+}
+
+static struct class uverbs_class = {
+	.name    = "infiniband_verbs",
+	.release = ib_uverbs_release_class_dev
+};
+
+static ssize_t show_abi_version(struct class *class, char *buf)
+{
+	return sprintf(buf, "%d\n", IB_USER_VERBS_ABI_VERSION);
+}
+static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
+
+static void ib_uverbs_add_one(struct ib_device *device)
+{
+	struct ib_uverbs_device *uverbs_dev;
+
+	if (!device->alloc_ucontext)
+		return;
+
+	uverbs_dev = kmalloc(sizeof *uverbs_dev, GFP_KERNEL);
+	if (!uverbs_dev)
+		return;
+
+	memset(uverbs_dev, 0, sizeof *uverbs_dev);
+
+	spin_lock(&map_lock);
+	uverbs_dev->devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
+	if (uverbs_dev->devnum >= IB_UVERBS_MAX_DEVICES) {
+		spin_unlock(&map_lock);
+		goto err;
+	}
+	set_bit(uverbs_dev->devnum, dev_map);
+	spin_unlock(&map_lock);
+
+	uverbs_dev->ib_dev   = device;
+	uverbs_dev->num_comp = 1;
+
+	if (device->mmap)
+		cdev_init(&uverbs_dev->dev, &uverbs_mmap_fops);
+	else
+		cdev_init(&uverbs_dev->dev, &uverbs_fops);
+	uverbs_dev->dev.owner = THIS_MODULE;
+	kobject_set_name(&uverbs_dev->dev.kobj, "uverbs%d", uverbs_dev->devnum);
+	if (cdev_add(&uverbs_dev->dev, IB_UVERBS_BASE_DEV + uverbs_dev->devnum, 1))
+		goto err;
+
+	uverbs_dev->class_dev.class = &uverbs_class;
+	uverbs_dev->class_dev.dev   = device->dma_device;
+	uverbs_dev->class_dev.devt  = uverbs_dev->dev.dev;
+	snprintf(uverbs_dev->class_dev.class_id, BUS_ID_SIZE, "uverbs%d", uverbs_dev->devnum);
+	if (class_device_register(&uverbs_dev->class_dev))
+		goto err_cdev;
+
+	if (class_device_create_file(&uverbs_dev->class_dev, &class_device_attr_ibdev))
+		goto err_class;
+
+	ib_set_client_data(device, &uverbs_client, uverbs_dev);
+
+	return;
+
+err_class:
+	class_device_unregister(&uverbs_dev->class_dev);
+
+err_cdev:
+	cdev_del(&uverbs_dev->dev);
+	clear_bit(uverbs_dev->devnum, dev_map);
+
+err:
+	kfree(uverbs_dev);
+	return;
+}
+
+static void ib_uverbs_remove_one(struct ib_device *device)
+{
+	struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
+
+	if (!uverbs_dev)
+		return;
+
+	class_device_unregister(&uverbs_dev->class_dev);
+}
+
+static struct super_block *uverbs_event_get_sb(struct file_system_type *fs_type, int flags,
+					       const char *dev_name, void *data)
+{
+	return get_sb_pseudo(fs_type, "infinibandevent:", NULL,
+			     INFINIBANDEVENTFS_MAGIC);
+}
+
+static struct file_system_type uverbs_event_fs = {
+	/* No owner field so module can be unloaded */
+	.name    = "infinibandeventfs",
+	.get_sb  = uverbs_event_get_sb,
+	.kill_sb = kill_litter_super
+};
+
+static int __init ib_uverbs_init(void)
+{
+	int ret;
+
+	spin_lock_init(&map_lock);
+
+	ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
+				     "infiniband_verbs");
+	if (ret) {
+		printk(KERN_ERR "user_verbs: couldn't register device number\n");
+		goto out;
+	}
+
+	ret = class_register(&uverbs_class);
+	if (ret) {
+		printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
+		goto out_chrdev;
+	}
+
+	ret = class_create_file(&uverbs_class, &class_attr_abi_version);
+	if (ret) {
+		printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
+		goto out_class;
+	}
+
+	ret = register_filesystem(&uverbs_event_fs);
+	if (ret) {
+		printk(KERN_ERR "user_verbs: couldn't register infinibandeventfs\n");
+		goto out_class;
+	}
+
+	uverbs_event_mnt = kern_mount(&uverbs_event_fs);
+	if (IS_ERR(uverbs_event_mnt)) {
+		ret = PTR_ERR(uverbs_event_mnt);
+		printk(KERN_ERR "user_verbs: couldn't mount infinibandeventfs\n");
+		goto out_fs;
+	}
+
+	ret = ib_register_client(&uverbs_client);
+	if (ret) {
+		printk(KERN_ERR "user_verbs: couldn't register client\n");
+		goto out_mnt;
+	}
+
+	return 0;
+
+out_mnt:
+	mntput(uverbs_event_mnt);
+
+out_fs:
+	unregister_filesystem(&uverbs_event_fs);
+
+out_class:
+	class_unregister(&uverbs_class);
+
+out_chrdev:
+	unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
+
+out:
+	return ret;
+}
+
+static void __exit ib_uverbs_cleanup(void)
+{
+	ib_unregister_client(&uverbs_client);
+	mntput(uverbs_event_mnt);
+	unregister_filesystem(&uverbs_event_fs);
+	class_unregister(&uverbs_class);
+	unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
+}
+
+module_init(ib_uverbs_init);
+module_exit(ib_uverbs_cleanup);
diff --git a/drivers/infiniband/core/uverbs_mem.c b/drivers/infiniband/core/uverbs_mem.c
new file mode 100644
index 000000000000..ed550f6595bd
--- /dev/null
+++ b/drivers/infiniband/core/uverbs_mem.c
@@ -0,0 +1,221 @@
+/*
+ * Copyright (c) 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  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.
+ *
+ * $Id: uverbs_mem.c 2743 2005-06-28 22:27:59Z roland $
+ */
+
+#include <linux/mm.h>
+#include <linux/dma-mapping.h>
+
+#include "uverbs.h"
+
+struct ib_umem_account_work {
+	struct work_struct work;
+	struct mm_struct  *mm;
+	unsigned long      diff;
+};
+
+
+static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
+{
+	struct ib_umem_chunk *chunk, *tmp;
+	int i;
+
+	list_for_each_entry_safe(chunk, tmp, &umem->chunk_list, list) {
+		dma_unmap_sg(dev->dma_device, chunk->page_list,
+			     chunk->nents, DMA_BIDIRECTIONAL);
+		for (i = 0; i < chunk->nents; ++i) {
+			if (umem->writable && dirty)
+				set_page_dirty_lock(chunk->page_list[i].page);
+			put_page(chunk->page_list[i].page);
+		}
+
+		kfree(chunk);
+	}
+}
+
+int ib_umem_get(struct ib_device *dev, struct ib_umem *mem,
+		void *addr, size_t size, int write)
+{
+	struct page **page_list;
+	struct ib_umem_chunk *chunk;
+	unsigned long locked;
+	unsigned long lock_limit;
+	unsigned long cur_base;
+	unsigned long npages;
+	int ret = 0;
+	int off;
+	int i;
+
+	if (!can_do_mlock())
+		return -EPERM;
+
+	page_list = (struct page **) __get_free_page(GFP_KERNEL);
+	if (!page_list)
+		return -ENOMEM;
+
+	mem->user_base = (unsigned long) addr;
+	mem->length    = size;
+	mem->offset    = (unsigned long) addr & ~PAGE_MASK;
+	mem->page_size = PAGE_SIZE;
+	mem->writable  = write;
+
+	INIT_LIST_HEAD(&mem->chunk_list);
+
+	npages = PAGE_ALIGN(size + mem->offset) >> PAGE_SHIFT;
+
+	down_write(&current->mm->mmap_sem);
+
+	locked     = npages + current->mm->locked_vm;
+	lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
+
+	if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	cur_base = (unsigned long) addr & PAGE_MASK;
+
+	while (npages) {
+		ret = get_user_pages(current, current->mm, cur_base,
+				     min_t(int, npages,
+					   PAGE_SIZE / sizeof (struct page *)),
+				     1, !write, page_list, NULL);
+
+		if (ret < 0)
+			goto out;
+
+		cur_base += ret * PAGE_SIZE;
+		npages   -= ret;
+
+		off = 0;
+
+		while (ret) {
+			chunk = kmalloc(sizeof *chunk + sizeof (struct scatterlist) *
+					min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK),
+					GFP_KERNEL);
+			if (!chunk) {
+				ret = -ENOMEM;
+				goto out;
+			}
+
+			chunk->nents = min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK);
+			for (i = 0; i < chunk->nents; ++i) {
+				chunk->page_list[i].page   = page_list[i + off];
+				chunk->page_list[i].offset = 0;
+				chunk->page_list[i].length = PAGE_SIZE;
+			}
+
+			chunk->nmap = dma_map_sg(dev->dma_device,
+						 &chunk->page_list[0],
+						 chunk->nents,
+						 DMA_BIDIRECTIONAL);
+			if (chunk->nmap <= 0) {
+				for (i = 0; i < chunk->nents; ++i)
+					put_page(chunk->page_list[i].page);
+				kfree(chunk);
+
+				ret = -ENOMEM;
+				goto out;
+			}
+
+			ret -= chunk->nents;
+			off += chunk->nents;
+			list_add_tail(&chunk->list, &mem->chunk_list);
+		}
+
+		ret = 0;
+	}
+
+out:
+	if (ret < 0)
+		__ib_umem_release(dev, mem, 0);
+	else
+		current->mm->locked_vm = locked;
+
+	up_write(&current->mm->mmap_sem);
+	free_page((unsigned long) page_list);
+
+	return ret;
+}
+
+void ib_umem_release(struct ib_device *dev, struct ib_umem *umem)
+{
+	__ib_umem_release(dev, umem, 1);
+
+	down_write(&current->mm->mmap_sem);
+	current->mm->locked_vm -=
+		PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
+	up_write(&current->mm->mmap_sem);
+}
+
+static void ib_umem_account(void *work_ptr)
+{
+	struct ib_umem_account_work *work = work_ptr;
+
+	down_write(&work->mm->mmap_sem);
+	work->mm->locked_vm -= work->diff;
+	up_write(&work->mm->mmap_sem);
+	mmput(work->mm);
+	kfree(work);
+}
+
+void ib_umem_release_on_close(struct ib_device *dev, struct ib_umem *umem)
+{
+	struct ib_umem_account_work *work;
+	struct mm_struct *mm;
+
+	__ib_umem_release(dev, umem, 1);
+
+	mm = get_task_mm(current);
+	if (!mm)
+		return;
+
+	/*
+	 * We may be called with the mm's mmap_sem already held.  This
+	 * can happen when a userspace munmap() is the call that drops
+	 * the last reference to our file and calls our release
+	 * method.  If there are memory regions to destroy, we'll end
+	 * up here and not be able to take the mmap_sem.  Therefore we
+	 * defer the vm_locked accounting to the system workqueue.
+	 */
+
+	work = kmalloc(sizeof *work, GFP_KERNEL);
+	if (!work)
+		return;
+
+	INIT_WORK(&work->work, ib_umem_account, work);
+	work->mm   = mm;
+	work->diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
+
+	schedule_work(&work->work);
+}
diff --git a/drivers/infiniband/core/verbs.c b/drivers/infiniband/core/verbs.c
index 7c08ed0cd7dd..506fdf1f2a26 100644
--- a/drivers/infiniband/core/verbs.c
+++ b/drivers/infiniband/core/verbs.c
@@ -4,6 +4,7 @@
  * Copyright (c) 2004 Intel Corporation.  All rights reserved.
  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
  * Copyright (c) 2004 Voltaire Corporation.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  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
@@ -40,6 +41,7 @@
 #include <linux/err.h>
 
 #include <ib_verbs.h>
+#include <ib_cache.h>
 
 /* Protection domains */
 
@@ -47,10 +49,11 @@ struct ib_pd *ib_alloc_pd(struct ib_device *device)
 {
 	struct ib_pd *pd;
 
-	pd = device->alloc_pd(device);
+	pd = device->alloc_pd(device, NULL, NULL);
 
 	if (!IS_ERR(pd)) {
-		pd->device = device;
+		pd->device  = device;
+		pd->uobject = NULL;
 		atomic_set(&pd->usecnt, 0);
 	}
 
@@ -76,8 +79,9 @@ struct ib_ah *ib_create_ah(struct ib_pd *pd, struct ib_ah_attr *ah_attr)
 	ah = pd->device->create_ah(pd, ah_attr);
 
 	if (!IS_ERR(ah)) {
-		ah->device = pd->device;
-		ah->pd     = pd;
+		ah->device  = pd->device;
+		ah->pd      = pd;
+		ah->uobject = NULL;
 		atomic_inc(&pd->usecnt);
 	}
 
@@ -85,6 +89,40 @@ struct ib_ah *ib_create_ah(struct ib_pd *pd, struct ib_ah_attr *ah_attr)
 }
 EXPORT_SYMBOL(ib_create_ah);
 
+struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, struct ib_wc *wc,
+				   struct ib_grh *grh, u8 port_num)
+{
+	struct ib_ah_attr ah_attr;
+	u32 flow_class;
+	u16 gid_index;
+	int ret;
+
+	memset(&ah_attr, 0, sizeof ah_attr);
+	ah_attr.dlid = wc->slid;
+	ah_attr.sl = wc->sl;
+	ah_attr.src_path_bits = wc->dlid_path_bits;
+	ah_attr.port_num = port_num;
+
+	if (wc->wc_flags & IB_WC_GRH) {
+		ah_attr.ah_flags = IB_AH_GRH;
+		ah_attr.grh.dgid = grh->dgid;
+
+		ret = ib_find_cached_gid(pd->device, &grh->sgid, &port_num,
+					 &gid_index);
+		if (ret)
+			return ERR_PTR(ret);
+
+		ah_attr.grh.sgid_index = (u8) gid_index;
+		flow_class = be32_to_cpu(grh->version_tclass_flow);
+		ah_attr.grh.flow_label = flow_class & 0xFFFFF;
+		ah_attr.grh.traffic_class = (flow_class >> 20) & 0xFF;
+		ah_attr.grh.hop_limit = grh->hop_limit;
+	}
+
+	return ib_create_ah(pd, &ah_attr);
+}
+EXPORT_SYMBOL(ib_create_ah_from_wc);
+
 int ib_modify_ah(struct ib_ah *ah, struct ib_ah_attr *ah_attr)
 {
 	return ah->device->modify_ah ?
@@ -122,7 +160,7 @@ struct ib_qp *ib_create_qp(struct ib_pd *pd,
 {
 	struct ib_qp *qp;
 
-	qp = pd->device->create_qp(pd, qp_init_attr);
+	qp = pd->device->create_qp(pd, qp_init_attr, NULL);
 
 	if (!IS_ERR(qp)) {
 		qp->device     	  = pd->device;
@@ -130,6 +168,7 @@ struct ib_qp *ib_create_qp(struct ib_pd *pd,
 		qp->send_cq    	  = qp_init_attr->send_cq;
 		qp->recv_cq    	  = qp_init_attr->recv_cq;
 		qp->srq	       	  = qp_init_attr->srq;
+		qp->uobject       = NULL;
 		qp->event_handler = qp_init_attr->event_handler;
 		qp->qp_context    = qp_init_attr->qp_context;
 		qp->qp_type	  = qp_init_attr->qp_type;
@@ -197,10 +236,11 @@ struct ib_cq *ib_create_cq(struct ib_device *device,
 {
 	struct ib_cq *cq;
 
-	cq = device->create_cq(device, cqe);
+	cq = device->create_cq(device, cqe, NULL, NULL);
 
 	if (!IS_ERR(cq)) {
 		cq->device        = device;
+		cq->uobject       = NULL;
 		cq->comp_handler  = comp_handler;
 		cq->event_handler = event_handler;
 		cq->cq_context    = cq_context;
@@ -245,8 +285,9 @@ struct ib_mr *ib_get_dma_mr(struct ib_pd *pd, int mr_access_flags)
 	mr = pd->device->get_dma_mr(pd, mr_access_flags);
 
 	if (!IS_ERR(mr)) {
-		mr->device = pd->device;
-		mr->pd     = pd;
+		mr->device  = pd->device;
+		mr->pd      = pd;
+		mr->uobject = NULL;
 		atomic_inc(&pd->usecnt);
 		atomic_set(&mr->usecnt, 0);
 	}
@@ -267,8 +308,9 @@ struct ib_mr *ib_reg_phys_mr(struct ib_pd *pd,
 				     mr_access_flags, iova_start);
 
 	if (!IS_ERR(mr)) {
-		mr->device = pd->device;
-		mr->pd     = pd;
+		mr->device  = pd->device;
+		mr->pd      = pd;
+		mr->uobject = NULL;
 		atomic_inc(&pd->usecnt);
 		atomic_set(&mr->usecnt, 0);
 	}
@@ -344,8 +386,9 @@ struct ib_mw *ib_alloc_mw(struct ib_pd *pd)
 
 	mw = pd->device->alloc_mw(pd);
 	if (!IS_ERR(mw)) {
-		mw->device = pd->device;
-		mw->pd     = pd;
+		mw->device  = pd->device;
+		mw->pd      = pd;
+		mw->uobject = NULL;
 		atomic_inc(&pd->usecnt);
 	}
 
diff --git a/drivers/infiniband/hw/mthca/mthca_av.c b/drivers/infiniband/hw/mthca/mthca_av.c
index 085baf393ca4..d58dcbe66488 100644
--- a/drivers/infiniband/hw/mthca/mthca_av.c
+++ b/drivers/infiniband/hw/mthca/mthca_av.c
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. 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
diff --git a/drivers/infiniband/hw/mthca/mthca_cmd.c b/drivers/infiniband/hw/mthca/mthca_cmd.c
index cd9ed958d92f..1557a522d831 100644
--- a/drivers/infiniband/hw/mthca/mthca_cmd.c
+++ b/drivers/infiniband/hw/mthca/mthca_cmd.c
@@ -431,6 +431,36 @@ static int mthca_cmd_imm(struct mthca_dev *dev,
 				      timeout, status);
 }
 
+int mthca_cmd_init(struct mthca_dev *dev)
+{
+	sema_init(&dev->cmd.hcr_sem, 1);
+	sema_init(&dev->cmd.poll_sem, 1);
+	dev->cmd.use_events = 0;
+
+	dev->hcr = ioremap(pci_resource_start(dev->pdev, 0) + MTHCA_HCR_BASE,
+			   MTHCA_HCR_SIZE);
+	if (!dev->hcr) {
+		mthca_err(dev, "Couldn't map command register.");
+		return -ENOMEM;
+	}
+
+	dev->cmd.pool = pci_pool_create("mthca_cmd", dev->pdev,
+					MTHCA_MAILBOX_SIZE,
+					MTHCA_MAILBOX_SIZE, 0);
+	if (!dev->cmd.pool) {
+		iounmap(dev->hcr);
+		return -ENOMEM;
+	}
+
+	return 0;
+}
+
+void mthca_cmd_cleanup(struct mthca_dev *dev)
+{
+	pci_pool_destroy(dev->cmd.pool);
+	iounmap(dev->hcr);
+}
+
 /*
  * Switch to using events to issue FW commands (should be called after
  * event queue to command events has been initialized).
@@ -489,6 +519,33 @@ void mthca_cmd_use_polling(struct mthca_dev *dev)
 	up(&dev->cmd.poll_sem);
 }
 
+struct mthca_mailbox *mthca_alloc_mailbox(struct mthca_dev *dev,
+					  unsigned int gfp_mask)
+{
+	struct mthca_mailbox *mailbox;
+
+	mailbox = kmalloc(sizeof *mailbox, gfp_mask);
+	if (!mailbox)
+		return ERR_PTR(-ENOMEM);
+
+	mailbox->buf = pci_pool_alloc(dev->cmd.pool, gfp_mask, &mailbox->dma);
+	if (!mailbox->buf) {
+		kfree(mailbox);
+		return ERR_PTR(-ENOMEM);
+	}
+
+	return mailbox;
+}
+
+void mthca_free_mailbox(struct mthca_dev *dev, struct mthca_mailbox *mailbox)
+{
+	if (!mailbox)
+		return;
+
+	pci_pool_free(dev->cmd.pool, mailbox->buf, mailbox->dma);
+	kfree(mailbox);
+}
+
 int mthca_SYS_EN(struct mthca_dev *dev, u8 *status)
 {
 	u64 out;
@@ -513,20 +570,20 @@ int mthca_SYS_DIS(struct mthca_dev *dev, u8 *status)
 static int mthca_map_cmd(struct mthca_dev *dev, u16 op, struct mthca_icm *icm,
 			 u64 virt, u8 *status)
 {
-	u32 *inbox;
-	dma_addr_t indma;
+	struct mthca_mailbox *mailbox;
 	struct mthca_icm_iter iter;
+	__be64 *pages;
 	int lg;
 	int nent = 0;
 	int i;
 	int err = 0;
 	int ts = 0, tc = 0;
 
-	inbox = pci_alloc_consistent(dev->pdev, PAGE_SIZE, &indma);
-	if (!inbox)
-		return -ENOMEM;
-
-	memset(inbox, 0, PAGE_SIZE);
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	memset(mailbox->buf, 0, MTHCA_MAILBOX_SIZE);
+	pages = mailbox->buf;
 
 	for (mthca_icm_first(icm, &iter);
 	     !mthca_icm_last(&iter);
@@ -546,19 +603,17 @@ static int mthca_map_cmd(struct mthca_dev *dev, u16 op, struct mthca_icm *icm,
 		}
 		for (i = 0; i < mthca_icm_size(&iter) / (1 << lg); ++i, ++nent) {
 			if (virt != -1) {
-				*((__be64 *) (inbox + nent * 4)) =
-					cpu_to_be64(virt);
+				pages[nent * 2] = cpu_to_be64(virt);
 				virt += 1 << lg;
 			}
 
-			*((__be64 *) (inbox + nent * 4 + 2)) =
-				cpu_to_be64((mthca_icm_addr(&iter) +
-					     (i << lg)) | (lg - 12));
+			pages[nent * 2 + 1] = cpu_to_be64((mthca_icm_addr(&iter) +
+							   (i << lg)) | (lg - 12));
 			ts += 1 << (lg - 10);
 			++tc;
 
-			if (nent == PAGE_SIZE / 16) {
-				err = mthca_cmd(dev, indma, nent, 0, op,
+			if (nent == MTHCA_MAILBOX_SIZE / 16) {
+				err = mthca_cmd(dev, mailbox->dma, nent, 0, op,
 						CMD_TIME_CLASS_B, status);
 				if (err || *status)
 					goto out;
@@ -568,7 +623,7 @@ static int mthca_map_cmd(struct mthca_dev *dev, u16 op, struct mthca_icm *icm,
 	}
 
 	if (nent)
-		err = mthca_cmd(dev, indma, nent, 0, op,
+		err = mthca_cmd(dev, mailbox->dma, nent, 0, op,
 				CMD_TIME_CLASS_B, status);
 
 	switch (op) {
@@ -585,7 +640,7 @@ static int mthca_map_cmd(struct mthca_dev *dev, u16 op, struct mthca_icm *icm,
 	}
 
 out:
-	pci_free_consistent(dev->pdev, PAGE_SIZE, inbox, indma);
+	mthca_free_mailbox(dev, mailbox);
 	return err;
 }
 
@@ -606,8 +661,8 @@ int mthca_RUN_FW(struct mthca_dev *dev, u8 *status)
 
 int mthca_QUERY_FW(struct mthca_dev *dev, u8 *status)
 {
+	struct mthca_mailbox *mailbox;
 	u32 *outbox;
-	dma_addr_t outdma;
 	int err = 0;
 	u8 lg;
 
@@ -625,12 +680,12 @@ int mthca_QUERY_FW(struct mthca_dev *dev, u8 *status)
 #define QUERY_FW_EQ_ARM_BASE_OFFSET    0x40
 #define QUERY_FW_EQ_SET_CI_BASE_OFFSET 0x48
 
-	outbox = pci_alloc_consistent(dev->pdev, QUERY_FW_OUT_SIZE, &outdma);
-	if (!outbox) {
-		return -ENOMEM;
-	}
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	outbox = mailbox->buf;
 
-	err = mthca_cmd_box(dev, 0, outdma, 0, 0, CMD_QUERY_FW,
+	err = mthca_cmd_box(dev, 0, mailbox->dma, 0, 0, CMD_QUERY_FW,
 			    CMD_TIME_CLASS_A, status);
 
 	if (err)
@@ -681,15 +736,15 @@ int mthca_QUERY_FW(struct mthca_dev *dev, u8 *status)
 	}
 
 out:
-	pci_free_consistent(dev->pdev, QUERY_FW_OUT_SIZE, outbox, outdma);
+	mthca_free_mailbox(dev, mailbox);
 	return err;
 }
 
 int mthca_ENABLE_LAM(struct mthca_dev *dev, u8 *status)
 {
+	struct mthca_mailbox *mailbox;
 	u8 info;
 	u32 *outbox;
-	dma_addr_t outdma;
 	int err = 0;
 
 #define ENABLE_LAM_OUT_SIZE         0x100
@@ -700,11 +755,12 @@ int mthca_ENABLE_LAM(struct mthca_dev *dev, u8 *status)
 #define ENABLE_LAM_INFO_HIDDEN_FLAG (1 << 4)
 #define ENABLE_LAM_INFO_ECC_MASK    0x3
 
-	outbox = pci_alloc_consistent(dev->pdev, ENABLE_LAM_OUT_SIZE, &outdma);
-	if (!outbox)
-		return -ENOMEM;
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	outbox = mailbox->buf;
 
-	err = mthca_cmd_box(dev, 0, outdma, 0, 0, CMD_ENABLE_LAM,
+	err = mthca_cmd_box(dev, 0, mailbox->dma, 0, 0, CMD_ENABLE_LAM,
 			    CMD_TIME_CLASS_C, status);
 
 	if (err)
@@ -733,7 +789,7 @@ int mthca_ENABLE_LAM(struct mthca_dev *dev, u8 *status)
 		  (unsigned long long) dev->ddr_end);
 
 out:
-	pci_free_consistent(dev->pdev, ENABLE_LAM_OUT_SIZE, outbox, outdma);
+	mthca_free_mailbox(dev, mailbox);
 	return err;
 }
 
@@ -744,9 +800,9 @@ int mthca_DISABLE_LAM(struct mthca_dev *dev, u8 *status)
 
 int mthca_QUERY_DDR(struct mthca_dev *dev, u8 *status)
 {
+	struct mthca_mailbox *mailbox;
 	u8 info;
 	u32 *outbox;
-	dma_addr_t outdma;
 	int err = 0;
 
 #define QUERY_DDR_OUT_SIZE         0x100
@@ -757,11 +813,12 @@ int mthca_QUERY_DDR(struct mthca_dev *dev, u8 *status)
 #define QUERY_DDR_INFO_HIDDEN_FLAG (1 << 4)
 #define QUERY_DDR_INFO_ECC_MASK    0x3
 
-	outbox = pci_alloc_consistent(dev->pdev, QUERY_DDR_OUT_SIZE, &outdma);
-	if (!outbox)
-		return -ENOMEM;
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	outbox = mailbox->buf;
 
-	err = mthca_cmd_box(dev, 0, outdma, 0, 0, CMD_QUERY_DDR,
+	err = mthca_cmd_box(dev, 0, mailbox->dma, 0, 0, CMD_QUERY_DDR,
 			    CMD_TIME_CLASS_A, status);
 
 	if (err)
@@ -787,15 +844,15 @@ int mthca_QUERY_DDR(struct mthca_dev *dev, u8 *status)
 		  (unsigned long long) dev->ddr_end);
 
 out:
-	pci_free_consistent(dev->pdev, QUERY_DDR_OUT_SIZE, outbox, outdma);
+	mthca_free_mailbox(dev, mailbox);
 	return err;
 }
 
 int mthca_QUERY_DEV_LIM(struct mthca_dev *dev,
 			struct mthca_dev_lim *dev_lim, u8 *status)
 {
+	struct mthca_mailbox *mailbox;
 	u32 *outbox;
-	dma_addr_t outdma;
 	u8 field;
 	u16 size;
 	int err;
@@ -860,11 +917,12 @@ int mthca_QUERY_DEV_LIM(struct mthca_dev *dev,
 #define QUERY_DEV_LIM_LAMR_OFFSET           0x9f
 #define QUERY_DEV_LIM_MAX_ICM_SZ_OFFSET     0xa0
 
-	outbox = pci_alloc_consistent(dev->pdev, QUERY_DEV_LIM_OUT_SIZE, &outdma);
-	if (!outbox)
-		return -ENOMEM;
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	outbox = mailbox->buf;
 
-	err = mthca_cmd_box(dev, 0, outdma, 0, 0, CMD_QUERY_DEV_LIM,
+	err = mthca_cmd_box(dev, 0, mailbox->dma, 0, 0, CMD_QUERY_DEV_LIM,
 			    CMD_TIME_CLASS_A, status);
 
 	if (err)
@@ -1020,15 +1078,15 @@ int mthca_QUERY_DEV_LIM(struct mthca_dev *dev,
 	}
 
 out:
-	pci_free_consistent(dev->pdev, QUERY_DEV_LIM_OUT_SIZE, outbox, outdma);
+	mthca_free_mailbox(dev, mailbox);
 	return err;
 }
 
 int mthca_QUERY_ADAPTER(struct mthca_dev *dev,
 			struct mthca_adapter *adapter, u8 *status)
 {
+	struct mthca_mailbox *mailbox;
 	u32 *outbox;
-	dma_addr_t outdma;
 	int err;
 
 #define QUERY_ADAPTER_OUT_SIZE             0x100
@@ -1037,23 +1095,24 @@ int mthca_QUERY_ADAPTER(struct mthca_dev *dev,
 #define QUERY_ADAPTER_REVISION_ID_OFFSET   0x08
 #define QUERY_ADAPTER_INTA_PIN_OFFSET      0x10
 
-	outbox = pci_alloc_consistent(dev->pdev, QUERY_ADAPTER_OUT_SIZE, &outdma);
-	if (!outbox)
-		return -ENOMEM;
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	outbox = mailbox->buf;
 
-	err = mthca_cmd_box(dev, 0, outdma, 0, 0, CMD_QUERY_ADAPTER,
+	err = mthca_cmd_box(dev, 0, mailbox->dma, 0, 0, CMD_QUERY_ADAPTER,
 			    CMD_TIME_CLASS_A, status);
 
 	if (err)
 		goto out;
 
-	MTHCA_GET(adapter->vendor_id, outbox, QUERY_ADAPTER_VENDOR_ID_OFFSET);
-	MTHCA_GET(adapter->device_id, outbox, QUERY_ADAPTER_DEVICE_ID_OFFSET);
+	MTHCA_GET(adapter->vendor_id, outbox,   QUERY_ADAPTER_VENDOR_ID_OFFSET);
+	MTHCA_GET(adapter->device_id, outbox,   QUERY_ADAPTER_DEVICE_ID_OFFSET);
 	MTHCA_GET(adapter->revision_id, outbox, QUERY_ADAPTER_REVISION_ID_OFFSET);
-	MTHCA_GET(adapter->inta_pin, outbox, QUERY_ADAPTER_INTA_PIN_OFFSET);
+	MTHCA_GET(adapter->inta_pin, outbox,    QUERY_ADAPTER_INTA_PIN_OFFSET);
 
 out:
-	pci_free_consistent(dev->pdev, QUERY_DEV_LIM_OUT_SIZE, outbox, outdma);
+	mthca_free_mailbox(dev, mailbox);
 	return err;
 }
 
@@ -1061,8 +1120,8 @@ int mthca_INIT_HCA(struct mthca_dev *dev,
 		   struct mthca_init_hca_param *param,
 		   u8 *status)
 {
+	struct mthca_mailbox *mailbox;
 	u32 *inbox;
-	dma_addr_t indma;
 	int err;
 
 #define INIT_HCA_IN_SIZE             	 0x200
@@ -1102,9 +1161,10 @@ int mthca_INIT_HCA(struct mthca_dev *dev,
 #define  INIT_HCA_UAR_SCATCH_BASE_OFFSET (INIT_HCA_UAR_OFFSET + 0x10)
 #define  INIT_HCA_UAR_CTX_BASE_OFFSET    (INIT_HCA_UAR_OFFSET + 0x18)
 
-	inbox = pci_alloc_consistent(dev->pdev, INIT_HCA_IN_SIZE, &indma);
-	if (!inbox)
-		return -ENOMEM;
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	inbox = mailbox->buf;
 
 	memset(inbox, 0, INIT_HCA_IN_SIZE);
 
@@ -1167,10 +1227,9 @@ int mthca_INIT_HCA(struct mthca_dev *dev,
 		MTHCA_PUT(inbox, param->uarc_base,   INIT_HCA_UAR_CTX_BASE_OFFSET);
 	}
 
-	err = mthca_cmd(dev, indma, 0, 0, CMD_INIT_HCA,
-			HZ, status);
+	err = mthca_cmd(dev, mailbox->dma, 0, 0, CMD_INIT_HCA, HZ, status);
 
-	pci_free_consistent(dev->pdev, INIT_HCA_IN_SIZE, inbox, indma);
+	mthca_free_mailbox(dev, mailbox);
 	return err;
 }
 
@@ -1178,8 +1237,8 @@ int mthca_INIT_IB(struct mthca_dev *dev,
 		  struct mthca_init_ib_param *param,
 		  int port, u8 *status)
 {
+	struct mthca_mailbox *mailbox;
 	u32 *inbox;
-	dma_addr_t indma;
 	int err;
 	u32 flags;
 
@@ -1199,9 +1258,10 @@ int mthca_INIT_IB(struct mthca_dev *dev,
 #define INIT_IB_NODE_GUID_OFFSET 0x18
 #define INIT_IB_SI_GUID_OFFSET   0x20
 
-	inbox = pci_alloc_consistent(dev->pdev, INIT_IB_IN_SIZE, &indma);
-	if (!inbox)
-		return -ENOMEM;
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	inbox = mailbox->buf;
 
 	memset(inbox, 0, INIT_IB_IN_SIZE);
 
@@ -1221,10 +1281,10 @@ int mthca_INIT_IB(struct mthca_dev *dev,
 	MTHCA_PUT(inbox, param->node_guid, INIT_IB_NODE_GUID_OFFSET);
 	MTHCA_PUT(inbox, param->si_guid,   INIT_IB_SI_GUID_OFFSET);
 
-	err = mthca_cmd(dev, indma, port, 0, CMD_INIT_IB,
+	err = mthca_cmd(dev, mailbox->dma, port, 0, CMD_INIT_IB,
 			CMD_TIME_CLASS_A, status);
 
-	pci_free_consistent(dev->pdev, INIT_HCA_IN_SIZE, inbox, indma);
+	mthca_free_mailbox(dev, mailbox);
 	return err;
 }
 
@@ -1241,8 +1301,8 @@ int mthca_CLOSE_HCA(struct mthca_dev *dev, int panic, u8 *status)
 int mthca_SET_IB(struct mthca_dev *dev, struct mthca_set_ib_param *param,
 		 int port, u8 *status)
 {
+	struct mthca_mailbox *mailbox;
 	u32 *inbox;
-	dma_addr_t indma;
 	int err;
 	u32 flags = 0;
 
@@ -1253,9 +1313,10 @@ int mthca_SET_IB(struct mthca_dev *dev, struct mthca_set_ib_param *param,
 #define SET_IB_CAP_MASK_OFFSET 0x04
 #define SET_IB_SI_GUID_OFFSET  0x08
 
-	inbox = pci_alloc_consistent(dev->pdev, SET_IB_IN_SIZE, &indma);
-	if (!inbox)
-		return -ENOMEM;
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	inbox = mailbox->buf;
 
 	memset(inbox, 0, SET_IB_IN_SIZE);
 
@@ -1266,10 +1327,10 @@ int mthca_SET_IB(struct mthca_dev *dev, struct mthca_set_ib_param *param,
 	MTHCA_PUT(inbox, param->cap_mask, SET_IB_CAP_MASK_OFFSET);
 	MTHCA_PUT(inbox, param->si_guid,  SET_IB_SI_GUID_OFFSET);
 
-	err = mthca_cmd(dev, indma, port, 0, CMD_SET_IB,
+	err = mthca_cmd(dev, mailbox->dma, port, 0, CMD_SET_IB,
 			CMD_TIME_CLASS_B, status);
 
-	pci_free_consistent(dev->pdev, INIT_HCA_IN_SIZE, inbox, indma);
+	mthca_free_mailbox(dev, mailbox);
 	return err;
 }
 
@@ -1280,20 +1341,22 @@ int mthca_MAP_ICM(struct mthca_dev *dev, struct mthca_icm *icm, u64 virt, u8 *st
 
 int mthca_MAP_ICM_page(struct mthca_dev *dev, u64 dma_addr, u64 virt, u8 *status)
 {
+	struct mthca_mailbox *mailbox;
 	u64 *inbox;
-	dma_addr_t indma;
 	int err;
 
-	inbox = pci_alloc_consistent(dev->pdev, 16, &indma);
-	if (!inbox)
-		return -ENOMEM;
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	inbox = mailbox->buf;
 
 	inbox[0] = cpu_to_be64(virt);
 	inbox[1] = cpu_to_be64(dma_addr);
 
-	err = mthca_cmd(dev, indma, 1, 0, CMD_MAP_ICM, CMD_TIME_CLASS_B, status);
+	err = mthca_cmd(dev, mailbox->dma, 1, 0, CMD_MAP_ICM,
+			CMD_TIME_CLASS_B, status);
 
-	pci_free_consistent(dev->pdev, 16, inbox, indma);
+	mthca_free_mailbox(dev, mailbox);
 
 	if (!err)
 		mthca_dbg(dev, "Mapped page at %llx to %llx for ICM.\n",
@@ -1338,69 +1401,26 @@ int mthca_SET_ICM_SIZE(struct mthca_dev *dev, u64 icm_size, u64 *aux_pages,
 	return 0;
 }
 
-int mthca_SW2HW_MPT(struct mthca_dev *dev, void *mpt_entry,
+int mthca_SW2HW_MPT(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
 		    int mpt_index, u8 *status)
 {
-	dma_addr_t indma;
-	int err;
-
-	indma = pci_map_single(dev->pdev, mpt_entry,
-			       MTHCA_MPT_ENTRY_SIZE,
-			       PCI_DMA_TODEVICE);
-	if (pci_dma_mapping_error(indma))
-		return -ENOMEM;
-
-	err = mthca_cmd(dev, indma, mpt_index, 0, CMD_SW2HW_MPT,
-			CMD_TIME_CLASS_B, status);
-
-	pci_unmap_single(dev->pdev, indma,
-			 MTHCA_MPT_ENTRY_SIZE, PCI_DMA_TODEVICE);
-	return err;
+	return mthca_cmd(dev, mailbox->dma, mpt_index, 0, CMD_SW2HW_MPT,
+			 CMD_TIME_CLASS_B, status);
 }
 
-int mthca_HW2SW_MPT(struct mthca_dev *dev, void *mpt_entry,
+int mthca_HW2SW_MPT(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
 		    int mpt_index, u8 *status)
 {
-	dma_addr_t outdma = 0;
-	int err;
-
-	if (mpt_entry) {
-		outdma = pci_map_single(dev->pdev, mpt_entry,
-					MTHCA_MPT_ENTRY_SIZE,
-					PCI_DMA_FROMDEVICE);
-		if (pci_dma_mapping_error(outdma))
-			return -ENOMEM;
-	}
-
-	err = mthca_cmd_box(dev, 0, outdma, mpt_index, !mpt_entry,
-			    CMD_HW2SW_MPT,
-			    CMD_TIME_CLASS_B, status);
-
-	if (mpt_entry)
-		pci_unmap_single(dev->pdev, outdma,
-				 MTHCA_MPT_ENTRY_SIZE,
-				 PCI_DMA_FROMDEVICE);
-	return err;
+	return mthca_cmd_box(dev, 0, mailbox ? mailbox->dma : 0, mpt_index,
+			     !mailbox, CMD_HW2SW_MPT,
+			     CMD_TIME_CLASS_B, status);
 }
 
-int mthca_WRITE_MTT(struct mthca_dev *dev, u64 *mtt_entry,
+int mthca_WRITE_MTT(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
 		    int num_mtt, u8 *status)
 {
-	dma_addr_t indma;
-	int err;
-
-	indma = pci_map_single(dev->pdev, mtt_entry,
-			       (num_mtt + 2) * 8,
-			       PCI_DMA_TODEVICE);
-	if (pci_dma_mapping_error(indma))
-		return -ENOMEM;
-
-	err = mthca_cmd(dev, indma, num_mtt, 0, CMD_WRITE_MTT,
-			CMD_TIME_CLASS_B, status);
-
-	pci_unmap_single(dev->pdev, indma,
-			 (num_mtt + 2) * 8, PCI_DMA_TODEVICE);
-	return err;
+	return mthca_cmd(dev, mailbox->dma, num_mtt, 0, CMD_WRITE_MTT,
+			 CMD_TIME_CLASS_B, status);
 }
 
 int mthca_SYNC_TPT(struct mthca_dev *dev, u8 *status)
@@ -1418,92 +1438,38 @@ int mthca_MAP_EQ(struct mthca_dev *dev, u64 event_mask, int unmap,
 			 0, CMD_MAP_EQ, CMD_TIME_CLASS_B, status);
 }
 
-int mthca_SW2HW_EQ(struct mthca_dev *dev, void *eq_context,
+int mthca_SW2HW_EQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
 		   int eq_num, u8 *status)
 {
-	dma_addr_t indma;
-	int err;
-
-	indma = pci_map_single(dev->pdev, eq_context,
-			       MTHCA_EQ_CONTEXT_SIZE,
-			       PCI_DMA_TODEVICE);
-	if (pci_dma_mapping_error(indma))
-		return -ENOMEM;
-
-	err = mthca_cmd(dev, indma, eq_num, 0, CMD_SW2HW_EQ,
-			CMD_TIME_CLASS_A, status);
-
-	pci_unmap_single(dev->pdev, indma,
-			 MTHCA_EQ_CONTEXT_SIZE, PCI_DMA_TODEVICE);
-	return err;
+	return mthca_cmd(dev, mailbox->dma, eq_num, 0, CMD_SW2HW_EQ,
+			 CMD_TIME_CLASS_A, status);
 }
 
-int mthca_HW2SW_EQ(struct mthca_dev *dev, void *eq_context,
+int mthca_HW2SW_EQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
 		   int eq_num, u8 *status)
 {
-	dma_addr_t outdma = 0;
-	int err;
-
-	outdma = pci_map_single(dev->pdev, eq_context,
-				MTHCA_EQ_CONTEXT_SIZE,
-				PCI_DMA_FROMDEVICE);
-	if (pci_dma_mapping_error(outdma))
-		return -ENOMEM;
-
-	err = mthca_cmd_box(dev, 0, outdma, eq_num, 0,
-			    CMD_HW2SW_EQ,
-			    CMD_TIME_CLASS_A, status);
-
-	pci_unmap_single(dev->pdev, outdma,
-			 MTHCA_EQ_CONTEXT_SIZE,
-			 PCI_DMA_FROMDEVICE);
-	return err;
+	return mthca_cmd_box(dev, 0, mailbox->dma, eq_num, 0,
+			     CMD_HW2SW_EQ,
+			     CMD_TIME_CLASS_A, status);
 }
 
-int mthca_SW2HW_CQ(struct mthca_dev *dev, void *cq_context,
+int mthca_SW2HW_CQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
 		   int cq_num, u8 *status)
 {
-	dma_addr_t indma;
-	int err;
-
-	indma = pci_map_single(dev->pdev, cq_context,
-			       MTHCA_CQ_CONTEXT_SIZE,
-			       PCI_DMA_TODEVICE);
-	if (pci_dma_mapping_error(indma))
-		return -ENOMEM;
-
-	err = mthca_cmd(dev, indma, cq_num, 0, CMD_SW2HW_CQ,
+	return mthca_cmd(dev, mailbox->dma, cq_num, 0, CMD_SW2HW_CQ,
 			CMD_TIME_CLASS_A, status);
-
-	pci_unmap_single(dev->pdev, indma,
-			 MTHCA_CQ_CONTEXT_SIZE, PCI_DMA_TODEVICE);
-	return err;
 }
 
-int mthca_HW2SW_CQ(struct mthca_dev *dev, void *cq_context,
+int mthca_HW2SW_CQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
 		   int cq_num, u8 *status)
 {
-	dma_addr_t outdma = 0;
-	int err;
-
-	outdma = pci_map_single(dev->pdev, cq_context,
-				MTHCA_CQ_CONTEXT_SIZE,
-				PCI_DMA_FROMDEVICE);
-	if (pci_dma_mapping_error(outdma))
-		return -ENOMEM;
-
-	err = mthca_cmd_box(dev, 0, outdma, cq_num, 0,
-			    CMD_HW2SW_CQ,
-			    CMD_TIME_CLASS_A, status);
-
-	pci_unmap_single(dev->pdev, outdma,
-			 MTHCA_CQ_CONTEXT_SIZE,
-			 PCI_DMA_FROMDEVICE);
-	return err;
+	return mthca_cmd_box(dev, 0, mailbox->dma, cq_num, 0,
+			     CMD_HW2SW_CQ,
+			     CMD_TIME_CLASS_A, status);
 }
 
 int mthca_MODIFY_QP(struct mthca_dev *dev, int trans, u32 num,
-		    int is_ee, void *qp_context, u32 optmask,
+		    int is_ee, struct mthca_mailbox *mailbox, u32 optmask,
 		    u8 *status)
 {
 	static const u16 op[] = {
@@ -1520,36 +1486,34 @@ int mthca_MODIFY_QP(struct mthca_dev *dev, int trans, u32 num,
 		[MTHCA_TRANS_ANY2RST]   = CMD_ERR2RST_QPEE
 	};
 	u8 op_mod = 0;
-
-	dma_addr_t indma;
+	int my_mailbox = 0;
 	int err;
 
 	if (trans < 0 || trans >= ARRAY_SIZE(op))
 		return -EINVAL;
 
 	if (trans == MTHCA_TRANS_ANY2RST) {
-		indma  = 0;
 		op_mod = 3;	/* don't write outbox, any->reset */
 
 		/* For debugging */
-		qp_context = pci_alloc_consistent(dev->pdev, MTHCA_QP_CONTEXT_SIZE,
-						  &indma);
-		op_mod = 2;	/* write outbox, any->reset */
+		if (!mailbox) {
+			mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+			if (!IS_ERR(mailbox)) {
+				my_mailbox = 1;
+				op_mod     = 2;	/* write outbox, any->reset */
+			} else
+				mailbox = NULL;
+		}
 	} else {
-		indma = pci_map_single(dev->pdev, qp_context,
-				       MTHCA_QP_CONTEXT_SIZE,
-				       PCI_DMA_TODEVICE);
-		if (pci_dma_mapping_error(indma))
-			return -ENOMEM;
-
 		if (0) {
 			int i;
 			mthca_dbg(dev, "Dumping QP context:\n");
-			printk("  opt param mask: %08x\n", be32_to_cpup(qp_context));
+			printk("  opt param mask: %08x\n", be32_to_cpup(mailbox->buf));
 			for (i = 0; i < 0x100 / 4; ++i) {
 				if (i % 8 == 0)
 					printk("  [%02x] ", i * 4);
-				printk(" %08x", be32_to_cpu(((u32 *) qp_context)[i + 2]));
+				printk(" %08x",
+				       be32_to_cpu(((u32 *) mailbox->buf)[i + 2]));
 				if ((i + 1) % 8 == 0)
 					printk("\n");
 			}
@@ -1557,55 +1521,39 @@ int mthca_MODIFY_QP(struct mthca_dev *dev, int trans, u32 num,
 	}
 
 	if (trans == MTHCA_TRANS_ANY2RST) {
-		err = mthca_cmd_box(dev, 0, indma, (!!is_ee << 24) | num,
-				    op_mod, op[trans], CMD_TIME_CLASS_C, status);
+		err = mthca_cmd_box(dev, 0, mailbox ? mailbox->dma : 0,
+				    (!!is_ee << 24) | num, op_mod,
+				    op[trans], CMD_TIME_CLASS_C, status);
 
-		if (0) {
+		if (0 && mailbox) {
 			int i;
 			mthca_dbg(dev, "Dumping QP context:\n");
-			printk(" %08x\n", be32_to_cpup(qp_context));
+			printk(" %08x\n", be32_to_cpup(mailbox->buf));
 			for (i = 0; i < 0x100 / 4; ++i) {
 				if (i % 8 == 0)
 					printk("[%02x] ", i * 4);
-				printk(" %08x", be32_to_cpu(((u32 *) qp_context)[i + 2]));
+				printk(" %08x",
+				       be32_to_cpu(((u32 *) mailbox->buf)[i + 2]));
 				if ((i + 1) % 8 == 0)
 					printk("\n");
 			}
 		}
 
 	} else
-		err = mthca_cmd(dev, indma, (!!is_ee << 24) | num,
+		err = mthca_cmd(dev, mailbox->dma, (!!is_ee << 24) | num,
 				op_mod, op[trans], CMD_TIME_CLASS_C, status);
 
-	if (trans != MTHCA_TRANS_ANY2RST)
-		pci_unmap_single(dev->pdev, indma,
-				 MTHCA_QP_CONTEXT_SIZE, PCI_DMA_TODEVICE);
-	else
-		pci_free_consistent(dev->pdev, MTHCA_QP_CONTEXT_SIZE,
-				    qp_context, indma);
+	if (my_mailbox)
+		mthca_free_mailbox(dev, mailbox);
+
 	return err;
 }
 
 int mthca_QUERY_QP(struct mthca_dev *dev, u32 num, int is_ee,
-		   void *qp_context, u8 *status)
+		   struct mthca_mailbox *mailbox, u8 *status)
 {
-	dma_addr_t outdma = 0;
-	int err;
-
-	outdma = pci_map_single(dev->pdev, qp_context,
-				MTHCA_QP_CONTEXT_SIZE,
-				PCI_DMA_FROMDEVICE);
-	if (pci_dma_mapping_error(outdma))
-		return -ENOMEM;
-
-	err = mthca_cmd_box(dev, 0, outdma, (!!is_ee << 24) | num, 0,
-			    CMD_QUERY_QPEE,
-			    CMD_TIME_CLASS_A, status);
-
-	pci_unmap_single(dev->pdev, outdma,
-			 MTHCA_QP_CONTEXT_SIZE,
-			 PCI_DMA_FROMDEVICE);
-	return err;
+	return mthca_cmd_box(dev, 0, mailbox->dma, (!!is_ee << 24) | num, 0,
+			     CMD_QUERY_QPEE, CMD_TIME_CLASS_A, status);
 }
 
 int mthca_CONF_SPECIAL_QP(struct mthca_dev *dev, int type, u32 qpn,
@@ -1635,11 +1583,11 @@ int mthca_CONF_SPECIAL_QP(struct mthca_dev *dev, int type, u32 qpn,
 }
 
 int mthca_MAD_IFC(struct mthca_dev *dev, int ignore_mkey, int ignore_bkey,
-		  int port, struct ib_wc* in_wc, struct ib_grh* in_grh,
+		  int port, struct ib_wc *in_wc, struct ib_grh *in_grh,
 		  void *in_mad, void *response_mad, u8 *status)
 {
-	void *box;
-	dma_addr_t dma;
+	struct mthca_mailbox *inmailbox, *outmailbox;
+	void *inbox;
 	int err;
 	u32 in_modifier = port;
 	u8 op_modifier = 0;
@@ -1653,11 +1601,18 @@ int mthca_MAD_IFC(struct mthca_dev *dev, int ignore_mkey, int ignore_bkey,
 #define MAD_IFC_PKEY_OFFSET   0x10e
 #define MAD_IFC_GRH_OFFSET    0x140
 
-	box = pci_alloc_consistent(dev->pdev, MAD_IFC_BOX_SIZE, &dma);
-	if (!box)
-		return -ENOMEM;
+	inmailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(inmailbox))
+		return PTR_ERR(inmailbox);
+	inbox = inmailbox->buf;
 
-	memcpy(box, in_mad, 256);
+	outmailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(outmailbox)) {
+		mthca_free_mailbox(dev, inmailbox);
+		return PTR_ERR(outmailbox);
+	}
+
+	memcpy(inbox, in_mad, 256);
 
 	/*
 	 * Key check traps can't be generated unless we have in_wc to
@@ -1671,97 +1626,65 @@ int mthca_MAD_IFC(struct mthca_dev *dev, int ignore_mkey, int ignore_bkey,
 	if (in_wc) {
 		u8 val;
 
-		memset(box + 256, 0, 256);
+		memset(inbox + 256, 0, 256);
 
-		MTHCA_PUT(box, in_wc->qp_num, 	  MAD_IFC_MY_QPN_OFFSET);
-		MTHCA_PUT(box, in_wc->src_qp, 	  MAD_IFC_RQPN_OFFSET);
+		MTHCA_PUT(inbox, in_wc->qp_num,     MAD_IFC_MY_QPN_OFFSET);
+		MTHCA_PUT(inbox, in_wc->src_qp,     MAD_IFC_RQPN_OFFSET);
 
 		val = in_wc->sl << 4;
-		MTHCA_PUT(box, val,               MAD_IFC_SL_OFFSET);
+		MTHCA_PUT(inbox, val,               MAD_IFC_SL_OFFSET);
 
 		val = in_wc->dlid_path_bits |
 			(in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
-		MTHCA_PUT(box, val,               MAD_IFC_GRH_OFFSET);
+		MTHCA_PUT(inbox, val,               MAD_IFC_GRH_OFFSET);
 
-		MTHCA_PUT(box, in_wc->slid,       MAD_IFC_RLID_OFFSET);
-		MTHCA_PUT(box, in_wc->pkey_index, MAD_IFC_PKEY_OFFSET);
+		MTHCA_PUT(inbox, in_wc->slid,       MAD_IFC_RLID_OFFSET);
+		MTHCA_PUT(inbox, in_wc->pkey_index, MAD_IFC_PKEY_OFFSET);
 
 		if (in_grh)
-			memcpy((u8 *) box + MAD_IFC_GRH_OFFSET, in_grh, 40);
+			memcpy(inbox + MAD_IFC_GRH_OFFSET, in_grh, 40);
 
 		op_modifier |= 0x10;
 
 		in_modifier |= in_wc->slid << 16;
 	}
 
-	err = mthca_cmd_box(dev, dma, dma + 512, in_modifier, op_modifier,
+	err = mthca_cmd_box(dev, inmailbox->dma, outmailbox->dma,
+			    in_modifier, op_modifier,
 			    CMD_MAD_IFC, CMD_TIME_CLASS_C, status);
 
 	if (!err && !*status)
-		memcpy(response_mad, box + 512, 256);
+		memcpy(response_mad, outmailbox->buf, 256);
 
-	pci_free_consistent(dev->pdev, MAD_IFC_BOX_SIZE, box, dma);
+	mthca_free_mailbox(dev, inmailbox);
+	mthca_free_mailbox(dev, outmailbox);
 	return err;
 }
 
-int mthca_READ_MGM(struct mthca_dev *dev, int index, void *mgm,
-		   u8 *status)
+int mthca_READ_MGM(struct mthca_dev *dev, int index,
+		   struct mthca_mailbox *mailbox, u8 *status)
 {
-	dma_addr_t outdma = 0;
-	int err;
-
-	outdma = pci_map_single(dev->pdev, mgm,
-				MTHCA_MGM_ENTRY_SIZE,
-				PCI_DMA_FROMDEVICE);
-	if (pci_dma_mapping_error(outdma))
-		return -ENOMEM;
-
-	err = mthca_cmd_box(dev, 0, outdma, index, 0,
-			    CMD_READ_MGM,
-			    CMD_TIME_CLASS_A, status);
-
-	pci_unmap_single(dev->pdev, outdma,
-			 MTHCA_MGM_ENTRY_SIZE,
-			 PCI_DMA_FROMDEVICE);
-	return err;
+	return mthca_cmd_box(dev, 0, mailbox->dma, index, 0,
+			     CMD_READ_MGM, CMD_TIME_CLASS_A, status);
 }
 
-int mthca_WRITE_MGM(struct mthca_dev *dev, int index, void *mgm,
-		    u8 *status)
+int mthca_WRITE_MGM(struct mthca_dev *dev, int index,
+		    struct mthca_mailbox *mailbox, u8 *status)
 {
-	dma_addr_t indma;
-	int err;
-
-	indma = pci_map_single(dev->pdev, mgm,
-			       MTHCA_MGM_ENTRY_SIZE,
-			       PCI_DMA_TODEVICE);
-	if (pci_dma_mapping_error(indma))
-		return -ENOMEM;
-
-	err = mthca_cmd(dev, indma, index, 0, CMD_WRITE_MGM,
-			CMD_TIME_CLASS_A, status);
-
-	pci_unmap_single(dev->pdev, indma,
-			 MTHCA_MGM_ENTRY_SIZE, PCI_DMA_TODEVICE);
-	return err;
+	return mthca_cmd(dev, mailbox->dma, index, 0, CMD_WRITE_MGM,
+			 CMD_TIME_CLASS_A, status);
 }
 
-int mthca_MGID_HASH(struct mthca_dev *dev, void *gid, u16 *hash,
-		    u8 *status)
+int mthca_MGID_HASH(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    u16 *hash, u8 *status)
 {
-	dma_addr_t indma;
 	u64 imm;
 	int err;
 
-	indma = pci_map_single(dev->pdev, gid, 16, PCI_DMA_TODEVICE);
-	if (pci_dma_mapping_error(indma))
-		return -ENOMEM;
-
-	err = mthca_cmd_imm(dev, indma, &imm, 0, 0, CMD_MGID_HASH,
+	err = mthca_cmd_imm(dev, mailbox->dma, &imm, 0, 0, CMD_MGID_HASH,
 			    CMD_TIME_CLASS_A, status);
-	*hash = imm;
 
-	pci_unmap_single(dev->pdev, indma, 16, PCI_DMA_TODEVICE);
+	*hash = imm;
 	return err;
 }
 
diff --git a/drivers/infiniband/hw/mthca/mthca_cmd.h b/drivers/infiniband/hw/mthca/mthca_cmd.h
index adf039b3c540..ed517f175dd6 100644
--- a/drivers/infiniband/hw/mthca/mthca_cmd.h
+++ b/drivers/infiniband/hw/mthca/mthca_cmd.h
@@ -37,8 +37,7 @@
 
 #include <ib_verbs.h>
 
-#define MTHCA_CMD_MAILBOX_ALIGN 16UL
-#define MTHCA_CMD_MAILBOX_EXTRA (MTHCA_CMD_MAILBOX_ALIGN - 1)
+#define MTHCA_MAILBOX_SIZE 4096
 
 enum {
 	/* command completed successfully: */
@@ -112,6 +111,11 @@ enum {
 	DEV_LIM_FLAG_UD_MULTI           = 1 << 21,
 };
 
+struct mthca_mailbox {
+	dma_addr_t dma;
+	void      *buf;
+};
+
 struct mthca_dev_lim {
 	int max_srq_sz;
 	int max_qp_sz;
@@ -235,11 +239,17 @@ struct mthca_set_ib_param {
 	u32 cap_mask;
 };
 
+int mthca_cmd_init(struct mthca_dev *dev);
+void mthca_cmd_cleanup(struct mthca_dev *dev);
 int mthca_cmd_use_events(struct mthca_dev *dev);
 void mthca_cmd_use_polling(struct mthca_dev *dev);
 void mthca_cmd_event(struct mthca_dev *dev, u16 token,
 		     u8  status, u64 out_param);
 
+struct mthca_mailbox *mthca_alloc_mailbox(struct mthca_dev *dev,
+					  unsigned int gfp_mask);
+void mthca_free_mailbox(struct mthca_dev *dev, struct mthca_mailbox *mailbox);
+
 int mthca_SYS_EN(struct mthca_dev *dev, u8 *status);
 int mthca_SYS_DIS(struct mthca_dev *dev, u8 *status);
 int mthca_MAP_FA(struct mthca_dev *dev, struct mthca_icm *icm, u8 *status);
@@ -270,41 +280,39 @@ int mthca_MAP_ICM_AUX(struct mthca_dev *dev, struct mthca_icm *icm, u8 *status);
 int mthca_UNMAP_ICM_AUX(struct mthca_dev *dev, u8 *status);
 int mthca_SET_ICM_SIZE(struct mthca_dev *dev, u64 icm_size, u64 *aux_pages,
 		       u8 *status);
-int mthca_SW2HW_MPT(struct mthca_dev *dev, void *mpt_entry,
+int mthca_SW2HW_MPT(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
 		    int mpt_index, u8 *status);
-int mthca_HW2SW_MPT(struct mthca_dev *dev, void *mpt_entry,
+int mthca_HW2SW_MPT(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
 		    int mpt_index, u8 *status);
-int mthca_WRITE_MTT(struct mthca_dev *dev, u64 *mtt_entry,
+int mthca_WRITE_MTT(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
 		    int num_mtt, u8 *status);
 int mthca_SYNC_TPT(struct mthca_dev *dev, u8 *status);
 int mthca_MAP_EQ(struct mthca_dev *dev, u64 event_mask, int unmap,
 		 int eq_num, u8 *status);
-int mthca_SW2HW_EQ(struct mthca_dev *dev, void *eq_context,
+int mthca_SW2HW_EQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
 		   int eq_num, u8 *status);
-int mthca_HW2SW_EQ(struct mthca_dev *dev, void *eq_context,
+int mthca_HW2SW_EQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
 		   int eq_num, u8 *status);
-int mthca_SW2HW_CQ(struct mthca_dev *dev, void *cq_context,
+int mthca_SW2HW_CQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
 		   int cq_num, u8 *status);
-int mthca_HW2SW_CQ(struct mthca_dev *dev, void *cq_context,
+int mthca_HW2SW_CQ(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
 		   int cq_num, u8 *status);
 int mthca_MODIFY_QP(struct mthca_dev *dev, int trans, u32 num,
-		    int is_ee, void *qp_context, u32 optmask,
+		    int is_ee, struct mthca_mailbox *mailbox, u32 optmask,
 		    u8 *status);
 int mthca_QUERY_QP(struct mthca_dev *dev, u32 num, int is_ee,
-		   void *qp_context, u8 *status);
+		   struct mthca_mailbox *mailbox, u8 *status);
 int mthca_CONF_SPECIAL_QP(struct mthca_dev *dev, int type, u32 qpn,
 			  u8 *status);
 int mthca_MAD_IFC(struct mthca_dev *dev, int ignore_mkey, int ignore_bkey,
-		  int port, struct ib_wc* in_wc, struct ib_grh* in_grh,
+		  int port, struct ib_wc *in_wc, struct ib_grh *in_grh,
 		  void *in_mad, void *response_mad, u8 *status);
-int mthca_READ_MGM(struct mthca_dev *dev, int index, void *mgm,
-		   u8 *status);
-int mthca_WRITE_MGM(struct mthca_dev *dev, int index, void *mgm,
-		    u8 *status);
-int mthca_MGID_HASH(struct mthca_dev *dev, void *gid, u16 *hash,
-		    u8 *status);
+int mthca_READ_MGM(struct mthca_dev *dev, int index,
+		   struct mthca_mailbox *mailbox, u8 *status);
+int mthca_WRITE_MGM(struct mthca_dev *dev, int index,
+		    struct mthca_mailbox *mailbox, u8 *status);
+int mthca_MGID_HASH(struct mthca_dev *dev, struct mthca_mailbox *mailbox,
+		    u16 *hash, u8 *status);
 int mthca_NOP(struct mthca_dev *dev, u8 *status);
 
-#define MAILBOX_ALIGN(x) ((void *) ALIGN((unsigned long) (x), MTHCA_CMD_MAILBOX_ALIGN))
-
 #endif /* MTHCA_CMD_H */
diff --git a/drivers/infiniband/hw/mthca/mthca_cq.c b/drivers/infiniband/hw/mthca/mthca_cq.c
index 2bf347b84c31..5687c3014522 100644
--- a/drivers/infiniband/hw/mthca/mthca_cq.c
+++ b/drivers/infiniband/hw/mthca/mthca_cq.c
@@ -1,5 +1,7 @@
 /*
  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
+ * Copyright (c) 2005 Cisco Systems, Inc. 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
@@ -171,6 +173,17 @@ static inline void set_cqe_hw(struct mthca_cqe *cqe)
 	cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
 }
 
+static void dump_cqe(struct mthca_dev *dev, void *cqe_ptr)
+{
+	__be32 *cqe = cqe_ptr;
+
+	(void) cqe;	/* avoid warning if mthca_dbg compiled away... */
+	mthca_dbg(dev, "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
+		  be32_to_cpu(cqe[0]), be32_to_cpu(cqe[1]), be32_to_cpu(cqe[2]),
+		  be32_to_cpu(cqe[3]), be32_to_cpu(cqe[4]), be32_to_cpu(cqe[5]),
+		  be32_to_cpu(cqe[6]), be32_to_cpu(cqe[7]));
+}
+
 /*
  * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
  * should be correct before calling update_cons_index().
@@ -280,16 +293,12 @@ static int handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
 	int dbd;
 	u32 new_wqe;
 
-	if (1 && cqe->syndrome != SYNDROME_WR_FLUSH_ERR) {
-		int j;
-
-		mthca_dbg(dev, "%x/%d: error CQE -> QPN %06x, WQE @ %08x\n",
-			  cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
-			  be32_to_cpu(cqe->wqe));
-
-		for (j = 0; j < 8; ++j)
-			printk(KERN_DEBUG "  [%2x] %08x\n",
-			       j * 4, be32_to_cpu(((u32 *) cqe)[j]));
+	if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
+		mthca_dbg(dev, "local QP operation err "
+			  "(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
+			  be32_to_cpu(cqe->my_qpn), be32_to_cpu(cqe->wqe),
+			  cq->cqn, cq->cons_index);
+		dump_cqe(dev, cqe);
 	}
 
 	/*
@@ -364,8 +373,12 @@ static int handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
 	 * If we're at the end of the WQE chain, or we've used up our
 	 * doorbell count, free the CQE.  Otherwise just update it for
 	 * the next poll operation.
+	 *
+	 * This does not apply to mem-free HCAs: they don't use the
+	 * doorbell count field, and so we should always free the CQE.
 	 */
-	if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
+	if (mthca_is_memfree(dev) ||
+	    !(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
 		return 0;
 
 	cqe->db_cnt   = cpu_to_be16(be16_to_cpu(cqe->db_cnt) - dbd);
@@ -377,15 +390,6 @@ static int handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
 	return 0;
 }
 
-static void dump_cqe(struct mthca_cqe *cqe)
-{
-	int j;
-
-	for (j = 0; j < 8; ++j)
-		printk(KERN_DEBUG "  [%2x] %08x\n",
-		       j * 4, be32_to_cpu(((u32 *) cqe)[j]));
-}
-
 static inline int mthca_poll_one(struct mthca_dev *dev,
 				 struct mthca_cq *cq,
 				 struct mthca_qp **cur_qp,
@@ -414,8 +418,7 @@ static inline int mthca_poll_one(struct mthca_dev *dev,
 		mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
 			  cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
 			  be32_to_cpu(cqe->wqe));
-
-		dump_cqe(cqe);
+		dump_cqe(dev, cqe);
 	}
 
 	is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
@@ -638,19 +641,19 @@ static void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq *cq)
 	int size;
 
 	if (cq->is_direct)
-		pci_free_consistent(dev->pdev,
-				    (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE,
-				    cq->queue.direct.buf,
-				    pci_unmap_addr(&cq->queue.direct,
-						   mapping));
+		dma_free_coherent(&dev->pdev->dev,
+				  (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE,
+				  cq->queue.direct.buf,
+				  pci_unmap_addr(&cq->queue.direct,
+						 mapping));
 	else {
 		size = (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE;
 		for (i = 0; i < (size + PAGE_SIZE - 1) / PAGE_SIZE; ++i)
 			if (cq->queue.page_list[i].buf)
-				pci_free_consistent(dev->pdev, PAGE_SIZE,
-						    cq->queue.page_list[i].buf,
-						    pci_unmap_addr(&cq->queue.page_list[i],
-								   mapping));
+				dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
+						  cq->queue.page_list[i].buf,
+						  pci_unmap_addr(&cq->queue.page_list[i],
+								 mapping));
 
 		kfree(cq->queue.page_list);
 	}
@@ -670,8 +673,8 @@ static int mthca_alloc_cq_buf(struct mthca_dev *dev, int size,
 		npages        = 1;
 		shift         = get_order(size) + PAGE_SHIFT;
 
-		cq->queue.direct.buf = pci_alloc_consistent(dev->pdev,
-							    size, &t);
+		cq->queue.direct.buf = dma_alloc_coherent(&dev->pdev->dev,
+							  size, &t, GFP_KERNEL);
 		if (!cq->queue.direct.buf)
 			return -ENOMEM;
 
@@ -709,7 +712,8 @@ static int mthca_alloc_cq_buf(struct mthca_dev *dev, int size,
 
 		for (i = 0; i < npages; ++i) {
 			cq->queue.page_list[i].buf =
-				pci_alloc_consistent(dev->pdev, PAGE_SIZE, &t);
+				dma_alloc_coherent(&dev->pdev->dev, PAGE_SIZE,
+						   &t, GFP_KERNEL);
 			if (!cq->queue.page_list[i].buf)
 				goto err_free;
 
@@ -743,10 +747,11 @@ err_out:
 }
 
 int mthca_init_cq(struct mthca_dev *dev, int nent,
+		  struct mthca_ucontext *ctx, u32 pdn,
 		  struct mthca_cq *cq)
 {
 	int size = nent * MTHCA_CQ_ENTRY_SIZE;
-	void *mailbox = NULL;
+	struct mthca_mailbox *mailbox;
 	struct mthca_cq_context *cq_context;
 	int err = -ENOMEM;
 	u8 status;
@@ -754,45 +759,49 @@ int mthca_init_cq(struct mthca_dev *dev, int nent,
 
 	might_sleep();
 
-	cq->ibcq.cqe = nent - 1;
+	cq->ibcq.cqe  = nent - 1;
+	cq->is_kernel = !ctx;
 
 	cq->cqn = mthca_alloc(&dev->cq_table.alloc);
 	if (cq->cqn == -1)
 		return -ENOMEM;
 
 	if (mthca_is_memfree(dev)) {
-		cq->arm_sn = 1;
-
 		err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
 		if (err)
 			goto err_out;
 
-		err = -ENOMEM;
+		if (cq->is_kernel) {
+			cq->arm_sn = 1;
 
-		cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
-						     cq->cqn, &cq->set_ci_db);
-		if (cq->set_ci_db_index < 0)
-			goto err_out_icm;
+			err = -ENOMEM;
 
-		cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
-						  cq->cqn, &cq->arm_db);
-		if (cq->arm_db_index < 0)
-			goto err_out_ci;
+			cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
+							     cq->cqn, &cq->set_ci_db);
+			if (cq->set_ci_db_index < 0)
+				goto err_out_icm;
+
+			cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
+							  cq->cqn, &cq->arm_db);
+			if (cq->arm_db_index < 0)
+				goto err_out_ci;
+		}
 	}
 
-	mailbox = kmalloc(sizeof (struct mthca_cq_context) + MTHCA_CMD_MAILBOX_EXTRA,
-			  GFP_KERNEL);
-	if (!mailbox)
-		goto err_out_mailbox;
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		goto err_out_arm;
 
-	cq_context = MAILBOX_ALIGN(mailbox);
+	cq_context = mailbox->buf;
 
-	err = mthca_alloc_cq_buf(dev, size, cq);
-	if (err)
-		goto err_out_mailbox;
+	if (cq->is_kernel) {
+		err = mthca_alloc_cq_buf(dev, size, cq);
+		if (err)
+			goto err_out_mailbox;
 
-	for (i = 0; i < nent; ++i)
-		set_cqe_hw(get_cqe(cq, i));
+		for (i = 0; i < nent; ++i)
+			set_cqe_hw(get_cqe(cq, i));
+	}
 
 	spin_lock_init(&cq->lock);
 	atomic_set(&cq->refcount, 1);
@@ -803,11 +812,14 @@ int mthca_init_cq(struct mthca_dev *dev, int nent,
 						  MTHCA_CQ_STATE_DISARMED |
 						  MTHCA_CQ_FLAG_TR);
 	cq_context->start           = cpu_to_be64(0);
-	cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24 |
-						  dev->driver_uar.index);
+	cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
+	if (ctx)
+		cq_context->logsize_usrpage |= cpu_to_be32(ctx->uar.index);
+	else
+		cq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
 	cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
 	cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
-	cq_context->pd              = cpu_to_be32(dev->driver_pd.pd_num);
+	cq_context->pd              = cpu_to_be32(pdn);
 	cq_context->lkey            = cpu_to_be32(cq->mr.ibmr.lkey);
 	cq_context->cqn             = cpu_to_be32(cq->cqn);
 
@@ -816,7 +828,7 @@ int mthca_init_cq(struct mthca_dev *dev, int nent,
 		cq_context->state_db = cpu_to_be32(cq->arm_db_index);
 	}
 
-	err = mthca_SW2HW_CQ(dev, cq_context, cq->cqn, &status);
+	err = mthca_SW2HW_CQ(dev, mailbox, cq->cqn, &status);
 	if (err) {
 		mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
 		goto err_out_free_mr;
@@ -840,22 +852,25 @@ int mthca_init_cq(struct mthca_dev *dev, int nent,
 
 	cq->cons_index = 0;
 
-	kfree(mailbox);
+	mthca_free_mailbox(dev, mailbox);
 
 	return 0;
 
 err_out_free_mr:
-	mthca_free_mr(dev, &cq->mr);
-	mthca_free_cq_buf(dev, cq);
+	if (cq->is_kernel) {
+		mthca_free_mr(dev, &cq->mr);
+		mthca_free_cq_buf(dev, cq);
+	}
 
 err_out_mailbox:
-	kfree(mailbox);
+	mthca_free_mailbox(dev, mailbox);
 
-	if (mthca_is_memfree(dev))
+err_out_arm:
+	if (cq->is_kernel && mthca_is_memfree(dev))
 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
 
 err_out_ci:
-	if (mthca_is_memfree(dev))
+	if (cq->is_kernel && mthca_is_memfree(dev))
 		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
 
 err_out_icm:
@@ -870,32 +885,31 @@ err_out:
 void mthca_free_cq(struct mthca_dev *dev,
 		   struct mthca_cq *cq)
 {
-	void *mailbox;
+	struct mthca_mailbox *mailbox;
 	int err;
 	u8 status;
 
 	might_sleep();
 
-	mailbox = kmalloc(sizeof (struct mthca_cq_context) + MTHCA_CMD_MAILBOX_EXTRA,
-			  GFP_KERNEL);
-	if (!mailbox) {
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox)) {
 		mthca_warn(dev, "No memory for mailbox to free CQ.\n");
 		return;
 	}
 
-	err = mthca_HW2SW_CQ(dev, MAILBOX_ALIGN(mailbox), cq->cqn, &status);
+	err = mthca_HW2SW_CQ(dev, mailbox, cq->cqn, &status);
 	if (err)
 		mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
 	else if (status)
-		mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n",
-			   status);
+		mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n", status);
 
 	if (0) {
-		u32 *ctx = MAILBOX_ALIGN(mailbox);
+		u32 *ctx = mailbox->buf;
 		int j;
 
 		printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
-		       cq->cqn, cq->cons_index, !!next_cqe_sw(cq));
+		       cq->cqn, cq->cons_index,
+		       cq->is_kernel ? !!next_cqe_sw(cq) : 0);
 		for (j = 0; j < 16; ++j)
 			printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
 	}
@@ -913,17 +927,18 @@ void mthca_free_cq(struct mthca_dev *dev,
 	atomic_dec(&cq->refcount);
 	wait_event(cq->wait, !atomic_read(&cq->refcount));
 
-	mthca_free_mr(dev, &cq->mr);
-	mthca_free_cq_buf(dev, cq);
-
-	if (mthca_is_memfree(dev)) {
-		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
-		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
-		mthca_table_put(dev, dev->cq_table.table, cq->cqn);
+	if (cq->is_kernel) {
+		mthca_free_mr(dev, &cq->mr);
+		mthca_free_cq_buf(dev, cq);
+		if (mthca_is_memfree(dev)) {
+			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
+			mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
+		}
 	}
 
+	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
 	mthca_free(&dev->cq_table.alloc, cq->cqn);
-	kfree(mailbox);
+	mthca_free_mailbox(dev, mailbox);
 }
 
 int __devinit mthca_init_cq_table(struct mthca_dev *dev)
diff --git a/drivers/infiniband/hw/mthca/mthca_dev.h b/drivers/infiniband/hw/mthca/mthca_dev.h
index e3d79e267dc9..5ecdd2eeeb0f 100644
--- a/drivers/infiniband/hw/mthca/mthca_dev.h
+++ b/drivers/infiniband/hw/mthca/mthca_dev.h
@@ -1,5 +1,7 @@
 /*
  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  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
@@ -46,8 +48,8 @@
 
 #define DRV_NAME	"ib_mthca"
 #define PFX		DRV_NAME ": "
-#define DRV_VERSION	"0.06-pre"
-#define DRV_RELDATE	"November 8, 2004"
+#define DRV_VERSION	"0.06"
+#define DRV_RELDATE	"June 23, 2005"
 
 enum {
 	MTHCA_FLAG_DDR_HIDDEN = 1 << 1,
@@ -98,6 +100,7 @@ enum {
 };
 
 struct mthca_cmd {
+	struct pci_pool          *pool;
 	int                       use_events;
 	struct semaphore          hcr_sem;
 	struct semaphore 	  poll_sem;
@@ -376,9 +379,15 @@ void mthca_unregister_device(struct mthca_dev *dev);
 int mthca_uar_alloc(struct mthca_dev *dev, struct mthca_uar *uar);
 void mthca_uar_free(struct mthca_dev *dev, struct mthca_uar *uar);
 
-int mthca_pd_alloc(struct mthca_dev *dev, struct mthca_pd *pd);
+int mthca_pd_alloc(struct mthca_dev *dev, int privileged, struct mthca_pd *pd);
 void mthca_pd_free(struct mthca_dev *dev, struct mthca_pd *pd);
 
+struct mthca_mtt *mthca_alloc_mtt(struct mthca_dev *dev, int size);
+void mthca_free_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt);
+int mthca_write_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt,
+		    int start_index, u64 *buffer_list, int list_len);
+int mthca_mr_alloc(struct mthca_dev *dev, u32 pd, int buffer_size_shift,
+		   u64 iova, u64 total_size, u32 access, struct mthca_mr *mr);
 int mthca_mr_alloc_notrans(struct mthca_dev *dev, u32 pd,
 			   u32 access, struct mthca_mr *mr);
 int mthca_mr_alloc_phys(struct mthca_dev *dev, u32 pd,
@@ -405,6 +414,7 @@ int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
 int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify notify);
 int mthca_arbel_arm_cq(struct ib_cq *cq, enum ib_cq_notify notify);
 int mthca_init_cq(struct mthca_dev *dev, int nent,
+		  struct mthca_ucontext *ctx, u32 pdn,
 		  struct mthca_cq *cq);
 void mthca_free_cq(struct mthca_dev *dev,
 		   struct mthca_cq *cq);
@@ -430,12 +440,14 @@ int mthca_alloc_qp(struct mthca_dev *dev,
 		   struct mthca_cq *recv_cq,
 		   enum ib_qp_type type,
 		   enum ib_sig_type send_policy,
+		   struct ib_qp_cap *cap,
 		   struct mthca_qp *qp);
 int mthca_alloc_sqp(struct mthca_dev *dev,
 		    struct mthca_pd *pd,
 		    struct mthca_cq *send_cq,
 		    struct mthca_cq *recv_cq,
 		    enum ib_sig_type send_policy,
+		    struct ib_qp_cap *cap,
 		    int qpn,
 		    int port,
 		    struct mthca_sqp *sqp);
diff --git a/drivers/infiniband/hw/mthca/mthca_doorbell.h b/drivers/infiniband/hw/mthca/mthca_doorbell.h
index 821039a49049..535fad7710fb 100644
--- a/drivers/infiniband/hw/mthca/mthca_doorbell.h
+++ b/drivers/infiniband/hw/mthca/mthca_doorbell.h
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. 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
diff --git a/drivers/infiniband/hw/mthca/mthca_eq.c b/drivers/infiniband/hw/mthca/mthca_eq.c
index f46d615d396f..cbcf2b4722e4 100644
--- a/drivers/infiniband/hw/mthca/mthca_eq.c
+++ b/drivers/infiniband/hw/mthca/mthca_eq.c
@@ -469,7 +469,7 @@ static int __devinit mthca_create_eq(struct mthca_dev *dev,
 		PAGE_SIZE;
 	u64 *dma_list = NULL;
 	dma_addr_t t;
-	void *mailbox = NULL;
+	struct mthca_mailbox *mailbox;
 	struct mthca_eq_context *eq_context;
 	int err = -ENOMEM;
 	int i;
@@ -494,17 +494,16 @@ static int __devinit mthca_create_eq(struct mthca_dev *dev,
 	if (!dma_list)
 		goto err_out_free;
 
-	mailbox = kmalloc(sizeof *eq_context + MTHCA_CMD_MAILBOX_EXTRA,
-			  GFP_KERNEL);
-	if (!mailbox)
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
 		goto err_out_free;
-	eq_context = MAILBOX_ALIGN(mailbox);
+	eq_context = mailbox->buf;
 
 	for (i = 0; i < npages; ++i) {
-		eq->page_list[i].buf = pci_alloc_consistent(dev->pdev,
-							    PAGE_SIZE, &t);
+		eq->page_list[i].buf = dma_alloc_coherent(&dev->pdev->dev,
+							  PAGE_SIZE, &t, GFP_KERNEL);
 		if (!eq->page_list[i].buf)
-			goto err_out_free;
+			goto err_out_free_pages;
 
 		dma_list[i] = t;
 		pci_unmap_addr_set(&eq->page_list[i], mapping, t);
@@ -517,7 +516,7 @@ static int __devinit mthca_create_eq(struct mthca_dev *dev,
 
 	eq->eqn = mthca_alloc(&dev->eq_table.alloc);
 	if (eq->eqn == -1)
-		goto err_out_free;
+		goto err_out_free_pages;
 
 	err = mthca_mr_alloc_phys(dev, dev->driver_pd.pd_num,
 				  dma_list, PAGE_SHIFT, npages,
@@ -548,7 +547,7 @@ static int __devinit mthca_create_eq(struct mthca_dev *dev,
 	eq_context->intr            = intr;
 	eq_context->lkey            = cpu_to_be32(eq->mr.ibmr.lkey);
 
-	err = mthca_SW2HW_EQ(dev, eq_context, eq->eqn, &status);
+	err = mthca_SW2HW_EQ(dev, mailbox, eq->eqn, &status);
 	if (err) {
 		mthca_warn(dev, "SW2HW_EQ failed (%d)\n", err);
 		goto err_out_free_mr;
@@ -561,7 +560,7 @@ static int __devinit mthca_create_eq(struct mthca_dev *dev,
 	}
 
 	kfree(dma_list);
-	kfree(mailbox);
+	mthca_free_mailbox(dev, mailbox);
 
 	eq->eqn_mask   = swab32(1 << eq->eqn);
 	eq->cons_index = 0;
@@ -579,17 +578,19 @@ static int __devinit mthca_create_eq(struct mthca_dev *dev,
  err_out_free_eq:
 	mthca_free(&dev->eq_table.alloc, eq->eqn);
 
- err_out_free:
+ err_out_free_pages:
 	for (i = 0; i < npages; ++i)
 		if (eq->page_list[i].buf)
-			pci_free_consistent(dev->pdev, PAGE_SIZE,
-					    eq->page_list[i].buf,
-					    pci_unmap_addr(&eq->page_list[i],
-							   mapping));
+			dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
+					  eq->page_list[i].buf,
+					  pci_unmap_addr(&eq->page_list[i],
+							 mapping));
+
+	mthca_free_mailbox(dev, mailbox);
 
+ err_out_free:
 	kfree(eq->page_list);
 	kfree(dma_list);
-	kfree(mailbox);
 
  err_out:
 	return err;
@@ -598,25 +599,22 @@ static int __devinit mthca_create_eq(struct mthca_dev *dev,
 static void mthca_free_eq(struct mthca_dev *dev,
 			  struct mthca_eq *eq)
 {
-	void *mailbox = NULL;
+	struct mthca_mailbox *mailbox;
 	int err;
 	u8 status;
 	int npages = (eq->nent * MTHCA_EQ_ENTRY_SIZE + PAGE_SIZE - 1) /
 		PAGE_SIZE;
 	int i;
 
-	mailbox = kmalloc(sizeof (struct mthca_eq_context) + MTHCA_CMD_MAILBOX_EXTRA,
-			  GFP_KERNEL);
-	if (!mailbox)
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
 		return;
 
-	err = mthca_HW2SW_EQ(dev, MAILBOX_ALIGN(mailbox),
-			     eq->eqn, &status);
+	err = mthca_HW2SW_EQ(dev, mailbox, eq->eqn, &status);
 	if (err)
 		mthca_warn(dev, "HW2SW_EQ failed (%d)\n", err);
 	if (status)
-		mthca_warn(dev, "HW2SW_EQ returned status 0x%02x\n",
-			   status);
+		mthca_warn(dev, "HW2SW_EQ returned status 0x%02x\n", status);
 
 	dev->eq_table.arm_mask &= ~eq->eqn_mask;
 
@@ -625,7 +623,7 @@ static void mthca_free_eq(struct mthca_dev *dev,
 		for (i = 0; i < sizeof (struct mthca_eq_context) / 4; ++i) {
 			if (i % 4 == 0)
 				printk("[%02x] ", i * 4);
-			printk(" %08x", be32_to_cpup(MAILBOX_ALIGN(mailbox) + i * 4));
+			printk(" %08x", be32_to_cpup(mailbox->buf + i * 4));
 			if ((i + 1) % 4 == 0)
 				printk("\n");
 		}
@@ -638,7 +636,7 @@ static void mthca_free_eq(struct mthca_dev *dev,
 				    pci_unmap_addr(&eq->page_list[i], mapping));
 
 	kfree(eq->page_list);
-	kfree(mailbox);
+	mthca_free_mailbox(dev, mailbox);
 }
 
 static void mthca_free_irqs(struct mthca_dev *dev)
@@ -709,8 +707,7 @@ static int __devinit mthca_map_eq_regs(struct mthca_dev *dev)
 		if (mthca_map_reg(dev, ((pci_resource_len(dev->pdev, 0) - 1) &
 					dev->fw.arbel.eq_arm_base) + 4, 4,
 				  &dev->eq_regs.arbel.eq_arm)) {
-			mthca_err(dev, "Couldn't map interrupt clear register, "
-				  "aborting.\n");
+			mthca_err(dev, "Couldn't map EQ arm register, aborting.\n");
 			mthca_unmap_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
 					dev->fw.arbel.clr_int_base, MTHCA_CLR_INT_SIZE,
 					dev->clr_base);
@@ -721,8 +718,7 @@ static int __devinit mthca_map_eq_regs(struct mthca_dev *dev)
 				  dev->fw.arbel.eq_set_ci_base,
 				  MTHCA_EQ_SET_CI_SIZE,
 				  &dev->eq_regs.arbel.eq_set_ci_base)) {
-			mthca_err(dev, "Couldn't map interrupt clear register, "
-				  "aborting.\n");
+			mthca_err(dev, "Couldn't map EQ CI register, aborting.\n");
 			mthca_unmap_reg(dev, ((pci_resource_len(dev->pdev, 0) - 1) &
 					      dev->fw.arbel.eq_arm_base) + 4, 4,
 					dev->eq_regs.arbel.eq_arm);
diff --git a/drivers/infiniband/hw/mthca/mthca_main.c b/drivers/infiniband/hw/mthca/mthca_main.c
index d40590356df8..2ef916859e17 100644
--- a/drivers/infiniband/hw/mthca/mthca_main.c
+++ b/drivers/infiniband/hw/mthca/mthca_main.c
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. 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
@@ -69,7 +70,7 @@ MODULE_PARM_DESC(msi, "attempt to use MSI if nonzero");
 #endif /* CONFIG_PCI_MSI */
 
 static const char mthca_version[] __devinitdata =
-	"ib_mthca: Mellanox InfiniBand HCA driver v"
+	DRV_NAME ": Mellanox InfiniBand HCA driver v"
 	DRV_VERSION " (" DRV_RELDATE ")\n";
 
 static struct mthca_profile default_profile = {
@@ -664,7 +665,7 @@ static int __devinit mthca_setup_hca(struct mthca_dev *dev)
 		goto err_pd_table_free;
 	}
 
-	err = mthca_pd_alloc(dev, &dev->driver_pd);
+	err = mthca_pd_alloc(dev, 1, &dev->driver_pd);
 	if (err) {
 		mthca_err(dev, "Failed to create driver PD, "
 			  "aborting.\n");
@@ -927,13 +928,13 @@ static int __devinit mthca_init_one(struct pci_dev *pdev,
 	 */
 	if (!(pci_resource_flags(pdev, 0) & IORESOURCE_MEM) ||
 	    pci_resource_len(pdev, 0) != 1 << 20) {
-		dev_err(&pdev->dev, "Missing DCS, aborting.");
+		dev_err(&pdev->dev, "Missing DCS, aborting.\n");
 		err = -ENODEV;
 		goto err_disable_pdev;
 	}
 	if (!(pci_resource_flags(pdev, 2) & IORESOURCE_MEM) ||
 	    pci_resource_len(pdev, 2) != 1 << 23) {
-		dev_err(&pdev->dev, "Missing UAR, aborting.");
+		dev_err(&pdev->dev, "Missing UAR, aborting.\n");
 		err = -ENODEV;
 		goto err_disable_pdev;
 	}
@@ -1004,25 +1005,18 @@ static int __devinit mthca_init_one(struct pci_dev *pdev,
 	    !pci_enable_msi(pdev))
 		mdev->mthca_flags |= MTHCA_FLAG_MSI;
 
-	sema_init(&mdev->cmd.hcr_sem, 1);
-	sema_init(&mdev->cmd.poll_sem, 1);
-	mdev->cmd.use_events = 0;
-
-	mdev->hcr = ioremap(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE, MTHCA_HCR_SIZE);
-	if (!mdev->hcr) {
-		mthca_err(mdev, "Couldn't map command register, "
-			  "aborting.\n");
-		err = -ENOMEM;
+	if (mthca_cmd_init(mdev)) {
+		mthca_err(mdev, "Failed to init command interface, aborting.\n");
 		goto err_free_dev;
 	}
 
 	err = mthca_tune_pci(mdev);
 	if (err)
-		goto err_iounmap;
+		goto err_cmd;
 
 	err = mthca_init_hca(mdev);
 	if (err)
-		goto err_iounmap;
+		goto err_cmd;
 
 	if (mdev->fw_ver < mthca_hca_table[id->driver_data].latest_fw) {
 		mthca_warn(mdev, "HCA FW version %x.%x.%x is old (%x.%x.%x is current).\n",
@@ -1070,8 +1064,8 @@ err_cleanup:
 err_close:
 	mthca_close_hca(mdev);
 
-err_iounmap:
-	iounmap(mdev->hcr);
+err_cmd:
+	mthca_cmd_cleanup(mdev);
 
 err_free_dev:
 	if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
@@ -1118,10 +1112,8 @@ static void __devexit mthca_remove_one(struct pci_dev *pdev)
 		iounmap(mdev->kar);
 		mthca_uar_free(mdev, &mdev->driver_uar);
 		mthca_cleanup_uar_table(mdev);
-
 		mthca_close_hca(mdev);
-
-		iounmap(mdev->hcr);
+		mthca_cmd_cleanup(mdev);
 
 		if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
 			pci_disable_msix(pdev);
@@ -1163,7 +1155,7 @@ static struct pci_device_id mthca_pci_table[] = {
 MODULE_DEVICE_TABLE(pci, mthca_pci_table);
 
 static struct pci_driver mthca_driver = {
-	.name		= "ib_mthca",
+	.name		= DRV_NAME,
 	.id_table	= mthca_pci_table,
 	.probe		= mthca_init_one,
 	.remove		= __devexit_p(mthca_remove_one)
diff --git a/drivers/infiniband/hw/mthca/mthca_mcg.c b/drivers/infiniband/hw/mthca/mthca_mcg.c
index 70a6553a588e..5be7d949dbf6 100644
--- a/drivers/infiniband/hw/mthca/mthca_mcg.c
+++ b/drivers/infiniband/hw/mthca/mthca_mcg.c
@@ -66,22 +66,23 @@ static const u8 zero_gid[16];	/* automatically initialized to 0 */
  * entry in hash chain and *mgm holds end of hash chain.
  */
 static int find_mgm(struct mthca_dev *dev,
-		    u8 *gid, struct mthca_mgm *mgm,
+		    u8 *gid, struct mthca_mailbox *mgm_mailbox,
 		    u16 *hash, int *prev, int *index)
 {
-	void *mailbox;
+	struct mthca_mailbox *mailbox;
+	struct mthca_mgm *mgm = mgm_mailbox->buf;
 	u8 *mgid;
 	int err;
 	u8 status;
 
-	mailbox = kmalloc(16 + MTHCA_CMD_MAILBOX_EXTRA, GFP_KERNEL);
-	if (!mailbox)
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
 		return -ENOMEM;
-	mgid = MAILBOX_ALIGN(mailbox);
+	mgid = mailbox->buf;
 
 	memcpy(mgid, gid, 16);
 
-	err = mthca_MGID_HASH(dev, mgid, hash, &status);
+	err = mthca_MGID_HASH(dev, mailbox, hash, &status);
 	if (err)
 		goto out;
 	if (status) {
@@ -103,7 +104,7 @@ static int find_mgm(struct mthca_dev *dev,
 	*prev  = -1;
 
 	do {
-		err = mthca_READ_MGM(dev, *index, mgm, &status);
+		err = mthca_READ_MGM(dev, *index, mgm_mailbox, &status);
 		if (err)
 			goto out;
 		if (status) {
@@ -129,14 +130,14 @@ static int find_mgm(struct mthca_dev *dev,
 	*index = -1;
 
  out:
-	kfree(mailbox);
+	mthca_free_mailbox(dev, mailbox);
 	return err;
 }
 
 int mthca_multicast_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 {
 	struct mthca_dev *dev = to_mdev(ibqp->device);
-	void *mailbox;
+	struct mthca_mailbox *mailbox;
 	struct mthca_mgm *mgm;
 	u16 hash;
 	int index, prev;
@@ -145,15 +146,15 @@ int mthca_multicast_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 	int err;
 	u8 status;
 
-	mailbox = kmalloc(sizeof *mgm + MTHCA_CMD_MAILBOX_EXTRA, GFP_KERNEL);
-	if (!mailbox)
-		return -ENOMEM;
-	mgm = MAILBOX_ALIGN(mailbox);
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	mgm = mailbox->buf;
 
 	if (down_interruptible(&dev->mcg_table.sem))
 		return -EINTR;
 
-	err = find_mgm(dev, gid->raw, mgm, &hash, &prev, &index);
+	err = find_mgm(dev, gid->raw, mailbox, &hash, &prev, &index);
 	if (err)
 		goto out;
 
@@ -170,7 +171,7 @@ int mthca_multicast_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 			goto out;
 		}
 
-		err = mthca_READ_MGM(dev, index, mgm, &status);
+		err = mthca_READ_MGM(dev, index, mailbox, &status);
 		if (err)
 			goto out;
 		if (status) {
@@ -195,7 +196,7 @@ int mthca_multicast_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 		goto out;
 	}
 
-	err = mthca_WRITE_MGM(dev, index, mgm, &status);
+	err = mthca_WRITE_MGM(dev, index, mailbox, &status);
 	if (err)
 		goto out;
 	if (status) {
@@ -206,7 +207,7 @@ int mthca_multicast_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 	if (!link)
 		goto out;
 
-	err = mthca_READ_MGM(dev, prev, mgm, &status);
+	err = mthca_READ_MGM(dev, prev, mailbox, &status);
 	if (err)
 		goto out;
 	if (status) {
@@ -217,7 +218,7 @@ int mthca_multicast_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 
 	mgm->next_gid_index = cpu_to_be32(index << 5);
 
-	err = mthca_WRITE_MGM(dev, prev, mgm, &status);
+	err = mthca_WRITE_MGM(dev, prev, mailbox, &status);
 	if (err)
 		goto out;
 	if (status) {
@@ -227,14 +228,14 @@ int mthca_multicast_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 
  out:
 	up(&dev->mcg_table.sem);
-	kfree(mailbox);
+	mthca_free_mailbox(dev, mailbox);
 	return err;
 }
 
 int mthca_multicast_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 {
 	struct mthca_dev *dev = to_mdev(ibqp->device);
-	void *mailbox;
+	struct mthca_mailbox *mailbox;
 	struct mthca_mgm *mgm;
 	u16 hash;
 	int prev, index;
@@ -242,15 +243,15 @@ int mthca_multicast_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 	int err;
 	u8 status;
 
-	mailbox = kmalloc(sizeof *mgm + MTHCA_CMD_MAILBOX_EXTRA, GFP_KERNEL);
-	if (!mailbox)
-		return -ENOMEM;
-	mgm = MAILBOX_ALIGN(mailbox);
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	mgm = mailbox->buf;
 
 	if (down_interruptible(&dev->mcg_table.sem))
 		return -EINTR;
 
-	err = find_mgm(dev, gid->raw, mgm, &hash, &prev, &index);
+	err = find_mgm(dev, gid->raw, mailbox, &hash, &prev, &index);
 	if (err)
 		goto out;
 
@@ -285,7 +286,7 @@ int mthca_multicast_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 	mgm->qp[loc]   = mgm->qp[i - 1];
 	mgm->qp[i - 1] = 0;
 
-	err = mthca_WRITE_MGM(dev, index, mgm, &status);
+	err = mthca_WRITE_MGM(dev, index, mailbox, &status);
 	if (err)
 		goto out;
 	if (status) {
@@ -304,7 +305,7 @@ int mthca_multicast_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 		if (be32_to_cpu(mgm->next_gid_index) >> 5) {
 			err = mthca_READ_MGM(dev,
 					     be32_to_cpu(mgm->next_gid_index) >> 5,
-					     mgm, &status);
+					     mailbox, &status);
 			if (err)
 				goto out;
 			if (status) {
@@ -316,7 +317,7 @@ int mthca_multicast_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 		} else
 			memset(mgm->gid, 0, 16);
 
-		err = mthca_WRITE_MGM(dev, index, mgm, &status);
+		err = mthca_WRITE_MGM(dev, index, mailbox, &status);
 		if (err)
 			goto out;
 		if (status) {
@@ -327,7 +328,7 @@ int mthca_multicast_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 	} else {
 		/* Remove entry from AMGM */
 		index = be32_to_cpu(mgm->next_gid_index) >> 5;
-		err = mthca_READ_MGM(dev, prev, mgm, &status);
+		err = mthca_READ_MGM(dev, prev, mailbox, &status);
 		if (err)
 			goto out;
 		if (status) {
@@ -338,7 +339,7 @@ int mthca_multicast_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 
 		mgm->next_gid_index = cpu_to_be32(index << 5);
 
-		err = mthca_WRITE_MGM(dev, prev, mgm, &status);
+		err = mthca_WRITE_MGM(dev, prev, mailbox, &status);
 		if (err)
 			goto out;
 		if (status) {
@@ -350,7 +351,7 @@ int mthca_multicast_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
 
  out:
 	up(&dev->mcg_table.sem);
-	kfree(mailbox);
+	mthca_free_mailbox(dev, mailbox);
 	return err;
 }
 
diff --git a/drivers/infiniband/hw/mthca/mthca_memfree.c b/drivers/infiniband/hw/mthca/mthca_memfree.c
index 637b30e35592..2a8646150355 100644
--- a/drivers/infiniband/hw/mthca/mthca_memfree.c
+++ b/drivers/infiniband/hw/mthca/mthca_memfree.c
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  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
@@ -47,6 +48,15 @@ enum {
 	MTHCA_TABLE_CHUNK_SIZE = 1 << 18
 };
 
+struct mthca_user_db_table {
+	struct semaphore mutex;
+	struct {
+		u64                uvirt;
+		struct scatterlist mem;
+		int                refcount;
+	}                page[0];
+};
+
 void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm)
 {
 	struct mthca_icm_chunk *chunk, *tmp;
@@ -179,9 +189,14 @@ out:
 
 void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
 {
-	int i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
+	int i;
 	u8 status;
 
+	if (!mthca_is_memfree(dev))
+		return;
+
+	i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
+
 	down(&table->mutex);
 
 	if (--table->icm[i]->refcount == 0) {
@@ -256,6 +271,9 @@ void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
 {
 	int i;
 
+	if (!mthca_is_memfree(dev))
+		return;
+
 	for (i = start; i <= end; i += MTHCA_TABLE_CHUNK_SIZE / table->obj_size)
 		mthca_table_put(dev, table, i);
 }
@@ -336,13 +354,133 @@ void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
 	kfree(table);
 }
 
-static u64 mthca_uarc_virt(struct mthca_dev *dev, int page)
+static u64 mthca_uarc_virt(struct mthca_dev *dev, struct mthca_uar *uar, int page)
 {
 	return dev->uar_table.uarc_base +
-		dev->driver_uar.index * dev->uar_table.uarc_size +
+		uar->index * dev->uar_table.uarc_size +
 		page * 4096;
 }
 
+int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
+		      struct mthca_user_db_table *db_tab, int index, u64 uaddr)
+{
+	int ret = 0;
+	u8 status;
+	int i;
+
+	if (!mthca_is_memfree(dev))
+		return 0;
+
+	if (index < 0 || index > dev->uar_table.uarc_size / 8)
+		return -EINVAL;
+
+	down(&db_tab->mutex);
+
+	i = index / MTHCA_DB_REC_PER_PAGE;
+
+	if ((db_tab->page[i].refcount >= MTHCA_DB_REC_PER_PAGE)       ||
+	    (db_tab->page[i].uvirt && db_tab->page[i].uvirt != uaddr) ||
+	    (uaddr & 4095)) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	if (db_tab->page[i].refcount) {
+		++db_tab->page[i].refcount;
+		goto out;
+	}
+
+	ret = get_user_pages(current, current->mm, uaddr & PAGE_MASK, 1, 1, 0,
+			     &db_tab->page[i].mem.page, NULL);
+	if (ret < 0)
+		goto out;
+
+	db_tab->page[i].mem.length = 4096;
+	db_tab->page[i].mem.offset = uaddr & ~PAGE_MASK;
+
+	ret = pci_map_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
+	if (ret < 0) {
+		put_page(db_tab->page[i].mem.page);
+		goto out;
+	}
+
+	ret = mthca_MAP_ICM_page(dev, sg_dma_address(&db_tab->page[i].mem),
+				 mthca_uarc_virt(dev, uar, i), &status);
+	if (!ret && status)
+		ret = -EINVAL;
+	if (ret) {
+		pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
+		put_page(db_tab->page[i].mem.page);
+		goto out;
+	}
+
+	db_tab->page[i].uvirt    = uaddr;
+	db_tab->page[i].refcount = 1;
+
+out:
+	up(&db_tab->mutex);
+	return ret;
+}
+
+void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
+			 struct mthca_user_db_table *db_tab, int index)
+{
+	if (!mthca_is_memfree(dev))
+		return;
+
+	/*
+	 * To make our bookkeeping simpler, we don't unmap DB
+	 * pages until we clean up the whole db table.
+	 */
+
+	down(&db_tab->mutex);
+
+	--db_tab->page[index / MTHCA_DB_REC_PER_PAGE].refcount;
+
+	up(&db_tab->mutex);
+}
+
+struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev)
+{
+	struct mthca_user_db_table *db_tab;
+	int npages;
+	int i;
+
+	if (!mthca_is_memfree(dev))
+		return NULL;
+
+	npages = dev->uar_table.uarc_size / 4096;
+	db_tab = kmalloc(sizeof *db_tab + npages * sizeof *db_tab->page, GFP_KERNEL);
+	if (!db_tab)
+		return ERR_PTR(-ENOMEM);
+
+	init_MUTEX(&db_tab->mutex);
+	for (i = 0; i < npages; ++i) {
+		db_tab->page[i].refcount = 0;
+		db_tab->page[i].uvirt    = 0;
+	}
+
+	return db_tab;
+}
+
+void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
+			       struct mthca_user_db_table *db_tab)
+{
+	int i;
+	u8 status;
+
+	if (!mthca_is_memfree(dev))
+		return;
+
+	for (i = 0; i < dev->uar_table.uarc_size / 4096; ++i) {
+		if (db_tab->page[i].uvirt) {
+			mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, uar, i), 1, &status);
+			pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
+			put_page(db_tab->page[i].mem.page);
+		}
+	}
+}
+
 int mthca_alloc_db(struct mthca_dev *dev, int type, u32 qn, u32 **db)
 {
 	int group;
@@ -399,7 +537,8 @@ int mthca_alloc_db(struct mthca_dev *dev, int type, u32 qn, u32 **db)
 	}
 	memset(page->db_rec, 0, 4096);
 
-	ret = mthca_MAP_ICM_page(dev, page->mapping, mthca_uarc_virt(dev, i), &status);
+	ret = mthca_MAP_ICM_page(dev, page->mapping,
+				 mthca_uarc_virt(dev, &dev->driver_uar, i), &status);
 	if (!ret && status)
 		ret = -EINVAL;
 	if (ret) {
@@ -453,7 +592,7 @@ void mthca_free_db(struct mthca_dev *dev, int type, int db_index)
 
 	if (bitmap_empty(page->used, MTHCA_DB_REC_PER_PAGE) &&
 	    i >= dev->db_tab->max_group1 - 1) {
-		mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, i), 1, &status);
+		mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
 
 		dma_free_coherent(&dev->pdev->dev, 4096,
 				  page->db_rec, page->mapping);
@@ -522,7 +661,7 @@ void mthca_cleanup_db_tab(struct mthca_dev *dev)
 		if (!bitmap_empty(dev->db_tab->page[i].used, MTHCA_DB_REC_PER_PAGE))
 			mthca_warn(dev, "Kernel UARC page %d not empty\n", i);
 
-		mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, i), 1, &status);
+		mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
 
 		dma_free_coherent(&dev->pdev->dev, 4096,
 				  dev->db_tab->page[i].db_rec,
diff --git a/drivers/infiniband/hw/mthca/mthca_memfree.h b/drivers/infiniband/hw/mthca/mthca_memfree.h
index fe7be2a6bc4a..4761d844cb5f 100644
--- a/drivers/infiniband/hw/mthca/mthca_memfree.h
+++ b/drivers/infiniband/hw/mthca/mthca_memfree.h
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  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
@@ -148,7 +149,7 @@ struct mthca_db_table {
 	struct semaphore      mutex;
 };
 
-enum {
+enum mthca_db_type {
 	MTHCA_DB_TYPE_INVALID   = 0x0,
 	MTHCA_DB_TYPE_CQ_SET_CI = 0x1,
 	MTHCA_DB_TYPE_CQ_ARM    = 0x2,
@@ -158,6 +159,17 @@ enum {
 	MTHCA_DB_TYPE_GROUP_SEP = 0x7
 };
 
+struct mthca_user_db_table;
+struct mthca_uar;
+
+int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
+		      struct mthca_user_db_table *db_tab, int index, u64 uaddr);
+void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
+			 struct mthca_user_db_table *db_tab, int index);
+struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev);
+void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
+			       struct mthca_user_db_table *db_tab);
+
 int mthca_init_db_tab(struct mthca_dev *dev);
 void mthca_cleanup_db_tab(struct mthca_dev *dev);
 int mthca_alloc_db(struct mthca_dev *dev, int type, u32 qn, u32 **db);
diff --git a/drivers/infiniband/hw/mthca/mthca_mr.c b/drivers/infiniband/hw/mthca/mthca_mr.c
index 8960fc2306be..cbe50feaf680 100644
--- a/drivers/infiniband/hw/mthca/mthca_mr.c
+++ b/drivers/infiniband/hw/mthca/mthca_mr.c
@@ -40,6 +40,12 @@
 #include "mthca_cmd.h"
 #include "mthca_memfree.h"
 
+struct mthca_mtt {
+	struct mthca_buddy *buddy;
+	int                 order;
+	u32                 first_seg;
+};
+
 /*
  * Must be packed because mtt_seg is 64 bits but only aligned to 32 bits.
  */
@@ -173,8 +179,8 @@ static void __devexit mthca_buddy_cleanup(struct mthca_buddy *buddy)
 	kfree(buddy->bits);
 }
 
-static u32 mthca_alloc_mtt(struct mthca_dev *dev, int order,
-			   struct mthca_buddy *buddy)
+static u32 mthca_alloc_mtt_range(struct mthca_dev *dev, int order,
+				 struct mthca_buddy *buddy)
 {
 	u32 seg = mthca_buddy_alloc(buddy, order);
 
@@ -191,14 +197,102 @@ static u32 mthca_alloc_mtt(struct mthca_dev *dev, int order,
 	return seg;
 }
 
-static void mthca_free_mtt(struct mthca_dev *dev, u32 seg, int order,
-			   struct mthca_buddy* buddy)
+static struct mthca_mtt *__mthca_alloc_mtt(struct mthca_dev *dev, int size,
+					   struct mthca_buddy *buddy)
 {
-	mthca_buddy_free(buddy, seg, order);
+	struct mthca_mtt *mtt;
+	int i;
 
-	if (mthca_is_memfree(dev))
-		mthca_table_put_range(dev, dev->mr_table.mtt_table, seg,
-				      seg + (1 << order) - 1);
+	if (size <= 0)
+		return ERR_PTR(-EINVAL);
+
+	mtt = kmalloc(sizeof *mtt, GFP_KERNEL);
+	if (!mtt)
+		return ERR_PTR(-ENOMEM);
+
+	mtt->buddy = buddy;
+	mtt->order = 0;
+	for (i = MTHCA_MTT_SEG_SIZE / 8; i < size; i <<= 1)
+		++mtt->order;
+
+	mtt->first_seg = mthca_alloc_mtt_range(dev, mtt->order, buddy);
+	if (mtt->first_seg == -1) {
+		kfree(mtt);
+		return ERR_PTR(-ENOMEM);
+	}
+
+	return mtt;
+}
+
+struct mthca_mtt *mthca_alloc_mtt(struct mthca_dev *dev, int size)
+{
+	return __mthca_alloc_mtt(dev, size, &dev->mr_table.mtt_buddy);
+}
+
+void mthca_free_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt)
+{
+	if (!mtt)
+		return;
+
+	mthca_buddy_free(mtt->buddy, mtt->first_seg, mtt->order);
+
+	mthca_table_put_range(dev, dev->mr_table.mtt_table,
+			      mtt->first_seg,
+			      mtt->first_seg + (1 << mtt->order) - 1);
+
+	kfree(mtt);
+}
+
+int mthca_write_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt,
+		    int start_index, u64 *buffer_list, int list_len)
+{
+	struct mthca_mailbox *mailbox;
+	u64 *mtt_entry;
+	int err = 0;
+	u8 status;
+	int i;
+
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	mtt_entry = mailbox->buf;
+
+	while (list_len > 0) {
+		mtt_entry[0] = cpu_to_be64(dev->mr_table.mtt_base +
+					   mtt->first_seg * MTHCA_MTT_SEG_SIZE +
+					   start_index * 8);
+		mtt_entry[1] = 0;
+		for (i = 0; i < list_len && i < MTHCA_MAILBOX_SIZE / 8 - 2; ++i)
+			mtt_entry[i + 2] = cpu_to_be64(buffer_list[i] |
+						       MTHCA_MTT_FLAG_PRESENT);
+
+		/*
+		 * If we have an odd number of entries to write, add
+		 * one more dummy entry for firmware efficiency.
+		 */
+		if (i & 1)
+			mtt_entry[i + 2] = 0;
+
+		err = mthca_WRITE_MTT(dev, mailbox, (i + 1) & ~1, &status);
+		if (err) {
+			mthca_warn(dev, "WRITE_MTT failed (%d)\n", err);
+			goto out;
+		}
+		if (status) {
+			mthca_warn(dev, "WRITE_MTT returned status 0x%02x\n",
+				   status);
+			err = -EINVAL;
+			goto out;
+		}
+
+		list_len    -= i;
+		start_index += i;
+		buffer_list += i;
+	}
+
+out:
+	mthca_free_mailbox(dev, mailbox);
+	return err;
 }
 
 static inline u32 tavor_hw_index_to_key(u32 ind)
@@ -237,91 +331,18 @@ static inline u32 key_to_hw_index(struct mthca_dev *dev, u32 key)
 		return tavor_key_to_hw_index(key);
 }
 
-int mthca_mr_alloc_notrans(struct mthca_dev *dev, u32 pd,
-			   u32 access, struct mthca_mr *mr)
+int mthca_mr_alloc(struct mthca_dev *dev, u32 pd, int buffer_size_shift,
+		   u64 iova, u64 total_size, u32 access, struct mthca_mr *mr)
 {
-	void *mailbox = NULL;
+	struct mthca_mailbox *mailbox;
 	struct mthca_mpt_entry *mpt_entry;
 	u32 key;
+	int i;
 	int err;
 	u8 status;
 
 	might_sleep();
 
-	mr->order = -1;
-	key = mthca_alloc(&dev->mr_table.mpt_alloc);
-	if (key == -1)
-		return -ENOMEM;
-	mr->ibmr.rkey = mr->ibmr.lkey = hw_index_to_key(dev, key);
-
-	if (mthca_is_memfree(dev)) {
-		err = mthca_table_get(dev, dev->mr_table.mpt_table, key);
-		if (err)
-			goto err_out_mpt_free;
-	}
-
-	mailbox = kmalloc(sizeof *mpt_entry + MTHCA_CMD_MAILBOX_EXTRA,
-			  GFP_KERNEL);
-	if (!mailbox) {
-		err = -ENOMEM;
-		goto err_out_table;
-	}
-	mpt_entry = MAILBOX_ALIGN(mailbox);
-
-	mpt_entry->flags = cpu_to_be32(MTHCA_MPT_FLAG_SW_OWNS     |
-				       MTHCA_MPT_FLAG_MIO         |
-				       MTHCA_MPT_FLAG_PHYSICAL    |
-				       MTHCA_MPT_FLAG_REGION      |
-				       access);
-	mpt_entry->page_size = 0;
-	mpt_entry->key       = cpu_to_be32(key);
-	mpt_entry->pd        = cpu_to_be32(pd);
-	mpt_entry->start     = 0;
-	mpt_entry->length    = ~0ULL;
-
-	memset(&mpt_entry->lkey, 0,
-	       sizeof *mpt_entry - offsetof(struct mthca_mpt_entry, lkey));
-
-	err = mthca_SW2HW_MPT(dev, mpt_entry,
-			      key & (dev->limits.num_mpts - 1),
-			      &status);
-	if (err) {
-		mthca_warn(dev, "SW2HW_MPT failed (%d)\n", err);
-		goto err_out_table;
-	} else if (status) {
-		mthca_warn(dev, "SW2HW_MPT returned status 0x%02x\n",
-			   status);
-		err = -EINVAL;
-		goto err_out_table;
-	}
-
-	kfree(mailbox);
-	return err;
-
-err_out_table:
-	if (mthca_is_memfree(dev))
-		mthca_table_put(dev, dev->mr_table.mpt_table, key);
-
-err_out_mpt_free:
-	mthca_free(&dev->mr_table.mpt_alloc, key);
-	kfree(mailbox);
-	return err;
-}
-
-int mthca_mr_alloc_phys(struct mthca_dev *dev, u32 pd,
-			u64 *buffer_list, int buffer_size_shift,
-			int list_len, u64 iova, u64 total_size,
-			u32 access, struct mthca_mr *mr)
-{
-	void *mailbox;
-	u64 *mtt_entry;
-	struct mthca_mpt_entry *mpt_entry;
-	u32 key;
-	int err = -ENOMEM;
-	u8 status;
-	int i;
-
-	might_sleep();
 	WARN_ON(buffer_size_shift >= 32);
 
 	key = mthca_alloc(&dev->mr_table.mpt_alloc);
@@ -335,75 +356,33 @@ int mthca_mr_alloc_phys(struct mthca_dev *dev, u32 pd,
 			goto err_out_mpt_free;
 	}
 
-	for (i = MTHCA_MTT_SEG_SIZE / 8, mr->order = 0;
-	     i < list_len;
-	     i <<= 1, ++mr->order)
-		; /* nothing */
-
-	mr->first_seg = mthca_alloc_mtt(dev, mr->order,
-				       	&dev->mr_table.mtt_buddy);
-	if (mr->first_seg == -1)
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox)) {
+		err = PTR_ERR(mailbox);
 		goto err_out_table;
-
-	/*
-	 * If list_len is odd, we add one more dummy entry for
-	 * firmware efficiency.
-	 */
-	mailbox = kmalloc(max(sizeof *mpt_entry,
-			      (size_t) 8 * (list_len + (list_len & 1) + 2)) +
-			  MTHCA_CMD_MAILBOX_EXTRA,
-			  GFP_KERNEL);
-	if (!mailbox)
-		goto err_out_free_mtt;
-
-	mtt_entry = MAILBOX_ALIGN(mailbox);
-
-	mtt_entry[0] = cpu_to_be64(dev->mr_table.mtt_base +
-				   mr->first_seg * MTHCA_MTT_SEG_SIZE);
-	mtt_entry[1] = 0;
-	for (i = 0; i < list_len; ++i)
-		mtt_entry[i + 2] = cpu_to_be64(buffer_list[i] |
-					       MTHCA_MTT_FLAG_PRESENT);
-	if (list_len & 1) {
-		mtt_entry[i + 2] = 0;
-		++list_len;
-	}
-
-	if (0) {
-		mthca_dbg(dev, "Dumping MPT entry\n");
-		for (i = 0; i < list_len + 2; ++i)
-			printk(KERN_ERR "[%2d] %016llx\n",
-			       i, (unsigned long long) be64_to_cpu(mtt_entry[i]));
-	}
-
-	err = mthca_WRITE_MTT(dev, mtt_entry, list_len, &status);
-	if (err) {
-		mthca_warn(dev, "WRITE_MTT failed (%d)\n", err);
-		goto err_out_mailbox_free;
-	}
-	if (status) {
-		mthca_warn(dev, "WRITE_MTT returned status 0x%02x\n",
-			   status);
-		err = -EINVAL;
-		goto err_out_mailbox_free;
 	}
-
-	mpt_entry = MAILBOX_ALIGN(mailbox);
+	mpt_entry = mailbox->buf;
 
 	mpt_entry->flags = cpu_to_be32(MTHCA_MPT_FLAG_SW_OWNS     |
 				       MTHCA_MPT_FLAG_MIO         |
 				       MTHCA_MPT_FLAG_REGION      |
 				       access);
+	if (!mr->mtt)
+		mpt_entry->flags |= cpu_to_be32(MTHCA_MPT_FLAG_PHYSICAL);
 
 	mpt_entry->page_size = cpu_to_be32(buffer_size_shift - 12);
 	mpt_entry->key       = cpu_to_be32(key);
 	mpt_entry->pd        = cpu_to_be32(pd);
 	mpt_entry->start     = cpu_to_be64(iova);
 	mpt_entry->length    = cpu_to_be64(total_size);
+
 	memset(&mpt_entry->lkey, 0,
 	       sizeof *mpt_entry - offsetof(struct mthca_mpt_entry, lkey));
-	mpt_entry->mtt_seg   = cpu_to_be64(dev->mr_table.mtt_base +
-					   mr->first_seg * MTHCA_MTT_SEG_SIZE);
+
+	if (mr->mtt)
+		mpt_entry->mtt_seg =
+			cpu_to_be64(dev->mr_table.mtt_base +
+				    mr->mtt->first_seg * MTHCA_MTT_SEG_SIZE);
 
 	if (0) {
 		mthca_dbg(dev, "Dumping MPT entry %08x:\n", mr->ibmr.lkey);
@@ -416,45 +395,70 @@ int mthca_mr_alloc_phys(struct mthca_dev *dev, u32 pd,
 		}
 	}
 
-	err = mthca_SW2HW_MPT(dev, mpt_entry,
+	err = mthca_SW2HW_MPT(dev, mailbox,
 			      key & (dev->limits.num_mpts - 1),
 			      &status);
-	if (err)
+	if (err) {
 		mthca_warn(dev, "SW2HW_MPT failed (%d)\n", err);
-	else if (status) {
+		goto err_out_mailbox;
+	} else if (status) {
 		mthca_warn(dev, "SW2HW_MPT returned status 0x%02x\n",
 			   status);
 		err = -EINVAL;
+		goto err_out_mailbox;
 	}
 
-	kfree(mailbox);
+	mthca_free_mailbox(dev, mailbox);
 	return err;
 
-err_out_mailbox_free:
-	kfree(mailbox);
-
-err_out_free_mtt:
-	mthca_free_mtt(dev, mr->first_seg, mr->order, &dev->mr_table.mtt_buddy);
+err_out_mailbox:
+	mthca_free_mailbox(dev, mailbox);
 
 err_out_table:
-	if (mthca_is_memfree(dev))
-		mthca_table_put(dev, dev->mr_table.mpt_table, key);
+	mthca_table_put(dev, dev->mr_table.mpt_table, key);
 
 err_out_mpt_free:
 	mthca_free(&dev->mr_table.mpt_alloc, key);
 	return err;
 }
 
-/* Free mr or fmr */
-static void mthca_free_region(struct mthca_dev *dev, u32 lkey, int order,
-			      u32 first_seg, struct mthca_buddy *buddy)
+int mthca_mr_alloc_notrans(struct mthca_dev *dev, u32 pd,
+			   u32 access, struct mthca_mr *mr)
 {
-	if (order >= 0)
-		mthca_free_mtt(dev, first_seg, order, buddy);
+	mr->mtt = NULL;
+	return mthca_mr_alloc(dev, pd, 12, 0, ~0ULL, access, mr);
+}
 
-	if (mthca_is_memfree(dev))
-		mthca_table_put(dev, dev->mr_table.mpt_table,
-				arbel_key_to_hw_index(lkey));
+int mthca_mr_alloc_phys(struct mthca_dev *dev, u32 pd,
+			u64 *buffer_list, int buffer_size_shift,
+			int list_len, u64 iova, u64 total_size,
+			u32 access, struct mthca_mr *mr)
+{
+	int err;
+
+	mr->mtt = mthca_alloc_mtt(dev, list_len);
+	if (IS_ERR(mr->mtt))
+		return PTR_ERR(mr->mtt);
+
+	err = mthca_write_mtt(dev, mr->mtt, 0, buffer_list, list_len);
+	if (err) {
+		mthca_free_mtt(dev, mr->mtt);
+		return err;
+	}
+
+	err = mthca_mr_alloc(dev, pd, buffer_size_shift, iova,
+			     total_size, access, mr);
+	if (err)
+		mthca_free_mtt(dev, mr->mtt);
+
+	return err;
+}
+
+/* Free mr or fmr */
+static void mthca_free_region(struct mthca_dev *dev, u32 lkey)
+{
+	mthca_table_put(dev, dev->mr_table.mpt_table,
+			arbel_key_to_hw_index(lkey));
 
 	mthca_free(&dev->mr_table.mpt_alloc, key_to_hw_index(dev, lkey));
 }
@@ -476,15 +480,15 @@ void mthca_free_mr(struct mthca_dev *dev, struct mthca_mr *mr)
 		mthca_warn(dev, "HW2SW_MPT returned status 0x%02x\n",
 			   status);
 
-	mthca_free_region(dev, mr->ibmr.lkey, mr->order, mr->first_seg,
-			  &dev->mr_table.mtt_buddy);
+	mthca_free_region(dev, mr->ibmr.lkey);
+	mthca_free_mtt(dev, mr->mtt);
 }
 
 int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd,
 		    u32 access, struct mthca_fmr *mr)
 {
 	struct mthca_mpt_entry *mpt_entry;
-	void *mailbox;
+	struct mthca_mailbox *mailbox;
 	u64 mtt_seg;
 	u32 key, idx;
 	u8 status;
@@ -522,31 +526,24 @@ int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd,
 		mr->mem.tavor.mpt = dev->mr_table.tavor_fmr.mpt_base +
 		       	sizeof *(mr->mem.tavor.mpt) * idx;
 
-	for (i = MTHCA_MTT_SEG_SIZE / 8, mr->order = 0;
-	     i < list_len;
-	     i <<= 1, ++mr->order)
-		; /* nothing */
-
-	mr->first_seg = mthca_alloc_mtt(dev, mr->order,
-				       	dev->mr_table.fmr_mtt_buddy);
-	if (mr->first_seg == -1)
+	mr->mtt = __mthca_alloc_mtt(dev, list_len, dev->mr_table.fmr_mtt_buddy);
+	if (IS_ERR(mr->mtt))
 		goto err_out_table;
 
-	mtt_seg = mr->first_seg * MTHCA_MTT_SEG_SIZE;
+	mtt_seg = mr->mtt->first_seg * MTHCA_MTT_SEG_SIZE;
 
 	if (mthca_is_memfree(dev)) {
 		mr->mem.arbel.mtts = mthca_table_find(dev->mr_table.mtt_table,
-						      mr->first_seg);
+						      mr->mtt->first_seg);
 		BUG_ON(!mr->mem.arbel.mtts);
 	} else
 		mr->mem.tavor.mtts = dev->mr_table.tavor_fmr.mtt_base + mtt_seg;
 
-	mailbox = kmalloc(sizeof *mpt_entry + MTHCA_CMD_MAILBOX_EXTRA,
-			  GFP_KERNEL);
-	if (!mailbox)
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
 		goto err_out_free_mtt;
 
-	mpt_entry = MAILBOX_ALIGN(mailbox);
+	mpt_entry = mailbox->buf;
 
 	mpt_entry->flags = cpu_to_be32(MTHCA_MPT_FLAG_SW_OWNS     |
 				       MTHCA_MPT_FLAG_MIO         |
@@ -571,7 +568,7 @@ int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd,
 		}
 	}
 
-	err = mthca_SW2HW_MPT(dev, mpt_entry,
+	err = mthca_SW2HW_MPT(dev, mailbox,
 			      key & (dev->limits.num_mpts - 1),
 			      &status);
 	if (err) {
@@ -585,19 +582,17 @@ int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd,
 		goto err_out_mailbox_free;
 	}
 
-	kfree(mailbox);
+	mthca_free_mailbox(dev, mailbox);
 	return 0;
 
 err_out_mailbox_free:
-	kfree(mailbox);
+	mthca_free_mailbox(dev, mailbox);
 
 err_out_free_mtt:
-	mthca_free_mtt(dev, mr->first_seg, mr->order,
-		       dev->mr_table.fmr_mtt_buddy);
+	mthca_free_mtt(dev, mr->mtt);
 
 err_out_table:
-	if (mthca_is_memfree(dev))
-		mthca_table_put(dev, dev->mr_table.mpt_table, key);
+	mthca_table_put(dev, dev->mr_table.mpt_table, key);
 
 err_out_mpt_free:
 	mthca_free(&dev->mr_table.mpt_alloc, mr->ibmr.lkey);
@@ -609,8 +604,9 @@ int mthca_free_fmr(struct mthca_dev *dev, struct mthca_fmr *fmr)
 	if (fmr->maps)
 		return -EBUSY;
 
-	mthca_free_region(dev, fmr->ibmr.lkey, fmr->order, fmr->first_seg,
-			  dev->mr_table.fmr_mtt_buddy);
+	mthca_free_region(dev, fmr->ibmr.lkey);
+	mthca_free_mtt(dev, fmr->mtt);
+
 	return 0;
 }
 
@@ -826,7 +822,8 @@ int __devinit mthca_init_mr_table(struct mthca_dev *dev)
 	if (dev->limits.reserved_mtts) {
 		i = fls(dev->limits.reserved_mtts - 1);
 
-		if (mthca_alloc_mtt(dev, i, dev->mr_table.fmr_mtt_buddy) == -1) {
+		if (mthca_alloc_mtt_range(dev, i,
+					  dev->mr_table.fmr_mtt_buddy) == -1) {
 			mthca_warn(dev, "MTT table of order %d is too small.\n",
 				  dev->mr_table.fmr_mtt_buddy->max_order);
 			err = -ENOMEM;
diff --git a/drivers/infiniband/hw/mthca/mthca_pd.c b/drivers/infiniband/hw/mthca/mthca_pd.c
index ea66847e4ea3..c2c899844e98 100644
--- a/drivers/infiniband/hw/mthca/mthca_pd.c
+++ b/drivers/infiniband/hw/mthca/mthca_pd.c
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  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
@@ -37,23 +38,27 @@
 
 #include "mthca_dev.h"
 
-int mthca_pd_alloc(struct mthca_dev *dev, struct mthca_pd *pd)
+int mthca_pd_alloc(struct mthca_dev *dev, int privileged, struct mthca_pd *pd)
 {
-	int err;
+	int err = 0;
 
 	might_sleep();
 
+	pd->privileged = privileged;
+
 	atomic_set(&pd->sqp_count, 0);
 	pd->pd_num = mthca_alloc(&dev->pd_table.alloc);
 	if (pd->pd_num == -1)
 		return -ENOMEM;
 
-	err = mthca_mr_alloc_notrans(dev, pd->pd_num,
-				     MTHCA_MPT_FLAG_LOCAL_READ |
-				     MTHCA_MPT_FLAG_LOCAL_WRITE,
-				     &pd->ntmr);
-	if (err)
-		mthca_free(&dev->pd_table.alloc, pd->pd_num);
+	if (privileged) {
+		err = mthca_mr_alloc_notrans(dev, pd->pd_num,
+					     MTHCA_MPT_FLAG_LOCAL_READ |
+					     MTHCA_MPT_FLAG_LOCAL_WRITE,
+					     &pd->ntmr);
+		if (err)
+			mthca_free(&dev->pd_table.alloc, pd->pd_num);
+	}
 
 	return err;
 }
@@ -61,7 +66,8 @@ int mthca_pd_alloc(struct mthca_dev *dev, struct mthca_pd *pd)
 void mthca_pd_free(struct mthca_dev *dev, struct mthca_pd *pd)
 {
 	might_sleep();
-	mthca_free_mr(dev, &pd->ntmr);
+	if (pd->privileged)
+		mthca_free_mr(dev, &pd->ntmr);
 	mthca_free(&dev->pd_table.alloc, pd->pd_num);
 }
 
diff --git a/drivers/infiniband/hw/mthca/mthca_provider.c b/drivers/infiniband/hw/mthca/mthca_provider.c
index 159f4e6c312d..81919a7b4935 100644
--- a/drivers/infiniband/hw/mthca/mthca_provider.c
+++ b/drivers/infiniband/hw/mthca/mthca_provider.c
@@ -1,5 +1,7 @@
 /*
  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
+ * Copyright (c) 2005 Cisco Systems. 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
@@ -33,9 +35,12 @@
  */
 
 #include <ib_smi.h>
+#include <linux/mm.h>
 
 #include "mthca_dev.h"
 #include "mthca_cmd.h"
+#include "mthca_user.h"
+#include "mthca_memfree.h"
 
 static int mthca_query_device(struct ib_device *ibdev,
 			      struct ib_device_attr *props)
@@ -52,7 +57,7 @@ static int mthca_query_device(struct ib_device *ibdev,
 	if (!in_mad || !out_mad)
 		goto out;
 
-	memset(props, 0, sizeof props);
+	memset(props, 0, sizeof *props);
 
 	props->fw_ver              = mdev->fw_ver;
 
@@ -283,7 +288,78 @@ static int mthca_query_gid(struct ib_device *ibdev, u8 port,
 	return err;
 }
 
-static struct ib_pd *mthca_alloc_pd(struct ib_device *ibdev)
+static struct ib_ucontext *mthca_alloc_ucontext(struct ib_device *ibdev,
+						struct ib_udata *udata)
+{
+	struct mthca_alloc_ucontext_resp uresp;
+	struct mthca_ucontext           *context;
+	int                              err;
+
+	memset(&uresp, 0, sizeof uresp);
+
+	uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
+	if (mthca_is_memfree(to_mdev(ibdev)))
+		uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
+	else
+		uresp.uarc_size = 0;
+
+	context = kmalloc(sizeof *context, GFP_KERNEL);
+	if (!context)
+		return ERR_PTR(-ENOMEM);
+
+	err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
+	if (err) {
+		kfree(context);
+		return ERR_PTR(err);
+	}
+
+	context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
+	if (IS_ERR(context->db_tab)) {
+		err = PTR_ERR(context->db_tab);
+		mthca_uar_free(to_mdev(ibdev), &context->uar);
+		kfree(context);
+		return ERR_PTR(err);
+	}
+
+	if (ib_copy_to_udata(udata, &uresp, sizeof uresp)) {
+		mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
+		mthca_uar_free(to_mdev(ibdev), &context->uar);
+		kfree(context);
+		return ERR_PTR(-EFAULT);
+	}
+
+	return &context->ibucontext;
+}
+
+static int mthca_dealloc_ucontext(struct ib_ucontext *context)
+{
+	mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
+				  to_mucontext(context)->db_tab);
+	mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
+	kfree(to_mucontext(context));
+
+	return 0;
+}
+
+static int mthca_mmap_uar(struct ib_ucontext *context,
+			  struct vm_area_struct *vma)
+{
+	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
+		return -EINVAL;
+
+	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
+
+	if (io_remap_pfn_range(vma, vma->vm_start,
+			       to_mucontext(context)->uar.pfn,
+			       PAGE_SIZE, vma->vm_page_prot))
+		return -EAGAIN;
+
+	return 0;
+}
+
+static struct ib_pd *mthca_alloc_pd(struct ib_device *ibdev,
+				    struct ib_ucontext *context,
+				    struct ib_udata *udata)
 {
 	struct mthca_pd *pd;
 	int err;
@@ -292,12 +368,20 @@ static struct ib_pd *mthca_alloc_pd(struct ib_device *ibdev)
 	if (!pd)
 		return ERR_PTR(-ENOMEM);
 
-	err = mthca_pd_alloc(to_mdev(ibdev), pd);
+	err = mthca_pd_alloc(to_mdev(ibdev), !context, pd);
 	if (err) {
 		kfree(pd);
 		return ERR_PTR(err);
 	}
 
+	if (context) {
+		if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
+			mthca_pd_free(to_mdev(ibdev), pd);
+			kfree(pd);
+			return ERR_PTR(-EFAULT);
+		}
+	}
+
 	return &pd->ibpd;
 }
 
@@ -337,8 +421,10 @@ static int mthca_ah_destroy(struct ib_ah *ah)
 }
 
 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
-				     struct ib_qp_init_attr *init_attr)
+				     struct ib_qp_init_attr *init_attr,
+				     struct ib_udata *udata)
 {
+	struct mthca_create_qp ucmd;
 	struct mthca_qp *qp;
 	int err;
 
@@ -347,41 +433,82 @@ static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
 	case IB_QPT_UC:
 	case IB_QPT_UD:
 	{
+		struct mthca_ucontext *context;
+
 		qp = kmalloc(sizeof *qp, GFP_KERNEL);
 		if (!qp)
 			return ERR_PTR(-ENOMEM);
 
-		qp->sq.max    = init_attr->cap.max_send_wr;
-		qp->rq.max    = init_attr->cap.max_recv_wr;
-		qp->sq.max_gs = init_attr->cap.max_send_sge;
-		qp->rq.max_gs = init_attr->cap.max_recv_sge;
+		if (pd->uobject) {
+			context = to_mucontext(pd->uobject->context);
+
+			if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
+				return ERR_PTR(-EFAULT);
+
+			err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
+						context->db_tab,
+						ucmd.sq_db_index, ucmd.sq_db_page);
+			if (err) {
+				kfree(qp);
+				return ERR_PTR(err);
+			}
+
+			err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
+						context->db_tab,
+						ucmd.rq_db_index, ucmd.rq_db_page);
+			if (err) {
+				mthca_unmap_user_db(to_mdev(pd->device),
+						    &context->uar,
+						    context->db_tab,
+						    ucmd.sq_db_index);
+				kfree(qp);
+				return ERR_PTR(err);
+			}
+
+			qp->mr.ibmr.lkey = ucmd.lkey;
+			qp->sq.db_index  = ucmd.sq_db_index;
+			qp->rq.db_index  = ucmd.rq_db_index;
+		}
 
 		err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
 				     to_mcq(init_attr->send_cq),
 				     to_mcq(init_attr->recv_cq),
 				     init_attr->qp_type, init_attr->sq_sig_type,
-				     qp);
+				     &init_attr->cap, qp);
+
+		if (err && pd->uobject) {
+			context = to_mucontext(pd->uobject->context);
+
+			mthca_unmap_user_db(to_mdev(pd->device),
+					    &context->uar,
+					    context->db_tab,
+					    ucmd.sq_db_index);
+			mthca_unmap_user_db(to_mdev(pd->device),
+					    &context->uar,
+					    context->db_tab,
+					    ucmd.rq_db_index);
+		}
+
 		qp->ibqp.qp_num = qp->qpn;
 		break;
 	}
 	case IB_QPT_SMI:
 	case IB_QPT_GSI:
 	{
+		/* Don't allow userspace to create special QPs */
+		if (pd->uobject)
+			return ERR_PTR(-EINVAL);
+
 		qp = kmalloc(sizeof (struct mthca_sqp), GFP_KERNEL);
 		if (!qp)
 			return ERR_PTR(-ENOMEM);
 
-		qp->sq.max    = init_attr->cap.max_send_wr;
-		qp->rq.max    = init_attr->cap.max_recv_wr;
-		qp->sq.max_gs = init_attr->cap.max_send_sge;
-		qp->rq.max_gs = init_attr->cap.max_recv_sge;
-
 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
 
 		err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
 				      to_mcq(init_attr->send_cq),
 				      to_mcq(init_attr->recv_cq),
-				      init_attr->sq_sig_type,
+				      init_attr->sq_sig_type, &init_attr->cap,
 				      qp->ibqp.qp_num, init_attr->port_num,
 				      to_msqp(qp));
 		break;
@@ -396,42 +523,115 @@ static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
 		return ERR_PTR(err);
 	}
 
-        init_attr->cap.max_inline_data = 0;
+	init_attr->cap.max_inline_data = 0;
+	init_attr->cap.max_send_wr     = qp->sq.max;
+	init_attr->cap.max_recv_wr     = qp->rq.max;
+	init_attr->cap.max_send_sge    = qp->sq.max_gs;
+	init_attr->cap.max_recv_sge    = qp->rq.max_gs;
 
 	return &qp->ibqp;
 }
 
 static int mthca_destroy_qp(struct ib_qp *qp)
 {
+	if (qp->uobject) {
+		mthca_unmap_user_db(to_mdev(qp->device),
+				    &to_mucontext(qp->uobject->context)->uar,
+				    to_mucontext(qp->uobject->context)->db_tab,
+				    to_mqp(qp)->sq.db_index);
+		mthca_unmap_user_db(to_mdev(qp->device),
+				    &to_mucontext(qp->uobject->context)->uar,
+				    to_mucontext(qp->uobject->context)->db_tab,
+				    to_mqp(qp)->rq.db_index);
+	}
 	mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
 	kfree(qp);
 	return 0;
 }
 
-static struct ib_cq *mthca_create_cq(struct ib_device *ibdev, int entries)
+static struct ib_cq *mthca_create_cq(struct ib_device *ibdev, int entries,
+				     struct ib_ucontext *context,
+				     struct ib_udata *udata)
 {
+	struct mthca_create_cq ucmd;
 	struct mthca_cq *cq;
 	int nent;
 	int err;
 
+	if (context) {
+		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
+			return ERR_PTR(-EFAULT);
+
+		err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
+					to_mucontext(context)->db_tab,
+					ucmd.set_db_index, ucmd.set_db_page);
+		if (err)
+			return ERR_PTR(err);
+
+		err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
+					to_mucontext(context)->db_tab,
+					ucmd.arm_db_index, ucmd.arm_db_page);
+		if (err)
+			goto err_unmap_set;
+	}
+
 	cq = kmalloc(sizeof *cq, GFP_KERNEL);
-	if (!cq)
-		return ERR_PTR(-ENOMEM);
+	if (!cq) {
+		err = -ENOMEM;
+		goto err_unmap_arm;
+	}
+
+	if (context) {
+		cq->mr.ibmr.lkey    = ucmd.lkey;
+		cq->set_ci_db_index = ucmd.set_db_index;
+		cq->arm_db_index    = ucmd.arm_db_index;
+	}
 
 	for (nent = 1; nent <= entries; nent <<= 1)
 		; /* nothing */
 
-	err = mthca_init_cq(to_mdev(ibdev), nent, cq);
-	if (err) {
-		kfree(cq);
-		cq = ERR_PTR(err);
+	err = mthca_init_cq(to_mdev(ibdev), nent,
+			    context ? to_mucontext(context) : NULL,
+			    context ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
+			    cq);
+	if (err)
+		goto err_free;
+
+	if (context && ib_copy_to_udata(udata, &cq->cqn, sizeof (__u32))) {
+		mthca_free_cq(to_mdev(ibdev), cq);
+		goto err_free;
 	}
 
 	return &cq->ibcq;
+
+err_free:
+	kfree(cq);
+
+err_unmap_arm:
+	if (context)
+		mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
+				    to_mucontext(context)->db_tab, ucmd.arm_db_index);
+
+err_unmap_set:
+	if (context)
+		mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
+				    to_mucontext(context)->db_tab, ucmd.set_db_index);
+
+	return ERR_PTR(err);
 }
 
 static int mthca_destroy_cq(struct ib_cq *cq)
 {
+	if (cq->uobject) {
+		mthca_unmap_user_db(to_mdev(cq->device),
+				    &to_mucontext(cq->uobject->context)->uar,
+				    to_mucontext(cq->uobject->context)->db_tab,
+				    to_mcq(cq)->arm_db_index);
+		mthca_unmap_user_db(to_mdev(cq->device),
+				    &to_mucontext(cq->uobject->context)->uar,
+				    to_mucontext(cq->uobject->context)->db_tab,
+				    to_mcq(cq)->set_ci_db_index);
+	}
 	mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
 	kfree(cq);
 
@@ -558,6 +758,7 @@ static struct ib_mr *mthca_reg_phys_mr(struct ib_pd       *pd,
 				  convert_access(acc), mr);
 
 	if (err) {
+		kfree(page_list);
 		kfree(mr);
 		return ERR_PTR(err);
 	}
@@ -566,6 +767,87 @@ static struct ib_mr *mthca_reg_phys_mr(struct ib_pd       *pd,
 	return &mr->ibmr;
 }
 
+static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, struct ib_umem *region,
+				       int acc, struct ib_udata *udata)
+{
+	struct mthca_dev *dev = to_mdev(pd->device);
+	struct ib_umem_chunk *chunk;
+	struct mthca_mr *mr;
+	u64 *pages;
+	int shift, n, len;
+	int i, j, k;
+	int err = 0;
+
+	shift = ffs(region->page_size) - 1;
+
+	mr = kmalloc(sizeof *mr, GFP_KERNEL);
+	if (!mr)
+		return ERR_PTR(-ENOMEM);
+
+	n = 0;
+	list_for_each_entry(chunk, &region->chunk_list, list)
+		n += chunk->nents;
+
+	mr->mtt = mthca_alloc_mtt(dev, n);
+	if (IS_ERR(mr->mtt)) {
+		err = PTR_ERR(mr->mtt);
+		goto err;
+	}
+
+	pages = (u64 *) __get_free_page(GFP_KERNEL);
+	if (!pages) {
+		err = -ENOMEM;
+		goto err_mtt;
+	}
+
+	i = n = 0;
+
+	list_for_each_entry(chunk, &region->chunk_list, list)
+		for (j = 0; j < chunk->nmap; ++j) {
+			len = sg_dma_len(&chunk->page_list[j]) >> shift;
+			for (k = 0; k < len; ++k) {
+				pages[i++] = sg_dma_address(&chunk->page_list[j]) +
+					region->page_size * k;
+				/*
+				 * Be friendly to WRITE_MTT command
+				 * and leave two empty slots for the
+				 * index and reserved fields of the
+				 * mailbox.
+				 */
+				if (i == PAGE_SIZE / sizeof (u64) - 2) {
+					err = mthca_write_mtt(dev, mr->mtt,
+							      n, pages, i);
+					if (err)
+						goto mtt_done;
+					n += i;
+					i = 0;
+				}
+			}
+		}
+
+	if (i)
+		err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
+mtt_done:
+	free_page((unsigned long) pages);
+	if (err)
+		goto err_mtt;
+
+	err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, shift, region->virt_base,
+			     region->length, convert_access(acc), mr);
+
+	if (err)
+		goto err_mtt;
+
+	return &mr->ibmr;
+
+err_mtt:
+	mthca_free_mtt(dev, mr->mtt);
+
+err:
+	kfree(mr);
+	return ERR_PTR(err);
+}
+
 static int mthca_dereg_mr(struct ib_mr *mr)
 {
 	struct mthca_mr *mmr = to_mmr(mr);
@@ -690,6 +972,8 @@ int mthca_register_device(struct mthca_dev *dev)
 	int i;
 
 	strlcpy(dev->ib_dev.name, "mthca%d", IB_DEVICE_NAME_MAX);
+	dev->ib_dev.owner                = THIS_MODULE;
+
 	dev->ib_dev.node_type            = IB_NODE_CA;
 	dev->ib_dev.phys_port_cnt        = dev->limits.num_ports;
 	dev->ib_dev.dma_device           = &dev->pdev->dev;
@@ -699,6 +983,9 @@ int mthca_register_device(struct mthca_dev *dev)
 	dev->ib_dev.modify_port          = mthca_modify_port;
 	dev->ib_dev.query_pkey           = mthca_query_pkey;
 	dev->ib_dev.query_gid            = mthca_query_gid;
+	dev->ib_dev.alloc_ucontext       = mthca_alloc_ucontext;
+	dev->ib_dev.dealloc_ucontext     = mthca_dealloc_ucontext;
+	dev->ib_dev.mmap                 = mthca_mmap_uar;
 	dev->ib_dev.alloc_pd             = mthca_alloc_pd;
 	dev->ib_dev.dealloc_pd           = mthca_dealloc_pd;
 	dev->ib_dev.create_ah            = mthca_ah_create;
@@ -711,6 +998,7 @@ int mthca_register_device(struct mthca_dev *dev)
 	dev->ib_dev.poll_cq              = mthca_poll_cq;
 	dev->ib_dev.get_dma_mr           = mthca_get_dma_mr;
 	dev->ib_dev.reg_phys_mr          = mthca_reg_phys_mr;
+	dev->ib_dev.reg_user_mr          = mthca_reg_user_mr;
 	dev->ib_dev.dereg_mr             = mthca_dereg_mr;
 
 	if (dev->mthca_flags & MTHCA_FLAG_FMR) {
diff --git a/drivers/infiniband/hw/mthca/mthca_provider.h b/drivers/infiniband/hw/mthca/mthca_provider.h
index 619710f95a87..1d032791cc8b 100644
--- a/drivers/infiniband/hw/mthca/mthca_provider.h
+++ b/drivers/infiniband/hw/mthca/mthca_provider.h
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  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
@@ -54,18 +55,26 @@ struct mthca_uar {
 	int           index;
 };
 
+struct mthca_user_db_table;
+
+struct mthca_ucontext {
+	struct ib_ucontext          ibucontext;
+	struct mthca_uar            uar;
+	struct mthca_user_db_table *db_tab;
+};
+
+struct mthca_mtt;
+
 struct mthca_mr {
-	struct ib_mr ibmr;
-	int order;
-	u32 first_seg;
+	struct ib_mr      ibmr;
+	struct mthca_mtt *mtt;
 };
 
 struct mthca_fmr {
-	struct ib_fmr ibmr;
+	struct ib_fmr      ibmr;
 	struct ib_fmr_attr attr;
-	int order;
-	u32 first_seg;
-	int maps;
+	struct mthca_mtt  *mtt;
+	int                maps;
 	union {
 		struct {
 			struct mthca_mpt_entry __iomem *mpt;
@@ -83,6 +92,7 @@ struct mthca_pd {
 	u32             pd_num;
 	atomic_t        sqp_count;
 	struct mthca_mr ntmr;
+	int             privileged;
 };
 
 struct mthca_eq {
@@ -167,6 +177,7 @@ struct mthca_cq {
 	int                    cqn;
 	u32                    cons_index;
 	int                    is_direct;
+	int                    is_kernel;
 
 	/* Next fields are Arbel only */
 	int                    set_ci_db_index;
@@ -236,6 +247,11 @@ struct mthca_sqp {
 	dma_addr_t      header_dma;
 };
 
+static inline struct mthca_ucontext *to_mucontext(struct ib_ucontext *ibucontext)
+{
+	return container_of(ibucontext, struct mthca_ucontext, ibucontext);
+}
+
 static inline struct mthca_fmr *to_mfmr(struct ib_fmr *ibmr)
 {
 	return container_of(ibmr, struct mthca_fmr, ibmr);
diff --git a/drivers/infiniband/hw/mthca/mthca_qp.c b/drivers/infiniband/hw/mthca/mthca_qp.c
index ca73bab11a02..f7126b14d5ae 100644
--- a/drivers/infiniband/hw/mthca/mthca_qp.c
+++ b/drivers/infiniband/hw/mthca/mthca_qp.c
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems. 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
@@ -46,7 +47,9 @@ enum {
 	MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
 	MTHCA_ACK_REQ_FREQ       = 10,
 	MTHCA_FLIGHT_LIMIT       = 9,
-	MTHCA_UD_HEADER_SIZE     = 72 /* largest UD header possible */
+	MTHCA_UD_HEADER_SIZE     = 72, /* largest UD header possible */
+	MTHCA_INLINE_HEADER_SIZE = 4,  /* data segment overhead for inline */
+	MTHCA_INLINE_CHUNK_SIZE  = 16  /* inline data segment chunk */
 };
 
 enum {
@@ -357,6 +360,9 @@ static const struct {
 				[UD]  = (IB_QP_PKEY_INDEX |
 					 IB_QP_PORT       |
 					 IB_QP_QKEY),
+				[UC]  = (IB_QP_PKEY_INDEX |
+					 IB_QP_PORT       |
+					 IB_QP_ACCESS_FLAGS),
 				[RC]  = (IB_QP_PKEY_INDEX |
 					 IB_QP_PORT       |
 					 IB_QP_ACCESS_FLAGS),
@@ -378,6 +384,9 @@ static const struct {
 				[UD]  = (IB_QP_PKEY_INDEX |
 					 IB_QP_PORT       |
 					 IB_QP_QKEY),
+				[UC]  = (IB_QP_PKEY_INDEX |
+					 IB_QP_PORT       |
+					 IB_QP_ACCESS_FLAGS),
 				[RC]  = (IB_QP_PKEY_INDEX |
 					 IB_QP_PORT       |
 					 IB_QP_ACCESS_FLAGS),
@@ -388,6 +397,11 @@ static const struct {
 		[IB_QPS_RTR]   = {
 			.trans = MTHCA_TRANS_INIT2RTR,
 			.req_param = {
+				[UC]  = (IB_QP_AV                  |
+					 IB_QP_PATH_MTU            |
+					 IB_QP_DEST_QPN            |
+					 IB_QP_RQ_PSN              |
+					 IB_QP_MAX_DEST_RD_ATOMIC),
 				[RC]  = (IB_QP_AV                  |
 					 IB_QP_PATH_MTU            |
 					 IB_QP_DEST_QPN            |
@@ -398,6 +412,9 @@ static const struct {
 			.opt_param = {
 				[UD]  = (IB_QP_PKEY_INDEX |
 					 IB_QP_QKEY),
+				[UC]  = (IB_QP_ALT_PATH     |
+					 IB_QP_ACCESS_FLAGS |
+					 IB_QP_PKEY_INDEX),
 				[RC]  = (IB_QP_ALT_PATH     |
 					 IB_QP_ACCESS_FLAGS |
 					 IB_QP_PKEY_INDEX),
@@ -413,6 +430,8 @@ static const struct {
 			.trans = MTHCA_TRANS_RTR2RTS,
 			.req_param = {
 				[UD]  = IB_QP_SQ_PSN,
+				[UC]  = (IB_QP_SQ_PSN            |
+					 IB_QP_MAX_QP_RD_ATOMIC),
 				[RC]  = (IB_QP_TIMEOUT           |
 					 IB_QP_RETRY_CNT         |
 					 IB_QP_RNR_RETRY         |
@@ -423,6 +442,11 @@ static const struct {
 			.opt_param = {
 				[UD]  = (IB_QP_CUR_STATE             |
 					 IB_QP_QKEY),
+				[UC]  = (IB_QP_CUR_STATE             |
+					 IB_QP_ALT_PATH              |
+					 IB_QP_ACCESS_FLAGS          |
+					 IB_QP_PKEY_INDEX            |
+					 IB_QP_PATH_MIG_STATE),
 				[RC]  = (IB_QP_CUR_STATE             |
 					 IB_QP_ALT_PATH              |
 					 IB_QP_ACCESS_FLAGS          |
@@ -442,6 +466,9 @@ static const struct {
 			.opt_param = {
 				[UD]  = (IB_QP_CUR_STATE             |
 					 IB_QP_QKEY),
+				[UC]  = (IB_QP_ACCESS_FLAGS          |
+					 IB_QP_ALT_PATH              |
+					 IB_QP_PATH_MIG_STATE),
 				[RC]  = (IB_QP_ACCESS_FLAGS          |
 					 IB_QP_ALT_PATH              |
 					 IB_QP_PATH_MIG_STATE        |
@@ -462,6 +489,10 @@ static const struct {
 			.opt_param = {
 				[UD]  = (IB_QP_CUR_STATE             |
 					 IB_QP_QKEY),
+				[UC]  = (IB_QP_CUR_STATE             |
+					 IB_QP_ALT_PATH              |
+					 IB_QP_ACCESS_FLAGS          |
+					 IB_QP_PATH_MIG_STATE),
 				[RC]  = (IB_QP_CUR_STATE             |
 					 IB_QP_ALT_PATH              |
 					 IB_QP_ACCESS_FLAGS          |
@@ -476,6 +507,14 @@ static const struct {
 			.opt_param = {
 				[UD]  = (IB_QP_PKEY_INDEX            |
 					 IB_QP_QKEY),
+				[UC]  = (IB_QP_AV                    |
+					 IB_QP_MAX_QP_RD_ATOMIC      |
+					 IB_QP_MAX_DEST_RD_ATOMIC    |
+					 IB_QP_CUR_STATE             |
+					 IB_QP_ALT_PATH              |
+					 IB_QP_ACCESS_FLAGS          |
+					 IB_QP_PKEY_INDEX            |
+					 IB_QP_PATH_MIG_STATE),
 				[RC]  = (IB_QP_AV                    |
 					 IB_QP_TIMEOUT               |
 					 IB_QP_RETRY_CNT             |
@@ -501,6 +540,7 @@ static const struct {
 			.opt_param = {
 				[UD]  = (IB_QP_CUR_STATE             |
 					 IB_QP_QKEY),
+				[UC]  = (IB_QP_CUR_STATE),
 				[RC]  = (IB_QP_CUR_STATE             |
 					 IB_QP_MIN_RNR_TIMER),
 				[MLX] = (IB_QP_CUR_STATE             |
@@ -552,7 +592,7 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
 	struct mthca_dev *dev = to_mdev(ibqp->device);
 	struct mthca_qp *qp = to_mqp(ibqp);
 	enum ib_qp_state cur_state, new_state;
-	void *mailbox = NULL;
+	struct mthca_mailbox *mailbox;
 	struct mthca_qp_param *qp_param;
 	struct mthca_qp_context *qp_context;
 	u32 req_param, opt_param;
@@ -609,10 +649,10 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
 		return -EINVAL;
 	}
 
-	mailbox = kmalloc(sizeof (*qp_param) + MTHCA_CMD_MAILBOX_EXTRA, GFP_KERNEL);
-	if (!mailbox)
-		return -ENOMEM;
-	qp_param = MAILBOX_ALIGN(mailbox);
+	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
+	if (IS_ERR(mailbox))
+		return PTR_ERR(mailbox);
+	qp_param = mailbox->buf;
 	qp_context = &qp_param->context;
 	memset(qp_param, 0, sizeof *qp_param);
 
@@ -652,7 +692,11 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
 
 	/* leave arbel_sched_queue as 0 */
 
-	qp_context->usr_page   = cpu_to_be32(dev->driver_uar.index);
+	if (qp->ibqp.uobject)
+		qp_context->usr_page =
+			cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
+	else
+		qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
 	qp_context->local_qpn  = cpu_to_be32(qp->qpn);
 	if (attr_mask & IB_QP_DEST_QPN) {
 		qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
@@ -683,7 +727,7 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
 	if (attr_mask & IB_QP_AV) {
 		qp_context->pri_path.g_mylmc     = attr->ah_attr.src_path_bits & 0x7f;
 		qp_context->pri_path.rlid        = cpu_to_be16(attr->ah_attr.dlid);
-		qp_context->pri_path.static_rate = (!!attr->ah_attr.static_rate) << 3;
+		qp_context->pri_path.static_rate = !!attr->ah_attr.static_rate;
 		if (attr->ah_attr.ah_flags & IB_AH_GRH) {
 			qp_context->pri_path.g_mylmc |= 1 << 7;
 			qp_context->pri_path.mgid_index = attr->ah_attr.grh.sgid_index;
@@ -724,9 +768,9 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
 	}
 
-	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
-		qp_context->params1 |= cpu_to_be32(min(attr->max_dest_rd_atomic ?
-						       ffs(attr->max_dest_rd_atomic) - 1 : 0,
+	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
+		qp_context->params1 |= cpu_to_be32(min(attr->max_rd_atomic ?
+						       ffs(attr->max_rd_atomic) - 1 : 0,
 						       7) << 21);
 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
 	}
@@ -764,10 +808,10 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
 		qp->atomic_rd_en = attr->qp_access_flags;
 	}
 
-	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
+	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
 		u8 rra_max;
 
-		if (qp->resp_depth && !attr->max_rd_atomic) {
+		if (qp->resp_depth && !attr->max_dest_rd_atomic) {
 			/*
 			 * Lowering our responder resources to zero.
 			 * Turn off RDMA/atomics as responder.
@@ -778,7 +822,7 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
 								MTHCA_QP_OPTPAR_RAE);
 		}
 
-		if (!qp->resp_depth && attr->max_rd_atomic) {
+		if (!qp->resp_depth && attr->max_dest_rd_atomic) {
 			/*
 			 * Increasing our responder resources from
 			 * zero.  Turn on RDMA/atomics as appropriate.
@@ -799,7 +843,7 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
 		}
 
 		for (rra_max = 0;
-		     1 << rra_max < attr->max_rd_atomic &&
+		     1 << rra_max < attr->max_dest_rd_atomic &&
 			     rra_max < dev->qp_table.rdb_shift;
 		     ++rra_max)
 			; /* nothing */
@@ -807,7 +851,7 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
 		qp_context->params2      |= cpu_to_be32(rra_max << 21);
 		qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
 
-		qp->resp_depth = attr->max_rd_atomic;
+		qp->resp_depth = attr->max_dest_rd_atomic;
 	}
 
 	qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
@@ -835,7 +879,7 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
 	}
 
 	err = mthca_MODIFY_QP(dev, state_table[cur_state][new_state].trans,
-			      qp->qpn, 0, qp_param, 0, &status);
+			      qp->qpn, 0, mailbox, 0, &status);
 	if (status) {
 		mthca_warn(dev, "modify QP %d returned status %02x.\n",
 			   state_table[cur_state][new_state].trans, status);
@@ -845,7 +889,7 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
 	if (!err)
 		qp->state = new_state;
 
-	kfree(mailbox);
+	mthca_free_mailbox(dev, mailbox);
 
 	if (is_sqp(dev, qp))
 		store_attrs(to_msqp(qp), attr, attr_mask);
@@ -917,6 +961,15 @@ static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
 
 	qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
 				    1 << qp->sq.wqe_shift);
+
+	/*
+	 * If this is a userspace QP, we don't actually have to
+	 * allocate anything.  All we need is to calculate the WQE
+	 * sizes and the send_wqe_offset, so we're done now.
+	 */
+	if (pd->ibpd.uobject)
+		return 0;
+
 	size = PAGE_ALIGN(qp->send_wqe_offset +
 			  (qp->sq.max << qp->sq.wqe_shift));
 
@@ -934,7 +987,8 @@ static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
 			mthca_dbg(dev, "Creating direct QP of size %d (shift %d)\n",
 				  size, shift);
 
-		qp->queue.direct.buf = pci_alloc_consistent(dev->pdev, size, &t);
+		qp->queue.direct.buf = dma_alloc_coherent(&dev->pdev->dev, size,
+							  &t, GFP_KERNEL);
 		if (!qp->queue.direct.buf)
 			goto err_out;
 
@@ -973,7 +1027,8 @@ static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
 
 		for (i = 0; i < npages; ++i) {
 			qp->queue.page_list[i].buf =
-				pci_alloc_consistent(dev->pdev, PAGE_SIZE, &t);
+				dma_alloc_coherent(&dev->pdev->dev, PAGE_SIZE,
+						   &t, GFP_KERNEL);
 			if (!qp->queue.page_list[i].buf)
 				goto err_out_free;
 
@@ -996,16 +1051,15 @@ static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
 
  err_out_free:
 	if (qp->is_direct) {
-		pci_free_consistent(dev->pdev, size,
-				    qp->queue.direct.buf,
-				    pci_unmap_addr(&qp->queue.direct, mapping));
+		dma_free_coherent(&dev->pdev->dev, size, qp->queue.direct.buf,
+				  pci_unmap_addr(&qp->queue.direct, mapping));
 	} else
 		for (i = 0; i < npages; ++i) {
 			if (qp->queue.page_list[i].buf)
-				pci_free_consistent(dev->pdev, PAGE_SIZE,
-						    qp->queue.page_list[i].buf,
-						    pci_unmap_addr(&qp->queue.page_list[i],
-								   mapping));
+				dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
+						  qp->queue.page_list[i].buf,
+						  pci_unmap_addr(&qp->queue.page_list[i],
+								 mapping));
 
 		}
 
@@ -1015,10 +1069,32 @@ static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
 	return err;
 }
 
-static int mthca_alloc_memfree(struct mthca_dev *dev,
+static void mthca_free_wqe_buf(struct mthca_dev *dev,
 			       struct mthca_qp *qp)
 {
-	int ret = 0;
+	int i;
+	int size = PAGE_ALIGN(qp->send_wqe_offset +
+			      (qp->sq.max << qp->sq.wqe_shift));
+
+	if (qp->is_direct) {
+		dma_free_coherent(&dev->pdev->dev, size, qp->queue.direct.buf,
+				  pci_unmap_addr(&qp->queue.direct, mapping));
+	} else {
+		for (i = 0; i < size / PAGE_SIZE; ++i) {
+			dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
+					  qp->queue.page_list[i].buf,
+					  pci_unmap_addr(&qp->queue.page_list[i],
+							 mapping));
+		}
+	}
+
+	kfree(qp->wrid);
+}
+
+static int mthca_map_memfree(struct mthca_dev *dev,
+			     struct mthca_qp *qp)
+{
+	int ret;
 
 	if (mthca_is_memfree(dev)) {
 		ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
@@ -1029,35 +1105,15 @@ static int mthca_alloc_memfree(struct mthca_dev *dev,
 		if (ret)
 			goto err_qpc;
 
-		ret = mthca_table_get(dev, dev->qp_table.rdb_table,
-				      qp->qpn << dev->qp_table.rdb_shift);
-		if (ret)
-			goto err_eqpc;
-
-		qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
-						 qp->qpn, &qp->rq.db);
-		if (qp->rq.db_index < 0) {
-			ret = -ENOMEM;
-			goto err_rdb;
-		}
+ 		ret = mthca_table_get(dev, dev->qp_table.rdb_table,
+ 				      qp->qpn << dev->qp_table.rdb_shift);
+ 		if (ret)
+ 			goto err_eqpc;
 
-		qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
-						 qp->qpn, &qp->sq.db);
-		if (qp->sq.db_index < 0) {
-			ret = -ENOMEM;
-			goto err_rq_db;
-		}
 	}
 
 	return 0;
 
-err_rq_db:
-	mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
-
-err_rdb:
-	mthca_table_put(dev, dev->qp_table.rdb_table,
-			qp->qpn << dev->qp_table.rdb_shift);
-
 err_eqpc:
 	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
 
@@ -1067,16 +1123,41 @@ err_qpc:
 	return ret;
 }
 
+static void mthca_unmap_memfree(struct mthca_dev *dev,
+				struct mthca_qp *qp)
+{
+	mthca_table_put(dev, dev->qp_table.rdb_table,
+			qp->qpn << dev->qp_table.rdb_shift);
+	mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
+	mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
+}
+
+static int mthca_alloc_memfree(struct mthca_dev *dev,
+			       struct mthca_qp *qp)
+{
+	int ret = 0;
+
+	if (mthca_is_memfree(dev)) {
+		qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
+						 qp->qpn, &qp->rq.db);
+		if (qp->rq.db_index < 0)
+			return ret;
+
+		qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
+						 qp->qpn, &qp->sq.db);
+		if (qp->sq.db_index < 0)
+			mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
+	}
+
+	return ret;
+}
+
 static void mthca_free_memfree(struct mthca_dev *dev,
 			       struct mthca_qp *qp)
 {
 	if (mthca_is_memfree(dev)) {
 		mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
 		mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
-		mthca_table_put(dev, dev->qp_table.rdb_table,
-				qp->qpn << dev->qp_table.rdb_shift);
-		mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
-		mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
 	}
 }
 
@@ -1108,13 +1189,28 @@ static int mthca_alloc_qp_common(struct mthca_dev *dev,
 	mthca_wq_init(&qp->sq);
 	mthca_wq_init(&qp->rq);
 
-	ret = mthca_alloc_memfree(dev, qp);
+	ret = mthca_map_memfree(dev, qp);
 	if (ret)
 		return ret;
 
 	ret = mthca_alloc_wqe_buf(dev, pd, qp);
 	if (ret) {
-		mthca_free_memfree(dev, qp);
+		mthca_unmap_memfree(dev, qp);
+		return ret;
+	}
+
+	/*
+	 * If this is a userspace QP, we're done now.  The doorbells
+	 * will be allocated and buffers will be initialized in
+	 * userspace.
+	 */
+	if (pd->ibpd.uobject)
+		return 0;
+
+	ret = mthca_alloc_memfree(dev, qp);
+	if (ret) {
+		mthca_free_wqe_buf(dev, qp);
+		mthca_unmap_memfree(dev, qp);
 		return ret;
 	}
 
@@ -1147,22 +1243,39 @@ static int mthca_alloc_qp_common(struct mthca_dev *dev,
 	return 0;
 }
 
-static void mthca_align_qp_size(struct mthca_dev *dev, struct mthca_qp *qp)
+static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
+			     struct mthca_qp *qp)
 {
-	int i;
-
-	if (!mthca_is_memfree(dev))
-		return;
+	/* Sanity check QP size before proceeding */
+	if (cap->max_send_wr  > 65536 || cap->max_recv_wr  > 65536 ||
+	    cap->max_send_sge > 64    || cap->max_recv_sge > 64)
+		return -EINVAL;
 
-	for (i = 0; 1 << i < qp->rq.max; ++i)
-		; /* nothing */
+	if (mthca_is_memfree(dev)) {
+		qp->rq.max = cap->max_recv_wr ?
+			roundup_pow_of_two(cap->max_recv_wr) : 0;
+		qp->sq.max = cap->max_send_wr ?
+			roundup_pow_of_two(cap->max_send_wr) : 0;
+	} else {
+		qp->rq.max = cap->max_recv_wr;
+		qp->sq.max = cap->max_send_wr;
+	}
 
-	qp->rq.max = 1 << i;
+	qp->rq.max_gs = cap->max_recv_sge;
+	qp->sq.max_gs = max_t(int, cap->max_send_sge,
+			      ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
+				    MTHCA_INLINE_CHUNK_SIZE) /
+			      sizeof (struct mthca_data_seg));
 
-	for (i = 0; 1 << i < qp->sq.max; ++i)
-		; /* nothing */
+	/*
+	 * For MLX transport we need 2 extra S/G entries:
+	 * one for the header and one for the checksum at the end
+	 */
+	if ((qp->transport == MLX && qp->sq.max_gs + 2 > dev->limits.max_sg) ||
+	    qp->sq.max_gs > dev->limits.max_sg || qp->rq.max_gs > dev->limits.max_sg)
+		return -EINVAL;
 
-	qp->sq.max = 1 << i;
+	return 0;
 }
 
 int mthca_alloc_qp(struct mthca_dev *dev,
@@ -1171,11 +1284,14 @@ int mthca_alloc_qp(struct mthca_dev *dev,
 		   struct mthca_cq *recv_cq,
 		   enum ib_qp_type type,
 		   enum ib_sig_type send_policy,
+		   struct ib_qp_cap *cap,
 		   struct mthca_qp *qp)
 {
 	int err;
 
-	mthca_align_qp_size(dev, qp);
+	err = mthca_set_qp_size(dev, cap, qp);
+	if (err)
+		return err;
 
 	switch (type) {
 	case IB_QPT_RC: qp->transport = RC; break;
@@ -1208,14 +1324,17 @@ int mthca_alloc_sqp(struct mthca_dev *dev,
 		    struct mthca_cq *send_cq,
 		    struct mthca_cq *recv_cq,
 		    enum ib_sig_type send_policy,
+		    struct ib_qp_cap *cap,
 		    int qpn,
 		    int port,
 		    struct mthca_sqp *sqp)
 {
-	int err = 0;
 	u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
+	int err;
 
-	mthca_align_qp_size(dev, &sqp->qp);
+	err = mthca_set_qp_size(dev, cap, &sqp->qp);
+	if (err)
+		return err;
 
 	sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
 	sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
@@ -1274,8 +1393,6 @@ void mthca_free_qp(struct mthca_dev *dev,
 		   struct mthca_qp *qp)
 {
 	u8 status;
-	int size;
-	int i;
 	struct mthca_cq *send_cq;
 	struct mthca_cq *recv_cq;
 
@@ -1305,31 +1422,22 @@ void mthca_free_qp(struct mthca_dev *dev,
 	if (qp->state != IB_QPS_RESET)
 		mthca_MODIFY_QP(dev, MTHCA_TRANS_ANY2RST, qp->qpn, 0, NULL, 0, &status);
 
-	mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn);
-	if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
-		mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn);
-
-	mthca_free_mr(dev, &qp->mr);
-
-	size = PAGE_ALIGN(qp->send_wqe_offset +
-			  (qp->sq.max << qp->sq.wqe_shift));
+	/*
+	 * If this is a userspace QP, the buffers, MR, CQs and so on
+	 * will be cleaned up in userspace, so all we have to do is
+	 * unref the mem-free tables and free the QPN in our table.
+	 */
+	if (!qp->ibqp.uobject) {
+		mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn);
+		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
+			mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn);
 
-	if (qp->is_direct) {
-		pci_free_consistent(dev->pdev, size,
-				    qp->queue.direct.buf,
-				    pci_unmap_addr(&qp->queue.direct, mapping));
-	} else {
-		for (i = 0; i < size / PAGE_SIZE; ++i) {
-			pci_free_consistent(dev->pdev, PAGE_SIZE,
-					    qp->queue.page_list[i].buf,
-					    pci_unmap_addr(&qp->queue.page_list[i],
-							   mapping));
-		}
+		mthca_free_mr(dev, &qp->mr);
+		mthca_free_memfree(dev, qp);
+		mthca_free_wqe_buf(dev, qp);
 	}
 
-	kfree(qp->wrid);
-
-	mthca_free_memfree(dev, qp);
+	mthca_unmap_memfree(dev, qp);
 
 	if (is_sqp(dev, qp)) {
 		atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
@@ -1529,6 +1637,26 @@ int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 
 			break;
 
+		case UC:
+			switch (wr->opcode) {
+			case IB_WR_RDMA_WRITE:
+			case IB_WR_RDMA_WRITE_WITH_IMM:
+				((struct mthca_raddr_seg *) wqe)->raddr =
+					cpu_to_be64(wr->wr.rdma.remote_addr);
+				((struct mthca_raddr_seg *) wqe)->rkey =
+					cpu_to_be32(wr->wr.rdma.rkey);
+				((struct mthca_raddr_seg *) wqe)->reserved = 0;
+				wqe += sizeof (struct mthca_raddr_seg);
+				size += sizeof (struct mthca_raddr_seg) / 16;
+				break;
+
+			default:
+				/* No extra segments required for sends */
+				break;
+			}
+
+			break;
+
 		case UD:
 			((struct mthca_tavor_ud_seg *) wqe)->lkey =
 				cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
@@ -1814,9 +1942,29 @@ int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 					sizeof (struct mthca_atomic_seg);
 				break;
 
+			case IB_WR_RDMA_READ:
+			case IB_WR_RDMA_WRITE:
+			case IB_WR_RDMA_WRITE_WITH_IMM:
+				((struct mthca_raddr_seg *) wqe)->raddr =
+					cpu_to_be64(wr->wr.rdma.remote_addr);
+				((struct mthca_raddr_seg *) wqe)->rkey =
+					cpu_to_be32(wr->wr.rdma.rkey);
+				((struct mthca_raddr_seg *) wqe)->reserved = 0;
+				wqe += sizeof (struct mthca_raddr_seg);
+				size += sizeof (struct mthca_raddr_seg) / 16;
+				break;
+
+			default:
+				/* No extra segments required for sends */
+				break;
+			}
+
+			break;
+
+		case UC:
+			switch (wr->opcode) {
 			case IB_WR_RDMA_WRITE:
 			case IB_WR_RDMA_WRITE_WITH_IMM:
-			case IB_WR_RDMA_READ:
 				((struct mthca_raddr_seg *) wqe)->raddr =
 					cpu_to_be64(wr->wr.rdma.remote_addr);
 				((struct mthca_raddr_seg *) wqe)->rkey =
diff --git a/drivers/infiniband/hw/mthca/mthca_user.h b/drivers/infiniband/hw/mthca/mthca_user.h
new file mode 100644
index 000000000000..3024c1b4547d
--- /dev/null
+++ b/drivers/infiniband/hw/mthca/mthca_user.h
@@ -0,0 +1,81 @@
+/*
+ * Copyright (c) 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  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 MTHCA_USER_H
+#define MTHCA_USER_H
+
+#include <linux/types.h>
+
+/*
+ * Make sure that all structs defined in this file remain laid out so
+ * that they pack the same way on 32-bit and 64-bit architectures (to
+ * avoid incompatibility between 32-bit userspace and 64-bit kernels).
+ * In particular do not use pointer types -- pass pointers in __u64
+ * instead.
+ */
+
+struct mthca_alloc_ucontext_resp {
+	__u32 qp_tab_size;
+	__u32 uarc_size;
+};
+
+struct mthca_alloc_pd_resp {
+	__u32 pdn;
+	__u32 reserved;
+};
+
+struct mthca_create_cq {
+	__u32 lkey;
+	__u32 pdn;
+	__u64 arm_db_page;
+	__u64 set_db_page;
+	__u32 arm_db_index;
+	__u32 set_db_index;
+};
+
+struct mthca_create_cq_resp {
+	__u32 cqn;
+	__u32 reserved;
+};
+
+struct mthca_create_qp {
+	__u32 lkey;
+	__u32 reserved;
+	__u64 sq_db_page;
+	__u64 rq_db_page;
+	__u32 sq_db_index;
+	__u32 rq_db_index;
+};
+
+#endif /* MTHCA_USER_H */
diff --git a/drivers/infiniband/include/ib_cm.h b/drivers/infiniband/include/ib_cm.h
new file mode 100644
index 000000000000..da650115e79a
--- /dev/null
+++ b/drivers/infiniband/include/ib_cm.h
@@ -0,0 +1,569 @@
+/*
+ * Copyright (c) 2004 Intel Corporation.  All rights reserved.
+ * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
+ * Copyright (c) 2004 Voltaire Corporation.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. 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.
+ *
+ * $Id: ib_cm.h 2730 2005-06-28 16:43:03Z sean.hefty $
+ */
+#if !defined(IB_CM_H)
+#define IB_CM_H
+
+#include <ib_mad.h>
+#include <ib_sa.h>
+
+enum ib_cm_state {
+	IB_CM_IDLE,
+	IB_CM_LISTEN,
+	IB_CM_REQ_SENT,
+	IB_CM_REQ_RCVD,
+	IB_CM_MRA_REQ_SENT,
+	IB_CM_MRA_REQ_RCVD,
+	IB_CM_REP_SENT,
+	IB_CM_REP_RCVD,
+	IB_CM_MRA_REP_SENT,
+	IB_CM_MRA_REP_RCVD,
+	IB_CM_ESTABLISHED,
+	IB_CM_DREQ_SENT,
+	IB_CM_DREQ_RCVD,
+	IB_CM_TIMEWAIT,
+	IB_CM_SIDR_REQ_SENT,
+	IB_CM_SIDR_REQ_RCVD
+};
+
+enum ib_cm_lap_state {
+	IB_CM_LAP_IDLE,
+	IB_CM_LAP_SENT,
+	IB_CM_LAP_RCVD,
+	IB_CM_MRA_LAP_SENT,
+	IB_CM_MRA_LAP_RCVD,
+};
+
+enum ib_cm_event_type {
+	IB_CM_REQ_ERROR,
+	IB_CM_REQ_RECEIVED,
+	IB_CM_REP_ERROR,
+	IB_CM_REP_RECEIVED,
+	IB_CM_RTU_RECEIVED,
+	IB_CM_USER_ESTABLISHED,
+	IB_CM_DREQ_ERROR,
+	IB_CM_DREQ_RECEIVED,
+	IB_CM_DREP_RECEIVED,
+	IB_CM_TIMEWAIT_EXIT,
+	IB_CM_MRA_RECEIVED,
+	IB_CM_REJ_RECEIVED,
+	IB_CM_LAP_ERROR,
+	IB_CM_LAP_RECEIVED,
+	IB_CM_APR_RECEIVED,
+	IB_CM_SIDR_REQ_ERROR,
+	IB_CM_SIDR_REQ_RECEIVED,
+	IB_CM_SIDR_REP_RECEIVED
+};
+
+enum ib_cm_data_size {
+	IB_CM_REQ_PRIVATE_DATA_SIZE	 = 92,
+	IB_CM_MRA_PRIVATE_DATA_SIZE	 = 222,
+	IB_CM_REJ_PRIVATE_DATA_SIZE	 = 148,
+	IB_CM_REP_PRIVATE_DATA_SIZE	 = 196,
+	IB_CM_RTU_PRIVATE_DATA_SIZE	 = 224,
+	IB_CM_DREQ_PRIVATE_DATA_SIZE	 = 220,
+	IB_CM_DREP_PRIVATE_DATA_SIZE	 = 224,
+	IB_CM_REJ_ARI_LENGTH		 = 72,
+	IB_CM_LAP_PRIVATE_DATA_SIZE	 = 168,
+	IB_CM_APR_PRIVATE_DATA_SIZE	 = 148,
+	IB_CM_APR_INFO_LENGTH		 = 72,
+	IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE = 216,
+	IB_CM_SIDR_REP_PRIVATE_DATA_SIZE = 136,
+	IB_CM_SIDR_REP_INFO_LENGTH	 = 72
+};
+
+struct ib_cm_id;
+
+struct ib_cm_req_event_param {
+	struct ib_cm_id		*listen_id;
+	struct ib_device	*device;
+	u8			port;
+
+	struct ib_sa_path_rec	*primary_path;
+	struct ib_sa_path_rec	*alternate_path;
+
+	u64			remote_ca_guid;
+	u32			remote_qkey;
+	u32			remote_qpn;
+	enum ib_qp_type		qp_type;
+
+	u32			starting_psn;
+	u8			responder_resources;
+	u8			initiator_depth;
+	unsigned int		local_cm_response_timeout:5;
+	unsigned int		flow_control:1;
+	unsigned int		remote_cm_response_timeout:5;
+	unsigned int		retry_count:3;
+	unsigned int		rnr_retry_count:3;
+	unsigned int		srq:1;
+};
+
+struct ib_cm_rep_event_param {
+	u64			remote_ca_guid;
+	u32			remote_qkey;
+	u32			remote_qpn;
+	u32			starting_psn;
+	u8			responder_resources;
+	u8			initiator_depth;
+	unsigned int		target_ack_delay:5;
+	unsigned int		failover_accepted:2;
+	unsigned int		flow_control:1;
+	unsigned int		rnr_retry_count:3;
+	unsigned int		srq:1;
+};
+
+enum ib_cm_rej_reason {
+	IB_CM_REJ_NO_QP				= __constant_htons(1),
+	IB_CM_REJ_NO_EEC			= __constant_htons(2),
+	IB_CM_REJ_NO_RESOURCES			= __constant_htons(3),
+	IB_CM_REJ_TIMEOUT			= __constant_htons(4),
+	IB_CM_REJ_UNSUPPORTED			= __constant_htons(5),
+	IB_CM_REJ_INVALID_COMM_ID		= __constant_htons(6),
+	IB_CM_REJ_INVALID_COMM_INSTANCE		= __constant_htons(7),
+	IB_CM_REJ_INVALID_SERVICE_ID		= __constant_htons(8),
+	IB_CM_REJ_INVALID_TRANSPORT_TYPE	= __constant_htons(9),
+	IB_CM_REJ_STALE_CONN			= __constant_htons(10),
+	IB_CM_REJ_RDC_NOT_EXIST			= __constant_htons(11),
+	IB_CM_REJ_INVALID_GID			= __constant_htons(12),
+	IB_CM_REJ_INVALID_LID			= __constant_htons(13),
+	IB_CM_REJ_INVALID_SL			= __constant_htons(14),
+	IB_CM_REJ_INVALID_TRAFFIC_CLASS		= __constant_htons(15),
+	IB_CM_REJ_INVALID_HOP_LIMIT		= __constant_htons(16),
+	IB_CM_REJ_INVALID_PACKET_RATE		= __constant_htons(17),
+	IB_CM_REJ_INVALID_ALT_GID		= __constant_htons(18),
+	IB_CM_REJ_INVALID_ALT_LID		= __constant_htons(19),
+	IB_CM_REJ_INVALID_ALT_SL		= __constant_htons(20),
+	IB_CM_REJ_INVALID_ALT_TRAFFIC_CLASS	= __constant_htons(21),
+	IB_CM_REJ_INVALID_ALT_HOP_LIMIT		= __constant_htons(22),
+	IB_CM_REJ_INVALID_ALT_PACKET_RATE	= __constant_htons(23),
+	IB_CM_REJ_PORT_CM_REDIRECT		= __constant_htons(24),
+	IB_CM_REJ_PORT_REDIRECT			= __constant_htons(25),
+	IB_CM_REJ_INVALID_MTU			= __constant_htons(26),
+	IB_CM_REJ_INSUFFICIENT_RESP_RESOURCES	= __constant_htons(27),
+	IB_CM_REJ_CONSUMER_DEFINED		= __constant_htons(28),
+	IB_CM_REJ_INVALID_RNR_RETRY		= __constant_htons(29),
+	IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID	= __constant_htons(30),
+	IB_CM_REJ_INVALID_CLASS_VERSION		= __constant_htons(31),
+	IB_CM_REJ_INVALID_FLOW_LABEL		= __constant_htons(32),
+	IB_CM_REJ_INVALID_ALT_FLOW_LABEL	= __constant_htons(33)
+};
+
+struct ib_cm_rej_event_param {
+	enum ib_cm_rej_reason	reason;
+	void			*ari;
+	u8			ari_length;
+};
+
+struct ib_cm_mra_event_param {
+	u8	service_timeout;
+};
+
+struct ib_cm_lap_event_param {
+	struct ib_sa_path_rec	*alternate_path;
+};
+
+enum ib_cm_apr_status {
+	IB_CM_APR_SUCCESS,
+	IB_CM_APR_INVALID_COMM_ID,
+	IB_CM_APR_UNSUPPORTED,
+	IB_CM_APR_REJECT,
+	IB_CM_APR_REDIRECT,
+	IB_CM_APR_IS_CURRENT,
+	IB_CM_APR_INVALID_QPN_EECN,
+	IB_CM_APR_INVALID_LID,
+	IB_CM_APR_INVALID_GID,
+	IB_CM_APR_INVALID_FLOW_LABEL,
+	IB_CM_APR_INVALID_TCLASS,
+	IB_CM_APR_INVALID_HOP_LIMIT,
+	IB_CM_APR_INVALID_PACKET_RATE,
+	IB_CM_APR_INVALID_SL
+};
+
+struct ib_cm_apr_event_param {
+	enum ib_cm_apr_status	ap_status;
+	void			*apr_info;
+	u8			info_len;
+};
+
+struct ib_cm_sidr_req_event_param {
+	struct ib_cm_id		*listen_id;
+	struct ib_device	*device;
+	u8			port;
+
+	u16	pkey;
+};
+
+enum ib_cm_sidr_status {
+	IB_SIDR_SUCCESS,
+	IB_SIDR_UNSUPPORTED,
+	IB_SIDR_REJECT,
+	IB_SIDR_NO_QP,
+	IB_SIDR_REDIRECT,
+	IB_SIDR_UNSUPPORTED_VERSION
+};
+
+struct ib_cm_sidr_rep_event_param {
+	enum ib_cm_sidr_status	status;
+	u32			qkey;
+	u32			qpn;
+	void			*info;
+	u8			info_len;
+
+};
+
+struct ib_cm_event {
+	enum ib_cm_event_type	event;
+	union {
+		struct ib_cm_req_event_param	req_rcvd;
+		struct ib_cm_rep_event_param	rep_rcvd;
+		/* No data for RTU received events. */
+		struct ib_cm_rej_event_param	rej_rcvd;
+		struct ib_cm_mra_event_param	mra_rcvd;
+		struct ib_cm_lap_event_param	lap_rcvd;
+		struct ib_cm_apr_event_param	apr_rcvd;
+		/* No data for DREQ/DREP received events. */
+		struct ib_cm_sidr_req_event_param sidr_req_rcvd;
+		struct ib_cm_sidr_rep_event_param sidr_rep_rcvd;
+		enum ib_wc_status		send_status;
+	} param;
+
+	void			*private_data;
+};
+
+/**
+ * ib_cm_handler - User-defined callback to process communication events.
+ * @cm_id: Communication identifier associated with the reported event.
+ * @event: Information about the communication event.
+ *
+ * IB_CM_REQ_RECEIVED and IB_CM_SIDR_REQ_RECEIVED communication events
+ * generated as a result of listen requests result in the allocation of a
+ * new @cm_id.  The new @cm_id is returned to the user through this callback.
+ * Clients are responsible for destroying the new @cm_id.  For peer-to-peer
+ * IB_CM_REQ_RECEIVED and all other events, the returned @cm_id corresponds
+ * to a user's existing communication identifier.
+ *
+ * Users may not call ib_destroy_cm_id while in the context of this callback;
+ * however, returning a non-zero value instructs the communication manager to
+ * destroy the @cm_id after the callback completes.
+ */
+typedef int (*ib_cm_handler)(struct ib_cm_id *cm_id,
+			     struct ib_cm_event *event);
+
+struct ib_cm_id {
+	ib_cm_handler		cm_handler;
+	void			*context;
+	u64			service_id;
+	u64			service_mask;
+	enum ib_cm_state	state;		/* internal CM/debug use */
+	enum ib_cm_lap_state	lap_state;	/* internal CM/debug use */
+	u32			local_id;
+	u32			remote_id;
+};
+
+/**
+ * ib_create_cm_id - Allocate a communication identifier.
+ * @cm_handler: Callback invoked to notify the user of CM events.
+ * @context: User specified context associated with the communication
+ *   identifier.
+ *
+ * Communication identifiers are used to track connection states, service
+ * ID resolution requests, and listen requests.
+ */
+struct ib_cm_id *ib_create_cm_id(ib_cm_handler cm_handler,
+				 void *context);
+
+/**
+ * ib_destroy_cm_id - Destroy a connection identifier.
+ * @cm_id: Connection identifier to destroy.
+ *
+ * This call blocks until the connection identifier is destroyed.
+ */
+void ib_destroy_cm_id(struct ib_cm_id *cm_id);
+
+#define IB_SERVICE_ID_AGN_MASK	__constant_cpu_to_be64(0xFF00000000000000ULL)
+#define IB_CM_ASSIGN_SERVICE_ID __constant_cpu_to_be64(0x0200000000000000ULL)
+
+/**
+ * ib_cm_listen - Initiates listening on the specified service ID for
+ *   connection and service ID resolution requests.
+ * @cm_id: Connection identifier associated with the listen request.
+ * @service_id: Service identifier matched against incoming connection
+ *   and service ID resolution requests.  The service ID should be specified
+ *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
+ *   assign a service ID to the caller.
+ * @service_mask: Mask applied to service ID used to listen across a
+ *   range of service IDs.  If set to 0, the service ID is matched
+ *   exactly.  This parameter is ignored if %service_id is set to
+ *   IB_CM_ASSIGN_SERVICE_ID.
+ */
+int ib_cm_listen(struct ib_cm_id *cm_id,
+		 u64 service_id,
+		 u64 service_mask);
+
+struct ib_cm_req_param {
+	struct ib_sa_path_rec	*primary_path;
+	struct ib_sa_path_rec	*alternate_path;
+	u64			service_id;
+	u32			qp_num;
+	enum ib_qp_type		qp_type;
+	u32			starting_psn;
+	const void		*private_data;
+	u8			private_data_len;
+	u8			peer_to_peer;
+	u8			responder_resources;
+	u8			initiator_depth;
+	u8			remote_cm_response_timeout;
+	u8			flow_control;
+	u8			local_cm_response_timeout;
+	u8			retry_count;
+	u8			rnr_retry_count;
+	u8			max_cm_retries;
+	u8			srq;
+};
+
+/**
+ * ib_send_cm_req - Sends a connection request to the remote node.
+ * @cm_id: Connection identifier that will be associated with the
+ *   connection request.
+ * @param: Connection request information needed to establish the
+ *   connection.
+ */
+int ib_send_cm_req(struct ib_cm_id *cm_id,
+		   struct ib_cm_req_param *param);
+
+struct ib_cm_rep_param {
+	u32		qp_num;
+	u32		starting_psn;
+	const void	*private_data;
+	u8		private_data_len;
+	u8		responder_resources;
+	u8		initiator_depth;
+	u8		target_ack_delay;
+	u8		failover_accepted;
+	u8		flow_control;
+	u8		rnr_retry_count;
+	u8		srq;
+};
+
+/**
+ * ib_send_cm_rep - Sends a connection reply in response to a connection
+ *   request.
+ * @cm_id: Connection identifier that will be associated with the
+ *   connection request.
+ * @param: Connection reply information needed to establish the
+ *   connection.
+ */
+int ib_send_cm_rep(struct ib_cm_id *cm_id,
+		   struct ib_cm_rep_param *param);
+
+/**
+ * ib_send_cm_rtu - Sends a connection ready to use message in response
+ *   to a connection reply message.
+ * @cm_id: Connection identifier associated with the connection request.
+ * @private_data: Optional user-defined private data sent with the
+ *   ready to use message.
+ * @private_data_len: Size of the private data buffer, in bytes.
+ */
+int ib_send_cm_rtu(struct ib_cm_id *cm_id,
+		   const void *private_data,
+		   u8 private_data_len);
+
+/**
+ * ib_send_cm_dreq - Sends a disconnection request for an existing
+ *   connection.
+ * @cm_id: Connection identifier associated with the connection being
+ *   released.
+ * @private_data: Optional user-defined private data sent with the
+ *   disconnection request message.
+ * @private_data_len: Size of the private data buffer, in bytes.
+ */
+int ib_send_cm_dreq(struct ib_cm_id *cm_id,
+		    const void *private_data,
+		    u8 private_data_len);
+
+/**
+ * ib_send_cm_drep - Sends a disconnection reply to a disconnection request.
+ * @cm_id: Connection identifier associated with the connection being
+ *   released.
+ * @private_data: Optional user-defined private data sent with the
+ *   disconnection reply message.
+ * @private_data_len: Size of the private data buffer, in bytes.
+ *
+ * If the cm_id is in the correct state, the CM will transition the connection
+ * to the timewait state, even if an error occurs sending the DREP message.
+ */
+int ib_send_cm_drep(struct ib_cm_id *cm_id,
+		    const void *private_data,
+		    u8 private_data_len);
+
+/**
+ * ib_cm_establish - Forces a connection state to established.
+ * @cm_id: Connection identifier to transition to established.
+ *
+ * This routine should be invoked by users who receive messages on a
+ * connected QP before an RTU has been received.
+ */
+int ib_cm_establish(struct ib_cm_id *cm_id);
+
+/**
+ * ib_send_cm_rej - Sends a connection rejection message to the
+ *   remote node.
+ * @cm_id: Connection identifier associated with the connection being
+ *   rejected.
+ * @reason: Reason for the connection request rejection.
+ * @ari: Optional additional rejection information.
+ * @ari_length: Size of the additional rejection information, in bytes.
+ * @private_data: Optional user-defined private data sent with the
+ *   rejection message.
+ * @private_data_len: Size of the private data buffer, in bytes.
+ */
+int ib_send_cm_rej(struct ib_cm_id *cm_id,
+		   enum ib_cm_rej_reason reason,
+		   void *ari,
+		   u8 ari_length,
+		   const void *private_data,
+		   u8 private_data_len);
+
+/**
+ * ib_send_cm_mra - Sends a message receipt acknowledgement to a connection
+ *   message.
+ * @cm_id: Connection identifier associated with the connection message.
+ * @service_timeout: The maximum time required for the sender to reply to
+ *   to the connection message.
+ * @private_data: Optional user-defined private data sent with the
+ *   message receipt acknowledgement.
+ * @private_data_len: Size of the private data buffer, in bytes.
+ */
+int ib_send_cm_mra(struct ib_cm_id *cm_id,
+		   u8 service_timeout,
+		   const void *private_data,
+		   u8 private_data_len);
+
+/**
+ * ib_send_cm_lap - Sends a load alternate path request.
+ * @cm_id: Connection identifier associated with the load alternate path
+ *   message.
+ * @alternate_path: A path record that identifies the alternate path to
+ *   load.
+ * @private_data: Optional user-defined private data sent with the
+ *   load alternate path message.
+ * @private_data_len: Size of the private data buffer, in bytes.
+ */
+int ib_send_cm_lap(struct ib_cm_id *cm_id,
+		   struct ib_sa_path_rec *alternate_path,
+		   const void *private_data,
+		   u8 private_data_len);
+
+/**
+ * ib_cm_init_qp_attr - Initializes the QP attributes for use in transitioning
+ *   to a specified QP state.
+ * @cm_id: Communication identifier associated with the QP attributes to
+ *   initialize.
+ * @qp_attr: On input, specifies the desired QP state.  On output, the
+ *   mandatory and desired optional attributes will be set in order to
+ *   modify the QP to the specified state.
+ * @qp_attr_mask: The QP attribute mask that may be used to transition the
+ *   QP to the specified state.
+ *
+ * Users must set the @qp_attr->qp_state to the desired QP state.  This call
+ * will set all required attributes for the given transition, along with
+ * known optional attributes.  Users may override the attributes returned from
+ * this call before calling ib_modify_qp.
+ */
+int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
+		       struct ib_qp_attr *qp_attr,
+		       int *qp_attr_mask);
+
+/**
+ * ib_send_cm_apr - Sends an alternate path response message in response to
+ *   a load alternate path request.
+ * @cm_id: Connection identifier associated with the alternate path response.
+ * @status: Reply status sent with the alternate path response.
+ * @info: Optional additional information sent with the alternate path
+ *   response.
+ * @info_length: Size of the additional information, in bytes.
+ * @private_data: Optional user-defined private data sent with the
+ *   alternate path response message.
+ * @private_data_len: Size of the private data buffer, in bytes.
+ */
+int ib_send_cm_apr(struct ib_cm_id *cm_id,
+		   enum ib_cm_apr_status status,
+		   void *info,
+		   u8 info_length,
+		   const void *private_data,
+		   u8 private_data_len);
+
+struct ib_cm_sidr_req_param {
+	struct ib_sa_path_rec	*path;
+	u64			service_id;
+	int			timeout_ms;
+	const void		*private_data;
+	u8			private_data_len;
+	u8			max_cm_retries;
+	u16			pkey;
+};
+
+/**
+ * ib_send_cm_sidr_req - Sends a service ID resolution request to the
+ *   remote node.
+ * @cm_id: Communication identifier that will be associated with the
+ *   service ID resolution request.
+ * @param: Service ID resolution request information.
+ */
+int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
+			struct ib_cm_sidr_req_param *param);
+
+struct ib_cm_sidr_rep_param {
+	u32			qp_num;
+	u32			qkey;
+	enum ib_cm_sidr_status	status;
+	const void		*info;
+	u8			info_length;
+	const void		*private_data;
+	u8			private_data_len;
+};
+
+/**
+ * ib_send_cm_sidr_rep - Sends a service ID resolution request to the
+ *   remote node.
+ * @cm_id: Communication identifier associated with the received service ID
+ *   resolution request.
+ * @param: Service ID resolution reply information.
+ */
+int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
+			struct ib_cm_sidr_rep_param *param);
+
+#endif /* IB_CM_H */
diff --git a/drivers/infiniband/include/ib_fmr_pool.h b/drivers/infiniband/include/ib_fmr_pool.h
index e8769657cbbb..6c9e24d6e144 100644
--- a/drivers/infiniband/include/ib_fmr_pool.h
+++ b/drivers/infiniband/include/ib_fmr_pool.h
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
+ * Copyright (c) 2005 Sun Microsystems, Inc. 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
@@ -29,7 +30,7 @@
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  *
- * $Id: ib_fmr_pool.h 1349 2004-12-16 21:09:43Z roland $
+ * $Id: ib_fmr_pool.h 2730 2005-06-28 16:43:03Z sean.hefty $
  */
 
 #if !defined(IB_FMR_POOL_H)
@@ -78,7 +79,7 @@ struct ib_pool_fmr {
 struct ib_fmr_pool *ib_create_fmr_pool(struct ib_pd             *pd,
 				       struct ib_fmr_pool_param *params);
 
-int ib_destroy_fmr_pool(struct ib_fmr_pool *pool);
+void ib_destroy_fmr_pool(struct ib_fmr_pool *pool);
 
 int ib_flush_fmr_pool(struct ib_fmr_pool *pool);
 
diff --git a/drivers/infiniband/include/ib_mad.h b/drivers/infiniband/include/ib_mad.h
index 4a6bf6763a97..491b6f25b3b8 100644
--- a/drivers/infiniband/include/ib_mad.h
+++ b/drivers/infiniband/include/ib_mad.h
@@ -33,12 +33,14 @@
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  *
- * $Id: ib_mad.h 1389 2004-12-27 22:56:47Z roland $
+ * $Id: ib_mad.h 2775 2005-07-02 13:42:12Z halr $
  */
 
 #if !defined( IB_MAD_H )
 #define IB_MAD_H
 
+#include <linux/pci.h>
+
 #include <ib_verbs.h>
 
 /* Management base version */
@@ -56,6 +58,8 @@
 #define IB_MGMT_CLASS_VENDOR_RANGE2_START	0x30
 #define IB_MGMT_CLASS_VENDOR_RANGE2_END		0x4F
 
+#define	IB_OPENIB_OUI				(0x001405)
+
 /* Management methods */
 #define IB_MGMT_METHOD_GET			0x01
 #define IB_MGMT_METHOD_SET			0x02
@@ -70,18 +74,37 @@
 
 #define IB_MGMT_MAX_METHODS			128
 
+/* RMPP information */
+#define IB_MGMT_RMPP_VERSION			1
+
+#define IB_MGMT_RMPP_TYPE_DATA			1
+#define IB_MGMT_RMPP_TYPE_ACK			2
+#define IB_MGMT_RMPP_TYPE_STOP			3
+#define IB_MGMT_RMPP_TYPE_ABORT			4
+
+#define IB_MGMT_RMPP_FLAG_ACTIVE		1
+#define IB_MGMT_RMPP_FLAG_FIRST			(1<<1)
+#define IB_MGMT_RMPP_FLAG_LAST			(1<<2)
+
+#define IB_MGMT_RMPP_NO_RESPTIME		0x1F
+
+#define	IB_MGMT_RMPP_STATUS_SUCCESS		0
+#define	IB_MGMT_RMPP_STATUS_RESX		1
+#define	IB_MGMT_RMPP_STATUS_T2L			118
+#define	IB_MGMT_RMPP_STATUS_BAD_LEN		119
+#define	IB_MGMT_RMPP_STATUS_BAD_SEG		120
+#define	IB_MGMT_RMPP_STATUS_BADT		121
+#define	IB_MGMT_RMPP_STATUS_W2S			122
+#define	IB_MGMT_RMPP_STATUS_S2B			123
+#define	IB_MGMT_RMPP_STATUS_BAD_STATUS		124
+#define	IB_MGMT_RMPP_STATUS_UNV			125
+#define	IB_MGMT_RMPP_STATUS_TMR			126
+#define	IB_MGMT_RMPP_STATUS_UNSPEC		127
+
 #define IB_QP0		0
 #define IB_QP1		__constant_htonl(1)
 #define IB_QP1_QKEY	0x80010000
-
-struct ib_grh {
-	u32		version_tclass_flow;
-	u16		paylen;
-	u8		next_hdr;
-	u8		hop_limit;
-	union ib_gid	sgid;
-	union ib_gid	dgid;
-} __attribute__ ((packed));
+#define IB_QP_SET_QKEY	0x80000000
 
 struct ib_mad_hdr {
 	u8	base_version;
@@ -94,7 +117,7 @@ struct ib_mad_hdr {
 	u16	attr_id;
 	u16	resv;
 	u32	attr_mod;
-} __attribute__ ((packed));
+};
 
 struct ib_rmpp_hdr {
 	u8	rmpp_version;
@@ -103,17 +126,41 @@ struct ib_rmpp_hdr {
 	u8	rmpp_status;
 	u32	seg_num;
 	u32	paylen_newwin;
+};
+
+typedef u64 __bitwise ib_sa_comp_mask;
+
+#define IB_SA_COMP_MASK(n) ((__force ib_sa_comp_mask) cpu_to_be64(1ull << n))
+
+/*
+ * ib_sa_hdr and ib_sa_mad structures must be packed because they have
+ * 64-bit fields that are only 32-bit aligned. 64-bit architectures will
+ * lay them out wrong otherwise.  (And unfortunately they are sent on
+ * the wire so we can't change the layout)
+ */
+struct ib_sa_hdr {
+	u64			sm_key;
+	u16			attr_offset;
+	u16			reserved;
+	ib_sa_comp_mask		comp_mask;
 } __attribute__ ((packed));
 
 struct ib_mad {
 	struct ib_mad_hdr	mad_hdr;
 	u8			data[232];
-} __attribute__ ((packed));
+};
 
 struct ib_rmpp_mad {
 	struct ib_mad_hdr	mad_hdr;
 	struct ib_rmpp_hdr	rmpp_hdr;
 	u8			data[220];
+};
+
+struct ib_sa_mad {
+	struct ib_mad_hdr	mad_hdr;
+	struct ib_rmpp_hdr	rmpp_hdr;
+	struct ib_sa_hdr	sa_hdr;
+	u8			data[200];
 } __attribute__ ((packed));
 
 struct ib_vendor_mad {
@@ -122,7 +169,70 @@ struct ib_vendor_mad {
 	u8			reserved;
 	u8			oui[3];
 	u8			data[216];
-} __attribute__ ((packed));
+};
+
+/**
+ * ib_mad_send_buf - MAD data buffer and work request for sends.
+ * @mad: References an allocated MAD data buffer.  The size of the data
+ *   buffer is specified in the @send_wr.length field.
+ * @mapping: DMA mapping information.
+ * @mad_agent: MAD agent that allocated the buffer.
+ * @context: User-controlled context fields.
+ * @send_wr: An initialized work request structure used when sending the MAD.
+ *   The wr_id field of the work request is initialized to reference this
+ *   data structure.
+ * @sge: A scatter-gather list referenced by the work request.
+ *
+ * Users are responsible for initializing the MAD buffer itself, with the
+ * exception of specifying the payload length field in any RMPP MAD.
+ */
+struct ib_mad_send_buf {
+	struct ib_mad		*mad;
+	DECLARE_PCI_UNMAP_ADDR(mapping)
+	struct ib_mad_agent	*mad_agent;
+	void			*context[2];
+	struct ib_send_wr	send_wr;
+	struct ib_sge		sge;
+};
+
+/**
+ * ib_get_rmpp_resptime - Returns the RMPP response time.
+ * @rmpp_hdr: An RMPP header.
+ */
+static inline u8 ib_get_rmpp_resptime(struct ib_rmpp_hdr *rmpp_hdr)
+{
+	return rmpp_hdr->rmpp_rtime_flags >> 3;
+}
+
+/**
+ * ib_get_rmpp_flags - Returns the RMPP flags.
+ * @rmpp_hdr: An RMPP header.
+ */
+static inline u8 ib_get_rmpp_flags(struct ib_rmpp_hdr *rmpp_hdr)
+{
+	return rmpp_hdr->rmpp_rtime_flags & 0x7;
+}
+
+/**
+ * ib_set_rmpp_resptime - Sets the response time in an RMPP header.
+ * @rmpp_hdr: An RMPP header.
+ * @rtime: The response time to set.
+ */
+static inline void ib_set_rmpp_resptime(struct ib_rmpp_hdr *rmpp_hdr, u8 rtime)
+{
+	rmpp_hdr->rmpp_rtime_flags = ib_get_rmpp_flags(rmpp_hdr) | (rtime << 3);
+}
+
+/**
+ * ib_set_rmpp_flags - Sets the flags in an RMPP header.
+ * @rmpp_hdr: An RMPP header.
+ * @flags: The flags to set.
+ */
+static inline void ib_set_rmpp_flags(struct ib_rmpp_hdr *rmpp_hdr, u8 flags)
+{
+	rmpp_hdr->rmpp_rtime_flags = (rmpp_hdr->rmpp_rtime_flags & 0xF1) |
+				     (flags & 0x7);
+}
 
 struct ib_mad_agent;
 struct ib_mad_send_wc;
@@ -168,6 +278,7 @@ typedef void (*ib_mad_recv_handler)(struct ib_mad_agent *mad_agent,
  * ib_mad_agent - Used to track MAD registration with the access layer.
  * @device: Reference to device registration is on.
  * @qp: Reference to QP used for sending and receiving MADs.
+ * @mr: Memory region for system memory usable for DMA.
  * @recv_handler: Callback handler for a received MAD.
  * @send_handler: Callback handler for a sent MAD.
  * @snoop_handler: Callback handler for snooped sent MADs.
@@ -176,16 +287,19 @@ typedef void (*ib_mad_recv_handler)(struct ib_mad_agent *mad_agent,
  *   Unsolicited MADs sent by this client will have the upper 32-bits
  *   of their TID set to this value.
  * @port_num: Port number on which QP is registered
+ * @rmpp_version: If set, indicates the RMPP version used by this agent.
  */
 struct ib_mad_agent {
 	struct ib_device	*device;
 	struct ib_qp		*qp;
+	struct ib_mr		*mr;
 	ib_mad_recv_handler	recv_handler;
 	ib_mad_send_handler	send_handler;
 	ib_mad_snoop_handler	snoop_handler;
 	void			*context;
 	u32			hi_tid;
 	u8			port_num;
+	u8			rmpp_version;
 };
 
 /**
@@ -219,6 +333,7 @@ struct ib_mad_recv_buf {
  * ib_mad_recv_wc - received MAD information.
  * @wc: Completion information for the received data.
  * @recv_buf: Specifies the location of the received data buffer(s).
+ * @rmpp_list: Specifies a list of RMPP reassembled received MAD buffers.
  * @mad_len: The length of the received MAD, without duplicated headers.
  *
  * For received response, the wr_id field of the wc is set to the wr_id
@@ -227,6 +342,7 @@ struct ib_mad_recv_buf {
 struct ib_mad_recv_wc {
 	struct ib_wc		*wc;
 	struct ib_mad_recv_buf	recv_buf;
+	struct list_head	rmpp_list;
 	int			mad_len;
 };
 
@@ -322,6 +438,16 @@ int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent);
  * @bad_send_wr: Specifies the MAD on which an error was encountered.
  *
  * Sent MADs are not guaranteed to complete in the order that they were posted.
+ *
+ * If the MAD requires RMPP, the data buffer should contain a single copy
+ * of the common MAD, RMPP, and class specific headers, followed by the class
+ * defined data.  If the class defined data would not divide evenly into
+ * RMPP segments, then space must be allocated at the end of the referenced
+ * buffer for any required padding.  To indicate the amount of class defined
+ * data being transferred, the paylen_newwin field in the RMPP header should
+ * be set to the size of the class specific header plus the amount of class
+ * defined data being transferred.  The paylen_newwin field should be
+ * specified in network-byte order.
  */
 int ib_post_send_mad(struct ib_mad_agent *mad_agent,
 		     struct ib_send_wr *send_wr,
@@ -334,15 +460,13 @@ int ib_post_send_mad(struct ib_mad_agent *mad_agent,
  *   referenced buffer should be at least the size of the mad_len specified
  *   by @mad_recv_wc.
  *
- * This call copies a chain of received RMPP MADs into a single data buffer,
+ * This call copies a chain of received MAD segments into a single data buffer,
  * removing duplicated headers.
  */
-void ib_coalesce_recv_mad(struct ib_mad_recv_wc *mad_recv_wc,
-			  void *buf);
+void ib_coalesce_recv_mad(struct ib_mad_recv_wc *mad_recv_wc, void *buf);
 
 /**
- * ib_free_recv_mad - Returns data buffers used to receive a MAD to the
- *   access layer.
+ * ib_free_recv_mad - Returns data buffers used to receive a MAD.
  * @mad_recv_wc: Work completion information for a received MAD.
  *
  * Clients receiving MADs through their ib_mad_recv_handler must call this
@@ -358,8 +482,18 @@ void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc);
  * MADs will be returned to the user through the corresponding
  * ib_mad_send_handler.
  */
-void ib_cancel_mad(struct ib_mad_agent *mad_agent,
-		   u64 wr_id);
+void ib_cancel_mad(struct ib_mad_agent *mad_agent, u64 wr_id);
+
+/**
+ * ib_modify_mad - Modifies an outstanding send MAD operation.
+ * @mad_agent: Specifies the registration associated with sent MAD.
+ * @wr_id: Indicates the work request identifier of the MAD to modify.
+ * @timeout_ms: New timeout value for sent MAD.
+ *
+ * This call will reset the timeout value for a sent MAD to the specified
+ * value.
+ */
+int ib_modify_mad(struct ib_mad_agent *mad_agent, u64 wr_id, u32 timeout_ms);
 
 /**
  * ib_redirect_mad_qp - Registers a QP for MAD services.
@@ -401,4 +535,43 @@ struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
 		      struct ib_wc *wc);
 
+/**
+ * ib_create_send_mad - Allocate and initialize a data buffer and work request
+ *   for sending a MAD.
+ * @mad_agent: Specifies the registered MAD service to associate with the MAD.
+ * @remote_qpn: Specifies the QPN of the receiving node.
+ * @pkey_index: Specifies which PKey the MAD will be sent using.  This field
+ *   is valid only if the remote_qpn is QP 1.
+ * @ah: References the address handle used to transfer to the remote node.
+ * @rmpp_active: Indicates if the send will enable RMPP.
+ * @hdr_len: Indicates the size of the data header of the MAD.  This length
+ *   should include the common MAD header, RMPP header, plus any class
+ *   specific header.
+ * @data_len: Indicates the size of any user-transferred data.  The call will
+ *   automatically adjust the allocated buffer size to account for any
+ *   additional padding that may be necessary.
+ * @gfp_mask: GFP mask used for the memory allocation.
+ *
+ * This is a helper routine that may be used to allocate a MAD.  Users are
+ * not required to allocate outbound MADs using this call.  The returned
+ * MAD send buffer will reference a data buffer usable for sending a MAD, along
+ * with an initialized work request structure.  Users may modify the returned
+ * MAD data buffer or work request before posting the send.
+ *
+ * The returned data buffer will be cleared.  Users are responsible for
+ * initializing the common MAD and any class specific headers.  If @rmpp_active
+ * is set, the RMPP header will be initialized for sending.
+ */
+struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
+					    u32 remote_qpn, u16 pkey_index,
+					    struct ib_ah *ah, int rmpp_active,
+					    int hdr_len, int data_len,
+					    unsigned int __nocast gfp_mask);
+
+/**
+ * ib_free_send_mad - Returns data buffers used to send a MAD.
+ * @send_buf: Previously allocated send data buffer.
+ */
+void ib_free_send_mad(struct ib_mad_send_buf *send_buf);
+
 #endif /* IB_MAD_H */
diff --git a/drivers/infiniband/include/ib_sa.h b/drivers/infiniband/include/ib_sa.h
index 00222285eb9a..6d999f7b5d93 100644
--- a/drivers/infiniband/include/ib_sa.h
+++ b/drivers/infiniband/include/ib_sa.h
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Voltaire, Inc.  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
@@ -29,7 +30,7 @@
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  *
- * $Id: ib_sa.h 1389 2004-12-27 22:56:47Z roland $
+ * $Id: ib_sa.h 2811 2005-07-06 18:11:43Z halr $
  */
 
 #ifndef IB_SA_H
@@ -41,9 +42,11 @@
 #include <ib_mad.h>
 
 enum {
-	IB_SA_CLASS_VERSION	= 2,	/* IB spec version 1.1/1.2 */
+	IB_SA_CLASS_VERSION		= 2,	/* IB spec version 1.1/1.2 */
 
-	IB_SA_METHOD_DELETE	= 0x15
+	IB_SA_METHOD_GET_TABLE		= 0x12,
+	IB_SA_METHOD_GET_TABLE_RESP	= 0x92,
+	IB_SA_METHOD_DELETE		= 0x15
 };
 
 enum ib_sa_selector {
@@ -87,10 +90,6 @@ static inline int ib_sa_rate_enum_to_int(enum ib_sa_rate rate)
 	}
 }
 
-typedef u64 __bitwise ib_sa_comp_mask;
-
-#define IB_SA_COMP_MASK(n)	((__force ib_sa_comp_mask) cpu_to_be64(1ull << n))
-
 /*
  * Structures for SA records are named "struct ib_sa_xxx_rec."  No
  * attempt is made to pack structures to match the physical layout of
@@ -195,6 +194,61 @@ struct ib_sa_mcmember_rec {
 	int          proxy_join;
 };
 
+/* Service Record Component Mask Sec 15.2.5.14 Ver 1.1	*/
+#define IB_SA_SERVICE_REC_SERVICE_ID			IB_SA_COMP_MASK( 0)
+#define IB_SA_SERVICE_REC_SERVICE_GID			IB_SA_COMP_MASK( 1)
+#define IB_SA_SERVICE_REC_SERVICE_PKEY			IB_SA_COMP_MASK( 2)
+/* reserved:								 3 */
+#define IB_SA_SERVICE_REC_SERVICE_LEASE			IB_SA_COMP_MASK( 4)
+#define IB_SA_SERVICE_REC_SERVICE_KEY			IB_SA_COMP_MASK( 5)
+#define IB_SA_SERVICE_REC_SERVICE_NAME			IB_SA_COMP_MASK( 6)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_0		IB_SA_COMP_MASK( 7)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_1		IB_SA_COMP_MASK( 8)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_2		IB_SA_COMP_MASK( 9)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_3		IB_SA_COMP_MASK(10)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_4		IB_SA_COMP_MASK(11)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_5		IB_SA_COMP_MASK(12)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_6		IB_SA_COMP_MASK(13)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_7		IB_SA_COMP_MASK(14)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_8		IB_SA_COMP_MASK(15)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_9		IB_SA_COMP_MASK(16)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_10		IB_SA_COMP_MASK(17)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_11		IB_SA_COMP_MASK(18)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_12		IB_SA_COMP_MASK(19)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_13		IB_SA_COMP_MASK(20)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_14		IB_SA_COMP_MASK(21)
+#define IB_SA_SERVICE_REC_SERVICE_DATA8_15		IB_SA_COMP_MASK(22)
+#define IB_SA_SERVICE_REC_SERVICE_DATA16_0		IB_SA_COMP_MASK(23)
+#define IB_SA_SERVICE_REC_SERVICE_DATA16_1		IB_SA_COMP_MASK(24)
+#define IB_SA_SERVICE_REC_SERVICE_DATA16_2		IB_SA_COMP_MASK(25)
+#define IB_SA_SERVICE_REC_SERVICE_DATA16_3		IB_SA_COMP_MASK(26)
+#define IB_SA_SERVICE_REC_SERVICE_DATA16_4		IB_SA_COMP_MASK(27)
+#define IB_SA_SERVICE_REC_SERVICE_DATA16_5		IB_SA_COMP_MASK(28)
+#define IB_SA_SERVICE_REC_SERVICE_DATA16_6		IB_SA_COMP_MASK(29)
+#define IB_SA_SERVICE_REC_SERVICE_DATA16_7		IB_SA_COMP_MASK(30)
+#define IB_SA_SERVICE_REC_SERVICE_DATA32_0		IB_SA_COMP_MASK(31)
+#define IB_SA_SERVICE_REC_SERVICE_DATA32_1		IB_SA_COMP_MASK(32)
+#define IB_SA_SERVICE_REC_SERVICE_DATA32_2		IB_SA_COMP_MASK(33)
+#define IB_SA_SERVICE_REC_SERVICE_DATA32_3		IB_SA_COMP_MASK(34)
+#define IB_SA_SERVICE_REC_SERVICE_DATA64_0		IB_SA_COMP_MASK(35)
+#define IB_SA_SERVICE_REC_SERVICE_DATA64_1		IB_SA_COMP_MASK(36)
+
+#define IB_DEFAULT_SERVICE_LEASE 	0xFFFFFFFF
+
+struct ib_sa_service_rec {
+	u64		id;
+	union ib_gid	gid;
+	u16 		pkey;
+	/* reserved */
+	u32		lease;
+	u8		key[16];
+	u8		name[64];
+	u8		data8[16];
+	u16		data16[8];
+	u32		data32[4];
+	u64		data64[2];
+};
+
 struct ib_sa_query;
 
 void ib_sa_cancel_query(int id, struct ib_sa_query *query);
@@ -202,7 +256,7 @@ void ib_sa_cancel_query(int id, struct ib_sa_query *query);
 int ib_sa_path_rec_get(struct ib_device *device, u8 port_num,
 		       struct ib_sa_path_rec *rec,
 		       ib_sa_comp_mask comp_mask,
-		       int timeout_ms, int gfp_mask,
+		       int timeout_ms, unsigned int __nocast gfp_mask,
 		       void (*callback)(int status,
 					struct ib_sa_path_rec *resp,
 					void *context),
@@ -213,13 +267,24 @@ int ib_sa_mcmember_rec_query(struct ib_device *device, u8 port_num,
 			     u8 method,
 			     struct ib_sa_mcmember_rec *rec,
 			     ib_sa_comp_mask comp_mask,
-			     int timeout_ms, int gfp_mask,
+			     int timeout_ms, unsigned int __nocast gfp_mask,
 			     void (*callback)(int status,
 					      struct ib_sa_mcmember_rec *resp,
 					      void *context),
 			     void *context,
 			     struct ib_sa_query **query);
 
+int ib_sa_service_rec_query(struct ib_device *device, u8 port_num,
+			 u8 method,
+			 struct ib_sa_service_rec *rec,
+			 ib_sa_comp_mask comp_mask,
+			 int timeout_ms, unsigned int __nocast gfp_mask,
+			 void (*callback)(int status,
+					  struct ib_sa_service_rec *resp,
+					  void *context),
+			 void *context,
+			 struct ib_sa_query **sa_query);
+
 /**
  * ib_sa_mcmember_rec_set - Start an MCMember set query
  * @device:device to send query on
@@ -248,7 +313,7 @@ static inline int
 ib_sa_mcmember_rec_set(struct ib_device *device, u8 port_num,
 		       struct ib_sa_mcmember_rec *rec,
 		       ib_sa_comp_mask comp_mask,
-		       int timeout_ms, int gfp_mask,
+		       int timeout_ms, unsigned int __nocast gfp_mask,
 		       void (*callback)(int status,
 					struct ib_sa_mcmember_rec *resp,
 					void *context),
@@ -290,7 +355,7 @@ static inline int
 ib_sa_mcmember_rec_delete(struct ib_device *device, u8 port_num,
 			  struct ib_sa_mcmember_rec *rec,
 			  ib_sa_comp_mask comp_mask,
-			  int timeout_ms, int gfp_mask,
+			  int timeout_ms, unsigned int __nocast gfp_mask,
 			  void (*callback)(int status,
 					   struct ib_sa_mcmember_rec *resp,
 					   void *context),
diff --git a/drivers/infiniband/include/ib_user_cm.h b/drivers/infiniband/include/ib_user_cm.h
new file mode 100644
index 000000000000..500b1af6ff77
--- /dev/null
+++ b/drivers/infiniband/include/ib_user_cm.h
@@ -0,0 +1,328 @@
+/*
+ * Copyright (c) 2005 Topspin Communications.  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.
+ *
+ * $Id: ib_user_cm.h 2576 2005-06-09 17:00:30Z libor $
+ */
+
+#ifndef IB_USER_CM_H
+#define IB_USER_CM_H
+
+#include <linux/types.h>
+
+#define IB_USER_CM_ABI_VERSION 1
+
+enum {
+	IB_USER_CM_CMD_CREATE_ID,
+	IB_USER_CM_CMD_DESTROY_ID,
+	IB_USER_CM_CMD_ATTR_ID,
+
+	IB_USER_CM_CMD_LISTEN,
+	IB_USER_CM_CMD_ESTABLISH,
+
+	IB_USER_CM_CMD_SEND_REQ,
+	IB_USER_CM_CMD_SEND_REP,
+	IB_USER_CM_CMD_SEND_RTU,
+	IB_USER_CM_CMD_SEND_DREQ,
+	IB_USER_CM_CMD_SEND_DREP,
+	IB_USER_CM_CMD_SEND_REJ,
+	IB_USER_CM_CMD_SEND_MRA,
+	IB_USER_CM_CMD_SEND_LAP,
+	IB_USER_CM_CMD_SEND_APR,
+	IB_USER_CM_CMD_SEND_SIDR_REQ,
+	IB_USER_CM_CMD_SEND_SIDR_REP,
+
+	IB_USER_CM_CMD_EVENT,
+};
+/*
+ * command ABI structures.
+ */
+struct ib_ucm_cmd_hdr {
+	__u32 cmd;
+	__u16 in;
+	__u16 out;
+};
+
+struct ib_ucm_create_id {
+	__u64 response;
+};
+
+struct ib_ucm_create_id_resp {
+	__u32 id;
+};
+
+struct ib_ucm_destroy_id {
+	__u32 id;
+};
+
+struct ib_ucm_attr_id {
+	__u64 response;
+	__u32 id;
+};
+
+struct ib_ucm_attr_id_resp {
+	__u64 service_id;
+	__u64 service_mask;
+	__u32 local_id;
+	__u32 remote_id;
+};
+
+struct ib_ucm_listen {
+	__u64 service_id;
+	__u64 service_mask;
+	__u32 id;
+};
+
+struct ib_ucm_establish {
+	__u32 id;
+};
+
+struct ib_ucm_private_data {
+	__u64 data;
+	__u32 id;
+	__u8  len;
+	__u8  reserved[3];
+};
+
+struct ib_ucm_path_rec {
+	__u8  dgid[16];
+	__u8  sgid[16];
+	__u16 dlid;
+	__u16 slid;
+	__u32 raw_traffic;
+	__u32 flow_label;
+	__u32 reversible;
+	__u32 mtu;
+	__u16 pkey;
+	__u8  hop_limit;
+	__u8  traffic_class;
+	__u8  numb_path;
+	__u8  sl;
+	__u8  mtu_selector;
+	__u8  rate_selector;
+	__u8  rate;
+	__u8  packet_life_time_selector;
+	__u8  packet_life_time;
+	__u8  preference;
+};
+
+struct ib_ucm_req {
+	__u32 id;
+	__u32 qpn;
+	__u32 qp_type;
+	__u32 psn;
+	__u64 sid;
+	__u64 data;
+	__u64 primary_path;
+	__u64 alternate_path;
+	__u8  len;
+	__u8  peer_to_peer;
+	__u8  responder_resources;
+	__u8  initiator_depth;
+	__u8  remote_cm_response_timeout;
+	__u8  flow_control;
+	__u8  local_cm_response_timeout;
+	__u8  retry_count;
+	__u8  rnr_retry_count;
+	__u8  max_cm_retries;
+	__u8  srq;
+	__u8  reserved[1];
+};
+
+struct ib_ucm_rep {
+	__u64 data;
+	__u32 id;
+	__u32 qpn;
+	__u32 psn;
+	__u8  len;
+	__u8  responder_resources;
+	__u8  initiator_depth;
+	__u8  target_ack_delay;
+	__u8  failover_accepted;
+	__u8  flow_control;
+	__u8  rnr_retry_count;
+	__u8  srq;
+};
+
+struct ib_ucm_info {
+	__u32 id;
+	__u32 status;
+	__u64 info;
+	__u64 data;
+	__u8  info_len;
+	__u8  data_len;
+	__u8  reserved[2];
+};
+
+struct ib_ucm_mra {
+	__u64 data;
+	__u32 id;
+	__u8  len;
+	__u8  timeout;
+	__u8  reserved[2];
+};
+
+struct ib_ucm_lap {
+	__u64 path;
+	__u64 data;
+	__u32 id;
+	__u8  len;
+	__u8  reserved[3];
+};
+
+struct ib_ucm_sidr_req {
+	__u32 id;
+	__u32 timeout;
+	__u64 sid;
+	__u64 data;
+	__u64 path;
+	__u16 pkey;
+	__u8  len;
+	__u8  max_cm_retries;
+};
+
+struct ib_ucm_sidr_rep {
+	__u32 id;
+	__u32 qpn;
+	__u32 qkey;
+	__u32 status;
+	__u64 info;
+	__u64 data;
+	__u8  info_len;
+	__u8  data_len;
+	__u8  reserved[2];
+};
+/*
+ * event notification ABI structures.
+ */
+struct ib_ucm_event_get {
+	__u64 response;
+	__u64 data;
+	__u64 info;
+	__u8  data_len;
+	__u8  info_len;
+	__u8  reserved[2];
+};
+
+struct ib_ucm_req_event_resp {
+	__u32                  listen_id;
+	/* device */
+	/* port */
+	struct ib_ucm_path_rec primary_path;
+	struct ib_ucm_path_rec alternate_path;
+	__u64                  remote_ca_guid;
+	__u32                  remote_qkey;
+	__u32                  remote_qpn;
+	__u32                  qp_type;
+	__u32                  starting_psn;
+	__u8  responder_resources;
+	__u8  initiator_depth;
+	__u8  local_cm_response_timeout;
+	__u8  flow_control;
+	__u8  remote_cm_response_timeout;
+	__u8  retry_count;
+	__u8  rnr_retry_count;
+	__u8  srq;
+};
+
+struct ib_ucm_rep_event_resp {
+	__u64 remote_ca_guid;
+	__u32 remote_qkey;
+	__u32 remote_qpn;
+	__u32 starting_psn;
+	__u8  responder_resources;
+	__u8  initiator_depth;
+	__u8  target_ack_delay;
+	__u8  failover_accepted;
+	__u8  flow_control;
+	__u8  rnr_retry_count;
+	__u8  srq;
+	__u8  reserved[1];
+};
+
+struct ib_ucm_rej_event_resp {
+	__u32 reason;
+	/* ari in ib_ucm_event_get info field. */
+};
+
+struct ib_ucm_mra_event_resp {
+	__u8  timeout;
+	__u8  reserved[3];
+};
+
+struct ib_ucm_lap_event_resp {
+	struct ib_ucm_path_rec path;
+};
+
+struct ib_ucm_apr_event_resp {
+	__u32 status;
+	/* apr info in ib_ucm_event_get info field. */
+};
+
+struct ib_ucm_sidr_req_event_resp {
+	__u32 listen_id;
+	/* device */
+	/* port */
+	__u16 pkey;
+	__u8  reserved[2];
+};
+
+struct ib_ucm_sidr_rep_event_resp {
+	__u32 status;
+	__u32 qkey;
+	__u32 qpn;
+	/* info in ib_ucm_event_get info field. */
+};
+
+#define IB_UCM_PRES_DATA      0x01
+#define IB_UCM_PRES_INFO      0x02
+#define IB_UCM_PRES_PRIMARY   0x04
+#define IB_UCM_PRES_ALTERNATE 0x08
+
+struct ib_ucm_event_resp {
+	__u32 id;
+	__u32 event;
+	__u32 present;
+	union {
+		struct ib_ucm_req_event_resp req_resp;
+		struct ib_ucm_rep_event_resp rep_resp;
+		struct ib_ucm_rej_event_resp rej_resp;
+		struct ib_ucm_mra_event_resp mra_resp;
+		struct ib_ucm_lap_event_resp lap_resp;
+		struct ib_ucm_apr_event_resp apr_resp;
+
+		struct ib_ucm_sidr_req_event_resp sidr_req_resp;
+		struct ib_ucm_sidr_rep_event_resp sidr_rep_resp;
+
+		__u32                             send_status;
+	} u;
+};
+
+#endif /* IB_USER_CM_H */
diff --git a/drivers/infiniband/include/ib_user_mad.h b/drivers/infiniband/include/ib_user_mad.h
index 06ad4a6075fa..a9a56b50aacc 100644
--- a/drivers/infiniband/include/ib_user_mad.h
+++ b/drivers/infiniband/include/ib_user_mad.h
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Voltaire, Inc. 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
@@ -29,7 +30,7 @@
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  *
- * $Id: ib_user_mad.h 1389 2004-12-27 22:56:47Z roland $
+ * $Id: ib_user_mad.h 2814 2005-07-06 19:14:09Z halr $
  */
 
 #ifndef IB_USER_MAD_H
@@ -42,7 +43,7 @@
  * Increment this value if any changes that break userspace ABI
  * compatibility are made.
  */
-#define IB_USER_MAD_ABI_VERSION	2
+#define IB_USER_MAD_ABI_VERSION	5
 
 /*
  * Make sure that all structs defined in this file remain laid out so
@@ -51,13 +52,13 @@
  */
 
 /**
- * ib_user_mad - MAD packet
- * @data - Contents of MAD
+ * ib_user_mad_hdr - MAD packet header
  * @id - ID of agent MAD received with/to be sent with
  * @status - 0 on successful receive, ETIMEDOUT if no response
  *   received (transaction ID in data[] will be set to TID of original
  *   request) (ignored on send)
  * @timeout_ms - Milliseconds to wait for response (unset on receive)
+ * @retries - Number of automatic retries to attempt
  * @qpn - Remote QP number received from/to be sent to
  * @qkey - Remote Q_Key to be sent with (unset on receive)
  * @lid - Remote lid received from/to be sent to
@@ -72,11 +73,12 @@
  *
  * All multi-byte quantities are stored in network (big endian) byte order.
  */
-struct ib_user_mad {
-	__u8	data[256];
+struct ib_user_mad_hdr {
 	__u32	id;
 	__u32	status;
 	__u32	timeout_ms;
+	__u32	retries;
+	__u32	length;
 	__u32	qpn;
 	__u32   qkey;
 	__u16	lid;
@@ -91,6 +93,17 @@ struct ib_user_mad {
 };
 
 /**
+ * ib_user_mad - MAD packet
+ * @hdr - MAD packet header
+ * @data - Contents of MAD
+ *
+ */
+struct ib_user_mad {
+	struct ib_user_mad_hdr hdr;
+	__u8	data[0];
+};
+
+/**
  * ib_user_mad_reg_req - MAD registration request
  * @id - Set by the kernel; used to identify agent in future requests.
  * @qpn - Queue pair number; must be 0 or 1.
@@ -103,6 +116,8 @@ struct ib_user_mad {
  *   management class to receive.
  * @oui: Indicates IEEE OUI when mgmt_class is a vendor class
  *   in the range from 0x30 to 0x4f. Otherwise not used.
+ * @rmpp_version: If set, indicates the RMPP version used.
+ *
  */
 struct ib_user_mad_reg_req {
 	__u32	id;
@@ -111,6 +126,7 @@ struct ib_user_mad_reg_req {
 	__u8	mgmt_class;
 	__u8	mgmt_class_version;
 	__u8    oui[3];
+	__u8	rmpp_version;
 };
 
 #define IB_IOCTL_MAGIC		0x1b
diff --git a/drivers/infiniband/include/ib_user_verbs.h b/drivers/infiniband/include/ib_user_verbs.h
new file mode 100644
index 000000000000..7c613706af72
--- /dev/null
+++ b/drivers/infiniband/include/ib_user_verbs.h
@@ -0,0 +1,389 @@
+/*
+ * Copyright (c) 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  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.
+ *
+ * $Id: ib_user_verbs.h 2708 2005-06-24 17:27:21Z roland $
+ */
+
+#ifndef IB_USER_VERBS_H
+#define IB_USER_VERBS_H
+
+#include <linux/types.h>
+
+/*
+ * Increment this value if any changes that break userspace ABI
+ * compatibility are made.
+ */
+#define IB_USER_VERBS_ABI_VERSION	1
+
+enum {
+	IB_USER_VERBS_CMD_QUERY_PARAMS,
+	IB_USER_VERBS_CMD_GET_CONTEXT,
+	IB_USER_VERBS_CMD_QUERY_DEVICE,
+	IB_USER_VERBS_CMD_QUERY_PORT,
+	IB_USER_VERBS_CMD_QUERY_GID,
+	IB_USER_VERBS_CMD_QUERY_PKEY,
+	IB_USER_VERBS_CMD_ALLOC_PD,
+	IB_USER_VERBS_CMD_DEALLOC_PD,
+	IB_USER_VERBS_CMD_CREATE_AH,
+	IB_USER_VERBS_CMD_MODIFY_AH,
+	IB_USER_VERBS_CMD_QUERY_AH,
+	IB_USER_VERBS_CMD_DESTROY_AH,
+	IB_USER_VERBS_CMD_REG_MR,
+	IB_USER_VERBS_CMD_REG_SMR,
+	IB_USER_VERBS_CMD_REREG_MR,
+	IB_USER_VERBS_CMD_QUERY_MR,
+	IB_USER_VERBS_CMD_DEREG_MR,
+	IB_USER_VERBS_CMD_ALLOC_MW,
+	IB_USER_VERBS_CMD_BIND_MW,
+	IB_USER_VERBS_CMD_DEALLOC_MW,
+	IB_USER_VERBS_CMD_CREATE_CQ,
+	IB_USER_VERBS_CMD_RESIZE_CQ,
+	IB_USER_VERBS_CMD_DESTROY_CQ,
+	IB_USER_VERBS_CMD_POLL_CQ,
+	IB_USER_VERBS_CMD_PEEK_CQ,
+	IB_USER_VERBS_CMD_REQ_NOTIFY_CQ,
+	IB_USER_VERBS_CMD_CREATE_QP,
+	IB_USER_VERBS_CMD_QUERY_QP,
+	IB_USER_VERBS_CMD_MODIFY_QP,
+	IB_USER_VERBS_CMD_DESTROY_QP,
+	IB_USER_VERBS_CMD_POST_SEND,
+	IB_USER_VERBS_CMD_POST_RECV,
+	IB_USER_VERBS_CMD_ATTACH_MCAST,
+	IB_USER_VERBS_CMD_DETACH_MCAST
+};
+
+/*
+ * Make sure that all structs defined in this file remain laid out so
+ * that they pack the same way on 32-bit and 64-bit architectures (to
+ * avoid incompatibility between 32-bit userspace and 64-bit kernels).
+ * In particular do not use pointer types -- pass pointers in __u64
+ * instead.
+ */
+
+struct ib_uverbs_async_event_desc {
+	__u64 element;
+	__u32 event_type;	/* enum ib_event_type */
+	__u32 reserved;
+};
+
+struct ib_uverbs_comp_event_desc {
+	__u64 cq_handle;
+};
+
+/*
+ * All commands from userspace should start with a __u32 command field
+ * followed by __u16 in_words and out_words fields (which give the
+ * length of the command block and response buffer if any in 32-bit
+ * words).  The kernel driver will read these fields first and read
+ * the rest of the command struct based on these value.
+ */
+
+struct ib_uverbs_cmd_hdr {
+	__u32 command;
+	__u16 in_words;
+	__u16 out_words;
+};
+
+/*
+ * No driver_data for "query params" command, since this is intended
+ * to be a core function with no possible device dependence.
+ */
+struct ib_uverbs_query_params {
+	__u64 response;
+};
+
+struct ib_uverbs_query_params_resp {
+	__u32 num_cq_events;
+};
+
+struct ib_uverbs_get_context {
+	__u64 response;
+	__u64 cq_fd_tab;
+	__u64 driver_data[0];
+};
+
+struct ib_uverbs_get_context_resp {
+	__u32 async_fd;
+	__u32 reserved;
+};
+
+struct ib_uverbs_query_device {
+	__u64 response;
+	__u64 driver_data[0];
+};
+
+struct ib_uverbs_query_device_resp {
+	__u64 fw_ver;
+	__u64 node_guid;
+	__u64 sys_image_guid;
+	__u64 max_mr_size;
+	__u64 page_size_cap;
+	__u32 vendor_id;
+	__u32 vendor_part_id;
+	__u32 hw_ver;
+	__u32 max_qp;
+	__u32 max_qp_wr;
+	__u32 device_cap_flags;
+	__u32 max_sge;
+	__u32 max_sge_rd;
+	__u32 max_cq;
+	__u32 max_cqe;
+	__u32 max_mr;
+	__u32 max_pd;
+	__u32 max_qp_rd_atom;
+	__u32 max_ee_rd_atom;
+	__u32 max_res_rd_atom;
+	__u32 max_qp_init_rd_atom;
+	__u32 max_ee_init_rd_atom;
+	__u32 atomic_cap;
+	__u32 max_ee;
+	__u32 max_rdd;
+	__u32 max_mw;
+	__u32 max_raw_ipv6_qp;
+	__u32 max_raw_ethy_qp;
+	__u32 max_mcast_grp;
+	__u32 max_mcast_qp_attach;
+	__u32 max_total_mcast_qp_attach;
+	__u32 max_ah;
+	__u32 max_fmr;
+	__u32 max_map_per_fmr;
+	__u32 max_srq;
+	__u32 max_srq_wr;
+	__u32 max_srq_sge;
+	__u16 max_pkeys;
+	__u8  local_ca_ack_delay;
+	__u8  phys_port_cnt;
+	__u8  reserved[4];
+};
+
+struct ib_uverbs_query_port {
+	__u64 response;
+	__u8  port_num;
+	__u8  reserved[7];
+	__u64 driver_data[0];
+};
+
+struct ib_uverbs_query_port_resp {
+	__u32 port_cap_flags;
+	__u32 max_msg_sz;
+	__u32 bad_pkey_cntr;
+	__u32 qkey_viol_cntr;
+	__u32 gid_tbl_len;
+	__u16 pkey_tbl_len;
+	__u16 lid;
+	__u16 sm_lid;
+	__u8  state;
+	__u8  max_mtu;
+	__u8  active_mtu;
+	__u8  lmc;
+	__u8  max_vl_num;
+	__u8  sm_sl;
+	__u8  subnet_timeout;
+	__u8  init_type_reply;
+	__u8  active_width;
+	__u8  active_speed;
+	__u8  phys_state;
+	__u8  reserved[3];
+};
+
+struct ib_uverbs_query_gid {
+	__u64 response;
+	__u8  port_num;
+	__u8  index;
+	__u8  reserved[6];
+	__u64 driver_data[0];
+};
+
+struct ib_uverbs_query_gid_resp {
+	__u8  gid[16];
+};
+
+struct ib_uverbs_query_pkey {
+	__u64 response;
+	__u8  port_num;
+	__u8  index;
+	__u8  reserved[6];
+	__u64 driver_data[0];
+};
+
+struct ib_uverbs_query_pkey_resp {
+	__u16 pkey;
+	__u16 reserved;
+};
+
+struct ib_uverbs_alloc_pd {
+	__u64 response;
+	__u64 driver_data[0];
+};
+
+struct ib_uverbs_alloc_pd_resp {
+	__u32 pd_handle;
+};
+
+struct ib_uverbs_dealloc_pd {
+	__u32 pd_handle;
+};
+
+struct ib_uverbs_reg_mr {
+	__u64 response;
+	__u64 start;
+	__u64 length;
+	__u64 hca_va;
+	__u32 pd_handle;
+	__u32 access_flags;
+	__u64 driver_data[0];
+};
+
+struct ib_uverbs_reg_mr_resp {
+	__u32 mr_handle;
+	__u32 lkey;
+	__u32 rkey;
+};
+
+struct ib_uverbs_dereg_mr {
+	__u32 mr_handle;
+};
+
+struct ib_uverbs_create_cq {
+	__u64 response;
+	__u64 user_handle;
+	__u32 cqe;
+	__u32 event_handler;
+	__u64 driver_data[0];
+};
+
+struct ib_uverbs_create_cq_resp {
+	__u32 cq_handle;
+	__u32 cqe;
+};
+
+struct ib_uverbs_destroy_cq {
+	__u32 cq_handle;
+};
+
+struct ib_uverbs_create_qp {
+	__u64 response;
+	__u64 user_handle;
+	__u32 pd_handle;
+	__u32 send_cq_handle;
+	__u32 recv_cq_handle;
+	__u32 srq_handle;
+	__u32 max_send_wr;
+	__u32 max_recv_wr;
+	__u32 max_send_sge;
+	__u32 max_recv_sge;
+	__u32 max_inline_data;
+	__u8  sq_sig_all;
+	__u8  qp_type;
+	__u8  is_srq;
+	__u8  reserved;
+	__u64 driver_data[0];
+};
+
+struct ib_uverbs_create_qp_resp {
+	__u32 qp_handle;
+	__u32 qpn;
+};
+
+/*
+ * This struct needs to remain a multiple of 8 bytes to keep the
+ * alignment of the modify QP parameters.
+ */
+struct ib_uverbs_qp_dest {
+	__u8  dgid[16];
+	__u32 flow_label;
+	__u16 dlid;
+	__u16 reserved;
+	__u8  sgid_index;
+	__u8  hop_limit;
+	__u8  traffic_class;
+	__u8  sl;
+	__u8  src_path_bits;
+	__u8  static_rate;
+	__u8  is_global;
+	__u8  port_num;
+};
+
+struct ib_uverbs_modify_qp {
+	struct ib_uverbs_qp_dest dest;
+	struct ib_uverbs_qp_dest alt_dest;
+	__u32 qp_handle;
+	__u32 attr_mask;
+	__u32 qkey;
+	__u32 rq_psn;
+	__u32 sq_psn;
+	__u32 dest_qp_num;
+	__u32 qp_access_flags;
+	__u16 pkey_index;
+	__u16 alt_pkey_index;
+	__u8  qp_state;
+	__u8  cur_qp_state;
+	__u8  path_mtu;
+	__u8  path_mig_state;
+	__u8  en_sqd_async_notify;
+	__u8  max_rd_atomic;
+	__u8  max_dest_rd_atomic;
+	__u8  min_rnr_timer;
+	__u8  port_num;
+	__u8  timeout;
+	__u8  retry_cnt;
+	__u8  rnr_retry;
+	__u8  alt_port_num;
+	__u8  alt_timeout;
+	__u8  reserved[2];
+	__u64 driver_data[0];
+};
+
+struct ib_uverbs_modify_qp_resp {
+};
+
+struct ib_uverbs_destroy_qp {
+	__u32 qp_handle;
+};
+
+struct ib_uverbs_attach_mcast {
+	__u8  gid[16];
+	__u32 qp_handle;
+	__u16 mlid;
+	__u16 reserved;
+	__u64 driver_data[0];
+};
+
+struct ib_uverbs_detach_mcast {
+	__u8  gid[16];
+	__u32 qp_handle;
+	__u16 mlid;
+	__u16 reserved;
+	__u64 driver_data[0];
+};
+
+#endif /* IB_USER_VERBS_H */
diff --git a/drivers/infiniband/include/ib_verbs.h b/drivers/infiniband/include/ib_verbs.h
index cf01f044a223..5d24edaa66e6 100644
--- a/drivers/infiniband/include/ib_verbs.h
+++ b/drivers/infiniband/include/ib_verbs.h
@@ -4,6 +4,7 @@
  * Copyright (c) 2004 Intel Corporation.  All rights reserved.
  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
  * Copyright (c) 2004 Voltaire Corporation.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  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
@@ -41,7 +42,10 @@
 
 #include <linux/types.h>
 #include <linux/device.h>
+
 #include <asm/atomic.h>
+#include <asm/scatterlist.h>
+#include <asm/uaccess.h>
 
 union ib_gid {
 	u8	raw[16];
@@ -285,6 +289,15 @@ struct ib_global_route {
 	u8		traffic_class;
 };
 
+struct ib_grh {
+	u32		version_tclass_flow;
+	u16		paylen;
+	u8		next_hdr;
+	u8		hop_limit;
+	union ib_gid	sgid;
+	union ib_gid	dgid;
+};
+
 enum {
 	IB_MULTICAST_QPN = 0xffffff
 };
@@ -544,7 +557,7 @@ struct ib_send_wr {
 	int			num_sge;
 	enum ib_wr_opcode	opcode;
 	int			send_flags;
-	u32			imm_data;
+	__be32			imm_data;
 	union {
 		struct {
 			u64	remote_addr;
@@ -562,6 +575,7 @@ struct ib_send_wr {
 			u32	remote_qpn;
 			u32	remote_qkey;
 			int	timeout_ms; /* valid for MADs only */
+			int	retries;    /* valid for MADs only */
 			u16	pkey_index; /* valid for GSI only */
 			u8	port_num;   /* valid for DR SMPs on switch only */
 		} ud;
@@ -618,29 +632,86 @@ struct ib_fmr_attr {
 	u8	page_size;
 };
 
+struct ib_ucontext {
+	struct ib_device       *device;
+	struct list_head	pd_list;
+	struct list_head	mr_list;
+	struct list_head	mw_list;
+	struct list_head	cq_list;
+	struct list_head	qp_list;
+	struct list_head	srq_list;
+	struct list_head	ah_list;
+	spinlock_t              lock;
+};
+
+struct ib_uobject {
+	u64			user_handle;	/* handle given to us by userspace */
+	struct ib_ucontext     *context;	/* associated user context */
+	struct list_head	list;		/* link to context's list */
+	u32			id;		/* index into kernel idr */
+};
+
+struct ib_umem {
+	unsigned long		user_base;
+	unsigned long		virt_base;
+	size_t			length;
+	int			offset;
+	int			page_size;
+	int                     writable;
+	struct list_head	chunk_list;
+};
+
+struct ib_umem_chunk {
+	struct list_head	list;
+	int                     nents;
+	int                     nmap;
+	struct scatterlist      page_list[0];
+};
+
+struct ib_udata {
+	void __user *inbuf;
+	void __user *outbuf;
+	size_t       inlen;
+	size_t       outlen;
+};
+
+#define IB_UMEM_MAX_PAGE_CHUNK						\
+	((PAGE_SIZE - offsetof(struct ib_umem_chunk, page_list)) /	\
+	 ((void *) &((struct ib_umem_chunk *) 0)->page_list[1] -	\
+	  (void *) &((struct ib_umem_chunk *) 0)->page_list[0]))
+
+struct ib_umem_object {
+	struct ib_uobject	uobject;
+	struct ib_umem		umem;
+};
+
 struct ib_pd {
-	struct ib_device *device;
-	atomic_t          usecnt; /* count all resources */
+	struct ib_device       *device;
+	struct ib_uobject      *uobject;
+	atomic_t          	usecnt; /* count all resources */
 };
 
 struct ib_ah {
 	struct ib_device	*device;
 	struct ib_pd		*pd;
+	struct ib_uobject	*uobject;
 };
 
 typedef void (*ib_comp_handler)(struct ib_cq *cq, void *cq_context);
 
 struct ib_cq {
-	struct ib_device *device;
-	ib_comp_handler   comp_handler;
-	void             (*event_handler)(struct ib_event *, void *);
-	void *            cq_context;
-	int               cqe;
-	atomic_t          usecnt; /* count number of work queues */
+	struct ib_device       *device;
+	struct ib_uobject      *uobject;
+	ib_comp_handler   	comp_handler;
+	void                  (*event_handler)(struct ib_event *, void *);
+	void *            	cq_context;
+	int               	cqe;
+	atomic_t          	usecnt; /* count number of work queues */
 };
 
 struct ib_srq {
 	struct ib_device	*device;
+	struct ib_uobject	*uobject;
 	struct ib_pd		*pd;
 	void			*srq_context;
 	atomic_t		usecnt;
@@ -652,6 +723,7 @@ struct ib_qp {
 	struct ib_cq	       *send_cq;
 	struct ib_cq	       *recv_cq;
 	struct ib_srq	       *srq;
+	struct ib_uobject      *uobject;
 	void                  (*event_handler)(struct ib_event *, void *);
 	void		       *qp_context;
 	u32			qp_num;
@@ -659,16 +731,18 @@ struct ib_qp {
 };
 
 struct ib_mr {
-	struct ib_device *device;
-	struct ib_pd     *pd;
-	u32		  lkey;
-	u32		  rkey;
-	atomic_t          usecnt; /* count number of MWs */
+	struct ib_device  *device;
+	struct ib_pd	  *pd;
+	struct ib_uobject *uobject;
+	u32		   lkey;
+	u32		   rkey;
+	atomic_t	   usecnt; /* count number of MWs */
 };
 
 struct ib_mw {
 	struct ib_device	*device;
 	struct ib_pd		*pd;
+	struct ib_uobject	*uobject;
 	u32			rkey;
 };
 
@@ -737,7 +811,14 @@ struct ib_device {
 	int		           (*modify_port)(struct ib_device *device,
 						  u8 port_num, int port_modify_mask,
 						  struct ib_port_modify *port_modify);
-	struct ib_pd *             (*alloc_pd)(struct ib_device *device);
+	struct ib_ucontext *       (*alloc_ucontext)(struct ib_device *device,
+						     struct ib_udata *udata);
+	int                        (*dealloc_ucontext)(struct ib_ucontext *context);
+	int                        (*mmap)(struct ib_ucontext *context,
+					   struct vm_area_struct *vma);
+	struct ib_pd *             (*alloc_pd)(struct ib_device *device,
+					       struct ib_ucontext *context,
+					       struct ib_udata *udata);
 	int                        (*dealloc_pd)(struct ib_pd *pd);
 	struct ib_ah *             (*create_ah)(struct ib_pd *pd,
 						struct ib_ah_attr *ah_attr);
@@ -747,7 +828,8 @@ struct ib_device {
 					       struct ib_ah_attr *ah_attr);
 	int                        (*destroy_ah)(struct ib_ah *ah);
 	struct ib_qp *             (*create_qp)(struct ib_pd *pd,
-						struct ib_qp_init_attr *qp_init_attr);
+						struct ib_qp_init_attr *qp_init_attr,
+						struct ib_udata *udata);
 	int                        (*modify_qp)(struct ib_qp *qp,
 						struct ib_qp_attr *qp_attr,
 						int qp_attr_mask);
@@ -762,8 +844,9 @@ struct ib_device {
 	int                        (*post_recv)(struct ib_qp *qp,
 						struct ib_recv_wr *recv_wr,
 						struct ib_recv_wr **bad_recv_wr);
-	struct ib_cq *             (*create_cq)(struct ib_device *device,
-						int cqe);
+	struct ib_cq *             (*create_cq)(struct ib_device *device, int cqe,
+						struct ib_ucontext *context,
+						struct ib_udata *udata);
 	int                        (*destroy_cq)(struct ib_cq *cq);
 	int                        (*resize_cq)(struct ib_cq *cq, int *cqe);
 	int                        (*poll_cq)(struct ib_cq *cq, int num_entries,
@@ -780,6 +863,10 @@ struct ib_device {
 						  int num_phys_buf,
 						  int mr_access_flags,
 						  u64 *iova_start);
+	struct ib_mr *             (*reg_user_mr)(struct ib_pd *pd,
+						  struct ib_umem *region,
+						  int mr_access_flags,
+						  struct ib_udata *udata);
 	int                        (*query_mr)(struct ib_mr *mr,
 					       struct ib_mr_attr *mr_attr);
 	int                        (*dereg_mr)(struct ib_mr *mr);
@@ -817,6 +904,7 @@ struct ib_device {
 						  struct ib_mad *in_mad,
 						  struct ib_mad *out_mad);
 
+	struct module               *owner;
 	struct class_device          class_dev;
 	struct kobject               ports_parent;
 	struct list_head             port_list;
@@ -852,6 +940,16 @@ void *ib_get_client_data(struct ib_device *device, struct ib_client *client);
 void  ib_set_client_data(struct ib_device *device, struct ib_client *client,
 			 void *data);
 
+static inline int ib_copy_from_udata(void *dest, struct ib_udata *udata, size_t len)
+{
+	return copy_from_user(dest, udata->inbuf, len) ? -EFAULT : 0;
+}
+
+static inline int ib_copy_to_udata(struct ib_udata *udata, void *src, size_t len)
+{
+	return copy_to_user(udata->outbuf, src, len) ? -EFAULT : 0;
+}
+
 int ib_register_event_handler  (struct ib_event_handler *event_handler);
 int ib_unregister_event_handler(struct ib_event_handler *event_handler);
 void ib_dispatch_event(struct ib_event *event);
@@ -902,6 +1000,21 @@ int ib_dealloc_pd(struct ib_pd *pd);
 struct ib_ah *ib_create_ah(struct ib_pd *pd, struct ib_ah_attr *ah_attr);
 
 /**
+ * ib_create_ah_from_wc - Creates an address handle associated with the
+ *   sender of the specified work completion.
+ * @pd: The protection domain associated with the address handle.
+ * @wc: Work completion information associated with a received message.
+ * @grh: References the received global route header.  This parameter is
+ *   ignored unless the work completion indicates that the GRH is valid.
+ * @port_num: The outbound port number to associate with the address.
+ *
+ * The address handle is used to reference a local or global destination
+ * in all UD QP post sends.
+ */
+struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, struct ib_wc *wc,
+				   struct ib_grh *grh, u8 port_num);
+
+/**
  * ib_modify_ah - Modifies the address vector associated with an address
  *   handle.
  * @ah: The address handle to modify.
diff --git a/drivers/infiniband/ulp/ipoib/ipoib_ib.c b/drivers/infiniband/ulp/ipoib/ipoib_ib.c
index 8238766746b2..eee82363167d 100644
--- a/drivers/infiniband/ulp/ipoib/ipoib_ib.c
+++ b/drivers/infiniband/ulp/ipoib/ipoib_ib.c
@@ -81,7 +81,7 @@ void ipoib_free_ah(struct kref *kref)
 
 	unsigned long flags;
 
-	if (ah->last_send <= priv->tx_tail) {
+	if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 		ipoib_dbg(priv, "Freeing ah %p\n", ah->ah);
 		ib_destroy_ah(ah->ah);
 		kfree(ah);
@@ -355,7 +355,7 @@ static void __ipoib_reap_ah(struct net_device *dev)
 
 	spin_lock_irq(&priv->lock);
 	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
-		if (ah->last_send <= priv->tx_tail) {
+		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
 			list_del(&ah->list);
 			list_add_tail(&ah->list, &remove_list);
 		}
@@ -486,7 +486,7 @@ int ipoib_ib_dev_stop(struct net_device *dev)
 			 * assume the HW is wedged and just free up
 			 * all our pending work requests.
 			 */
-			while (priv->tx_tail < priv->tx_head) {
+			while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
 				tx_req = &priv->tx_ring[priv->tx_tail &
 							(IPOIB_TX_RING_SIZE - 1)];
 				dma_unmap_single(priv->ca->dma_device,
diff --git a/drivers/infiniband/ulp/ipoib/ipoib_main.c b/drivers/infiniband/ulp/ipoib/ipoib_main.c
index 6f60abbaebd5..fa00816a3cf7 100644
--- a/drivers/infiniband/ulp/ipoib/ipoib_main.c
+++ b/drivers/infiniband/ulp/ipoib/ipoib_main.c
@@ -600,9 +600,10 @@ static int ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
 
 			ipoib_mcast_send(dev, (union ib_gid *) (phdr->hwaddr + 4), skb);
 		} else {
-			/* unicast GID -- should be ARP reply */
+			/* unicast GID -- should be ARP or RARP reply */
 
-			if (be16_to_cpup((u16 *) skb->data) != ETH_P_ARP) {
+			if ((be16_to_cpup((__be16 *) skb->data) != ETH_P_ARP) &&
+			    (be16_to_cpup((__be16 *) skb->data) != ETH_P_RARP)) {
 				ipoib_warn(priv, "Unicast, no %s: type %04x, QPN %06x "
 					   IPOIB_GID_FMT "\n",
 					   skb->dst ? "neigh" : "dst",